(updated 9.18.19)

9.18.19 @7:45am - (Replying to @yotambarnoy) I am aspiring to the same ending, hoping that acting as equal will end their separate and hostile existence.

9.18.19 @4:32am - (Replying to @EsserHartmut @ingorohlfing and @tomdrabowicz) In social science, following Blalock Duncan et al, the ideas of testing and identifications are not new. But if you were a student of Suppes or Lewis, these ideas are new, b/c you cannot identify the effect of something (eg action) that you cannot represent. #Bookofwhy

9.18.19 @4:17am - While we are at it, here is another book on experimental philosophy, also dounlaudable, with a chapter (11) on "causal reasoning". #Bookofwhy

9.17.19 @11:58pm - An unexpected opportunity emerging from the Israeli election.

9.17.19 @11:32pm - Philosophers of science are beginning to take interest in causal inference. This new book (downloadable) shows how traditional problems in philosophy come to light through the new lens #Bookofwhy

9.17.19 @3:08pm - Sharing a nostalgic photo. Some place in my army days (1953-1956) I remember driving a truck, collecting ballots from remote army units, to be counted. A most heroic endeavor on the roads of those days. Our mission: Not one soldier left uncounted. It worked!

9.17.19 @2:36pm - (Replying to @jprwg @strangecosmos and 2 others) Knowledge (or "information") is what constrains your answers and drives them from "maybe" to yes/no or probable. I am reluctant to use "information" because people tend to confuse it with Shannon's information - a purely probabilistic notion. "Mechanism" is too narrow #Bookofwhy

9.17.19 @2:10pm - (Replying to @on_clusters @ABravoBiosca and @IGLglobal) The walls of NBER are taller than the Himalaya. No access to anyone but club members.

9.17.19 @1:47am - (Replying to @hofbeezy) Not only doesn't it hurt, it is also a beautiful poetry; like chanting Kiddush or singing Hatikva. Its not the words but the melody.

9.17.19 @1:28am - (Replying to @strangecosmos @GaryMarcus and @ylecun) If by "ML techniques" you mean any future algorithm, then you are justified in saying "I dont know". But if by "ML techniques" you mean algorithms based on data only, we can tell even today that the answer is NO. We can't compute 3-dim volume from a 2-dim shadow. #Bookofwhy

9.17.19 @1:17am - (Replying to @RivasElenaRivas) Let's do it together. Fit a 100-layer NN to data coming from ice-cream sales and crimes. Interpret the fitted NN as structural causal model and ask it: "Would crime increase if we ban ice-cream?" What answer would we get? #Bookofwhy

9.17.19 @12:54am - (Replying to @strangecosmos @GaryMarcus and @ylecun) Why speculate on what Sutton means or meant? Do you @strangecosmos believe that any ML technique can solve any of the toy problems in #Bookofwhy or Primer, given all the data in the world, and no information beside data?

9.17.19 @12:43am - Today, it is an election day in Israel. Public transportation is free, from anywhere to everywhere. I wish I could take one of those trains and fulfill my duty (I have dual citizenship). For a deep understanding of issues and moods see

9.16.19 @10:18pm - (Replying to @womensmarch) @womensmarch has just replaced a fake feminist with a confessed Zionophobe. Is there anything more weird, fraudulent and contradictory than an anti-Zionist feminist?

9.16.19 @10:11pm - (Replying to @SethAMandel)

9.16.19 @10:10pm - (Replying to @Marissa_Jae)

9.16.19 @9:57pm - (Replying to @haileybanack @EpiEllie and @AmJEpi) Just in case you wish to inflame the paradox with more fuel, here is a gallon from my own brewery In #Bookofwhy we took the liberty of calling the birth-weight phonomenon a "paradox", though it has been explained. Paradoxes enjoy cultural immortality.

9.16.19 @9:22pm - Retweeting my comment on a controversial change in the leadership of the Women March movement. An interesting thread.

9.16.19 @8:40pm - (Replying to @ZionessMovement) I just can't agree with your last sentence, implying that "anti-Zionism" is less racist than antisemitism. The former is not only bigoted but also eliminationist and borders on genocidal. Are you implying that anti-Zionism is the lesser of the two evils?

9.16.19 @8:25pm - (Replying to @ZionessMovement) Women March, Inc. has just replaced a fake feminist with a confessed Zionophobe. Is there anything more weird, fraudulent and contradictory than an anti-Zionist feminist?

9.16.19 @6:09pm - (Replying to @strangecosmos @GaryMarcus and @ylecun) There are theoretical impediments that even "general methods" cannot circumvent. If you search for a common point of two parallel lines, you can have the best search-and-learning method in the world and you won't find one. ML confronts such impediments to reaching GAI. #Bookofway

9.16.19 @4:22am - (Replying to @ewerlopes) Yes. Causality is heavier, with all the proofs, and histories, and arguments with philosophers and economists and statisticians. Primer is game-like: "Look Ma, I can do today what I couldn't yesterday, and it makes so much sense!" You dont want to miss it. #Bookofwhy

9.15.19 @6:38pm - So, perhaps I was too naive in assuming that my colleagues in the stat dpt have been doing non-parametric estimation for the past two centuries. But now, that they are awaken to the importance of causal estimands, do they need outside help? #Bookofwhy

9.15.19 @6:04pm - Agree. But why do we have to "think hard" if the task of estimation has been the sole target of super smart statisticians for over a century, backed by an empire that produced thousands of PhD's. Are we "smarter" then they? or are CI estimands new to them? #Bookofwhy

9.15.19 @4:32pm - (Replying to @LauraBBalzer and @GilmanStephenE) By all means, assuming someone younger does the writing, and someone older does the diplomacy. #Bookofwhy

9.15.19 @2:43pm - (Replying to @GarridoWainer) No official document, just a long discussion on Twitter, followed by a survey whether it is ethical for an author to publish the review he/she received. I believe majority voted YES. #Bookofwhy

9.15.19 @2:35pm - After #Bookofwhy, the next entry into causal inference is the PRIMER It was praised already by so many readers, so I won't add, except to note that Wiley is coming up with a clean version next month. In the meantime, the corrected chapters are accessible.

9.15.19 @2:23pm - (Replying to @GilmanStephenE and @LauraBBalzer) By the HUGE gap, do you mean going from a finite sample to an estimate of the estimand?

9.15.19 @4:36pm - (Replying to @manuelbaltieri) To make the Ladder of Causation more connected to "thinking" and cognitive functions, I was considering labeling the rungs: 1. Foresight, 2. Control. 3 Understanding. It rings better with Toulmin, 1961, "Forecast and Understanding" #Bookofwhy

9.14.19 @7:52pm - (Replying to @vardi) I wonder what Melinda Baldwin's opinion would be on our proposal to make all reviews public, 5 years after decision (anonymously if requested), so as to remind reviewers of the higher judgment of history. #Bookofwhy

9.14.19 @7:52pm - (Replying to @vardi) [You Retweeted] I wonder what Melinda Baldwin's opinion would be on our proposal to make all reviews public, 5 years after decision (anonymously if requested), so as to remind reviewers of the higher judgment of history. #Bookofwhy

9.14.19 @4:05pm - (Replying to @PHuenermund) I see nothing wrong in self-citing, especially if no one else has articulated the idea that you deem relevant. The problem I do see is that most self-citings point to irrelevant publications, bearing little relationship to the discussion in the text. #Bookofwhy

9.14.19 @3:50pm - Some symbolic gifts fill your heart with gratitude, but thinking what this gift would do to the spirit of the rockets-stricken children of Sderot, stops your heart from beating.

9.14.19 @2:29pm - In ancient Greece, "over-Democracy" led to the invention of formal logic; they had to reign-in the endless arguments. What invention will Israel's democracy lead to? A crucial election will take place on Tuesday. Stay tuned.

9.14.19 @10:56am - In this paper on Off Policy Evaluation I was particularly interested in Appendix C: Bridging the Gap between Reinforcement Learning and Causal Inference. #Bookofwhy

9.14.19 @10:43am - Another causal discovery paper . This one applied to diabetes data. #Bookofwhy

9.14.19 @10:33am - This Economics Bulletin paper applies causal discovery to future prices in the Chinese Stock Index 300. It seems someone is about to get rich soon. #Bookofwhy

9.13.19 @1:26pm - (Replying to @manuelbaltieri) I trust readers of #Bookofwhy are equipped with night-vision glasses that tell them right away that no definition of causality can be constructed in probabilistic vocabulary, no matter how sophisticated or skillful.

9.12.19 @7:12pm - (Replying to @katchwreck) By enlarge, the literature on "dimensionality reduction" stands orthogonal to causality, by virtue of being statistical. There is however a point of contact when we minimize the number of covariates we need to adjust for. Beautiful algorithms exist for this task. #Bookofwhy

9.12.19 @4:23pm - Every time a reader praises Primer I take a minute and read a paragraph or two, and come back with an urge to reply: "You are so right!". Today I succumb to this urge and recommend it to all readers, especially free Ch4, #Bookofwhy

9.12.19 @4:46am - (Replying to @harrydq and @TelegraphTech) Many thanks for this pointer. For the life of me, the last things we contemplated was getting embroiled in the Brexit debate. On the other had, if the WHY rules the world, why should this debate be excused. #Bookofwhy

9.12.19 @3:50am - (Replying to @TelegraphTech and @harrydq) Too bad I was blocked by the Pay Wall as it became interesting. #Bookofwhy

9.11.19 @11:07pm - As 9/11 day comes to an end, I am sharing one of the most profound requiems I have ever heard: "Kaddish para Daniel" Written by Benjamin Lapidus, it combines Hebrew, Aramaic and Spanish in a rythm that shakes the foundations of our souls. Kaddish for 9/11

9.11.19 @10:02pm - (Replying to @BarryOSullivan and @UCC) A milestone in the history of AI.

9.11.19 @8:31pm - (Replying to @BarryOSullivan and @UCC) Wow! And my copy of Boole's Laws of Thoughts (1854) has the signature of James William Warren, AM, MRIA, Sep. 1864. Was he a professor at Cork? We, book collectors, form a bond of ownership. A weak form of immortality.

9.11.19 @8:08pm - (Replying to @BarryOSullivan) UCC!! Cork Ireland!! Thats where George Boole wrote The Laws of Thoughts (1854) and Boolean Algebra was born!! Glad you are still marking the pre-Christian calendar. I do too!! Though mine is in pre-Christian Hebrew, going (logically) from right to left. Cheers, and many springs

9.11.19 @5:27pm - It is still 9/11, and it is LA, my home town, where a Temple of ex-Morroco Jews has been vandalized. The populist slogan "Free Palestine" has become a license for every disgruntled group to spread its "My-Grievance-Above-All" mentality wherever it clicks.

9.11.19 @4:44pm - The first victim of 9/11, Danny Lewin (Z'L), was an internet innovator and a graduate of my alma mater, the Technion in Haifa, Israel. He was on board AA Flight 11 from Boston to LA, and was murdered as he struggled to advance toward the cockpit. Watch

9.11.19 @12:43am - An opportunity to learn DAGs from Felix Elwert, one of causal-inference top teachers: . #Bookofwhy

9.11.19 @12:14am - It's 12 midnight, September 11, in LA, an hour I can't forget, b/c my son was murdered by the same people who made this day unforgettable. I salute the people of Israel who remember my fellow Americans through these columns of lights in the "9/11 Living Memorial", in Jerusalem.

9.10.19 @5:49am - This paper on "counterfactual fairness" has reached my desk: and reinforced my conviction that "fairness" is a counterfactual notion, and must hence be managed by structural models - the breeding grounds of counterfactuals. #Bookofwhy

9.9.19 @10:04pm - History enthusiasts will probably find the discovery of this colorful mosaic to be a proof that man is a history-seeking machine. The place, Tabgha, is were I spent some of my army days, laden with sweet nostalgic memories of the sea of Galilee, Kinneret in Hebrew.

9.9.19 @5:04pm - I have added a link to Maudlin's review of #Bookofwhy on and, following our lively discussion here, I've added comments to clarify some not-so-obvious points in the book, especially the difference between Rung Two and Rung Three in the Ladder of Causation

9.9.19 @2:42pm - Our AI-minded readers should find this NYT interview with Gary Marcus and Ernest Davis to be illuminating: #Bookofwhy

9.8.19 @4:49am - (Replying to @PHuenermund and @nickchk) An important distinction that makes the difference between those who can take a causal problem and bring it to a stage where it can be estimated by random trees and other statistical tools and those who forget the first stage and assume it was prepared by someone else. #Bookofwhy

9.7.19 @4:46pm - (Replying to @autoregress) Sorry it I am/was going after the wrong economists, but think about what an enlightened econ. student feels upon hearing from leaders in his/her field: "We are waiting to be shown the money". Is this the spirit of Haavelmo, Marschack and Arrow? #Bookofwhy

9.7.19 @4:30pm - (Replying to @soupvector and @aregenberg) Thanks for posting, and welcome to this Twitter Zoo, where you can meet other readers enjoying their ability to do things they always wanted to do, including convincing economists that they, too, can do things today they always wanted to do. #Bookofwhy

9.7.19 @4:23pm - (Replying to @autoregress) "I like my flashlights demonstrated in the real world!" but I won't try them myself, not even near my faithful lamppost, and if someone demonstrates them elsewhere, I say: "We are different, for us .... for us "real world" is what we read in good econometric journals.#Bookofwhy

9.7.19 @4:00pm - (Replying to @autoregress) The "flashlight" works around your IV lamppost as well, see Fig. 5.1 in Causality (2000), you just need to press on the right button. #Bookofwhy

9.7.19 @3:53pm - (Replying to @autoregress) Imagine how many lost wallets and "good experimental studies" are awaiting their owners in train stations while economists cling to one lamppost, and won't try a flashlight, even one that performs well in epidemiology and other train stations. #Bookofwhy

9.7.19 @2:47pm - To all my Brazilian readers, colleagues and students - Happy Independence Day.

9.7.19 @2:38pm - (Replying to @HananyaNaftali) Thank you @HananyaNaftali for helping me start the day on a positive note. ps. I am trying to contact you by email:

9.7.19 @2:15pm - (Replying to @jon_y_huang @AdanZBecerra1 and 7 others) "will never accept" is a pretty bold statement for a discipline that prides itself on inventing structural approaches. As a student of history, I am fascinated by what happened to those economists, and whether modern day econ. students will shake them away from the new lamppost.

9.7.19 @1:50pm - (Replying to @DorotheaBaur) Not only "he doesn't entirely agree with the ladder" he actually misses the key transition from Rung-2 to Rung-3, as I explain in an earlier thread:

9.7.19 @12:58pm - (Replying to @DorotheaBaur) Thanks for posting, but note that "Pearl seems to think they are loaded with philosophical significance" is too humble. The #Bookofwhy actually claims that they are essential in science, & of practical significance in legal, medical and policy decisions.

9.7.19 @3:33am - (Replying to @sapinker) My first reaction to Maudlin's review was: The profound separation between interventions and counterfactuals is shown in (App. I), and by "Mute!" I mean: It would surprise me to find an idea from pre-1990 philosophy that I missed.

9.6.19 @12:38pm - (Replying to @djinnome) You are right, thanks for noting.

9.6.19 @12:35pm - (Replying to @_Srijit) Correct. Like "token" vs. "type", individual vs. population, actual vs. average.

9.6.19 @5:58am - (Replying to @_Srijit) What you and Maudlin are missing is remembering that interventional studies are non-deterministic estimating averages over populations (or unknown factors.)The statement "knowing that I am going to win" is not what the study gives you. I discuss it here:

9.6.19 @5:36am - (Replying to @yudapearl @arturlsc and 6 others) I am surprised that the computational aspects of DAGs is so underestimated. DAGs permit us to answer questions which otherwise are intractable. E.g.,"Tell me if the partial correlation R_{XY.Z} is zero", or "Tell me which parameter is estimable by OLS" #Bookofwhay

9.6.19 @4:08am - (Replying to @arturlsc @Jabaluck and 5 others) You must be kidding. Can you name another representation scheme, out of the many others, which allows you to see the testable implications of your causal assumptions ? #Bookofwhy

9.6.19 @3:02am - (Replying to @asweinmann) Appreciating your kind words. After a week of arguing with on-lookers, it is refreshing to hear from an unspoiled reader, interested in science, not in arguments. #Bookofwhy

9.6.19 @2:27am - Echoing our discussion of interventions and counterfactuals, I have summarized part of it in this paper submitted to the next issue of JCI. Comments are welcome, pointing to omissions, disagreements and improvements; my deadline = Sep 12 #Bookofwhy

9.6.19 @12:58am - (Replying to @Jabaluck @PHuenermund and 4 others) You have very low opinion of experimentalists, assuming they are incapable of generalizing from a toy problem to patterns of impediments they see in their substantive works. #Bookofwhy is written for enlightened experimentalists who, once aware, will dismantle those impediments.

9.6.19 @12:42am - (Replying to @Jabaluck @PHuenermund and 4 others) Every modeling task assumes away many difficult things, but at least we can represent them explicitly, and reason about ways of overcoming those "most difficult parts." I have not seen this ability demonstrated in "mostly harmless". #Bookofwhy

9.6.19 @12:33am - (Replying to @Jabaluck @PHuenermund and 4 others) Beauty! Please teach us the vocabulary of that "own language" and its internal logic, perhaps it is more effective than DAGs, since quasi-exps. think it is more "reliable", and has resulted in a "credibility revolution". I am truly curious, as computer scientist should #Bookofwhy

9.6.19 @12:19am - (Replying to @yudapearl @Jabaluck and 5 others) And what's the added wisdom? Are the impediments to fitting identification strategies in "real-life studies" different than 1. confounding,2. non-exclusion, 3.selection bias etc. all of which are representable in "make up" examples if one is serious about handling them.#Bookofwhy

9.6.19 @12:09am - (Replying to @Jabaluck @PHuenermund and 4 others) Please show us one challenge that cannot be represented in a made-up example and that only reveals itself in "real-life studies" where, instead of symbols, variables are decorated with: "return to school" "return to prison" "Years in service" What's the big deal?#Bookofwhy

9.5.19 @10:36pm - (Replying to @PHuenermund @Jabaluck and 4 others) @PHuenermund please remind me what a tDAG is. Is it a mental representation of knowledge which "quasi-experimentalists" consult when fitting an identification template, and which they refuse to commit to paper for fear of appearing traditional, and exposing assumptions?#Bookofwhy

9.5.19 @10:26pm - (Replying to @Jabaluck @PHuenermund and 4 others) Your skepticism reveals how powerful DAGs are in unveiling assumptions that DAG-averse cultures hide under the rug, ensuring that no one ever questions the confidence with which exogeneity or exclusions are overstated in quasi-experiment. The hidden invites no question#Bookofwhy

9.5.19 @10:01pm - (Replying to @Jabaluck @PHuenermund and 4 others) What about "DAG-invented" strategies like "conditional IV" or "Instrumental set" or "bow-free" or even Fig. 5.10 ?? Some are "DAG-invented" and some are "DAG-synthesized". The jury is not stupid, it's just kept ignorant by tribal zealots. #Bookofwhy

9.5.19 @8:48pm - (Replying to @Jabaluck @PHuenermund and 4 others) I surely believe that DAG as a good working hypothesis about the process (with Friedman's critics of secondary importance) and, more importantly, as a model for a huge number of real-life cases in which the mediator is physically "shielded" from the confounder. #bookofwhy

9.5.19 @8:35pm - (Replying to @Jabaluck @PHuenermund and 4 others) And to clarify, I do it too. If you look at the text below Fig. 5.10, I search the DAG for patterns of identification and, when not found, I repair the ones that are repairable. Except, I do it on a drawn DAG, not on my mental DAG, which I believe to be less reliable. #Bookofwhy

9.5.19 @8:21pm - (Replying to @Jabaluck @PHuenermund and 4 others) This is how Phillip and Sewall Wright "stumbled upon" IV's, now a celebrated "research design". They started with 4-5 variables, causal relations among them, Sewall's trick of converting them to covariance constraints and faint hope of inverting them back, no "design". #Bookofwhy

9.5.19 @8:07pm - (Replying to @RashidaTlaibz) Strangely! I do not see any Jewish Stars among the marchers! What happened to my old comrades? Sad! @EWilf

9.5.19 @7:31pm - (Replying to @y2silence) Interesting! Does Rosenbaum use the words "cause of crying"? Rubin proclaimed such questions to be "more of a cocktail conversation topic than a scientific inquiry", thus purging the word "why" from the vocabulary of his disciples. I thought his verdict still reigns. #Bookofwhy.

9.5.19 @7:14pm - (Replying to @_Srijit) Causal Bayesian Networks are called "Causal" because, unlike ordinary Bayesian Networks, which are purely associational, they provide answer to all interventional questions: What if we raise taxes? ie., all policy related questions - just what economists should rejoice #Bookofwhy

9.5.19 @7:02pm - The coin-game (below) exemplifies this kind of contradiction. Win-on-correct-guess is one model of the world, random winning (ignoring your guess) is another. Both are compatible with experiments, yet the second says:"No, you would't have lost had you acted differently"#Bookofwhy

9.5.19 @5:43pm - I owe readers an explication of what I mean by: "Intervention studies CANNOT ANSWER counterfactual questions". "Cannot answer" means that two different world models, both compatible with the studies, can generate two contradictory answers to the same question. #Bookofwhy

9.5.19 @2:19pm - Glad you re-posted this explanation of the difference between rung-2 (intervention) and rung-3 (counterfactuals) in the Ladder of Causation. Many researchers still find it hard to swallow (eg Maudlin) especially RL folks, for whom the world is just "interventions"#Bookofwhy

9.5.19 @5:59am - For a simple example consider a game where we win upon guessing the outcome of a fair coin and lose otherwise. The action "guess head" has no effect on winning, neither has "guess tail". Yet, upon winning, we can assert: "Had we acted differently we would have lost".#Bookofwhy

9.5.19 @5:47am - To elaborate, Causal Bayesian Networks [Causality Ch. 3] enable us to compute the effects of all possible actions, compound actions and actions conditioned on observed covariates and, still, none can answer the couterfactual:"What if we have done things differently? #Bookofwhy

9.4.19 @10:16pm - (Replying to @Jabaluck @PHuenermund and 4 others) "Real-world" again? Take any one of those revered "quasi experiment" successes in which exogeneity and exclusion were contested. Done. But I thought the discussion revolved around what comes first, the DAG or the ident. strategy. So now we start with an oracle, no DAG.#Bookofwhy

9.4.19 @9:42pm - (Replying to @yudapearl @PHuenermund and 5 others) "Starting with nothing" means no DAG, no SEM, no IV, just an oracle that can tells you, for every variable that comes to your mind, what the "sources of variations" are for that variable. #Bookofwhy

9.4.19 @9:20pm - (Replying to @PHuenermund @Jabaluck and 4 others) I tried to stay out of this discussion because it went over my head. I think @Jabaluck methodology could become clearer if he tells us how one should handle the problem in Causality Fig. 5.10 (below), step by step, starting with nothing but the desire to estimate beta.#Bookofwhy

9.4.19 @1:05pm - I just read Maudlin review. It's largely sensible, save for two mistakes: (1) ".. it is not possible to think causally but not counterfactually." Causal Bayesian Networks demonstrate that it is possible. (2) I was intimately familiar with the Tetrad project of the 1980's. Mute!

9.4.19 @5:53am - (Replying to @stephenpollard) This precious lady is so saintly innocent, that I begin to believe she really does not understand why people would call her a racist for supporting a racist movement. Many BDS supporters can't stomach it: Me? A racist? See

9.4.19 @3:00am - (Replying to @PHuenermund @pierre_azoulay and 4 others) What is the canonical example(s) economists used to demonstrate "selection on observables"? What are students told to do in such cases? #Bookofwhy

9.4.19 @2:54am - (Replying to @saurabh_jha21) Causal inference goes beyong ml/dl models, so I do not think you can implement the former with the latter.

9.4.19 @2:52am - (Replying to @FelixThoemmes and @y2silence) I heard about the Venn diagrams used in the context of "variance explained". Can you explicate what intuition they support?

9.4.19 @1:26am - (Replying to @y2silence) I interpret Horst's surprise (upon finding a suppressor in the data, 1941) as evidence that regression analysts expect correlations to behave like separation in graphs, that is, if a node Y is separated from X and from Z, it must also be separated from the pair (X,Z). #Bookofwhy

9.3.19 @11:11pm - Should have used it as a trailer for #Bookofwhy. Don't knock it, the script writer was a thoughtful philosopher.

9.3.19 @2:42pm - (Replying to @y2silence) And the winner is Yongnam Kim @y2silence !!! The example I had in mind had S and X interchanged, which works just as well, and tells us that either S or X need to be a collider for a suppressor to play tricks on us. Now, back to Horst (1941): Why was he surprised? #Bookofwhy

9.3.19 @1:40pm - (Replying to @FelixThoemmes) Keen observation. But the example I have in mind has no cancellation, and is familiar to every UG student of probability or statistics. #Bookofwhy

9.3.19 @1:36pm - (Replying to @djinnome) The independencies stated are not shown in the graph. eg X and S are shown dependent. So they need to rest on some compelling process.

9.2.19 @11:45pm - Hate to keep you in suspense. Yes! A super-suppressor does exist! Its a variable S, uncorrelated with X and Y, that, if added to the regression, turns X from a useless to a perfect predictor of Y. Can readers guess who S is? His name will tell us what suppression is. #Bookofwhy

9.2.19 @2:09am - (Replying to @matt_vowels) On the other hand Horst (1941) was interested in prediction, something I did not realize. Evidently, prediction-minded people also have intuition. It comes, I surmise, from causal assumptions that sneak secretly into intuition about predictions. Deserves some thought. #Bookofwhy

9.2.19 @12:39am - For readers who wrote to me last month about anti-Semitism and anti-Zionism, this latest article by Gil Troy is the best I've ever read: And Weiss's book that Troy reviews: is an insightful, eye-opening microscope of our generation.

9.1.19 @11:46pm - Social Scientist: Look what I found! A suppressor! Statistician: Big deal, it shows in regression analysis. Computer Scientist: Why were you surprised? Statistician: This is a question for psychologists. Social Scientist: No, its a question for all scientists Why was I surprised?

9.1.19 @11:35pm - To understand "suppressors", it is instructive to examine a "super-suppressor": A variable S that is uncorrelated with the regressor X and with outcome Y, yet, when added to the regression equation, turns X from useless to perfect predictor of Y. Puzzle: Does S exist?#Bookofwhy

9.1.19 @1:41pm - (Replying to @NunezKant) Agree. We sometimes forget what science is all about, and it is amazing that 12 years after, I still need "words" and "caps and gowns" to be reminded. Thanks for re-posting. #Bookofwhy

9.1.19 @4:59am - An unprecedented development at the UN. First time this world body addresses the inner core of the Arab-Israeli conflict, which was discussed earlier on this Tweeter. Kudos to the Brazilian delegate.

9.1.19 @3:25am - Today, September 1st, marks the 10-year anniversary of the publication of Causality (2009, 2nd ed.) I am proud to see that the book has stimulated 14,850 citations on Google Scholar and that, oddly, I am still agreeing with everything it says. #Bookofwhy

9.1.19 @2:40am - For readers who are tired of listening to my Israeli accent, here is my incredible co-author, Dana Mackenzie, introducing #Bookofwhy in plain English and lucid eloquence.

8.31.19 @5:05pm - (Replying to @nbarrowman) "Contributing cause" stands between "sufficient" and "necessary" cause. Note this interesting Wash. Post. definition of "Analysis": Interpretation of the news based on evidence, including data, as well as anticipating how events might unfold based on past events. All predictive!

8.31.19 @12:19pm - (Replying to @yudapearl @ang_hermann and 3 others) Specifically, the probabilities that annotate arrows emanating from "action nodes" in a decision tree are P(y|do(x)), not P(y|x), as classical textbooks might suggest. The former need DAGs to be estimated. #Bookofwhy

8.31.19 @11:55am - (Replying to @ang_hermann @furtadobb and 2 others) DAGs ARE used in the theory of decision. The reason we estimate P(y|do(x)) is to "Find x that maximizes E[U(y)|do(x)]" where U(y) is the utility of outcome y. For use in decision trees, see

8.31.19 @11:43am - (Replying to @TimFooler) No, I haven't given any such thought, worth looking into, while keeping in mind what we want to know and what we do know, ie, input--> output.

8.31.19 @1:45am - (1/3) This is an excellent paper, that every regression analyst should read. Primarily, to appreciate how problems that have lingered in decades of confusion can be untangled today using CI tools. What I learned from it was that the "suppressor surprise" is surprising even when
8.31.19 @1:45am - (2/3) cast in a purely predictive context: "How can adding a second lousy predictor make the first a better predictor?" Evidently, what people expect from predictors clashes with the logic of regression slopes. The explanation I offered here (Section 3)
8.31.19 @1:45am - (3/3) shows how the phenomenon comes about, but the reason for the clash is still puzzling: What exactly do people expect from predictors, and why? #Bookofwhy

8.30.19 @12:45am - (Replying to @analisereal @reflecmec and 4 others) One should add here that the second kind of intervention is identifiable whenever ETT is (ie, effect of treatment on the treated), as demonstrated in Primer (p. 109-111) primer-ch4:, #Bookofwhy

8.30.19 @3:58am - (Replying to @ClaudeAGarcia) And they dare tell us they have hard time recruiting subjects for randomized treatment ..... Hoping you have a great summer.

8.30.19 @1:24am - (Replying to @Moshe_Hoffman) Very interesting thread, touching on a long debated concept "the actual cause" [Causality ch. 10]. An important distinction may illuminate your analysis: "necessary vs. sufficient" causes. A recent post demonstrates it in the Oxygen-Match story #Bookofwhy

8.29.19 @11:28pm - (Replying to @boredyannlecun) According to Von Neumann we should all research computers. Thermostat controls are an xxx trillion $$$ industry. Counter that fact! #Bookofwhy

8.29.19 @9:45pm - (Replying to @analisereal @thosjleeper and 3 others) Great, you just proved that "linear models" imply homogeneous effects, but not the other way around. Linear combinations of nonlinear functions also guarantee effect-homogeneity. [Assuming that by "effects" we mean differences eg E[Y|do(x1)]-E[Y|do(x2)] @Bookofwhy

8.29.19 @8:02pm - (Replying to @thosjleeper @DanielNevo and 3 others) This has been my consistent usage, Yes. Although I would not be surprised if someone discovers a "nonlinear model" exhibiting the "effect homogeneity" property. #Bookofwhy

8.29.19 @7:56pm - (Replying to @ildiazm and @mgaldino) Agree. Parametric regression models comes in two varieties: 1. Carriers of statistical assumptions, eg. "E[Y|x] is linear in x" and 2. Statements of estimation strategies, eg. "Find the best linear estimate of Y, given x, regardless of the actual shape of E[Y|x]". #Bookofwhy

8.29.19 @1:57pm - (Replying to @DanielNevo @ildiazm and 2 others) I believe @analisereal summarized the "linear model" issues fairly well here: A distinction between "linear in parameters" and "linear in variables" is highly warranted. #Bookofwhy

8.29.19 @5:16am - (Replying to @bwundervald @nickdaleburns and 3 others) Good try, but the nonlinear function x3=x1*x2 makes the model nonlinear. It is not a matter of convention; it is substantive. In linear models, all causal effects are the same for all units (ie, all values of the error terms) We cant change this property by renaming. #Bookofwhy

8.29.19 @2:59am - (Replying to @nickdaleburns @RduvalH and 2 others) A "linear function of the separate predictors" is nice and explicit. But the words "linear model" may lead to some confusion, for the reasons I mentioned.

8.29.19 @2:48am - (Replying to @davekarpf) OK, now that you are a national celebrity, can you explain to us mortals what were you bedbugs a metaphor for? What did you mean to say in that comparison? Who are your heroes and your bedbugs?

8.29.19 @2:07am - (Replying to @thosjleeper and @mgaldino) Interesting. I was not aware of this confusion. Any reference to a Social Science book using this nomenclature? (preferably by authors who know the difference between regression eqs. and structural eqs.)

8.29.19 @1:50am - (Replying to @thosjleeper and @mgaldino) I am eager to learn, which community is it that labels a regression equation with product terms a "linear model"? Who are the careless authors who would do so? Why would they do it? #Bookofwhy

8.29.19 @1:41am - (Replying to @causalinf and @Undercoverhist) As tribute to the great fun we had today, I would like to dedicate a day each month to meet with potential authors of economics textbooks. Authors better be: 1) Tenured, ie free of peer pressures 2) Aspiring to brighten up the dark sides of econometric education #Bookofwhy

8.29.19 @12:27am - (Replying to @mgaldino) Sorry if any confusion, but I naively assumed that a regression equation containing a product term would not be classified as a "linear model". Two reasons: 1. it is not linear. (2)It is not a "model" of reality (ie a carrier of assumptions), but a tool of estimation. #Bookofwhy

8.28.19 @5:18am - (Replying to @emc2G @DanzigMD and 3 others) And you buy into this racist propaganda? Would you use the "separate but equal" metaphor on any other 2-states, say US & Canada? or US & Mexico? This comparison to racial segregation was manufactured by enemies of coexistence, and people in your sphere of information buy it?? SAD

8.28.19 @5:05am - (Replying to @emc2G @DanzigMD and 3 others) You make it sounds like "subjugation" is an Israeli pastime recreation, rather than a predicament forced upon them by neighbors who openly declare their intentions. Don't you read what Palestinians tell their children?

8.28.19 @2:55am - (Replying to @tdietterich) I would be very interested in your collection, when done. I think scientific creativity and theory formation are structurally similar to improvisational problem solving, as both invoke "modular template breaking" like the Lion Man in #Bookofwhy Chapter 1.

8.28.19 @1:24am - (Replying to @DanzigMD @emc2G and 3 others) I, for one, never understood what people are trying to achieve by saying the conflict is "COMPLICATED". An excuse from solving it? A diversion from addressing its core? To me, it's baby simple: A clash between two legitimate national movements, one says WE, the other says ME.

8.27.19 @9:45pm - (Replying to @emc2G @DanzigMD and 3 others) The comparison may sound inaccurate if you do not consider Zionophobia a form of racism, with genocidal intentions. I do, for the reasoned explained here:

8.27.19 @8:06pm - (Replying to @emc2G @DanzigMD and 3 others) @DanzigMD compared David Duke racism to the racist rhetoric and activities of Rashida Tlaib. Ashrawi and Miftah are more sophisticated, they know how to apologize once the damage is done. I have not heard Tlaib apologize for her Zionophobic outbursts. She can't! betray her base.

8.27.19 @2:57pm - (Replying to @emc2G @DanzigMD and 3 others) You would never get a racist to admit to what he/she is. The litmus test: Charge them with Zionophobia and see how proud they sing.

8.27.19 @5:40am - (1/2) A bunch of new papers have reached my screen which seem related to discussions we have had here on tweeter.
[PDF] Transcriptomic Causal Networks identified patterns of differential gene regulation in human brain from Schizophrenia cases versus controls [Three more...]
8.27.19 @5:40am - (2/2) [PDF] Data Management for Causal Algorithmic Fairness
Counterfactual Reasoning for Process Optimization Using Structural Causal Models
[PDF] Reinforcement Learning is not a Causal problem #Bookofwhy

8.26.19 @11:33pm - @RepSchneider has stood up to @IfNotNow with "shades of gray", but I prefer Joe Biden's answer to the same bullies (paraphrased): "Occupation! Occupation! I have not met a single Palestinian leader who is willing to accept Israel's right to exist". Its black and white! No gray!

8.26.19 @9:51pm - (Replying to @f2harrell @EpiEllie and 6 others) This article was written 2008 and, yet, I see no sign of causal definition of "diagnosis". I wonder if probabilistic notions of diagnosis are still ruling the field? Or have they been replaced by Diagnosis = Best Explanation, as in ??#Bookofwhy

8.25.19 @10:47pm - I had the privilege of knowing Danny Cohen in the 1990-1980's and of watching his brilliant mind at work. Unlike us, back-seat academics, he was an adventurer in real life - a pilot, a fighter, a system builder and a tough skeptic of AI. He will be missed. #Bookofwhy

8.25.19 @10:18pm - (Replying to @ayusharms) In more advanced studies one need indeed to accommodate cycles, (See for example Causality p.215 ). DAGs however allow us to leverage the full power of do-calculus, which has not been matched yet in cyclic systems. #Bookofwhy

8.24.19 @5:23pm - (1/ ) (Replying to @Chris_Auld) 1/I guess what you take as "formally similar" I take as vastly dissimilar. In one case (X-rest.) I can immediately write down the OLS estimand of EVERY parameter and in the other (cov-rest.) it is still an open question whether some parameters are identified, awaiting a decision
8.24.19 @5:38pm - (1/ ) (Replying to @yudapearl and @Chris_Auld) of whether those parameters can be solved uniquely from the covariance matrix [!!! decision] The mystery may be dissolved if you can just walk my students by the hand in Fig.5.10, & starting with the 3 eqs., show them why adding W-->Z spoils beta and W<--Z does not. #Bookofwhy

8.24.19 @5:01pm - (Replying to @Chris_Auld) Where discussion has gone? I am trying to extract the set of principles that leads econ. students towards the solution of Fig. 5.10. So far, a failure. Along the way, you said "cov. restrictions are just extension of exclusion restrictions" which blew me off 2 miles. #Bookofwhy

8.24.19 @4:39pm - (Replying to @Chris_Auld) My glitch! I meant the effect of Z on X is not identified. The same as delta in Fig. 5.10. Same as beta, if arrow W-->Z is added. In contrast, with exclusion restrictions alone (& all eps's uncorrelated), all parameters are OLS identified; not 2SLS, but straight OLS. #Bookofwhy

8.24.19 @12:39pm - (Replying to @Chris_Auld) Here is something it cannot do:
with Eps(Z) correlated with Eps(X), and Eps(Z) uncorrelated with Eps(Y) By "cannot do" I mean "it cannot identify effect of Z on Y." Moreover, it cannot tell us in general which system of eqs. is identifiable. #Bookofwhy

8.24.19 @6:52am - (Replying to @Chris_Auld) I dont see why covariance restrictions are straight forward extension of exclusion restrictions. The latters permit all effect to be identified by OLS, the former are still an open problem. Why is Z exogenous? Dont we need to examine the Z eq.? eg. what if we add W-->Z #Bookofwhy

8.24.19 @1:48am - European readers may have interest in this workshop on causality by @RonKenett . I hope he will share the slide with us, so we can learn what "fishbone diagrams" can do for CI, and how business applications can benefit from the causal revolution. #Bookofwhy

8.23.19 @6:33pm - (Replying to @Chris_Auld) I am not disputing the necessity or advantage of doing algebra vs. other methods. I am just asking: What are econ. students taught to do in cases like Fig. 5.10 ? No traps to my question. Just trying to learn something I could not get from the econ. literature. Help! #Bookofwhy

8.23.19 @3:39pm - (Replying to @Chris_Auld) Agree. But how is this "determination" done? By symbolic algebra? (super exponential) or, by step by step reasoning, as in Causality p. 153? #Bookofwhy

8.23.19 @3:28pm - (Replying to @analisereal) So, for my understanding, given model 5 below, DR will instruct us to adjust for both U and Z, correct?

8.23.19 @3:13pm - (Replying to @oacarah @AndersHuitfeldt and 4 others) The content probably add clarity, but the abstract speaks of protecting "consistency", not "precision" which, to me, adds to the confusion. #Bookofwhy

8.23.19 @10:04am - (Replying to @Jsevillamol) Thank you for making confounders simple (is it an oxymoron?) in this post. I would suggest another warning, against proxies of mediators, as demonstrated here: #Bookofwhy

8.23.19 @8:26am - (Replying to @AndersHuitfeldt @AdanZBecerra1 and 3 others) Thanks for making this distinction clear. Curious, why has "double robustness" developed in the context of causal inference tasks and not in classical statistics? Or has it? #Bookofwhy

8.23.19 @2:31am - (Replying to @jon_y_huang @AdanZBecerra1 and 2 others) What is the simplest model to demonstrate this preference? Doesn't it violate the heuristic advocated in: "A Crash Course in Good and Bad Control" #Bookofwhy

8.23.19 @1:42am - (Replying to @AdanZBecerra1 @jon_y_huang and 2 others) I understand that "double robustness" provides protection against misspecification in your model. Thus, naturally, your model (ie DAG) needs to be consulted before deciding if protection is needed and, if so, what protection would be adequate. #Bookofwhy

8.23.19 @12:41am - Watch this video of my friend @DanzigMD about Congresswomen @Ilhan & @RashidaTlaib & how they were caught partnering w Miftah, a Zionophobic NGO that accused us of using "the blood of Christians in the Jewish Passover."

8.22.19 @11:47pm - (Replying to @mehdirhasan) One thing you would never get a Zionophobe to accept: that anyone else has a right to dignity and self determination. They win debates on anti-semitism, but will never debate the ugliness of Zionophobia.

8.22.19 @10:46pm - I believe my guiding mantra would not be inappropriate here: "Only by taking models seriously we learn when they are not needed". #Bookofwhy

8.22.19 @10:23pm - (Replying to @josephpapptheat @ADL and 2 others) I believe by "our trauma" they mean the Zionophobic bigotry of Rep. Rashida Tlaib and the way some Democrats embrace this bigotry.

8.22.19 @2:47pm - (Replying to @FJnyc @mehdirhasan and 2 others) The new Orientalism: Mehdi Hasan is defining Jewish identity. And CNN pretends he knows what he is talking about.

8.22.19 @12:54pm - (Replying to @Chris_Auld) I am unfamiliar with the evaluation method you mention. How would your students tackle Fig.5.10 ?? What would be the first step? Input --> output? #Bookofwhy

8.22.19 @12:14am - (Replying to @JohannesTextor @PHuenermund and 2 others) Interesting paper on ranking efficiency of adjustment sets in linear models. I presume this coincides with the ranking in for non-parametric models. A skeptical economist may still ask: What about 2SLS? My answer: Where do econs. see 2SLS? #Bookofwhy

8.21.19 @10:41pm - (Replying to @n_iccolo and @FriedrichHayek) This is the crucial first step. Next, all you need is to read Primer and, believe me, you will be way ahead of most ML folks. #Bookofwhy

8.21.19 @9:46pm - (Replying to @yudapearl and @Chris_Auld) A question to economists and other folks curious about the role of "reduced form equation" (RFE). Q. If I were to ask 100 econ students what the RFE's are in this model: would I get one answer? three answers, or ten answers? Can we see one? #Bookofwhy

8.21.19 @9:29pm - (Replying to @FriedrichHayek) This is indeed an awesomely important work, that requires at least a few days/weeks to digest, with the aim of sorting out what primitive causal templates infants possess, and how advanced causal structures evolve from those templates. #Bookofwhy

8.21.19 @7:07pm - (Replying to @PHuenermund @djvanness and @Chris_Auld) True, DAGs are not expected to talk about efficiency, nor does any model of the world, they nevertheless do, to the maximum extent that efficiency considerations are dictated by the world. Are the alternatives more informative about efficiency? @Bookofwhy

8.21.19 @6:58pm - (Replying to @analisereal and @Chris_Auld) My earlier tweet seems to have gotten lost from this thread. Here is is And I wish a seasoned economist would tell us how his/her students are taught to solve this problem, step by step. #Bookofwhy

8.21.19 @6:18pm - (Replying to @tytung2020) The two laws are described concisely on page 168 here

8.21.19 @6:47am - Ten years ago I wrote this survey paper, which provides a panoramic view of the various approaches to causal inference. Its aim was to unify, rather than differentiate. I believe it was successful in showing how they all emerge from two laws. #Bookofwhy

8.21.19 @4:18am - (1/ ) It was not the mean-spirited tone of "uninformed" that triggered my reaction, but the realization that so many well-intentioned people can psych themselves into believing that slogans are knowledge. One of the reasons that I often fear left tyranny as much as right tyranny is
8.21.19 @4:18am - (2/ ) that the former strives to base its claims on "universal knowledge". "Everyone knows that there was no Temple in Jerusalem" said Arafat to Clinton. "Everyone knows that Israel is 'apartheid' state" chant BDS cronies, even intellectuals. I flip by the sound of those chants.

8.21.19 @1:39am - I haven't seen this intriguing thought experiment before. It is amazing how much truth can be unveiled through a mechanical transposition of just two words Jewish<---> Muslim. It highlights how crucial counterfactuals are to scientific thought. #Bookofwhy

8.21.19 @12:41am - (Replying to @lewbel) I thought non-parametric models include non-additive and non-separable errors. If true, then exog. ensures the OLS identifiability of RFE [Assuming, of course, we agree on what exogeneity and identifiability mean]. Oh God, when will we have a consensual glossary??? #Bookofwhy

8.20.19 @11:50pm - Thanks for noting the connection between Causality and the search for truth. I have explained here why I decided to continue both on @yudapearl . Someone has to counter the sirens of deceit that continue to dishonor the floor of the US Congress.

8.20.19 @11:28pm - (Replying to @DSPonFPGA and @SoniaCuff) Always a thrill to read ancient papers, like reading Greek mythology, Gee, what we used to believe in those days! #Bookofwhy

8.20.19 @3:45pm - (Replying to @Chris_Auld) Do we have a list of requirements that would guide my students towards an econ. solution of this problem? Something in the form: 1. examine the equations, 2. check if there exists...such that... 3. next check... #Bookofwhy

8.20.19 @3:38pm - (Replying to @Chris_Auld) This is a very interesting observation, deserving a mention in 3rd edition. The DAG only says: "use either beta1 or beta, both are consistent". The efficiency comes from noticing that beta1 invites 2sls est., which is not in the DAG, but in the mind of the modeller. #Bookofwhy

8.20.19 @5:23am - (Replying to @Chris_Auld) According to the definition in Phil Haile slides, Z is not exogeneous, as it is correlated with W. Can you outline just the conceptual steps of how LII is achieved. #Bookofwhy

8.20.19 @5:04am - Retweeting Elias talk on "Causal Data Science." If I were in Boston, I would not miss it. But given the circumstances, we will wait for the utube video. #Bookofwhy

8.20.19 @3:20am - (Replying to @CalumDavey) Appreciate your applaud but, for me, this is not "politics." It is a matter of being true to myself and my identity. It is paying back to a community that has invested dearly in my education and is now unable to fend for itself under this new barrage of populist slogans.

8.20.19 @2:39am - (Replying to @melvinwevers) Would "singularities" be more acceptable? But please do not ask me to change "morally deformed" - this is the lens through which people must judge the consequence of their words. This is my first book in traditional Chinese. Just the thought that students in remote areas of China are learning to speak cause and effect sends shivers in my spine. I hope the govt tolerates this revolution. #Bookofwhy

8.20.19 @2:29am - (Replying to @Chris_Auld) RF isn't useful here because X is not exog. by RF definition. As to precision, DAGs are not very helpful here, with two exceptions: 1. the partial order defined in 2. Ratios of correlation coefficients can be estimated by 2SLS. #Bookofwhy @PHuenermund

8.20.19 @2:13am - (Replying to @Chris_Auld) In Fig.5.10 (Causality p.153) Z is not exogenous, and X is not a valid IV in the traditional econ sense (exclusion is violated). So, I am not sure it can be solved using traditional econ methods, and would be eager to learn otherwise. #Bookofwhy

8.20.19 @1:08am - (Replying to @melvinwevers) I would replace "contaminate" with "stain". But "weed" connotes "undesirable" and "unintended" odd balls. I don't believe my fellow Democrats contemplated this kind of embarrassment, and most of them would rather see it disappear, if it didn't appear as Trump's victory.

8.20.19 @12:54am - (Replying to @thehuntinghouse) Thanks for granting my people right to a homeland, something Rashida can never do. But your use of the word "apartheid" proves how easy it is for decent people to fall victim to deceitful propaganda. Did you really fall for it?

8.20.19 @12:23am - (Replying to @pythiccoder) Thanks for the song. Phil Ochs was my favorite singer in the late 1960's, when I was part of the counterculture revolution. Yes, the fear of getting Blacklisted by Zionophobic big-mouths is what prevents my fellow Democrats from calling out Rashida's racism. Someone has to do it.

8.19.19 @11:42pm - (1/ ) Hating to insult or disappoint any of my followers, I was seriously considering your suggestion to create a new twitter handle. But one word you said made me change my mind: "uninformed". I have been reading, writing and researching the Middle East for the past 83 years. I was
8.19.19 @11:42pm - (2/ ) there when Azam Pasha declared (Oct. 11, 1947) "a war of extermination and momentous massacre" on a nation of refugees of which I was a son. And I was here at UCLA (2014) when BDS's Omar Barghouti re-denied my people right to self determination: A new
8.19.19 @11:42pm - (3/ ) twitter handle will give people of your persuasion the illusion that it is impossible for an "informed" person to disagree with their bubble of self righteousness and that "informed" people must be blind to the genocidal aims of BDS and its spokeswomen Rashida and Ilhan.
8.19.19 @11:42pm - (4/ ) I can't do it. I feel an obligation to truth and to history to let followers in your bubble know that "well informed" people exist who view them as gullible instruments in the service of a racist movement called BDS. Many of my colleagues feel same, but keep silent. I cant.

8.19.19 @8:22pm - (Replying to @StevePittelli) I would like very much to learn from a native English speaker what "time in history" these words harken, and why they seem insulting for some people. I chose them as carefully and as informedly as I could, given what I know about these two ladies stand for. Lifelong learning.

8.19.19 @7:42pm - (Replying to @Jacobb_Douglas @PHuenermund and 2 others) Can you elaborate your "how far" question? Please use several tweets and be pedantic about the references. By 3.2, do you mean Section 7.3.2? I cannot parse: "If Y=y and U=u, then X's PO=x." Please help

8.19.19 @6:49am - (Replying to @JonAMichaels) I honestly thought I was charitable, given that these two ladies have been dehumanizing a whole nation on a daily basis. Do you think their hatred is less than "sickly"? Is it controllable? Is it less dangerous if treated as "healthy"? Truly perplexed.

8.19.19 @4:08am - (Replying to @jmtroos) I tried to be charitable.

8.19.19 @3:32am - (Replying to @mom2phd) I am seriously worried about it. However, given that it can also be used for identifying patients "most in need" of a given treatment, I hope the net benefit to society will be positive. BTW, was the technique you mentioned based on counterfactual bounds? #Bookofwhy

8.19.19 @3:18am - (Replying to @BenWinegard @dabblingfrancis and @clairlemon) Interesting!. Do you have the source? Was I right about the correlation+plausibility combination?

8.19.19 @3:07am - The task of identifying individuals who are "susceptible to persuasion" (or "gullible"), has an enormous range of applications. Ang's slides tell you how it can be done using counterfactual logic. The technical paper, with proofs, is here: #bookofwhy

8.19.19 @1:28am - (1/ ) (Replying to @yudapearl and @Chris_Auld) 1/I did some further reading in Mann and Wald (MW) 1943, and I am fairly convinced now that their motivation was to facilitate identification of structural parameter and, not knowing S. Wright's, nor any other method, they identified the RF and tried to solve for the parameters.
8.19.19 @1:41am - (2/2) (Replying to @yudapearl and @Chris_Auld) This brings us to the question of whether today, that we know many other methods, should RF's be as revered as they were in the past? See what we can do today with other methods, and where RF will fail. Causality page 153. #Bookofwhy

8.19.19 @1:16am - Another new arrival, for those who do not speak Portuguese:

8.19.19 @1:01am - New arrival: My first Portuguese translation in paperback #Bookofwhy = O Livro Do Porque

8.19.19 @12:53am - Omar and Tlaib are already a "national scandal". Two morally deformed weeds in my party that have contaminated the US Congress with their sickly hatred of a certain country, and will continue to embarrass American democracy till someone (their voters?) say: Enough!

8.19.19 @12:31am - (Replying to @dabblingfrancis and @clairlemon) In the case of smoking, to be precise, it was a combination of correlations and "plausibility judgement", which is a type of causal assumptions. #Bookofwhy Chapter 5

8.18.19 @11:23pm - (Replying to @stephensenn @AlexJohnLondon and 3 others) Perhaps you can also tell us what functions were assumed for the arrows of Fig. 6.9(b), before you ran the simulation. How you made sure that the ellipses would be aligned as in Fig. 6.9(a) and, most importantly, what did you expect to learn from the simulation.#Bookofwhy

8.18.19 @8:16pm - A bit of history. The first BDS-style campaign started April 1, 1933, when the Nazi's boycotted Jewish businesses. It followed 3 years later, 1936, by the Palestinians, in their efforts to prevent European Jews from escaping - the most inhumane immigration policy in human history

8.18.19 @4:58pm - (Replying to @AlexJohnLondon @el_hult and 3 others) Don't miss Lord's Paradox, in its unmolested version. Just two plans, A and B, just two statisticians, just an innocent story, no ghosts, no red herrings. #Bookofwhy.

8.18.19 @3:11pm - (Replying to @stephensenn) The quoted passage has suffered the wrath of several misinterpretations, some suggesting a dining Hall serving several diets, and other complications. Fig. 6.9(b) disambiguate the data generation process. Were your two figures generated by this process? #Bookofwhy

8.18.19 @2:44pm - (Replying to @stephensenn @RonKenett and 2 others) If I have not provided explanation it must be that I do not understand those figures or, more specifically, how the data were generated, and whether both were generated in accordance with Fig. 6.9(b) -- the data generating model of Lord's Paradox. #Bookofwhy

8.18.19 @1:53pm - (Replying to @RonKenett @stephensenn and 2 others) Disagree! None of the versions separates Diet from Hall-of-Residence from Dining-Hall. Why complicate things? WHY? As summarized here: Lord Paradox is simple, and the decision between the two analysts is a simple exercise in causal analysis #Bookofwhy

8.18.19 @2:50am - (Replying to @broudsov and @rkarmani) Much of what was debated for centuries should be re-debated in our century, because AI brings to the table the first operationalization of ideas that were debated in the abstract. #Bookofwhy

8.18.19 @1:05am - (Replying to @broudsov and @rkarmani) I feel uncomfortable mixing reality "learning the cause" with fantasy: "rationalization". If you learn it, then it is there. The word "rationalization" weakens the necessity of such an understanding, as if there an alternative way of explaining observed regularities. #Bookofwhy

8.17.19 @10:26pm - (Replying to @heymanitshayden @stephensenn and 2 others) The Primer delves into the mathematics, and has many illustrating examples, thus empowering you to DO causal inference, as opposed to TALK ABOUT it. #Bookofwhy

8.17.19 @9:56pm - (Replying to @rkarman) My take: "Statistics is a torturous workaround until we learn the cause, then it begins to make sense."#Bookofwhy

8.17.19 @9:49pm - (Replying to @stephensenn @MadelynTheRose and @NP_Jewell) To read Primer is an irreversible decision. Like #Bookofwhy, it is going to be painful and, like #Bookofwhy, it is going to be transformative. Let me know when I can tweet our comrades in the trenches: "Stephen Senn has joined the revolution."

8.17.19 @4:44am - (Replying to @EngineerDiet @optempirics and @FatWhiteFamily) The words "mathematically convenient" are misleading when we compare designing digital circuits using Boolean algebra vs. equations of electrons and holes in semiconductors. The difference is between the doable and the undoable, and can only be appreciated by doing #Bookofwhy

8.17.19 @2:06am - I retweet this reply because I get many such inquiries: "Everything you can do with DAGs you can do with ...." The analogy is "Everything you can do with computers you can do with the theory of semiconductors, yet Boolean algebra helps, and so does programming language"#Bookofwhy

8.16.19 @6:26pm - (Replying to @optempirics and @FatWhiteFamily) I confirm your suspicion that everything you can represent in a DAG you can also represent in math equations. The former is an abstraction of the latter. Now, to appreciate the former, take any 3 variables in the latter and check if X is independent of Y given Z. #Bookofwhy

8.16.19 @12:55pm - (Replying to @mattshomepage) A quick scan warns me of something fishy here: all terms are probabilistic, I see no causal model, no causal assumptions, hence - no causal conclusions. Unless I am missing the key, ie. the input model, and then the question arises, why hide it?. #Bookofwhy

8.15.19 @2:57pm - (Replying to @Chris_Auld) Thanks so much Chris. I was about to post a query for the origin of the name. Now we can examine what Mann and Wald had in mind before the concept got distorted. #Bookofwhy.

8.15.19 @3:04am - (Replying to @stephensenn and @RonKenett) What's wrong with assuming " the diet being varied between Halls." or as #Bookofwhy says it: "the students eat in one of two dining halls with different diets". Each hall serves its own diet. What's so "WRONG" in assuming it, and moving forward to the paradox.

8.15.19 @2:34am - This wonderful road map on "good and bad controls" reminds me of a paper I wrote with S. Greenland on "Adjustments and their Consequences" Here the same issues are discussed in epi-vocabulary -- good for our dictionary #Bookofwhy

8.15.19 @2:18am - Replying to @stephensenn and @RonKenett) I am incapable of such offense. First, because terms such as "varying treatment within centers" is not in my vocabulary and, second, because I don't see such variations in Lord's story, nor in the "clean" data generation process of fig. 6.9. #Bookofwhy

8.15.19 @12:03am - (Replying to @stephensenn and @RonKenett) The adjustment equation is this: P(Y|do(Diet)) = SUM W_I P(Y|Diet,WI) P(WI) taken from, and telling us precisely how things are estimated. No weaknesses, no "two cases", no complications -- straight causal analysis and a paradox dissolved. #Bookofwhy

8.14.19 @11:47pm - (Replying to @jdramirezc) I am familiar with the examples cited. But here we are trying to understand the logic behind them. i.e., what information is ADDED when an economist says "This is a REDUCED FORM EQ." Or, What would we miss if we haven't heard him say it? Anything useful? #Bookofwhy

8.14.19 @11:33pm - (Replying to @jdramirezc) The supply-demand example was also in Haile's slides, but he insisted on all arguments being exogeneous. Relaxing this, turns RF into a regression equation - impossible. Our question: What would science miss if, suddenly, economists forget such a concept ever existed? #Bookofwhy

8.14.19 @10:41pm - (Replying to @jdramirezc) Your interpretation of Reduced Form is much broader than anything we heard here from economists, and would fit almost anything. Low and Maghir, likewise, are not providing a definition, they just talk around it. One day, we outsiders will get it too, I am sure. #Bookofwhy

8.14.19 @9:05pm - Continuing our efforts to improve communication with economists, Carlos and Andrew have posted: "A Crash Course in Good and Bad Control" -A road map guide for the perplexed traveler, novice and seasoned. Enjoy! #Bookofwhy

8.14.19 @8:03pm - (Replying to @jdramirezc) Thanks for the reference. We are exploring how economists think, so every such source is valuable. Oh, almost forgot, what does REF mean to you? #Bookofwhy

8.14.19 @6:04pm - (Replying to @Chris_Auld) Great!! So, if RF is useful "because they can be solved on a computer", so can many other equations, especially those that are OLS identifiable. Agree? And you and I (not sure about Imbens) can easily tell who those equations are. #Bookofwhy

8.14.19 @5:59pm - (Replying to @Chris_Auld) I do not view RF as arbitrary. On the contrary, it is very well defined, once you have a structural model. I am still exploring though how economists think about it, and why they chose (historically) to give those sort of equation a special name "REDUCED FORM". #Bookofwhy

8.14.19 @5:54pm - (Replying to @jdramirezc) Not really "simpler". Imagine a system with 100 exogenous variables, in which one structural equation has only three variables. The latter is "simple", "simpler than any reduced form one can write, yet we are fobidden from calling it "reduced form". #Bookofwhy

8.14.19 @5:05pm - (Replying to @stephensenn and @RonKenett) i) 'controlling for the initial weight' is achieved through the adjustment equation in my tweet. ii) fig. 6.9b would look the same if diets had been independently varied at the level of student. 6.9b assumes some students deviate (randomly) from the dictated diet. #Bookofwhy

8.14.19 @4:44pm - (Replying to @Chris_Auld) Are we in agreement then that identification considerations played a role in econometricians deciding which "solution" deserve the RF name and which do not? Else, why are we insisting on "exogeneity" of ALL arguments? #Bookofwhy

8.14.19 @3:57pm - (Replying to @Chris_Auld) This is my point. We have a "solution" to the system of equations and, yet, we DO NOT call this solution "reduced form". Why, because all observed variables in the RFE must be exogenous. I have no opinion of my own, just trying to understand how economists think about RFE.

8.14.19 @2:29pm - (Replying to @Chris_Auld) There are many solutions to the system of equations which were not baptized with a NAME; why RF? E.g., consider the front door: X-->Z--->Y, with X<-U->Y . Y=f(Z,U) is a solution, and so is Y=g(X,U). Behold, no name and no OLS. Further, why do u say "sometimes"? #Bookofwhy

8.14.19 @6:23am - (Replying to @stephensenn and @RonKenett) To reiterate. #Bookofwhy aims only to resolve the paradox under the assumptions stated in Fig. 6.9 (b). Here all assumptions are stated, because it is a data-generating process. Still, the paradox persists in our minds, then resolved., mission accomplished.

8.14.19 @5:38am - (Replying to @JadePinkSameera) Great! Please tell us if you find Lord's paradox to be paradoxical and if #Bookofwhy left anything undone in the way it explains away the paradox.

8.14.19 @5:28am - Foreign Affairs Magazine just published an excellent review of "Possible Minds" where the future of AI is discussed from 25 angles. (link: . My chapter can be viewed here (link: proposing model-based ML to overcome Deep-Learning limitations.

8.14.19 @5:21am - (Replying to @stephensenn and @RonKenett) #Bookofwhy may fail to do many things under the sun, but one thing it does not fail to do -- resolve a clash between the two intuitions, a clash that has baffled many analysts before, and still baffles them today, even under the "clean" data generation conditions of fig. 6.9(b)

8.14.19 @3:43am - (Replying to @RonKenett and @stephensenn) Different tasks do not make differences in conclusions. Accounting for "blocking structure" and "study design" and perhaps "quantum uncertainties" are unrelated the task of resolving a clash between two intuitions that persist EVEN in the "assumed causality links" #Bookofwhy

8.14.19 @2:05am - (Replying to @yudapearl @Chris_Auld and 7 others) Continuing our explorations of "Reduced Form Equations" (RFE) and what they mean to economists, I have tweeted this thread: (link: . I hope RFE experts approve of the way I explain it to my students. #Bookofwhy

8.14.19 @1:55am - (1/ ) Continuing our exploration of "Reduced Form Equations" (RFE) and what they mean to economists, let me address some hard questions that CI analysts frequently ask. Q1: Isn't a RFE just a regression equation? A1. Absolutely Not! A RFE carries causal information, a regression
8.14.19 @1:55am - (2/ ) equation does not. Q2: Isn't a RFE just a structural equation? A1. No! Although a RFE carries causal information (much like a structural equation) the RFE may not appear as such in the structural model; it is derived from many such equations though functional composition.
8.1 .19 @1:55am - (3/ ) (The output-instrument in the IV setting is a typical example). Q3: One may derive many equations from a structural model; what makes a RFE so special to deserve its own name. A3: It is exceptional because it comes with a license of identification by OLS. This is not usually
8.1 .19 @1:55am - (4/ ) the case for other derivable equations, say those relating two endogenous variables. Q4: But some of those other equations ARE identified by OLS; why haven't they been baptized with a name? A4:Because traditionally, economists did not have an easy way of telling which equation
8.14.19 @1:55am - (5/5) enjoys identification by OLS and which does not. Now they do, so it is quite likely that next generation econ. texts will introduce new names. For example, every structural equation identifiable by OLS should be so recognized. #Bookofwhy @PHuenermund @causalinf @analisereal

8.13.19 @11:02pm - (Replying to @EpiEllie) Intrigued by your "hotter take". Are descriptive analysts using the term "adjustment" in lieu of "stratification"? How would you explain to them what the difference is? #Bookofwhy

8.13.19 @10:45pm - (Replying to @RonKenett and @stephensenn) Why assume that there are "differences" between us? I am trying to explain a clash between two strongly held intuitions (no building systems, no automation) and Senn is trying to do something else (no intuitions no causal assumptions). I do not see "differences" #Bookofwhy

8.13.19 @9:33pm - (Replying to @Prof_Livengood and @OUPPhilosophy) John to painter: "Paint the wall either green or purple". Painter to John: "Your wife will get pretty angry." John was imperative, Painter was indicative. Both used a disjunctive do-operator, as interpreted here: (link: #Bookofwhy

8.13.19 @5:05pm - (Replying to @Prof_Livengood and @OUPPhilosophy) I will be happy to have a long conversation and to learn what the distinction is between "imperative" vs. 'indicative" in the context of the do-operator. Do they make different claims? #Bookofwhy

8.13.19 @3:53pm - My latest offering on "Lord's Paradox and the Power of Causal Thinking" with a few goodies for the curious and empowered (link: #Bookofwhy

8.13.19 @8:48am - (Replying to @NikosTzagarakis and @intoolab) Glad another person beside me sees how important vocabulary is to thinking straight.

8.13.19 @8:39am - (Replying to @BruceTedesco) There is something addictive in the word "Baysianism" (as in the word "induction"), perhaps because it seems to capture so many cognitive functions that blinds out from seeing what it can't. #Bookofwhy

8.13.19 @2:31am - (Replying to @imleslahdin and @spyrosmakrid) I commented on this paper a few days ago, and so did other ML folks.

8.13.19 @2:05am - Very interesting post. I wonder what it would take for a philosopher to read (link: and change the title of Section 2 to read: "What a Great Many Phenomena Bayesian Decision Theory CANNOT Model" #Bookofwhy @OUPPhilosophy

8.12.19 @4:06pm - (Replying to @AndrewPGrieve) When I see a quote I dont remember, I begin to realize how old I must be. 105 !

8.12.19 @1:39pm - (Replying to @yudapearl @Chris_Auld and 7 others) Do experts on "reduced-form" approve of the licenses I am attributing to Y=f(X,Z,U) ? If so, we can continue with what f says about the world, or perhaps with more licenses. #Bookofwhy

8.11.19 @9:46pm - (Replying to @analisereal @autoregress and 7 others) I bet there is some added value to saying "reduced forms", perhaps a pedagogical value, but, since I've never used it, its up to those who do to explain: "what is added?" Perhaps it is just a verbal assurance that someone already taken care of exogeneity. Hard to guess #Bookofwhy

8.11.19 @9:07pm - (Replying to @Chris_Auld @autoregress and 6 others) Sorry, I thought you said that (quoting): "equation 3.46 in Causality is a ratio of two reduced forms", so I looked at 3.46 and found there a ratio of two cond. expectations. I will try to be less pedantic. (Unless you would you like me to be more pedantic?) #Bookofwhy

8.11.19 @8:55pm - (Replying to @yudapearl @Chris_Auld and 7 others) 2/ infer: E[Y|do(X)] = E[Y|X], E[Y|do(Z)] = E[Y|Z], and E[Y|do(X),do(Z)] = E[Y|X,Z], regardless of the form of f. License #2. Y=f(X,Z,U) is RFE then so is W=g(X,Z,U) for some g, where W is ANY variable not in {X,Z}. Hence, E[W|do(X)] = E[W|X]. Will continue if allowed. #Bookofwhy

8.11.19 @8:39pm - (Replying to @yudapearl @Chris_Auld and 7 others) 1/ What is a Reduced Form Equation (RFE) - a layman interpretation. Introduction: an RFE is a causal, not statistical statement. It makes causal claims about the world, and provides licenses for causal inference. License #1- If Y=f(X,Z,U) is RFE, then thou have the license to

8.11.19 @8:13pm - (Replying to @Chris_Auld @autoregress and 6 others) Sorry, but Eq. 3.46 in Causality reads: b = E[Y|z]/E[X|z] It is a ratio of two conditional expectations, not "two reduced forms". The former are descriptive, the latters are causal -- like water and oil. IOW, E[Y|x] conveys no causal information, RF's do! Can't equate.#Bookofwhy

8.11.19 @7:19pm - (1/ ) (Replying to @Chris_Auld @autoregress and 6 others) I am the last one to claim that a concept is not "useful" its usefulness however needs to be clear, else it turns into "abusefulness." I am happy to finally hear from someone what it means to that someone, not to another person, or to some textbook, and I will try to refine it,
8.11.19 @7:33pm - (Replying to @yudapearl @Chris_Auld and 7 others) so as to explain it to my students, for whom "meaning" comes either as a claim about the world, or as a license to conclude something about the world. I will try translating your "meaning" of RF to their language, starting with the licences. Pls check for promiscuity #Bookofwhy.

8.11.19 @6:01pm - (Replying to @Chris_Auld @autoregress and 6 others) So, @Chris_Auld perhaps you can tell us in plain language what the sentence "Y=f(X,Z,U) is a reduced form equation" means to YOU, so I can tell my students how to communicate with economists. Let's forget MHE and ancient history. What does the label RF add? #Bookofwhy

8.11.19 @5:45pm - (Replying to @steventberry @autoregress and 5 others) So, if we agree that economists today are not clear on what is meant by RF equation, perhaps we should give them a few hints on how to explain RF when communicating with DAG-minded folks. Right now, they dont really know what the explanation should entail. Ready? #Bookofwhy

8.11.19 @3:52pm - (Replying to @steventberry @MariaGlymour and 5 others) We shouldn't argue about ancient history. I am trying to inform my students and other DAG-minded researchers how to communicate with economists, i.e., when an economist tells you "This is a reduced form equation" what does he/she mean? What does the equation claim etc.#Bookofwhy

8.11.19 @3:05pm - (Replying to @omaclaren @fuzzydunlop123 and 7 others) When my grandson insisted on understanding what makes a water molecule accelerate when it goes through a constriction, I said: its neighbors from behind are pushing it stronger than the neighbors in front, they are closer, hence more pressing. #Bookofwhy

8.11.19 @2:33pm - (Replying to @lewbel) I am not aware of the notions of "incompleteness" or "incoherence". Is there a simple example for the uninitiated?

8.11.19 @2:12pm - (Replying to @onnlucky @dchackethal and 2 others) I am convinced that, if Popper was alive, he would buy the Ladder of Causation from #Bookofwhy, after realizing that "induction" is too general a term to be operationalized.

8.11.19 @3:36am - (1/3) Econ. readers asked if they can get hold of those magical night-vision goggles that tell us which causal effects in an econ. model are identifiable by OLS (and how). The answer is embarrassingly simple: Consider Model 2 in p. 163 of (link:
8.11.19 @3:36am - (2/3) Take any two variables, say Z_3 and Y. If you can find a set S of observed variables (e.g., S={W_1,W_2}), non-descendants of Z_3, that block all back doors paths from Z_3 to Y, you are done; the coefficient a in the regression Y = a*Z_3 + b1*W_1 +b2*W_2 gives you the right
8.11.19 @3:36am - (3/3) answer. A similar goggle works for a single structural parameter (See page 84 of Primer (link: Duck soup. This re-begs the question whether the restricted notion of "reduced form" is still needed in 2019. @EconBookClub @lewbel #Bookofwhy.

8.11.19 @2:25am - (Replying to @mnjp) Glad you took the firing squad seriously. Some people think it is just a "toy problem" and never learn what counterfactuals mean. #Bookofwhy

8.11.19 @2:20am - (Replying to @Jabaluck @fuzzydunlop123 and 6 others) Adding my interpretation of Haile's interpretation of "reduced form". I hope I am right, or at least faithful:

8.10.19 @6:09pm - (Replying to @omaclaren @fuzzydunlop123 and 7 others) You are confirming my observation that people tend to cite the name of the conservation law as an explanation of observed phenomena. Yet naive me always wanted to know: why would a fluid particle accelerate when getting into the narrow part of the tube. Now I know why. #Bookofwhy

8.10.19 @5:25pm - (Replying to @omaclaren @fuzzydunlop123 and 7 others) How would Dowe "explain" Venturi's effect (pressure measured in a fluid flowing in a pipe of varying cross section).?

8.10.19 @5:07pm - (Replying to @fuzzydunlop123 @omaclaren and 7 others) Conservation laws are non-causal emergent properties of dynamic systems. The latters are causal, because each particle is "responding" to the forces in its neighborhood. Interestingly, people tend to accept the name of the conservation law as an explanation of events #Bookofwhy

8.10.19 @4:39pm - (1/ ) (Replying to @MariaGlymour @Jabaluck and 5 others) 1/ The translation of Phil's definition to DAG language is simple: Y=f(X,Z,U) is "reduced form" iff X, Z and U are ALL the root nodes that are ancestors of Y, both observed and unobserved. The question remains why would an analyst write down such an equation when all its claims
8.10.19 @4:44pm - (2/ ) (Replying to @yudapearl @MariaGlymour and 6 others) are obvious from the graph? For example, the claim that the causal effect of X (and Z) on Y can be estimated by OLS. The answer I believe has to do with the fact that the notion of "reduced form" emerged in a period when economists where desperate for conditions to justify
8.10.19 @4:51pm - (3/3) (Replying to @yudapearl @MariaGlymour and 6 others) identification by OLS. Today they are fortunate to have night-vision goggles that tell them immediately which effects are estimable by OLS, and which regressors to include/exclude (link: This was not always the case in the history of economics. #Bookofwhy

8.10.19 @6:48am - (Replying to @mathtick) I dont recall commenting on it. Will add to my "to do" pile.

8.10.19 @6:44am - (Replying to @tarinziyaee) No changes except for the many errata shown here: (link: (link: (link: and marked in red. The new edition will incorporate them in a clean text.

8.10.19 @2:13am - (Replying to @RonKenett and @ShalitUri) I do not recommend going in Hastie's direction (no pun) of starting with black-box and then asking: when would some statistical estimate have a "causal interpretation". #Bookofwhy goes the other way: start with what you know and ask if you can estimate what you want to know.

8.10.19 @2:07am - Correcting a link in a previous post. The new errata for the Causal Inference Primer book can be accessed here: (link: (marked in red). #Bookofwhy

8.9.19 @10:57pm - (Replying to @fuzzydunlop123 @lewbel and 6 others) Here we are, trying to out-guess each other what Haile meant by "reduced form", and MHE, and so and so... Something is terribly wrong in 2019 science if we can't find a red-blooded economist willing to take a stand and say: TO ME, "reduced form" means... How about it? #Bookofwhy

8.9.19 @10:37pm - (Replying to @fuzzydunlop123 @IvanWerning and 6 others) I agree here, "model" cannot replace "structural model," because statisticians think that the assumption of normality is a "model" and some MHE students think that a "regression equation" is a "model". The word "structural" tells us what assumptions it carries. #Bookofwhy

8.9.19 @10:18pm - (Replying to @lewbel @steventberry and 5 others) I have no doubt that "reduced form" exists or can be concocted. But the question is: What useful information does it provide to @lewbel (not MHE). Put differently, what would science miss if, suddenly, all economists get amnesia and forget such a concept ever existed? #Bookofwhy

8.9.19 @9:29pm - And while we are speaking of Causal Inference - Primer, (link:, good news comes from Wiley: They are preparing a revised edition to hit the shelves soon. It will include new errata collected and composed by Scott Meuller. See (link: #Bookofwhy

8.9.19 @8:19pm - (Replying to @thanhnguyentang) You can start with #Bookofwhy for fun,history and philosophy, but if you want to delve straight to the mathematics, and still have fun, there is no better introduction than Primer (link:

8.9.19 @8:01pm - (Replying to @IvanWerning @Jabaluck and 5 others) @IvanWening, the more I read your Tweets the more I agree with you (r u sure u r an economist?). The rearrangements you are permitting are permissible as long as they preserve information. Algebraic rearrangements are NOT information preserving. Nor are "Reduced forms" #Bookofwhy

8.9.19 @7:41pm - (Replying to @lewbel @steventberry and 5 others) "Reduced form" vs. "Confused form": Can anyone define what it means to him/her (not to Phil or MHE) and why it should not be purged from econometric discourse w/o loss of information? #Bookofwhy

8.9.19 @7:33pm - (Replying to @lewbel @steventberry and 5 others) I must be missing something important. If the "reduced form" is descriptive, how can it impose untestable assumptions? More basic: Does the "reduce form" preserve the structural assumptions carried by the structural model of which it is a "reduced form"? #Bookofwhy

8.9.19 @7:16pm - (Replying to @IvanWerning @Jabaluck and 5 others) Totally agree that "reduced form" is a source of much confusion. It gives the impression that it is merely a syntactic transformation of "form" with no loss of information. But if it is "descriptive", then we loose all the causal information that SEM provides. Shun! #Bookofwhy

8.9.19 @4:55pm - Machine Learning enthusiasts will be interested in George Lawton's new post titled "Causal Deep Learning Teaches AI to ask why" (link: I am not familiar with all the actors mentioned in the story, but I am glad ML is moving beyond curve fitting #Bookofwhy

8.9.19 @6:19am - (Replying to @Jacobb_Douglas @maximananyev and 6 others) The draft lottery, the price of beans in China, the court decision (see @Bookofwhy Chapter 1), etc. its a variable that only sick/creative stretch of imagination would deem it relevant to the context of interest.

8.9.19 @6:19am - (Replying to @Jacobb_Douglas @maximananyev and 6 others) The draft lottery, the price of beans in China, the court decision (see @Bookofwhy Chapter 1), etc. its a variable that only sick/creative stretch of imagination would deem it relevant to the context of interest.

8.9.19 @4:39am - (Replying to @sarahmrose @SHamiltonian and 2 others) As I said in my confession (link: I spend time on such debates knowing that there are dozens of silent and bright students out there, listening to the conversation, and gathering ammunition for future defense of commonsense. #Bookofwhy

8.9.19 @1:14am - (Replying to @fuzzydunlop123 @maximananyev and 6 others) One should add though that this is not done with the intention to "fool." When you don't SEE your assumptions you tend to believe that your "experiment" takes care of them. The Talmud says: If our eye were given the power to see our assumptions, we won't get out of bed #Bookofwhy

8.9.19 @1:04am - (Replying to @paulgp @maximananyev and 4 others) But, in addition to saving those answers from oblivion, the DAG also permits you to COMBINE them, and thus answer questions that DAG-averse folks CAN'T. E.g.,"Is the treatment ignorable?" or "Are there any testable implications". That's why these folks rarely ask them. #Bookofwhy

8.9.19 @12:10am - (Replying to @maximananyev @Jabaluck and 5 others) A DAG is none other but a collection of answers to the question: "What is the source of variation in variable X?" recorded as an arrow into X. What is often called "The DAG Approach" is consulting those answers, instead of re-asking when they're needed in estimation. #Bookofwhy

8.8.19 @11:44pm - A question on Quora read: ML is becoming too competitive! Should a person wishing to become a ML researcher give up? As an aspiring ML researcher, I had to tell them the truth:

8.8.19 @7:54pm - (Replying to @DonskerClass @Jabaluck and 5 others) This is one way of handling potential misspecifications, (good paper) another is sensitivity analysis (if done correctly) which passes the burden on to a Plausibility judgment on the likelihood that certain edges can attain a certain strength. #Bookofwhy

8.8.19 @6:54pm - (Replying to @causalinf @steventberry and 3 others) Lewbel's taxonomy notwithstanding, "Design" cannot replace "identification". "Design" connotes an option an analyst may or may not choose, "identification" is a property of your model; a causal query is either identifiable or not (given the model) no matter what you do.#Bookofwhy

8.8.19 @4:41pm - (Replying to @causalinf @steventberry and 2 others) @causalinf, Are you really going to replace "identification" with "design"???. Clarity with escapism?? "Identification" is perhaps the clearest notion developed by economists --"Design" is the foggiest . @Undercoverhist, as historian, watch the last days of clarity.#Bookofwhy

8.8.19 @4:17pm - (Replying to @fuzzydunlop123 @Jabaluck and 5 others) If someone succeeded in eliminating the word "design" from the literature (perhaps leave it in RCT contexts) clarity will shine brightly on causality-land. A catch-all for "I wish to be more specific". I hope it's avoided in #Bookofwhy. Just checked, it's used mostly harmlessly.

8.8.19 @2:38pm - (Replying to @Jabaluck @Jacobb_Douglas and 2 others) Please do not put words in my mouth. Conditional ignorabitlity assumptions are inscrutable because they are far removed from what we know, as proven by (1) PO textbooks and (2) PO folks refusing the test: Given a story (no DAG), tell me if "Y_x is ind. of X given Z" #Bookofwhy

8.8.19 @2:24pm - (Replying to @fuzzydunlop123 @Jabaluck and 6 others) I can't tell if the Yale folks stand behind the slides, nor can I tell whether the slides reflect outdated terminology or advocated terminology. Its healthy to see them going through some sort of soul-searching catharsis. Let's summarize what we learned by email

8.8.19 @1:55pm - (Replying to @fuzzydunlop123 @Jabaluck and 5 others) @fuzzydunlop123, I cant follow the barrage of tweets in the wake of Phil Haile slides, but you make consistent sense. Bringing up (link: (link: was timely -- we haven't seen a drastic improvement in Econ. texts yet. Is it coming? #Bookofwhy

8.8.19 @11:42am - (Replying to @steventberry @Jabaluck and 4 others) Very interesting and helpful. Thanks for posting. Is this set of slides representative of the current thinking at Yale-Economics ? Or are they still undergoing internal debate regarding the precise meanings of terms? #Bookofwhy

8.8.19 @6:05am - (Replying to @georgemsavva @stephensenn and @statsepi) But if Hall is perfectly correlated with Diet, why won't the effect of Diet on Gain not be the same as the effect of Hall on Gain. Once we decide what effect we need, the diagram delivers it for you, since we do not have unobserved confounders in the model. #Bookofwhy

8.8.19 @5:37am - (Replying to @_amirrahnama) I agree with your general statement above, though I have not read the paper by @tmiller_unimelb . I would be delighted to find out that causality and explanation are well understood in AI. #Bookofwhy

8.8.19 @3:01am - (Replying to @georgemsavva @stephensenn and @statsepi @georgemsavva, you hit it on the nail, thanks for making it explicit, and for stressing that #Bookofwhy deals with Lord's paradox, not with experimental design. Moreover, if we suspect Hall and Diet have separate effects on Weight, another diagram would resolve it just as well.

8.8.19 @2:39am - For those concerned with issues of Free Press and East-West relations, a Panel on Aug 15 in LA would be most illuminating. See (link: RSVP required.

8.8.19 @2:17am - (Replying to @stephensenn and @statsepi) You have to trust my honesty if I say: I don't have the slightest idea how this is related to Lord's dilemma with TWO HALLS, each serving ONE diet, and students are SHOWN what diet is served in each. Please start here, and tell us what factor is neglected in #Bookofwhy

8.8.19 @2:08am - (Replying to @statsepi and @stephensenn) Great news indeed. Humble advice: Try to focus on why YOU (not Nedler) are surprised by Lord's dilemma, and see for yourself whether #Bookofwhy does not pacify your surprise to the point where no such story can ever surprise you again.

8.8.19 @2:01am - (Replying to @stephensenn and @RonKenett) "reasonable under common circumstances " is insufficient. Resolution comes from understanding why two analyses, which were reasonable, even compelling under "common circumstances" suddenly cease to be reasonable under the current circumstances. Watch it explicated in #Bookofwhy

8.8.19 @1:50am - (Replying to @zaffama) You just gave a clear example. In nature, the sun's motion determines its angle, generating sun rays that are reflected from the atmosphere, sensed by the rooster, and make him crow. The sun's motion also determine the sun's angle next hour, which we call "Sunrise". #Bookofwhy

8.8.19 @1:35am - (Replying to @zaffama) My point was to show that the direction of causal effect can be opposite to the times of which events are observed or reported, but (God forbids!) not opposite to time's in the data-generating process. #Bookofwhy

8.8.19 @1:28am - (Replying to @yudapearl @RonKenett and @stephensenn) To be consistent with my request, I believe I did an honest job in confessing my "surprises" in my 3rd comment on Senn's post. Here: (link: . And note my conclusion: Yes, causal analysis does dissolve the clash of intuitions in Lord's paradox. #Bookofwhy

8.8.19 @1:11am - (Replying to @RonKenett and @stephensenn) Another comment: I find it hard to discuss this topic without knowing where my discussant stands on the issue of "surprise", or whether he/she is at all surprised by Lord's story. So, I beg you, and future discussants to start with an "analysis of surprise" #Bookofwhy

8.8.19 @12:58am - (Replying to @RonKenett and @stephensenn) Never mind engineering challenges and robotics. The #Bookofwhy chapter on Lord's paradox is an exercise in scientific psychology: Why are scientists surprised and vexed by the story? Can causal logic dissolve this surprise? The rest is irrelevant to our discussion. #Bookofwhy

8.8.19 @12:45am - (Replying to @DanielNevo @Jacobb_Douglas and @PHuenermund) The differences between the (in)dependencies of Y(a) and Y(not-a) do not show in the structure of the SWIG, this information comes from outside the graphical model. #Bookofwhy

8.8.19 @12:38am - (Replying to @PHuenermund @MariaGlymour and @Jacobb_Douglas) OK, Yielding to Maria and Paul. But please try to make the context clear to an engineer (like me), who can only think in terms of (1) what we know, (2) what we want to know and (3)what data we have available. Input-Output. #Bookofwhy

8.7.19 @11:18pm - (Replying to @DanielNevo @Jacobb_Douglas and @PHuenermund) Agree on "alleviated". But you do not need SWIGs for that task, an ordinary DAG can tell you all about d-separation and back-door tells you all about ignorability and, if you really need to, it shows it to you explicitly, see Causality p.343, (link:, #Bookofwhy

8.7.19 @7:06pm - (Replying to @Jacobb_Douglas and @PHuenermund) Best place to ask is by email. A "typical case" for PO is "we assume, as usual, that D is conditionally ignorable given X". A "typical case" in science is: "Given a story on 3-4 variables, lets see if measuring X would help us get the effect of D on Y . #Bookofwhy

8.7.19 @4:53pm - (Replying to @roieki and @SamHarrisOrg) Eize Perek?

8.7.19 @2:01pm - (1/2) (Replying to @PHuenermund and @Jacobb_Douglas) That is why the answer to the question "Can it be done in PO ?" is plain NO. If we say YES, the impression is created that it is just a matter of a few more computations. It is hard for people to appreciate the difference between tractable and intractable. And that is why PO
8.7.19 @2:07pm - (2/2) (Replying to @yudapearl @PHuenermund and @Jacobb_Douglas) PO folks dread toy problems, where the intractability shows immediately. It is like "solving" a polynomial assuming that someone else already computed its roots. #Bookofwhy

8.7.19 @4:38am - (1/4) I would like to welcome the 500 new followers who have joined us on Tweeter since @SamHarrisOrg posted our conversation on his podcast. Welcome to the wonderful land of WHY and, please, be aware of what you are getting yourself into. Our main theme is the #Bookofwhy and
8.7.19 @4:38am - (2/4) the way it attempts to democratize the science of cause and effect and apply it in artificial intelligence, philosophy, and the health and social sciences, see (link: We alert each other to new advances in causal reasoning and new methods of answering causal
8.7.19 @4:38am - (3/4) questions when all we have are data, assumptions and the logic of causation. We also debate detractors and nitpickers who mistrust fire descending from Mt. Olympus for use by ordinary mortals. I spend time on such debates knowing that for every nitpicker there are dozens of
8.7.19 @4:38am - (4/4) silent and bright students out there, listening to the conversation, and gathering ammunition for future defense of commonsense.
Overall, I hope you find this forum entertaining, challenging, and idea driven.

8.6.19 @10:49pm - Thank you @SamHarrisOrg for having me on your podcast and for our lovely discussion on cause and effect, counterfactuals, free will, and the future of AI. I believe it was your podcast that caused the # of followers on this Tweeter to cross the 20K mark. They'r welcome!#Bookofwhy

8.6.19 @5:13am - (Replying to @stephensenn) The answer is simple: The #Bookofwhy produces a consistent answer to problems defined in the #Bookofwhy, not to problems defined elsewhere, involving several dining halls, or dining halls shifting their diets, or other variations. One thing at a time. The idea of "surgery estimators" is ingenious, it would not occur to me that you can get extra mileage on top of "pruned estimators". However, is Figure 2(b) the simplest example to show this extra mileage?? Would c-equivalence help here? (link: #Bookofwhy

8.5.19 @11:18pm - (Replying to @jaketapper and @RashidaTlaib) @Jaketapper is 100% right, Congresswoman @RashidaTlaib lies. She knows that Palestinians want one more thing, in addition to posing as human right advocates -- dismantling their neighbors. If I am wrong, let us say so publically. She can't!

8.5.19 @7:14am - (Replying to @doinkboy @GRich_Cinci and 2 others) This is what causal inference is all about: "interpret it causal, given a set of assumptions (i.e., a causal model)." Except that the "interpretation" is no longer whimsical, as it used to be, it must obey the logic of causation. And this is what makes it "causal" #Bookofwhy Their numbers is still rising, I hope one of you continues --an amazing phenomenon.

8.5.19 @4:16am - I have just posted my 3rd comment on Lord Paradox (link: Here, I returned to my original goal of empowering readers with an understanding of the origin of the paradox and what we can learn from it. Red herrings have been taking too much of our time. #Bookofwhy
8.5.19 @4:16am - It is important to add that the huge literature on Simpson's and Lord's paradoxes attests to a century of scientific frustration with a simple causal problem, deeply entrenched in out intuition, yet helplessly begging for a formal language to get resolved. I can count dozens of
8.5.19 @4:16am - red herrings thrown at this stubborn problem, to deflect attention from its obvious resolution. Why? Because the latter requires the acceptance of a new language - the hardest transition for adults. I am grateful for the opportunity to talk to thousands of young followers here
8.5.19 @4:16am - on Twitter and convey to them my honest conviction: Chapter 6 of #Bookofwhy (paradoxes galore) is a condense summary of a century of confusion, and a powerful recipe for deliverance from that confusion. Additional references are: (link: (Lord's paradox),
8.5.19 @4:16am - (Simpson's paradox) (link: (Sure-thing Principle) And Chapter 6 of Causality (link: , where I document dozens of red-herrings 1900-2000.

8.5.19 @1:16am - Replying to @mendel_random and @oacarah) The point is: "discovered using a causal model, later depicted as diagrams". Red herrings do not stop us from discussing an issue. OK. We get your point: DAGs dont handle cycles and calculus does't handle non-differentiable functions. Can we get on with the discussion? #Bookofwhy

8.4.19 @7:22pm - (Replying to @mendel_random) I dont think it was worked out in folks' heads too well, they had to use some mathematics, and the mathematics they used came as "equations" which, at the time looked like algebraic but, in time, came to be recognized as non-algebraic, asymmetric, or "structural", #Bookofwhy

8.4.19 @3:13pm - (Replying to @stephensenn) Oh, @stephensenn, you never told us what's surprising you in Lord's story. In other words, must we go to finite sample before you can describe to us the reason for your surprise? #Bookofwhy

8.4.19 @3:09pm - (Replying to @stephensenn) Not wrong. Just irrelevant to Lord's paradox where we assume (using Wainer's model) that students are allowed to choose their own Hall and that each Hall serves ONE diet. If I am surprised with this simple story, I'll try to resolve it HERE, before complicating it. #Bookofwhy

8.4.19 @2:25pm - (Replying to @stephensenn) We differ here. The notion of "probability distribution function" is not meaningless is statistics. Most stat texts start with distributions, then go to "samples from distributions". Lord introduced his paradox at that level, and managed to surprise us; why quit? #Bookofwhy

8.4.19 @2:09pm - Now, speaking specifically about Lord's paradox, the paradox was introduced to us in "asymptotic" terms (ie, using distributions, not samples) and we were surprised. Is it likely that we can resolve our surprise by going to finite samples? or to "block design"? #Bookofwhy

8.4.19 @1:54pm - That is why I am begging folks: "Please, do not tell me 'I am not entirely satisfied' before you tell me why you are surprised (by the paradox) ". I am proud that #Bookofwhy addresses this question (of "surprise") head on, before offering "a resolution".

8.4.19 @1:36pm - (Replying to @mendel_random) No big deal. Replace "using Wright's DAGs" with "using Wright's equations, later depicted as DAGs" and the rest of the Tweet follows, especially "using what you know". The Tweeter discussion was about "using a model", which some folks shun. #Bookofwhy

8.4.19 @1:09pm - Any discussion of "paradoxes" is really an exercise in psychology. Yet we, quantitative analysts, are trying to avoid psychology at all cost. We can't. We must explicate why two strong intuitions seem to clash, and the conditions under which our intuitions fail. See #Bookofwhy

8.4.19 @6:06am - (Replying to @f2harrell @stephensenn and @Lester_Domes) I will expand, in a day or two. But it would help if you reconstruct Lord's paradox in your own way and pin point: What was paradoxical in the story? What was surprising there that deserved the word "paradox"? #Bookofwhy

8.3.19 @7:06pm - (Replying to @EpiEllie) I tried to look into it, but I am missing your research question, ie, the query.

8.3.19 @6:43pm - Good news for missing-data analysts. Karthika Mohan @karthica is joining the Editorial Board of Journal of Causal Inference jttps:// It's a welcoming invitation for articles on modern ways of recovering what we thought to be missing. #Bookofwhy

8.3.19 @4:42pm - I've just posted a comment (link: on S. Senn's n-th attempt to deconstruct Lord's paradox. It ends: "I hope we can now enjoy the power of causal analysis to resolve a paradox that generations of statisticians have found intriguing, if not vexing." #Bookofwhy

8.2.19 @9:25pm - Coming from my fellow statisticians it reminds me of King Solomon's saying: "Let a stranger appraise your work, not your mouth"(Proverbs 27:2). And the Mishna saying: "The baker does not judge his own bread" (Tosefta Yom Tov 3:7). I hope my bread tastes well to others. #Bookofwhy

8.2.19 @9:25pm - Many thanks! Carlos Cinelli came back from JSM-2019 and brought me a gift that I would cherish dearly:

8.2.19 @8:03pm - (Replying to @jasonintrator @glogauer_jakob and @TheIHRA) A group HAS a self. What makes an individual have self is memory, telling him: your experience yesterday informs your decision today. Same with groups, except here the pertinent experience spans centuries, and will easily get lost unless members attain self-determination

8.2.19 @7:26pm - (Replying to @jos_b_mahoney @jasonintrator and @glogauer_jakob) Unlike Jason, I would feel very very lonely, to know that I do not have Canaanite role models who spoke my language, who left my kids beautiful legends and who encoded their experience in a culture baked over 3,000 years. Woefully lonesome!

8.2.19 @5:15pm - (Replying to @jos_b_mahoney @jasonintrator and @glogauer_jakob) Charging people with anti-semitism is stone-age. Edward Said is guilty of a worse offense: Zionophobia, denying Jews the right to define their collective identity after writing volumes on "Orientalism"-the right of Arabs to define themselves. I'm sensitive to logical consistency.

8.2.19 @5:05pm - (Replying to @jasonintrator and @glogauer_jakob) The HE in that prayer (Birkat Hamazon) is God himself, called Ha'Rachaman (the merciful one). True, it was not about "now", but it reflects the continuous aspiration of the Jewish people to eventually regain sovereignty in their historical birth place, i.e., Zionism

8.2.19 @4:54pm - (Replying to @jasonintrator and @glogauer_jakob) Universal liberalism is central indeed to Jewish life. But how: "Thou shall not oppress the stranger, because YOU know what it means, YOU were once a stranger..." Namely, if you forget YOUR collective memory, gone is your universal liberalism. Can't have one without the other.

8.2.19 @4:44pm - (Replying to @jasonintrator and @glogauer_jakob) Their rejectionism started in 1920's, there was no loss of homes or political rights. Not even FEAR of losing homes or rights, as is documented in the Arab newspaper Carmel See "Early Zionists and Arabs"

8.2.19 @4:36pm - (Replying to @jasonintrator and @glogauer_jakob) Herzl, Jews in the diaspora have been divided about the role of Israel as a practical solution to the immediate problems Jews faced then. That role was never debated in my (and your?) grandfather house, where they prayed 3 times a day: "He will walk us in sovereignty to our land"

8.2.19 @4:27pm - (Replying to @jasonintrator and @glogauer_jakob) Your friends from JVP are dangerous indeed, not because they wish us death, but because they recklessly blind themselves to our death, ostensibly in pursuit "universal liberalism." I call them "Jews of Discomfort" here

8.2.19 @4:00pm - (Replying to @jasonintrator and @glogauer_jakob) Yes, I work on this. What is it that I highlighted "over" other facts? Isn't Palestinian rejectionism a fact that cannot be "over highlighted" when it comes to prospects for peace, lifting the occupation, and almost every issue youngsters care about. Do you talk to them about it?

8.2.19 @3:49pm - (Replying to @jasonintrator and @glogauer_jakob) I would never accuse JVP of being motivated by antisemitism; they are motivated by more dangerous forces which I have outlined here:"Our New Maranos"

8.2.19 @2:54pm - (Replying to @jasonintrator and @glogauer_jakob) What is it about the way I represent Palestinians that you think is (1) not factual and (2) not effective with younger American Jews? I hope you are not suggesting young people are not moved by facts? These youngsters were: (link: What about those you meet?

8.2.19 @1:45pm - (Replying to @glogauer_jakob) Truth has its secret way of prevailing, despite BDS's key slogan: "if you repeat a lie long enough, people will fall for it."

8.2.19 @1:32pm - (Replying to @AdanZBecerra1) Agree. And my reason: If we introduce regression before DAGs, students are likely to get trapped in the "regressional confusion" of the 20th century, unable to distinguish structural from regression equations. (link: #Bookofwhy

8.2.19 @5:29am - (Replying to @GivingTools and @causalinf) In my efforts to make causal diagrams palatable to economists I am trying build as much as possible on identification strategies devised by empirical economists. Can you point us to an author or two who came close to outlining diagramatic ideas informally? Thanks. #Bookofwhy

8.1.19 @9:13pm - (1/ ) I am recommending this paper to every data analyst educated by traditional textbooks, which start with regression equations, add and delete regressors, estimate and compare coefficients before and after deletion, and then ask which coefficient has "causal interpretation"
8.1.19 @9:13pm - (2/ ) I was shocked to realize that the majority of data analysts today are products of this culture, trapped in endless confusion, with little chance to snap out of it, since journal editors, reviewers and hiring committees are trapped in the same culture. The new PO framework does
8.1.19 @9:13pm - (3/ ) offer a theoretical escape route from this culture, through the assumption of "conditional ignorability" but, since it is congnitively formidable, practicing analysts must rely on regression arguments. Keele et al examine a causal model (Fig.2) and ask: suppose we regress Y
8.1.19 @9:13pm - (4/ ) on all observed variables; which of the coefficients has any causal interpretation. I have alerted economists to such questions here (link: (3.2.7) so, we can assume they have mastered the techniques by now. Students of #Bookofwhy seeking a gentle way to
8.1.19 @9:13pm - (5/5) approach their mentors or professors or their peers, this is a great channel to motivate them.

8.1.19 @8:35pm - (Replying to @VenkatNagaswamy) I was over-intoxicated by scientific poetry

8.1.19 @8:34pm - (Replying to @yudapearl @Jabaluck and @eliasbareinboim) But, lets deal with what they CAN DO. Let's ask their students to examine Fig. 2 of Keel etal (link: (link: and decide which regression coefficient coincides with a structural parameter. I challenged them here (link: #Bookofwhy

8.1.19 @8:12pm - (Replying to @Jabaluck and @eliasbareinboim) I happened to go over MHE last night. Read: "we follow convention and refer to the difference between the included coefficients in a long regression and a short regression as being determined by the OVB formula (p.44). You say: "They think you should write down a model and then
8.1.19 @8:23pm - (Replying to @yudapearl @Jabaluck and @eliasbareinboim) figure out an estimation strategy ..." I have not seen "a model" written down of what one believe about the world. Even Eq. (3.2.8), which is supposed to be a structural equation, is written after CIA is assumed (conditional ignorability) which we know is cognitively impossible.

8.1.19 @7:58pm - Too bad they cut off my song, just before the tenor!! Not too late. If any reader of #Bookofwhy has ANY unresolved question on DAGs, please check if that question is not answered here (link: If still unresolved, I am here on Tweeter, with an army of resolvers.

8.1.19 @7:48pm - Extremely inviting. "With a book of verse, and thou, beside me, singing in the wilderness, and wilderness is paradise anow." #Bookofwhy

7.31.19 @4:53pm - (1/2) As I was re-reading "Wagged by the DAG" (link: I've found a few collectibles that are relevant to our other discussions: (1) "It would be a disaster if models were allowed to produce information unintended by the modeler." (2) Tools ....
7.31.19 @4:53pm - (2/2) (2) "Tools that are indispensable in solving simple problems are unlikely to become dispensable when problems become more complex." (3) "[In some unnamed cultures,] selecting covariates for confounding control is still a black magic," #Bookofwhy

7.31.19 @2:25pm - (Replying to @EpiEllie and @BUSPHEpi) No one is really anti-DAG. But many propose extensions, enrichment and complementing methodologies. I was wondering if Dr. Krieger still believes the extensions she proposed in "Wagged by the DAG" are worth pursuing. (link: #Bookofwhy

7.31.19 @1:54pm - (Replying to @EpiEllie and @BUSPHEpi) Curious. Does Nancy Kreiger still upholds the views she expressed in: "The Tale Waggs the DAG?" (link: #Bookofwhy

7.31.19 @4:00am - Why I call it "upside-down culture"? Because the logical way to start is with what you KNOW, eg. structural equations, then use regression to estimate what you WISH TO KNOW. Confusion erupts when people think regression equations represent what you know. #Bookofwhy

7.31.19 @3:45am - (Replying to @TariqTaha123 @CNNSotu and @jaketapper) Straight from the book of slogans of BDS - the racist movement that shows no respect for truth or other people's identity. See (link:

7.31.19 @3:23am - (Replying to @willjharrison and @svenohl) Causation precedes manipulation. In #Bookofwhy we describe manipulation as a way of interrogating nature to reveal the causal forces that tie variables together.

7.31.19 @1:43am - (Replying to @isli_amar) Glad #Bookofwhy reached Algeria, probably from UK, swimming. I hope you start teaching it before young minds get frozen into "regression". The temporal constraints paper almost escaped my memory - so much has happened. Enjoy!

7.31.19 @1:30am - Link broken? Please try this one: (link:

7.31.19 @1:11am - In contrast, here is a new paper from political scienists (link: who address the upside-down culture of starting with "regression" and "controlling for" (eg Angrist's MHE) and then asking: Do our findings have any "causal interpretation?" #Bookofwhy

7.31.19 @12:53am - (Replying to @JamesLNuzzo and @NicoleBarbaro) Cofield's findings can be partially excused, for this was published 2010, only 10 years after epidemiologists started using graphs. But how come students in quantitative behavioral science do not rebel? #Bookofwhy

7.31.19 @12:17am - Very interesting paper. Published in 2019; 25 years after causal language has been mathematized and separated from statistical language! The author seems unaware of the causal revolution. No wonder! Has Psychometrika ever published an article on modern causal modeling?#Bookofwhy

7.30.19 @11:27pm - (Replying to @TariqTaha123 @CNNSotu and @jaketapper) My father was one of those Jews who came peacefully to his historical homeland. He was not silenced at all, I was there. He actually offered our Arab neighbors peaceful co-existence. Do you know what answer he got? I hate to embarrass you in public but, beware, I was there!

7.30.19 @9:44pm - (Replying to @JennieBrand1) Got it, thanks. Was I right in summarizing FE as: "An assumption of equality of two or more structural parameters which, in certain DAGs, leads to identification that otherwise will not be achieved." Q. The X's in your figs have no parents, what does it mean? #Bookofwhy

7.30.19 @9:04pm - (Replying to @JennieBrand1) Thanks Jennie, can you send an active link? Evidently, Oxford took UCLA off their university list, and they won't give me access to your paper.

7.30.19 @9:02pm - (Replying to @TariqTaha123 @CNNSotu and @jaketapper) I an going to show your Tweet when people ask me: What kind of neighbors Israel has? what kind of mentality drives them? and why it is so hard to reason with them? Dont they see that the two nations are equally indigenous to the land?My answer: see above "Me, Me, Me!"

7.30.19 @1:40am - (Replying to @maximananyev @Jabaluck and 3 others) Apropos:The more explicit the assumption, the more criticism it invites, for it triggers a richer space of alternative scenarios in which the assumption may fail. Researchers prefer therefore to declare threats in public and make assumptions in private. (link:

7.29.19 @10:41pm - (Replying to @Psylocke42356 @ZachWritesStuff and 2 others) This is precisely what I tell them in almost every piece I write. Can you reciprocate? Can you repeat after me (and ask Rashida to join): "Two states for two peoples, equally legitimate and equally indigenous". No ifs, no buts, just say "equally indigenous". Can you?

7.29.19 @10:10pm - (Replying to @Psylocke42356 @ZachWritesStuff and 2 others) Nations have a right to freedom and existence to the extent that they confirm such rights to their neighbors.

7.29.19 @9:58pm - One of the best quotes of the century. I hope it changes at least one heart.

7.29.19 @9:46pm - (Replying to @Jabaluck @fuzzydunlop123 and 2 others) That is why it is so important to write "The history of bad-control in pre-optics econometrics", to see precisely if the fumbling came from doubting the validity of models, or from inability to handle even a simple and valid model. Anyone writing? I'll help (anonymous) #Bookofwhy

7.29.19 @7:17pm - (Replying to @Jabaluck @fuzzydunlop123 and 2 others) And you insist you could have run a similar conversation with Angrist using the language of potential outcomes, where even "bad controls" are subject to embarrassment. BTW, is anyone writing the history of "bad controls" in pre-Telescopic econometric literature? #Bookofwhy

7.29.19 @3:13pm - (Replying to @fuzzydunlop123 @Jabaluck and 2 others) What am I watching? I thought DAGs are good for pedagogical purposes only. Now I see discussions on whether to condition on a variable or not...Dont tell me DAGs are good for discussions too, or for evaluating ID strategies, etc. etc. Beware of harsh consequences. #Bookofwhy

7.29.19 @2:31pm - Good news for causality research! Congratulations to Elias and co-authors for best paper award at the UAI-2019 conference, straight from Tel-Aviv marina, where I got my wind-surfing diploma in 1980 (framed in my office). They told me creative surfing is the secret to success.

7.29.19 @7:40am - (Replying to @ZachWritesStuff @CNNSotu and @jaketapper) What you teach your child is what you intend to do. Not one teacher, not two, but EVERY teacher. This is what you, as a journalist should decry and labor to change.

7.29.19 @7:31am - (Replying to @ZachWritesStuff @CNNSotu and @jaketapper) No one says Israel should not exist? You hav'nt heard what Palestinian teachers say repeatedly, nor what Omar Barghouti said at UCLA in 2014. See (link: .

7.29.19 @7:15am - (Replying to @glarange72 @CNNSotu and @jaketapper) The asymmetries on the ground are grim consequences of asymmetries in intention, with one side dreaming "We, We, We" and the other threatening "Me, Me, Me". Quite stark.

7.29.19 @6:19am - (Replying to @y2silence @stephensenn and 14 others) What is the simplest, canonical example of MLM showing that if you ignore the hierarchy you will not answer your research question properly. I am asking because the MLM papers I've chanced to read start and end with no explicit research questions. #Bookofwhy

7.29.19 @5:59am - (Replying to @ZachWritesStuff @CNNSotu and @jaketapper) States, indeed, just "exist", they have no rights. But when you tell your child: "Our neighbor has no right to exist" you are telling him: "We are not going to honor any peace agreement, forever". And you telling Israelis, you have compelling reasons to control those territories.

7.29.19 @4:50am - (Replying to @Vic1Nobody @ZachWritesStuff and 2 others) I am not a right-winger and I do not think it is fair to mention right-wing habits to shut down debates on core issues: Can Palestinians claim rights which they deny their neighbors. Current affairs are surface manifestations of this core issue.

7.29.19 @4:35am - (Replying to @shanbhardwaj @CNNSotu and @jaketapper) How? By tuning in, day and night, to what Palestinian leaders, clergy, and educators are saying (to their constituents, in Arabic) as I do, and my friends in the Israeli peace camp do, hoping to detect a seed of acceptance. Thus far - Nada.

7.29.19 @4:25am - (Replying to @glarange72 @CNNSotu and @jaketapper) Hypothetical questions have practical implication. In our case the implication is that one cannot demand rights to Palestinians that they deny their neighbors.

7.29.19 @2:51am - (Replying to @glarange72 @CNNSotu and @jaketapper) My position is that Jake Tapper should have asked her: "What should Israelis do if Palestinian leaders tell them what they do (I hope you heard them the past 75 years)???"

7.29.19 @2:40am - (Replying to @yudapearl @CNNSotu and @jaketapper) Poor Rashida, she just lost 90% of her voting base by saying: "Of course, but..." when Jake Tapper's asked her: "Does Israel have the right to exist?" It's tough to be a Zionophobe on CNN, torn between viewers norms of justice and voters push for elimination.

7.29.19 @1:40am - (Replying to @CNNSotu and @jaketapper) The punchline was missing: "What if the overwhelming majority of the Palestinians, leaders, educators and clergy, deny the right of Israel to exist, and openly declare that, occupation or not, they will to continue their arm struggle against Israel, in any borders?"

7.29.19 @1:32am - (Replying to @ZachWritesStuff @CNNSotu and @jaketapper) Zachary, Serious Person, associates a people's right to self-determination with "right-wing talking points". A serious person indeed.

7.29.19 @12:19am - (Replying to @AdanZBecerra1 @Lester_Domes and 14 others) I fail to see the relevance of MLM in this paper. I see constancy of effects across time, fine, but I do not see clusters or Set-Subset relationships in the examples. My blindness?

7.29.19 @12:03am - (Replying to @stephensenn @Lester_Domes and 13 others) I see no "strong undeclared assumption" in #Bookofwhy, perhaps because I read assumptions from DAGs and, looking at Fig.6.9, I see no assumption left out. Nada! The point is: {Data + DAG} determine which analyst was correct. No need to recruit Nelder to make this simple point.

7.28.19 @9:57pm - (Replying to @Lester_Domes @melb4886 and 13 others) Glad we are in agreement. I could never understand why certain folks, who are reluctant to learn causal inference, always excuse themselves with "we can do it in MLM". What is it in MLM that gives people the illusion that it can answer causal questions? #Bookofwhy

7.28.19 @9:09pm - (1/ ) Great paper. First time I understand what "fixed effect" is. I used to confuse it with homogeneity, but Fig. 1 tells us it is an assumption of equality of two structural parameters which, in certain DAGs, lead to identification that otherwise will not be achieved. So simple,
7.28.19 @9:09pm - (2/ ) "Why didn't they tell us?" I should complain, but I won't, because it was probably all there, in the papers and the lengthy motivations, and the indexed regression equations that I was too lazy to digest. Glad it is over, one figure did it. #Bookofwhy

7.28.19 @4:21pm - You are right. In case other readers of #Bookofwhy got stuck on Eq. 7.2 page 227, replacing Z with U would make it less mysterious in the context of the Z=Tar story. I hope we can make the change before the paperback edition hits the shelves. Thanks @the_aiju

7.28.19 @4:00pm - (1/2) I love your Advisor, and I think his "hide it" was an honest expression of the prevailing culture. I am sure that if he and colleagues start using DAGs in hiding, under the cover of "only for education," they will eventually use them everywhere, from seeing assumptions, to
7.28.19 @4:00pm - (2/2) to testing assumptions, to validating ID strategies, to discovering new ID strategies. Recall, IV was discovered using Wright's DAGs. Why? Because "using DAGs" simply means "using what you know". I presume the Church lifted its ban on telescopes "only for education"#Boodofwhy

7.28.19 @2:06pm - (Replying to @nyarlathotepesq) d-separation is valid for linear CYCLIC models as well. So it is easy to identify bad controls in simultaneous eq-ns. #Bookofwhy

7.28.19 @5:41am - FYI, I could not resist answering a new Quora question: "How does the Rubin causal model differ from graph-theoretic approaches like Pearl's do-calculus?" (link: There is nothing new there that our readers do not know; it is just compiled succintly #Bookofwhy

7.28.19 @5:34am - (Replying to @thosjleeper @PHuenermund and 4 others) No surprise that this is how IV is (still) taught in econ. They are talking about "structural equations" and call it "regression" and the variables "regressors". It is OK, as long as: (1) Econ. students can endure the confusion, and (2) a Glossary is in the making.#Bookofwhy

7.28.19 @3:51am - (1/ ) (Replying to @PHuenermund @dlmillimet and 3 others) In looking over other tweets in this thread I got the feeling that many of the questions are answered in the #causalinference literature, but the jargons are almost incompatible. I hope you plan on writing a glossary one day. For example, 2sls is an estimation method, hence it
7.28.19 @4:00am - (2/ ) (Replying to @yudapearl @PHuenermund and 4 others) has nothing to do with interpretation or with coefficients. Also, in the IV model the IV variable is the only "exogenous" variable. Finally, in a regression model "exogeneity" is not defined. The glossary seems like an endless job, but where would we be w/o it?#Bookofwhy

7.28.19 @2:20am - (Replying to @PHuenermund @dlmillimet and 3 others) @PHenermund I ventured to read your post and met your question: "How can we be sure that what we're estimating for the compliers is representative for the whole population?" Why don't Balke's bounds provide a general answer to your question? eg: (link: #Bookofwhy

7.28.19 @12:08am - (Replying to @stephensenn @AdanZBecerra1 and 12 others) Is there room for single-level reasoning anywhere in science? If not, I'll scrap 99% of my science books. If yes, I would first cast Lord's and Simpson's dilemmas in single-level context, see if I can solve them, then proceed to multilevel if needed, but only if needed.#Bookofwhy

7.27.19 @8:30pm - (Replying to @AdanZBecerra1 @stephensenn and 12 others) I am not questioning the need to do multilevel modelling when needed, be it in causal or predictive tasks. I am questioning the wisdom of forcing multilevel modeling on single-level causal questions (eg Lord's and Simpson's examples) w/o the tools of causal modeling. #Bookofwhy

7.27.19 @5:08pm - (Replying to @FJnyc @intelligence2 and 11 others) Mehdi Hasan will continue to stage these Kangaroo debates until someone charges him with "Zionophobic bigotry" in front of an audience, which would corner him to compare his perpetual denial of Jewish identity as people with his perpetual whining of "Islamophobic bigotry."

7.27.19 @3:49pm - (Replying to @AdanZBecerra1 @stephensenn and 12 others) @AdanZBecerra1 I do not think one should delve into multilevel models before acquiring the tools of handling single-level problems, like the one posed by Lord, or Simpson, or #Bookofwhy. Multilevel modeling, if imposed on every single-level problem can become counterproductive.

7.27.19 @3:11pm - (Replying to @matt_vowels) We are now hearing a 9th myth "A causal model is a special case of a predictive model". This one is particularly misleading, for it tells stat. students: "Dont bother to change your thinking, what we have been doing this past century is sufficient". No it ain't. #Bookofwhy

7.27.19 @3:02pm - (Replying to @marcelogelati) I can't decipher what they are trying to say, but if the're saying you can't infer causality from observationial studies alone, w/o a model, they are right. If they deny the power and ubiquity of qualitative models -- they are wrong. #Bookofwhy

7.27.19 @2:23pm - (Replying to @intelligence2 and @EWilf) Thanks for posting. I would never debate under such title, which pre-fixes its verdict. Show me one titled: "Is Anti-Zionism RACISM on its own merit" and I will show you a YES verdict earned. That is why I always use "Zionophobia", not anti-Semitism, e.g.

7.27.19 @4:37am - (Replying to @laurencepearl) It used to be Perl (on my grandfather visa, as he arrived at the Holy Land in 1924, from Poland). Legend says Jews bought their surnames centuries ago, from their European masters, based on their professions and the money they could raise. Seems Jewelry was a good profession.

7.26.19 @7:10pm - (Replying to @_asubbaswamy) Now I see it. But why did it take me 24 hours? Because I was missing one word: "E[T|do(a), c] is stable and IDENTIFIABLE". Naive me thought you are recommending an experiment with do(a). I suggest you stress this point explicitly in future papers.#Bookofwhy

7.26.19 @3:49pm - (Replying to @laurencepearl) Don't understand!! ??? You must be kidding! How do you estimate causal effects? Or, how do you find stable predictors? More importantly, how do your colleagues do that? You'r kidding!!! #Bookofwhy

7.26.19 @3:41pm - (Replying to @_asubbaswamy) Thanks for the refinement! Agree. S goes into X. A question to you: Why didn't you use a simpler example than Fig. 2(b) to show advantage over "pruned estimators" ?? It took me an hour to believe that such examples exist. #Bookofwhy

7.26.19 @3:06pm - Just occurred to me: Is't the front-door a simple example of getting an unbiased and stable estimator of Y, given observations on X and Z (with U varying)? [ X--->Z--->Y with confounder U, X<-U->Y. ] E[Y|x,z] is unstable, whereas the front-door formula is. #Bookofwhy

7.26.19 @2:43pm - (1/ ) Saying that "a causal model is a special case of a predictive model" is like saying "sailing is a special case of swimming, since it is conditional of something floating". In general, saying that task-A is a special case of task-B depends on what the sayer is trying to get:
7.26.19 @2:43pm - (2/2) an excuse for not doing A, or a license for doing A using methods used in B. The latter would be justified, if it was possible. Unfortunately, causal models require information (+methods) not available in traditional prediction modeling. So why say "special case"? #Bookofwhy

7.26.19 @2:19pm - You are right, the word "Zionophobia" is gaining traction, including the Justice Department Summit, see (link: It is becoming "the ugliest word in town", at least among people of conscience. It's the most effective defense weapon I know.

7.26.19 @6:00am - (Replying to @suchisaria @KordingLab and 4 others) The idea of "surgery estimators" is ingenious, it would not occur to me that you can get extra mileage on top of "pruned estimators". However, is Figure 2(b) the simplest example to show this extra mileage?? Would c-equivalence help here? (link: #Bookofwhy

7.26.19 @5:44am - (Replying to @nyarlathotepesq) You are right, the bet is not a proof, that is why I suggested an Appendix with some proofs. But the two polar poles is not a good analogy; going from linear to nonparametric amounts to monotonically removing constraints. Are you considering writing it? #Bookofwhy

7.26.19 @4:12pm - To me, the key difference between the stats-centric traditional approaches and a #DAG-centric approach: the former shun causal problems whose solutions can be derived from basic principles, the latter seek such problems. Smart folks should not wait for agreement, #Bookofwhy

7.26.19 @3:06pm - Just occurred to me: Is't the front-door a simple example of getting an unbiased and stable estimator of Y, given observations on X and Z (with U varying)? [ X--->Z--->Y with confounder U, X<-U->Y. ] E[Y|x,z] is unstable, whereas the front-door formula is. #Bookofwhy

7.26.19 @3:10am - (1/2) Great question! Three answers. 1) In practice, we are really seeking "safe control", not "bad control" and nonparametric analysis gives it to us. 2) If something is bad in both nonparametric and linear analyses, you can bet it is bad in between. 3)
7.26.19 @3:10am - (2/2) (3) All the fumbling and stumbling I have seen in the econometric literature occur already in linear systems, where the proof of "badness" is easy, see (link: Plus, you can discuss the question in the Appendix. #Bookofwhy

7.26.19 @2:07am - (1/2) BAD-CONTROL. If I were a young economist, seeking visibility and impact in my field, I would sit down and write an article titled: "Bad-Control - A lingering challenge and its resolution." I got this thought upon reading (link: noting that Angrist (2017)
7.26.19 @2:07am - (2/2) still calls this elementary econ. exercise: a "difficult problem". Such article will be a highly appreciated eye-opener to many of your peers, but it must be written diplomatically -- lot's of professional honor involved. #Bookofwhy

7.25.19 @12:34am - (Replying to @neurosutras and @JonAMichaels) Causal thinking does not mean purging predictive and associative relations; it means using such relationships properly whenever they emerge from a causal model of the world. #Bookofwhy

7.25.19 @6:38am - (Replying to @KordingLab @fhuszar and 4 others) Everything Econs did is correct, but what they did is of limited scope. I once asked: how many economists can do X, or Y etc. and I lost all my econ friends. See Causality p.216 fn10 - I will not repeat this mistake - do not force me, please. #Bookofwhy

7.25.19 @6:23am - (Replying to @brunofmr and @suchisaria) "Stability" under environmental change should not be confused with "stable distribution" which is a purely probabilistic notion, and has nothing to do with environmental changes. I hope no confusion results. #Bookofwhy

7.25.19 @5:03am - (Replying to @suchisaria @KordingLab and 4 others) Will do.

7.25.19 @5:02am - Not many people realize that the strength of a DAG comes from building ALL its logic on ONE primitive question: "Why does a variable vary?" All the rest is mechanically derived, demanding no further judgment. Poetically, I crowned it: "going where knowledge resides." #Bookofwhy

7.25.19 @4:44am - (Replying to @KordingLab @fhuszar and 4 others) I dont see it this way, since I hardly touched on "estimation" (ie, going from finite sample to distribution); I call the stat dept once I get an estimand. Econometrics has been doing some limited identification correctly. How limited? See: (link:, #Bookofwhay

7.25.19 @1:26am - (1/ ) I just read (link: and I agree with @suchisaria . Anyone concerned with stability (or invariance) should start with this paper to get a definition of we are looking for and why. Arjovsky etal paper should be read with this perspective in mind, as an attempt to
7.25.19 @1:26am - (2/ ) secure this sort of stability w/o having a model, but having a collection of varying datasets instead. The former informs us when the latter's attempts will succeed or fail. It is appropriate here to repeat my old slogan: "It is only by taking models seriously that we learn
7.25.19 @1:26am - (3/ ) when they are not needed". I wish I could quote from Aristotle but, somehow, the Greeks did not argue with their Babylonian rivals, the curve-fitters. They just went ahead and measured the radius of the earth AS IF their model was correct, and the earth was round. #Bookofwhy

7.24.19 @4:03pm - (Replying to @fhuszar @yudapearl and 3 others) It makes sense to me to start from what is it that we know must hold true. Discovering which invariances you want to guarantee based on your data is OK but the data are only a sample and it's important to declare which invariances are desired and why.

7.25.19 @12:44am - (Replying to @RichmanRonald) This paper (link: .is an interesting window into statistics education of 2019. Contrary to causal logic, statistics students start with data visualization routines and then ask: when do these have causal interpretation. I hope #Bookofwhy will change this order.

7.25.19 @12:19am - (Replying to @KordingLab @fhuszar and 4 othersz) There has always been an understand between ci and ml -- I do identification and you do estimation. Nothing has changed with all the talk about "causal ML", except perhaps ML folks internalizing the limits of the Ladder of Causation #Bookofwhy

7.24.19 @6:20pm - (Replying to @Jabaluck @PHuenermund and @fuzzydunlop123) I happened to pass by this thread. There is no testable implication for IV from observational studies unless treatment is discrete. See (link: #Bookofwhy

7.24.19 @6:03pm - (Replying to @suchisaria @tdietterich and 2 others) Yes, these papers are in scientific language. Not because they are using DAGs, but because they provide theoretical guarantees under meaningful assumptions. I've vowed to read them and comments. #Bookofwhy

7.24.19 @4:22pm - (Replying to @RichmanRonald) Thanks for the pointer. I will try reading it tonight. It is always a valuable learning experience to see how statisticians think about causal problems.

7.24.19 @3:40pm - (Replying to @fhuszar @tdietterich and 3 others) How about adding it as appendix to your blog? "From ppp to SSS - a declarative summary"

7.24.19 @2:35pm - (Replying to @yudapearl @fhuszar and 4 others) Most importantly, having read Arjovsky etal paper, do we understand what ppp and SSS are? Or, at least what the claim is? Such translation will help evaluate the claim under the light of existing theories. #Bookofwhy

7.24.19 @2:25pm - (Replying to @yudapearl @fhuszar and 4 others) "If a pattern ppp is seen in the data then something (SSS) must hold true in the world". The language of "my algorithm tries" of "my algorithm optimizes" etc reminds me of AI in the 1970's "MYCIN tries" "ELIZA optimizes" which was later replaced with declarative writing style.

7.24.19 @2:09pm - (Replying to @fhuszar @tdietterich and 3 others) Agree, but at this point we are trying, not to convince ML folks but to understand what their new method provides, and under what assumptions. To understand what a ML paper offers we need someone to translate the paper into a declarative language, that goes:

7.24.19 @5:23am - For those interested in the history of UAI, I wrote a personal memorandum of those days (link: #Bookofwhy

7.24.19 @4:56am - (Replying to @yudapearl @fhuszar and 4 others) we would say: "If we observe x,y pairs from multiple E's, and we find Y||E|X,W and NOT-X||E|Y,W, then something must hold true in the world." Now, let's continue from here: What is it that must hold true? Can you get it from the paper? #Bookofwhy

7.24.19 @4:50am - (Replying to @fhuszar @tdietterich and 3 others) Your summary is ALMOST Causal language, but not quite. In Causal language we do not invoke man-made algorithms (IRM) to describe environmental properties such as invariants; we speak declaratively. For example, instead of saying: ""IRM observes" x,y pairs from multiple E's,"

7.24.19 @4:34am - (Replying to @sobu_18) No, No, the word Zionophobia has a unique magic to it, unmatched by antisemitism. I've never met an anti-Semite who admitted to being one, and I've never met a Zionophobe who denied being one. Quite a difference for a mentality that denies a homeland to a people.

7.23.19 @11:38pm - (Replying to @suchisaria @tdietterich and 2 others) Great! I was't aware of these two papers. Now we know that searches for "stability" and "invariants" are on firm scientific grounds: When we suspect a certain relation is "stable" we can check and see that it is truly stable and what conditions will make it unstable.#Bookofwhy

7.23.19 @8:59pm - Sorry to have missed it, hoping a video was taken. I always regret not having a video from the first UAI, 34 years ago, at UCLA-1985, when probabilities first infiltrated the forests of AI. I see causality following a similar route. #Bookofwhy @eliasbareinboim

7.23.19 @7:02pm - (Replying to @tdietterich @zittrain and 2 others) I understand that IRM leverages the availability of multiple data sets, usually absent from CI. So let's represent the multiple data sets in the language of CI and see what must hold in the world for IRM to find the invariants it is searching for. Anyone done it? #Bookofwhy

7.23.19 @3:54pm - (Replying to @sobu_18) Careful reading of (link: reveals that #BDS is not charged with antisemitism, but with Zionophobia: The irrational animosity toward a homeland for the Jewish people. Populist slogans like "apartheid" tend to tarnish the credibility of their chanters.

7.23.19 @3:08pm - (Replying to @tdietterich @zittrain and 2 others) @tdietterich, Since you are familiar with both CI and Invariant Risk Mininization (IRM), can you (or anyone else) explain in input-output terms how IRM extracts from non-experimental data information that CI thought must be obtained from either a model or intervention. #Bookofwhy

7.23.19 @3:59am - (Replying to @pcastr) If you take seriously my experience with BDS activists (link: you will see that "boycott" is the last thing on their agenda. Their aim is to silence pro-coexistence voices. Is it really tenuous to see that absent such voices Israel's demise is almost certain?

7.23.19 @3:12am - I, likewise, just want to voice my conviction that support of BDS may have serious unintended consequences, a genocidal demise of Israel is one, as I describe here: (link: From there to the demise of the Jewish people takes another leap of logical deduction.

7.23.19 @12:34am - If you know a BDS promoter who bought into its rhetoric without checking the destructive aims of its leadership, you have found someone who has not studied BDS as thoroughly as I have. See (link:

7.22.19 @8:25am - (Replying to @azuur @acastroaraujo @Jabaluck) I beg to differ. It discards whatever does not have a MODEL, which need not be a DAG. Since every CI study must rely on a model, not having one means keeping one in your head and giving the impression that you operate model-free. Both are ill-advised. #Bookofwhy

7.22.19 @8:16am - (Replying to @azuur @acastroaraujo @Jabaluck) Completamente de acuerdo!!

7.22.19 @7:47am - (Replying to @bzaharatos @learnfromerror) I am also going to miss JSM this year (Carlos Cinelli will represent me at the Fellow Reception). As to Philosophy+Statistics seminar, glad someone will be there who can translate back to us, down in the trenches, speaking cause and effect. #Bookofwhy

7.22.19 @6:16am - (1/ ) (Replying to @RonKenett @learnfromerror) 1/ I am not sure that we are talking about the same notions of "generalizability". For me, this word means taking experimental results from one population and applying it to another, potentially different. I am not sure @learnfromerror means the same thing; I would be
7.22.19 @6:22am - (2/ ) (Replying to @yudapearl @RonKenett @learnfromerror) curious to know if she does. Why I am not sure? Because the way the two populations may differ may be non-statistical, namely they may have the same joint distribution functions on all variables, and differ ONLY in the causal forces holding the variables together #Bookofwhy
7.22.19 @6:30am - (3/ ) (Replying to @yudapearl @RonKenett @learnfromerror) Under such circumstances statistical methods cannot deliver remedy unless they are guided by a causal model of those underlying forces. Such models are absent from standard statistical writings, they are introduced back-door in the PO literature, as in

7.22.19 @1:20am - (1/ ) (Replying to @PHuenermund) Thanks Paul for starting this important discussion on @EconBookClub of Imbens paper. Since the paper is full of objectionable macro and mini-statements, I believe it is wise to focus on its core, summarizes precisely in the paragraph you posted. "Little is said about what 7.22.19 @1:30am - (2/ ) (Replying to @yudapearl @PHuenermund) comes before the identification question and what comes after the identification question" . The first reflects misunderstanding of what structural economics is about, while the second decries a "before vs. after" distinction that should be welcome with joy. The rest of Imben 7.22.19 @1:39am - (3/ ) (Replying to @yudapearl @PHuenermund) paper follows as a corollary of these two basic misunderstandings. Once we illuminate these two, it should be easy to clear the rest, especially the identification stage itself, which he admits to be virtually absent from his tool set. (if # of variables exceeds 3) #Bookofwhy

7.21.19 @11:47pm - An old saying goes: "When mathematicians notice an interesting problem it becomes SCIENCE". Today, Notices of The American Mathematical Society took notice of #Bookofwhy, thanks to Lisa Goldberg : ... I hope #causalinference gets enriched with new insights.

7.21.19 @11:18pm - This is funny! Cornel West is "standing in moral solidarity with four sisters?" Last I wrote about him and his moral deformity, it was he who needed moral solidarity ...

7.21.19 @10:45pm - (Replying to @bzaharatos) In what way?

7.21.19 @9:34pm - (Replying to @oacarah) On the mark! Which proves the algebra is for the birds.

7.21.19 @6:02pm - Your daughter, Isadora, is adorable. If she could only get that partial differential equation right, we could have had a perfect afternoon -- next time!

7.21.19 @4:49pm - Look Ma! I'm a statistician! They tell me that at the upcoming JSM meeting in Denver I'll be ordained as an ASA Fellow. I assume it means that, starting July 30, statisticians can treat #Bookofwhy as a homegrown tomatoe, and each of its toy problem as 10,000 "real-life" examples.

7.20.19 @3:01pm - (Replying to @AdanZBecerra1 @PHuenermund) Do you know what blog or forum do thoughtful statisticians use to communicate ideas about the philosophy of statistics, including causal inference ??

7.20.19 @4:41am - (Replying to @AdanZBecerra1 @PHuenermund) Another remarkable observation: none of the discussants had any clue on how to handle conditioning on post-treatment variables. The advice they got was: "Control for as many pre-treatment variables as you can." Is this the best place to learn how statisticians think? #Bookofwhy

7.19.19 @11:37pm - (Replying to @yudapearl @AdanZBecerra1 @PHuenermund) @AdanZBecerra1, I remember you from Gelman's blog. You reminded the discussants of DAGs and got bullied by one of the big guys who told you to go solve a "real life" problem. I was tempted to come to your rescue but those guys knew so much about "real-life"- got scared #Bookofwhy

7.19.19 @8:11pm - (Replying to @PHuenermund) Totally agree. Discussions should start with the question: Why are DAGs "pedagogically" more "transparent". Is it just "taste"? Or is it something universal in our minds that makes DAGs good "displayers of assumptions"? The rest follows from this cognitive phenomenon.#Bookofwhy

7.19.19 @7:37pm - (Replying to @PHuenermund) Interesting. Guido may have changed his mind - I did not. I truly believe that if any of my economist colleagues actually roles sleeves and solves a couple of toy problems, instead of talking ABOUT them, he/she will never go back to talking ABOUT them. #Bookofwhy

7.19.19 @12:28pm - Each toy problem is 10,000 "real life" examples, in which you cannot hide behind messy data, unobserved confounders, or other excuses and, moreover, you can check your method against ground truth. Those who shun toy problems do have something to hide, watch them. #Bookofwhy

7.19.19 @2:23am - (Replying to @aqsaqal) Sounds like a beautiful example of missing-data. And I bet it is recoverable, since the missingness of Weight is not caused by Weight itself. See a graphical approach (and stay away from imputation) #Bookofwhy @Carthica

7.19.19 @1:39am - (Replying to @thosjleeper @matloff) My My! Good point! It is better that they leash out on me than to let truth reduced to "gossip". Thanks, #Bookofwhy

7.18.19 @11:32pm - (Replying to @yudapearl @thosjleeper @matloff) Why is quoting unnamed statisticians a bad thing? Would naming them make less hostile? Recall, they are still alive, deciding on promotions, and perhaps repenting for what they once thought. Why embarrass them? #Bookofwhy

7.18.19 @11:27pm - (Replying to @thosjleeper @matloff) Comments on your thread. The statement P(Y|do(x)) is causal but not counterfactual, because it implies no contradiction, and it can be estimated from RCT. P(Y|see(x)), on the other hand, is statistical, not causal, because we can evaluate it without experiments. #Bookofwhy

7.18.19 @11:12pm - (Replying to @anirudhacharya1 @oacarah) How is this: ?? Any luck? #Bookofwhy

7.18.19 @10:56pm - (Replying to @klausmiller @PHuenermund and 3 others) Judea Pearl Retweeted Judea Pearl As I Tweeted here: ... this is the greatest thing that happened to DAGs: Sunrays are the best de-confounders. #Bookofwhy @quantadan @analisereal

7.18.19 @10:39pm - I like this title: "When in Doubt, DAG it Out. " @oacarah new commentary on "Analyzing Selection Bias for Credible Causal Inference" ... It confirms my mantra: "DAG goes where knowledge resides," which PO folks are about to internalize. #Bookofwhy

7.18.19 @8:11pm - (Replying to @vauhinivara @WSJ) Delighted to read your thread and to see that the WSJ Internship set up in memory of my son Daniel was instrumental in lifting you forward. Pitching story, so it seems, is much like pitching scientific articles - enlightened editors are scarce. Good luck #Bookofwhy @craigmatsuda

7.18.19 @2:41pm - Thanks for pointing to Imben's new paper on PO and CI. We finally have a window into the thinking of leading PO researchers. It will give econ. students a chance to compare alternatives and ask: Do we really want to think that way? #Bookofwhy @StatsPapers @causalinf @PHuenermund

7.18.19 @12:16am - (1/2) (Replying to @omaclaren @djvanness and 2 others) 1/2 We have this world; it is called classical mathematical modeling. What this world was missing (when I last fiddled in it) was the EXTREME case called DAGs. Namely, the miracle of how much can be accomplished with so few and weak assumptions. I might go back to classics, once
7.18.19 @12:22am - (2/2) (Replying to @yudapearl @omaclaren and 3 others) we fully understand the full potentials of this new miracle. But it seems to go on and surprise us with new capabilities: external validity, missing data, sensitivity analysis, fairness,...Its pouring! And you want us to quit? #Bookofwhy

7.17.19 @10:53pm - (Replying to @omaclaren @djvanness and 2 others) You have this strength in SCM if you want to express something you are really sure about. But if you are only sure about "who listens to whom" use DAGs. Perhaps you are after a calculus that works in between these two extremes? Happy sailing; watch the two extremes #Bookofwhy

7.17.19 @10:43pm - (Replying to @omaclaren @djvanness and 2 others) Easy! You write Y = OR(X,Z) or Y=AND(X,Z). This is how you express it if you think you can. If you can't commit to one function or another you just write Y=f(X,Z) and you let data decide. If your data must be taken in isolation, it is mathematical impossibility.#Bookofwhy

7.17.19 @10:35pm - (Replying to @omaclaren @djvanness and 2 others) You are repeating what I said, so we are settled. Recall the DAGS were devised to minimize the amount of information (ie assumptions) demanded from researchers, and limit it to those relations only that reside in the scientist's comfort zone --variable listening #Bookofwhy

7.17.19 @10:20pm - (Replying to @LMCheongTS) RL can handle some interventional problems, like predicting effects of actions tried before. As to counterfactuals, RL can potentially produce bounds on counterfactuals, the same as what Rungs 1 & 2 produce. A miracle shown here: . #Bookofwhy

7.17.19 @10:11pm - (Replying to @omaclaren @djvanness and 2 others) Yes, `interaction' a concept that is formalised in SCMs . Compute P(Y|do(x),z)) and check if it depends on z. Note that this quantity is computable from full SCM specification. Data is needed only to supplement missing specifications, say P(u), or functional forms #Bookofwhy

7.17.19 @10:04pm - (Replying to @omaclaren @djvanness and 2 others) To state directly that interaction does not exist SCM must resort to some parametric specification, say "linear family". Or, indirectly, SCM can tell us: "Check for yourself. I have given you part of me, a DAG, which implies a recipe for interrogating the data; use it. #Bookofwhy

7.17.19 @9:56pm - (Replying to @ntabari8221 @fulhack) The #Bookofwhy argues that too much time is spent on "confounders," What we should be looking for are the "de-conounders".

7.17.19 @9:52pm - (Replying to @djvanness @ConiByera and 2 others) Yes, it is "legal" with or without the U. And, remarkably, one cannot tell who is the modifier and who is the "cause". This distinction still demands a "penalty" definition: What would I lose if I incorrectly switch the two. #Bookofwhy

7.17.19 @9:45pm - (Replying to @djvanness @ConiByera and 2 others) Not really. The topology of the DAG (with or w/o U) tells us how to get the Z-specific causal effects. Whether or not those effects depend on Z would be revealed from the data, once we estimate P(Y|do(x), z) correctly. #Bookofwhy

7.17.19 @8:33pm - (Replying to @djvanness @ConiByera and 2 others) The last warning is given to you by the DAG, coupled with a method of quantifying the penalty. So, what is missing? #Bookofwhy

7.17.19 @8:30pm - (Replying to @djvanness @ConiByera and 2 others) I am still missing what that "additional extra-statistical structural information" is that you need "to get the policy estimates right." What kind of answer you expect the system to deliver to you if you had this extra information encoded in the DAG #Bookofwhy

7.17.19 @8:25pm - (Replying to @djvanness @ConiByera and 2 others) If your target quantity is P(Y|do(x),z) for different z's, you can identify it by controlling for C. The DAG with U gives you the same information, ie "interaction is possible". You probably want an early warning saying: "dont bother, no interaction here". Right? #Bookofwhy

7.17.19 @8:16pm - (Replying to @djvanness @ConiByera and 2 others) If all you wish is P(Y|do(X=1)) then you do not need to account for Z. But you really wish to know more. You must start by telling the inference-engine what it is. Perhaps you seek the difference P(Y|do(X=1), Z=0) - P(Y|do(X=1), Z=1)? Or ratio? or something else? #Bookofwhy

7.17.19 @8:06pm - (Replying to @djvanness @ConiByera and 2 others) I will try to answer slowly each one of your tweets. With DAGs, no one is left confused. The DAG you drew just says: I don't rule out interactions between Z and X. And he ( Mr. DAG) is now waiting for you to specify you research question. What is it? #Bookofwhy

7.17.19 @7:51pm - (Replying to @omaclaren @djvanness and 2 others) SCM assumes the existence of functions Y=f(x,z,w...), with the help of which we define other concepts such as causal effects, explanation etc. Whether or not we can identify those concepts from data is a separate issue. Those who wonder whether DAG has a concept C need first to
7.17.19 @7:59pm - (2 )(Replying to @yudapearl @omaclaren and 3 others) define what they mean by C. They can use mathematics, or examples, or the method of "penalty", which goes: "If I know C, I can escape from the penalty that I would suffer not knowing C." So far I have not seen a penalty definition of the concept sought by Dave. #Bookofwhy

7.17.19 @7:43pm - (Replying to @mgaldino) I care less about those "some economists" than I care about our "current and living economists". Do they tell their students: "We rely on models"? Are they proud or ashamed of using models? Are they reading models or hiding them? @Bookofwhy

7.17.19 @6:39pm - Every time I read Quora I wonder: Do empirical economists in Angrist etal school consider themselves "model-based" or "model-free". Some say: "unlike statisticians, we use models". Others say: "unlike traditional econs, we are not tied to models". I wish someone would clarify.

7.17.19 @6:17pm - (Replying to @djvanness @ConiByera and 2 others) I tried to follow this thread, and still can't see what "logical interaction" you wish to represent. DAG in itself just says "there can be an interaction". SCM says more: it tells us the function Y=f(X,Z,U,W...). But I sense that you want more, what is it? and what for?#Bookofwhy

7.17.19 @5:12am - Apropos "Zionophobic Thuggery", I am delighted to learn that the words are gaining traction all the way to the Department of Justice. They were mentioned 5-6 times in this Summit meeting of DOJ , and they are making my students vow: No more of this racism.

7.17.19 @4:50am - And I need to check my blog, to make sure I ever wrote it. Who knows how many useful nuggets are buried in that blog, written before Twitter, when I had time to write long answers, and when people expected in-depth answers to questions that textbooks tended to ignore. #Bookofwhy

7.17.19 @3:41am - Funny, 1/4 of a century later, the introduction to "Graphical Models, Causality, and Intervention" (1993) sounds like it was taken straight from #Bookofwhy

7.17.19 @2:33am - The first paper using do() notation was "A probabilistic calculus of actions" (1994) , though the idea was used earlier (1993), using set(X=x) notation, in #Bookofwhy

7.16.19 @12:39am - (Replying to @djvanness @y2silence and 2 others) Good example where most people agree that Z is a modifier of the effect of X on Y, and not that X is a modifier of the effect of Z on Y. So, back to penalty analysis. What will go wrong if I mislabel them and I mistake X to be a modifier of the effect of Z on Y? #Bookofwhy

7.16.19 @12:26am - (Replying to @VladZamfir @lostinio and 3 others) Curious, what is your favorite epistemology of uncertainty? Plus, what is a good way to assess the merit of such epistemology? #Bookofwhy

7.15.19 @9:56pm - (Replying to @VladZamfir @oliverbeige and 2 others) Two remarks. (1) There are hordes of identities one can extract out of probability theory. Why did Bayes' receive such prominence? (2) I am still searching for a definition of Bayesianism, not for what's wrong with its practitioners. #Bookofwhy

7.15.19 @8:41pm - (Replying to @djvanness) The way I usually get to the "essence" of a distinction is to ask myself: How would I be penalized if I did not know the distinction. In our case: What would I do wrong if I were to mistake an "effect modifier" for a "cause"?? From the penalty comes the "essence." #Bookofwhy

7.15.19 @6:20pm - (Replying to @djvanness) Many feel that DAGs and other models do not capture the "essence" of "effect modifier", as opposed to just "cause". But few have taken time to explicate, even semi-formally what that "essence" is, and how they would benefit if a model does capture it. Care to try? #Bookofwhy

7.15.19 @3:28pm - (Replying to @VladZamfir @oliverbeige and 2 others) The frequentist interpretation of Bayes rule is a one-line high-school algebra trick. This is not what made Bayes famous, controversial and revered. #Bookofwhy (p. 102-3) proposes an explanation of what made Bayes rule justifiably revered.

7.15.19 @2:48pm - (Replying to @ipyadev @_MiguelHernan @JohannesTextor) Thanks for posting this paper, by Clarice, which skipped my awareness. The key sentence is: "Any 2 direct causes of D are effect modifiers for each other on at least 2 scales, which can make a reasonable person question the utility of the concept. " I'll try to clarify #Bookofwhy

7.15.19 @5:46am - (1/ ) (Replying to @oliverbeige @VladZamfir and 2 others) For some people, Bayesianism is spraying priors on parameters and waiting for the posteriors to peak. For others, especially those who read Bayes (1763), Bayesianism means 1) leveraging subjective knowledge and 2) processing evidence by Bayes' Rule. I was in the second camp
7.15.19 @5:55am - (2/ ) (Replying to @yudapearl @oliverbeige and 3 others) When I coined the name "Bayesian networks" (1985), but I am camp-friendly and would not mind switching if you can tell me what I am switching to. In the meantime I also wrote "Why I am only half Bayesian" which will affect my next camp #Bookofwhy

7.15.19 @12:42am - (Replying to @brettjgall @BrendanNyhan @NateSilver538) This is an under-understatement. Causal inference is actually a mathematical impossibility without a model, more accurately, without a "causal model", namely one that cannot be articulated in the language of statistics. #Bookofwhy

7.14.19 @9:51pm - (Replying to @ZionessMovement @CoryBooker) Thank you @ZionessMovement for joining my Twitter and for informing me of how @CoryBooker and VC Biden responded to IfNotNow squeeks. I neve say the Israeli Palestinian conflict is "COMPLEX". It is baby simple: One side says "we, we, we" and the other "me, me, me". Baby simple.

7.14.19 @7:54pm - (Replying to @ipyadev @_MiguelHernan @JohannesTextor) In DAGs all variables are presumed to be "effect modifiers" by the nonparametric nature of the assumptions. You probably meant to ask: How do we mark variables that are NOT effect modifiers. My question: for what purpose? ie What would those marks enable you to do? #Bookofwhy

7.14.19 @1:55am - (Replying to @ERMANigeria) In Quora we are witnessing another phenomenon, not parametric addiction. Angrist for example, ... talks as if economics has a unique & exclusive ownership on all CI questions, and everyone else is doing data-fitting. It's a remarkable phenomenon. #Bookofwhy

7.14.19 @12:19am - (Replying to @ERMANigeria) Please share. I have not seen that. Perhaps because, immersed in non-parametric modeling, I failed to notice what my colleagues have been doing. #Bookofwhy

7.14.19 @12:17am - (Replying to @henning_lars) I agree. But how do they intend to "aggregate data across a variety of sources," about which "good old statistical methodology" is totally helpless. [And I include Meta Analysis in this state of helplessness]. I am afraid they will fall irreversibly for same dead-ends #Bookofwhy

7.14.19 @12:03am - Sharing. I have noticed that all CI-related questions on quora are answered by econometricians who, as we have noticed on this Twitter, are not too familiar with SCM (yet). I therefore posted a new answer: , in the hope of stimulating awareness #Bookofwhy

7.13.19 @6:43pm - (Replying to @akelleh) Beautiful introduction to Inverse Probability Weighing of do-adjustment. I hope you have more users than argumentative skeptics. #Bookofwhy

7.13.19 @6:25pm - (Replying to @stuartbuck1 @asacarny and 3 others) I do not see any condescending comments around. I see a request to accept the premise that "graphs go where the knowledge resides". If accepted, the rest is corollaries. If not, lets discuss what relationships the mind may find easier to discern than "Y listens to X" #Bookofwhy

7.13.19 @5:30pm - Thanks for posting these quotes from Karlin and Pearson. Interestingly, today we hear a slightly different song: "Yes, the process generating the data is important," - end of song. No! It's the beginning: How do we represent that "process"? How do we operationalize it? #Bookofwhy

7.13.19 @4:39pm - (1/ ) Interesting News: Google sister-company Verily is teaming with big pharma on clinical trials ... Quotes: 1) Will try to find ways to modernize their clinical trials and speed up the time it takes to bring a new drug to market. 2) aggregate data across a
7.13.19 @4:39pm - (2/ ) variety of sources, 3) Clinical trials have historically been expensive processes that rely on outdated technologies. Does anyone know the technical leaders of this initiative? e.g., Are the versed in the less outdated literature on "aggregating data"? #Bookofwhy

7.13.19 @3:05pm - (Replying to @dkipb) It runs deeper, agree. It's a product of screwed up education in all fields. Today's urgency is to get over those tribal narratives and get things right. And to get things right we need accept that business is not as usual -- those distinctions must be operationalized #Bookofwhy

7.13.19 @5:56am - (Replying to @jack_bowdenjack @RobinGomila) Sure, but this does not make the difference a "causal effect". What makes it a "causal effect" is the model and what the model says about the ignorability condition, which is here taken to hold apriori, thus assuming away the causal component of the problem. #Bookofwhy

7.13.19 @4:13am - (Replying to @NoahHaber @economeager and 2 others) @NoahHaber, I overlooked your thoughtful thread on the comparison of epi to econ. Allow me to reiterate another element: Epi use DAGs because "DAGs go to where knowledge resides". Econ. do not use DAGs because it is culturally prudent to do things in your head. #Bookofwhy

7.13.19 @2:54am - I've just posted an answer on quora ... which ends with a catchy phrase of what #Bookofwhy is trying to do: It tries to "dispel the myth that AI is about to be colonized by statistics. " -- Sharing

7.13.19 @2:35am - I still can't figure out what makes statistics leadership so eager to tell their constituency: Don't panic, there is no confusion, our ancestors have been here before, they touched, they recognized and pointed out, they discussed and distinguished, business is as usual #Bookofwhy

7.13.19 @2:09am - (1/ ) (Replying to @RobinGomila) I think the sentence: "I draw on econometric theory and established statistical findings to demonstrate that linear regression (OLS) is generally the best strategy to estimate causal effects on binary outcomes" will make some economists rethink if "Almost Harmless" is truly
7.13.19 @2:20am - (2/ ) (Replying to @yudapearl @RobinGomila) as harmless as some claim. It seems to me that you are dealing with strategies to estimate conditional expectations, not causal effects. This potential confusion is prevalent in the new culture schooled by structure-less econometrics.

7.12.19 @2:57pm - This "Mediation Formula" always restores my confidence in the power of very simple mathematics to produce very meaningful results. #Bookofwhy

7.12.19 @2:38pm - (1/ ) (Replying to @asacarny @paulgp and 2 others) I am not an NBER member, but ocassionally producing results that are of interest to economists. The exclusive policy of NBER prevents its members from learning these results. This in itself may not be sufficient reason to change policies, but the reputation of econometrics
7.12.19 @2:44pm - (2/ ) (Replying to @yudapearl @asacarny and 3 others) as an insular and outdated field does hurt its members, and increasingly so in the age of data-science. I therefore welcome your proposal to open another channel of communication between economics and the rest of the scientific community. #Bookofwhy

7.12.19 @3:31am - (Replying to @yudapearl @AndroSabashvili @_MiguelHernan) Gee, I think I know why they discarded explanations; it goes back to the RCT's roots and the philosophy of "well defined interventions". You cannot define "undoing past events" unless you have manipulations; plain "events" cannot explain things, only manipulations can. #Bookofwhy

7.12.19 @3:21am - (Replying to @AndroSabashvili) I am not sure "explanation" is part of @_MiguelHernan taxonomy, because explanations demand retrospective counterfactuals, not "predictive counterfactuals". Thus, scientists quest for explanation remains outside data science. I wish I knew why they changed the ladder. #Bookofwhy

7.12.19 @3:02am - (Replying to @AndroSabashvili) You are very right. I wrote to @miguelHernan: 5.7.19 @1:47pm - I am questioning the benefit of separating "description" from "prediction", skipping "diagnosis" and lumping together "intervening" and "retrospecting" under one opaque category "causal inference". No-Ans. #Bookofwhy

7.11.19 @10:21pm - (1/ ) (Replying to @PavlosMsaouel) 1/ Glad I found another fan. Shmueli's paper had many (1,200) citations which is encouraging. Statisticians hit the roof when I tell readers how causality was neglected in 20-century statistics. Glad they have accepted Shmuei's story. It seems that David Hand wrote his new paper
7.11.19 @10:40pm - (2/ ) (Replying to @yudapearl @PavlosMsaouel) to refute Shmueli's conclusions but, in my opinion, he actually reinforces them with additional evidence provided by the refs to Box, Breiman (2001) and Cox (1990). I hope #Bookofwhy sparks renewed discussion on the nature of modeling.

7.11.19 @8:43pm - (1/3) While struggling to find a ladder or demarcation lines between rungs, Hand's paper called my attention to a comprehensive article by Shmueli (2010) "To Explain or to Predict?", rich with references and quotes, which states: "Although not explicitly stated in the methodology
7.11.19 @8:43pm - (2/2) literature, applied statisticians instinctively sense that predicting and explaining are different." ... Going through the quotes and references in Hand's paper, I believe they confirm Shmueli's verdict: "the statistical literature lacks a thorough
7.11.19 @8:43pm - (3/3) discussion of the many differences that arise in the process of modeling for an explanatory versus a predictive goal." One may also note that, adding to a "thorough" discussion, the Ladder of Causation provides "formal distinctions" between those differences. #Bookofwhy

7.11.19 @4:49pm - (Replying to @namalhotra @matt_blackwell and 18 others) Whence do you think comes the urge to condition, despite "Didn't we know this already". Is there some textbook guideline that makes good people forget what they know? #Bookofwhy

7.11.19 @3:53pm - (Replying to @Scientific_Bird @elizpingree @BridgetPhetasy) You will not regret it. PRIMER is ideal, especially to Inquisitive-minded Birds.

7.11.19 @3:47pm - (Replying to @cutearguments @rlmcelreath) I was not aware of "Virtual History" when writing #Bookofwhy. Thanks. Gelman, on the other hand, is wrong; Rubin denies any meaning to "undoing past events". That's why he replaced "counterfactuals" with "potential outcomes".

7.11.19 @3:35pm - (Replying to @namalhotra @matt_blackwell and 18 others) This is quite shocking: "Overall, we find that 46.7% of the experimental studies published in APSR, AJPS, and JOP from 2012 to 2014 engaged in posttreatment conditioning (35 of 75 studies). " Something to keep in mind in scrutiniaing "real-life" experimental studies. #Bookofwhy

7.11.19 @6:48am - (Replying to @cdsamii @ingorohlfing and 14 others) What are the cultural roots of path models in political science? Duncan and LISREL? or some other source? When and how did PO enter into this field, as we can see in the works of King and Imai?

7.11.19 @3:56am - (Replying to @nasim_rahaman @MPI_IS) Gee, I recognize the books behind me, but not the Balcony. I hope it is not a "cross world" illusion. #Bookofwhy

7.11.19 @3:51am - For the physicists among us, especially those fascinated by quantum mechanics and Bell Inequality, DAGs appear to enlightened the conversation when we allow communication among observers, see

7.11.19 @1:34am - (Replying to @ingorohlfing @thosjleeper and 13 others) This is a good paper, thanks for posting. DAG's are used here not merely to convey assumptions, but also to visualize violations of assumptions and how these correspond to PO expressions. I can't imagine any paper on racial discrimination skipping DAGs in those roles #Bookofwhy

7.11.19 @1:11am - (Replying to @yskout @Jabaluck and 11 others) Robustness comes from making assumptions in the language of stored knowledge. And logical equivalence is not "modeling equivalence. Ask an economist to tell you if HIS latest model has testable implications. Can he do it? Sure! But compare his effort to students of DAGs#Bookofwhy

7.11.19 @12:49am - (Replying to @thosjleeper @PHuenermund and 12 others) I am an optimist. I imagine 18K out of our 19.1K followers to be silent observers today and vicious rebels tomorrow; they are not dumb. Yes, they have to worry about tenure, publications and other suppressants of academia, but they are no dumb. They see the potentials. #Bookofwhy

7.11.19 @12:31am - (Replying to @thosjleeper @PHuenermund and 12 others) I believe Imai and his students are recent converts and, in social science, we have Morgan, Winship and Elwert. #Bookofwhy

7.11.19 @12:21am - (Replying to @JDHaltigan @WiringTheBrain @PeterJungX) Agree. Mediation analysis is one of those areas where regression culture is deeply entrenched, and you can still find PhD theses written in stone-age style. I noted it in pp.324-5 of #Bookofwhy,

7.11.19 @12:12am - (Replying to @WiringTheBrain @PeterJungX) I am not sure to what extent your concern is valid today, 30 years after the Causal Revolution. I know there are pockets of regression analysts who refuse to elevate themselves from statistical thinking, but most of those in my circle are aware of ANCOVA's dangers #Bookofwhy

7.10.19 @11:05pm - (1/4) Happy Anniversary! About a year ago, I started Twitting and posted this: 6.27.18 - Hi everybody, the intense discussion over The Book of Why drove me to add my two cents. I will not be able to comment on every tweet, but I will try to squeak where it makes a difference....
7.10.19 @11:05pm - (2/4) A year later, I can hardly believe the 2,300 Tweets behind me, 19.1K followers, and a pleasant sense of comradeship with the many inquisitive minds that have been helping me demystify the science of cause and effect. I have benefited immensely from seeing how causality is
7.10.19 @11:05pm - (3/4) bouncing from your angle and how it should be presented to improve on past faults. To celebrate our anniversary, I am providing a link to a search-able file with all our conversations (my voice) sorted chronologically ... Today I re-read selected sections
7.10.19 @11:05pm - (4/4) and, believe me, there are quite a few non-trivial nuggets out there that should be retrieved, re-weaponized and reused. Finally, access to individual chapters of PRIMER is now available by clicking here Cheers, #Bookofwhy.

7.10.19 @6:38pm - (Replying to @MatthewMOConnel @GoldsteinBrooke @TuckerCarlson) We are shifting from defense to attack and, with your help, will make Zionophobia the ugliest word in town. Use it!

7.10.19 @5:21pm - (1/ ) (Replying to @fuzzydunlop123 @autoregress and 11 others) Economics? "Open to new methods"? The last open-minded economist I met was Hal White (1950-2012). When he passed away, his former students could no longer publish in top journals, and had to revert back to old methods (my interpretation). Now observe how hard Heckman and Pinto
7.10.19 @5:26pm - (2/ ) (Replying to @yudapearl @fuzzydunlop123 and 12 others) labor to preserve old methods ; pages after pages of derivations just to refrain from using d-separation. "Open to new methods"? Let's see how long it will take to Editors of top Econ Journal to invite ONE review paper on what graphical models can offer
7.10.19 @5:41pm - (3/ ) (Replying to @yudapearl @fuzzydunlop123 and 12 others) econometric researchers. This is 2019, almost 40 years after graphical models came into being and impacted almost every data-driven discipline, except economics. It's hard to believe, I agree, and surely our silent rebellious students will not call it "open minded" #Bookofwhy

7.10.19 @5:14pm - (Replying to @autoregress @fuzzydunlop123 and 11 others) Having been burned before by ungrounded theory is no excuse for refusing an eye-glass that helps you navigate your OWN theory.

7.10.19 @4:36pm - (1/ ) (Replying to @Jabaluck @yskout and 11 others) I agree, but would phrase it a bit differently. Any economist NOT familiar with DAGs would rejoice knowing that his/her most intimate chunks of economic knowledge can now be expressed in a scientifically prudent language, uncontaminated by parametric or statistical baggage
7.10.19 @4:51pm - (2/ ) (Replying to @yudapearl @Jabaluck and 12 others) ready to be submitted for mathematical or algorithmic analysis, in which each step is meaningful, that is, scrutinizable by the knowledge-providing economist, and which readily delivers answers to questions that otherwise take hours to answer. How about this phrase? #Bookofwhy

7.10.19 @3:36pm - (1/ ) (Replying to @Jabaluck @yskout and 11 others) 1/ Allow me to use your favorite: "You are missing my point". The key point that I am trying to make is that there is such a thing as "human representation of knowledge" and it has a cognitive library of primitive relationships from which more complex relationships are composed.
7.10.19 @3:44pm - (2/ ) (Replying to @yudapearl @Jabaluck and 12 others) The compound relation "prehospital behavior that might independently effect outcomes", is composed of many primitives relationships, each of the type "X directly affect Y." DAGs tap those primitive relationshiop directly, hence reliably. Compound relations require mental
7.10.19 @3:52pm - (3/ ) (Replying to @yudapearl @Jabaluck and 12 others) construction effort and are more vulnerable therefore to human error. This is the key to the clarity of DAGs, not "familiarity". "DAGs go to where knowledge resides" I once said. Any discussion of DAGs should start with this key observation; the rest are corollaries #Bookofwhy

7.10.19 @2:54pm - (1/2) (Replying to @Jabaluck @yskout and 11 others) 1/2 @Jabaluck Your resistance to DAG betrays your cultural upbringing (Rubin? Angrist? Imbens?) and refutes your own words: "We control for differences in prehospital behavior that might independently effect outcomes." Anyone who can judge if a difference "INDEPENDENTLY AFFECT
7.10.19 @3:04pm - (2/2) (Replying to @yudapearl @Jabaluck and 12 others) THINGS" can surely judge if "one variable directly affects another". The latter type of judgments is all that is required for constructing a DAG, hence it could not be as "incredibly challenging" as you describe it. This is universally true for any representation of knowledge
7.10.19 @3:11pm - (2/4) (Replying to @yudapearl @Jabaluck and 12 others) 2/4 and it does NOT depend "on whether people are familiar with the terminology." To semi-prove my point, notice that your recipe "We control for differences in prehospital behavior that might independently effect outcomes" is deficient, for it misses controls for variables that 7.10.19 @3:30pm - (4/4) (Replying to @yudapearl @Jabaluck and 12 others) 4/4 affect treatment, not outcome. So, it is hardly the case that "The language economists use seems efficient and unambiguous here." The language is ridden with ambiguities, which calls into question the credibility of key judgments issued by DAG-avoiding economists. #Bookofwhy

7.10.19 @2:13pm - (1/2) (Replying to @PeterJungX @WiringTheBrain) Oh, how I wish to see responses of statisticians to this question! How they interpret the words "works", "control for" "confounders" etc. How their answers vary depending on what rival camp they declare allegiance to, and more. Can we get a summary? My answer: Yes, if you are
7.10.19 @2:26pm - (2/2) (Replying to @yudapearl @PeterJungX @WiringTheBrain) versed in causal inference (CI) and its eye-glasses - graphical models. No, if you are a mainstream statistician, believing that "confounding" is a statistical notion. For a more detailed answer, see "confounding bias" pp. 53-60 of PRIMER #Bookofwhy

7.10.19 @5:12am - (Replying to @EpiSconroy @EpiEllie) Stat theory include probability theory, regression analysis, hypothesis testing, confidence intervals, etc. all are theories of the joint distribution functions that govern the observed data. #Bookofwhy

7.10.19 @3:58am - (1/ ) (Replying to @Jabaluck @autoregress and 10 others) I did not say I'll never convince you. I said: the reason I am spending time on tweeter is not in hope of convincing you or Angrist. I do it to empower the curious yet silent rebels among econ. students, what seems like an easier task. You (not sure about Angrist) will be
7.10.19 @4:15am - (2/ ) (Replying to @yudapearl @Jabaluck and 11 others) convinced (I hope) when one of these students asks in class: "Why can't we test the exclusion restriction by checking if E(Y|x, z) depends on z?" which will make you sorry for not teaching d-separation. You asked: what I have learned from economists, a question bothered me a
7.10.19 @4:36am - (3/ ) (Replying to @yudapearl @Jabaluck and 11 others) a lot, because I was hoping to tell my students: Economists have developed methods of solving problem A, R and T. But, aside from LATE, I have hard time giving A, R and T content that I can easily describe and comprehend. When I ask colleagues they send me to fancy articles
7.10.19 @5:03am - (4/ ) (Replying to @yudapearl @Jabaluck and 11 others) without telling me what nuggets of wisdom I can expect to find there if I try really hard. I made this plea on Twitter with not much success. But I have not given up; I know those nuggets exist and are waiting to be excavated. Perhaps by tomorrow's rebels #Bookofwhy

7.9.19 @10:56pm - (Replying to @yudapearl @EpiSconroy @EpiEllie) a carrier of scientific assumptions about the world outside the data (say populations, or individuals reactions to exposures) and should be used to exemplify ideas about scientific methods, hypotheses, evidence, predictions, abduction, inductions, etc etc..

7.9.19 @10:51pm - (Replying to @EpiSconroy @EpiEllie) The theories we learn in psych and stat are different indeed from the epi conception of a "theory". The first, because it was mainly verbal, the second because it was about the data, not the process generating the data. A DAG is a good embodiment of what we mean by a "theory" ,ie

7.9.19 @10:01pm - (1/ ) (Replying to @eliasbareinboim) As a strong advocate of "reality first, algorithm second" I should note that the level at which we model reality is sometime chosen to enable an algorithm. For example, an inference engine may issue the output: Sorry, your model does not allow for the identification of query Q
7.9.19 @10:05pm - (2/ ) (Replying to @yudapearl @eliasbareinboim) However, if you can only think of a variable Z that lies on the arrow X---->Y and satisfy additional properties, I would be able to identify Q using Algorithm-1. Likewise, if you can only think of W that...using Algorithm-2. Thus, the model is refined as we go along.#Bookofwhy

7.9.19 @3:41pm - (1/2) (Replying to @pablogerbas @Jabaluck and 9 others) I think you should get this published someplace, for the sake of people really interested in this applied research, so that they can see the layout of the problem clearly, and discuss substantive issues if any. Alternatively, you can publish it as an educational device to
7.9.19 @3:56pm - (2/2) (Replying to @yudapearl @pablogerbas and 10 others) enlighten X-econs with the way CI folks think about a problem, w/o dismissive calls for "real life problems". Or, perhaps a joint paper by all tweeting discussants. It will turn into a classics, perhaps even an underground "bubble-burster." I'll support it fully. #Bookofwhy

7.9.19 @3:30pm - (Replying to @autoregress @PHuenermund and 10 others) Before Galileo pointed his telescope towards the moon, he tried it on a tree, 2 km away, and saw nothing new, just the familiar old tree, but smiling in freshness. This story is fiction, and I am no Galileo, but DAGs are the eyeglasses of CI, no devils behind them. #Bookofway

7.9.19 @1:17pm - (1/2) (Replying to @autoregress @PHuenermund and 10 others) Sorry, I did not say "incapable". I curiously asked "what % of students can solve it?" Why? Because one can read "Harmless Economics" and "Mastering Metrics" 10 times over and find no clue on how to solve it. Moreover, I am sure that there are many secret rebels among those
7.9.19 @1:30pm - (2/2) (Replying to @yudapearl @autoregress and 11 others) students and readers, you may be one of them, who could not sit still seeing CI advancing to new heights and continue to act as if these advances have no bearing on economic problems. So, I am not surprised that % is rising and will continue to rise. I am tweeting here
7.9.19 @1:46pm - (3/3) (Replying to @yudapearl @autoregress and 11 others) not to convince Angrist or Jason, but to empower their curious and capable students to see through the X-Eco-bubble. And, yes, I am confident, very confident, that the bubble will burst as soon as one of them tells the others: Hey, look at this microscope! #Bookofwhy

7.9.19 @1:03pm - (Replying to @Jabaluck @PHuenermund and 9 others) You are again distorting my words, perhaps for realizing their truth. I never uttered the word "inferior". I said "suspect, for lack of ground truth". A doctor who insists on operating on patients w/o studying anatomy may be a great surgeon, but still highly suspect. #Bookofwhy

7.9.19 @6:30am - (Replying to @stephensenn @omaclaren @eliasbareinboim) Let's forget then the iid assumption and estimate E[Y|X=x] by OLS regression. Do I need to know about block design? I just randomized people to treatment and control by a fair coin. Where will I go wrong? #Bookofwhy

7.9.19 @6:10am - (Replying to @stephensenn @omaclaren @eliasbareinboim) Consider: (1) I randomize a treatment X and record data in the form of x,y pairs. (2) I forget that these pairs came from RCT and imagine that they are iid samples from some distribution P(x,y) (3) I estimate E(Y|X=x) under this illusion. Will the illusion hurt me? #Bookofwhy

7.9.19 @5:08am - A new explanation-seeking paper that hit my screen:

7.9.19 @4:27am - (1/2) (Replying to @omaclaren @eliasbareinboim @stephensenn) Yes, now that we find one reader thinking that the complete version makes a difference, we will try to include it in the upcoming paperback version. I personally think that it was his "reduction of data" mantra that defined 20th Cent. statistics agenda. Recall that Fisher
7.9.19 @4:37am - (2/3) (Replying to @yudapearl @omaclaren and 2 others) did not have notation for "causal effects". Even in the context of RCT, his concern was the reduction of data obtained from a randomized trial, not to assure correctness or unbiasedness. When he tried that (in mediation context), he blundered (ref = Rubin), #Bookofwhy. He
7.9.19 @4:44am - (3/3) (Replying to @yudapearl @omaclaren and 2 others) definitely did not have the concept of "causal assumption", which is essential for every task of modern "causal inference." Going from RCT-data to "causal effect" is indeed a matter of data reduction, the causal part is already prepared for you in the design. #Bookofwhy

7.9.19 @1:54am - Thanks you @PHuenermund for reading a "real life" study, and showing us that it is made up of the same biological tissues as "toy problems", only more of them. #Bookofwhy

7.9.19 @12:20am - (1/ ) (Replying to @Jabaluck @fuzzydunlop123 and 7 others) Please do not distort my words. You make it very unpleasant to interact with you when you do so. In our exchange I did not mention "economists" but X-econs, namely, model-avoiding economists of the quasi-experimental school. Nor have I mentioned DAGs, I spoke "models", which
7.9.19 @12:29am - (2/ ) (Replying to @yudapearl @Jabaluck and 8 others) includes structural economics. Finally, I never said: "economist do things wrong". I said there is no way of knowing if X-econs do things wrong, in the absence of ground truth, but it sounds very funny when a whole field prides itself on solving huge "real life" problems, but
7.9.19 @1:07am - (3/ ) (Replying to @yudapearl @Jabaluck and 8 others) ONLY huge "real life" problems, not "toy problems" which have ground truth and where everyone can see if you solved it correctly or not. Such problems are avoided like a plague, labeled "toy", "made-up" and worse, but never discussed in good company. I said "it sounds funny"
7.9.19 @1:14am - (4/ ) (Replying to @yudapearl @Jabaluck and 8 others) not in the hope of convincing you to take a step back and see how funny it is, but in the hope of confirming the feelings of hundreds, perhaps thousands of econ. students who I know are listening silently to this twitter exchange, perhaps after reading #Bookofwhy in hiding,
7.9.19 @1:22am - (5/ ) (Replying to @yudapearl @Jabaluck and 8 others) and asking themselves 3 times a day: Isn't funny that my professor can solve "real life" problems of such magnitude and importance, and he/she cant tell which scenario (scenario! not DAG) contains a legitimate IV? I tweet here to tell this silent student: You are not alone,
7.9.19 @1:38am - (6/ ) (Replying to @yudapearl @Jabaluck and 8 others) it is mighty funny, but it is going to change in a few years, and you can prepare yourself for the day when your field will change from "funny" to "well informed". Just make sure you spend the 2-3 hours it takes to acquire the art of causal modeling. I'll be there for you.

7.8.19 @11:50pm - (Replying to @Jabaluck @fuzzydunlop123 and 7 others) Show me one simple case where the exclusion requirement need not be justified. Now lets go to any "real life" problem and examine HOW it was justified, if at all. I cant find a model (not a DAG, a model) in the papers you ask me to read, so how can one tell if the reported
7.9.19 @12:03am - (2/ ) (Replying to @yudapearl @Jabaluck and 8 others) results are not biased substantially by violations of exclusion? You know that there is objective validity test to the results reported, and you are asking me to go through the numbers and show that they can do better with DAGs. The 4 scenarios we discussed tell us more about
7.9.19 @12:11am - (3/ ) (Replying to @yudapearl @Jabaluck and 8 others) problems with exclusion that 100-page "real life" study where those problems are not modeled (forget DAGs, by ANY model). If you think I could learn something from a "real life" article, please tell me what principle I can gain from it that I may not know already. #Bookofwhy

7.8.19 @11:09pm - (Replying to @Jabaluck @autoregress and 6 others) Retweeting: Sorry, these are not "toy problems"; they haunt each & every IV exercise, albeit suppressed by practitioners to the mercy of intuition, so as to escape "struggle". Each one represents 10,000 "real world" problems in which "exclusion" is/was justified. #Bookofwhy

7.8.19 @9:28pm - (Replying to @Jabaluck @autoregress and 6 others) Wrong. I answered it head on: "econs use too many in lieu of other approaches". I even proposed an explanation for this imbalance: X-econs avoid models because they do not know how to handle them mathematically. Care to estimate of % of Angrist's students who know how?#Bookofwhy

7.8.19 @9:11pm - (Replying to @Jabaluck @autoregress and 6 others) Sorry @Jabaluck, these are not "toy problems", because they haunt each & every IV exercise, albeit suppressed by practitioners to the mercy of intuition, so as to escape "struggle". Each one represents 10,000 "real world" problems in which "exclusion" is/was justified. #Bookofwhy

7.8.19 @8:59pm - (Replying to @omaclaren @eliasbareinboim @stephensenn) I wish someone would plot the frequency of the words "cause" or "causal" vs. time, from 1900 to 2019 and tell us, straight face, "No! There was no causal revolution in 1990s". (There were 13 such papers in JSM 2003) Why are u resisting the idea of a paradigm shift? #Bookofwhy

7.8.19 @8:29pm - (Replying to @omaclaren @eliasbareinboim @stephensenn) Statistics as a field is defined in two arenas. (1) What its leaders say and do. (2) What its textbook say and do. For (1) I read the presidential addresses in the past 20 years. For (2) I look at the index. #Bookofwhy

7.8.19 @8:24pm - (Replying to @omaclaren @eliasbareinboim @stephensenn) I dont think #Bookofwhy states that 1)Statistics doesn't deal with causality. It quotes Fisher's definition: "Statistics is summarizing data," and it decries the vacuum in stat textbook, and it places Fisher's DoE precisely where it belongs in history of CI. What would you add?

7.8.19 @4:59pm - (1/ ) @autoregress I am not an extremist, so I am not worried about econs missing out on complex DAGs. I am concerned about econs missing out on their simple IV models like the ones discussed here: ... 4 scenarios, 4 variables and we havn't found an X-eco who
7.8.19 @5:28pm - (2/ ) (Replying to @yudapearl @autoregress and 6 others) 2/ was able to tell us which of the 4 scenarios has a legitimate IV. Bad sample? Perhaps. So let me ask: How many of Angrist's @metrics52 students can do it? And if this does not jolt X-econs to do some honest soul-searching, what will? Their credibility is at stakes. #Bookofwhy

7.8.19 @1:37pm - (Replying to @Jabaluck @autoregress and 5 others) @Jabaluck I believe you misunderstood me. I have not used the words "back-door" in any of my tweets (to u). I used "sources of variation" which was your term. My main point is: X-Eco refrain from using a model for two reasons: (1) The mistrust the assumptions behind the model and
7.8.19 @2:00pm - (2/ ) (Replying to @yudapearl @Jabaluck and 6 others) (2) They mistrust models period, even when they trust a set of assumptions. I am only concerned about reason (2). If I am wrong, I will convert back to X-Eco. How can you prove me wrong? Show me one X-Eco paper that explicates and combines assumptions formally #Bookofwhy

7.8.19 @1:32am - (Replying to @vijayant_k @Canadian_JACD and 2 others) We are concerned with the effects of BP not with its causes. If it has an effect on cardiac stress we should be allowed to include the arrow BP--->Cardiac Stress in our model, and not to ask for supreme court permission.

7.8.19 @12:57am - (1/ ) (Replying to @mc_hankins @stephensenn @eliasbareinboim) 1/ Since we started working on external validity, transportability and data fusion in 2010 (eg, ) we have been hearing the whisper: "This sounds like meta-analysis." Yes we are yet to find a meta-analyst expert who can to tell us how to handle the simplest
7.8.19 @1:07am - (2/ ) (Replying to @yudapearl @mc_hankins and 2 others) example. (and I went to the very top). Why? Because we can show that the answer depends on the causal relationships between X1,X2,Y and other factors in the problem. Meta-analysis is a statistical pooling method that is oblivious to those relationships. Therefore, I take the
7.8.19 @1:17am - (3/ ) (Replying to @yudapearl @mc_hankins and 2 others) liberty to assume that meta-analytic efforts are orthogonal to the kind of problems we are trying to solve. However, I will immediately change my mind if I find a Meta-Analyst who solves the examples of or Fig. 1 of #Bookofwhy

7.7.19 @11:02pm - (Replying to @yudapearl @Jabaluck and 6 others) 4/ explicitly what I understand about each of those sources separately, and try to combine those understandings mathematically to decide if Z is a good IV. I was tempted to do (2) but, then, I realized to my horror that I am re-committing DAG-heresy. What shall I do? #Bookofwhy

7.7.19 @10:36pm - (1/ ) (Replying to @Jabaluck @MariaGlymour and 5 others) @Jabaluck your arguments are so persuasive that I decided to quit DAGs and convert to X-Eco (short for Experimental Economist). The moment of conversion was truly enlightening and, as if by divine revelation, I began to "understand the sources of variations" of variables,
7.7.19 @10:44pm - (2/ ) (Replying to @yudapearl @Jabaluck and 6 others) something DAG folks never understand. Fired by enlightenment, I chose Z as my potential IV, and I understood the "source of variations" of many other variables, X,Y,Z,W,S, T, V, all related to Z, to X and to Y. I truly understood those "sources of variations, and I felt elated
7.7.19 @10:51pm - (3/ ) (Replying to @yudapearl @Jabaluck and 6 others) Now I had to decide whether Z, which looked like a promising IV before, is still a good IV given what I understood about all the other background variables. I had two options: (1) Decide yes or no based on my understanding of all those sources of variations or (2) articulate

7.7.19 @10:22pm - (1/2) Sailors and passengers on this voyage of CI research should be interested in these new results that just reached my screen, , by @eliasbareinboim & team. Suppose we randomize treatments X1 and X2 in two separate studies, can we estimate the causal
7.7.19 @10:22pm - (2/2) effect of their conjunction (X1=x1, X2=x2) ?? I see dozens of pharmaceutical companies rushing to join our voyage. #Bookofwhy

7.7.19 @9:39pm - (Replying to @eliasbareinboim @EpiEllie and 2 others) I am willing to adopt the new construction method if it helps science. But we are still in need of a rule on when to include an arrow BP--->Y and when not to, assuming that we sear never to ask for the effect of BP on anything. What should we think about to decide? #Bookofwhy

7.7.19 @8:56pm - (1/ ) This is a brilliant idea, to use the causal hierarchy in "critical thinking" classes: Given an English expression, classify it into rung-1, 2 or 3. Some of the top researchers of our time should take it. I remember when our kids were young we got them a game called "propaganda"
7.7.19 @8:56pm - (2/ ) Each card had a false argument and players had to classify it into one of several types of falsehood. Great!! As to a simpler/gentler version for high schoolers? The #bookofwhy is all we have, but it should be fun to do one, with a creative cartoonist. Great!!

7.7.19 @5:59pm - (Replying to @yudapearl @raymondshpeley @f2harrell) certain aspect of the distribution, eg P(label|features), which is no more of model than say the estimated regression coefficient R, fitted to a cloud of samples. Would you call R a model? Does it carry any assumptions? Only if you havn't seen any of the data. #Bookofwhy

7.7.19 @5:51pm - (Replying to @raymondshpeley @f2harrell) Mathematically, models are carriers of assumptions. Stat models carry assumptions about the distribution, eg normality, binomial,.. and causal models carry assumptions about causal relationships, eg. who listens to whom. A learned neural net provides efficient representation to a

7.7.19 @4:50pm - Answer to What are the differences between the "experimentalists" and "structuralists" approaches to econometrics? by Judea Pearl

7.7.19 @2:57pm - (Replying to @vijayant_k @Canadian_JACD and 2 others) According to all experts I talk to, BP is not just an indicator of some risk factor, but it actually causing bad things, eg. heart stress. We do not want to go to extremes and claim that we are only seeing the display on the measuring device, not the BP itself. #Bookofwhy

7.7.19 @12:44pm - (Replying to @f2harrell @raymondshpeley) A point which I should have added (based on @f2harrell tweet) is that whereas statistics can give us an (estimated) distribution, DL gives us one aspect of the distribution, eg P(label|features). DL folks may argue that, in principle, they can learn the P as a function X-->(0,1)

7.7.19 @12:33pm - (Replying to @MariaGlymour @PHuenermund and 5 others) Who sold us "empirically validated" ointment? When I say a "valid instrument" I mean one that satisfies the the IV requirement according to the model that was proposed. Plus, that model may have testable implications.

7.7.19 @5:48am - (1/2) (Replying to @PHuenermund @Jabaluck and 5 others) I guess what the experimentalists are arguing is as follows: If you are committed NOT to put down any model on paper, and to work purely by intuition, then it is easier to start by asking yourself: "Is there an exogenous variable around that is somehow related to X and Y"? as
7.7.19 @5:58am - (1/3) (Replying to @yudapearl @PHuenermund and 6 others) as opposed to asking yourself: "Is there a way to control ALL confounders of X and Y?". Next comes the handling of "somehow related". If you are still committed to model-blind thinking, you will have to intuit handwaving about the exclusion restriction, which experimentalists
7.7.19 @6:07am - (3/4) (Replying to @yudapearl @PHuenermund and 6 others) are willing to stomach given the ease of having found SOME relevant exogenous variable in the mental forest. What they forget is that there is a middle ground: Find your favorite exogenous variable, model what you know about the relationships with X,Y and other factors, then
7.7.19 @6:19am - (4/5) (Replying to @yudapearl @PHuenermund and 6 others) use your modeling tools to decide if you have a valid IV, if not, can you repair it, can you find another, can you test your assumptions etc etc, all the nice thing that models give you, rather than remaining in the intuitive world. The trade off will be settled only when
7.7.19 @6:28am - (5/5) (Replying to @yudapearl @PHuenermund and 6 others) experimentalists agree to learn how to model problems with 5-10 variables, and see for themselves what they have missed by resisting it. #Bookofwhy

7.7.19 @4:51am - (Replying to @pentagoniac) Evidently I violated some cosmic rule of Quora protocol. No idea what was violated nor who vowed for my innocence.

7.7.19 @4:16am - Victory! My appeal was accepted. My answer on "econometrics, statistics, and machine learning" was reinstated on Quora, and my honor restored. I bet they were impressed by the clean life I live. #Bookofwhy

7.6.19 @7:51pm - (Replying to @mattshomepage) It is a mystery which I tried to answer in chap. 10 of #Bookofwhy. Millions of years of playful experiments with bows and arrows are encapsulated in our culture, our language, our elderly wisdom, our books etc. The practical question is: how to represent it and how to exploit it.

7.6.19 @7:28pm - (Replying to @y2silence @PWGTennant and 3 others) No one knows what they say. All we hear is: If A is not randomized (or randomizable) then "it all depends".

7.6.19 @4:49pm - (Replying to @imleslahdin) Answer deleted!! Thanks for telling me. I guess the vice-squad just found out about me. I appealed. Lets see how they handle "appealers".

7.6.19 @4:18pm - (1/2) (Replying to @AngeloDalli @EpiEllie and 2 others) An arrow BP--->Y has two interpretations: (1) Y listens to and responds to changes in BP. (2) Manipulating BP changes Y. (1) implies (2), but some may argue that (2) is all we can observe, hence it is "scientific" while (1) is "meta-physics". Now, since BP is not manipulable,
7.6.19 @4:26pm - (2/2) (Replying to @yudapearl @AngeloDalli and 3 others) adding an arrow BP--->Y to your model makes you suspect of membership in the "meta-physics" camp, which carries harsh consequences. This is why the BP issue is foundational, and that is why we see such hesitations and "it all depends" from the manipulationist camp. #Bookofwhy

7.6.19 @4:04pm - (Replying to @f2harrell) Do you think statistics offers more than "taking us from samples to properties of distribution functions" ?? If you do, then I confess to underplaying the role of #statistics. But did I? #Bookofwhy

7.6.19 @2:16pm - (Replying to @neuro_data @danilobzdok) Progress in plain geometry was very very advanced before someone said: lets add a 3rd dimension. If we want to remind people that the world is 3-dimensional it is helpful, I think, to label 3D-geometry "advanced".

7.6.19 @2:07pm - (Replying to @yudapearl @JaapAbbring) Here is a brilliant idea. As co-Editor of the Journal of Causal Inference I will invite you to tell our readers how they can benefit from tools developed in Econ. and you will reciprocate by convincing a mainstream Econ. journal to do likewise. Do we have a deal? #Bookofwhy

7.6.19 @1:59pm - (Replying to @yudapearl @JaapAbbring) involves graphical models, do-calculus, counterfactual logic etc. which do not rule out tools produced in econ. to deal with part (1). I see wide awakening on both sides to examine and evaluate each other tools. But it is only a Twitter-awakening, not in journals or NBER. Wait!!

7.6.19 @1:49pm - (Replying to @JaapAbbring) The words "relatively simple" are Imben's words, with which he justifies why he can avoid graphical models. See ... SCM has two parts: (1) a model of reality and (2) tools for dealing with (1). The first is identical to structural economics. The second

7.6.19 @1:29pm - (Replying to @danilobzdok @f2harrell) Neural network is a man-made artifact, crafted to capture some aspects of the data, say the relationship between symptoms & disease. By a "model" we mean a picture of reality, from which the data emerged. e.g. X--->Y and X<---Y are two different models that may yield same NN.

7.6.19 @1:21pm - (Replying to @f2harrell) I agree. Most neural networks do not involve models of the data-generation process. Agree. I was afraid you have found points of disagreement.

7.6.19 @2:36am - (Replying to @robertwplatt @Canadian_JACD and 3 others) Would adding or not adding an arrow from BP change whatever you are interested in doing? If it does, Eureka! We just discovered the scientific meaning of an arrow going out of nonmanipulable variable. Now we ask: How would you decide whether to add or not to add? #Bookofwhy

7.6.19 @1:49am - (1/2) (Replying to @JaapAbbring) I do not recall describing econometric models as "relatively simple," especially not structural econometrics, which can be viewed as a subset of SCM, lacking graphical tools and do-calculus. I may have used the phrase "relatively simple" to describe the kind of problems
7.6.19 @1:58am - (2/2) (Replying to @yudapearl @JaapAbbring) that can be handled by IV methods Angrist style, where no model is laid down for analysis and scrutiny, or PO methods Imbens-Rubin style, where ignorability assumptions beg to be discerned. Structural models + graphical tools is a winning combination #Bookofwhy

7.5.19 @11:49pm - (Replying to @davidmanheim @kareem_carr @EpiEllie) True, basic econometrics was meant to do CI, but look at the tools they are still using in the 21st century, not even realizing what tools are missing from the Econ-Echo-Chamber. As to DNN's etc. they were "advanced" five years ago; times they are achangin. #Bookofwhy @causalinf

7.5.19 @10:32pm - (Replying to @dkipb @EpiEllie and 3 others) I have not received comments on "Q as a limit" or on ... (which is a beautiful paper, no bias), so we do not know if it helped resolve the hesitations. But "first pass" is a cop out; this is a foundational issue: Listening vs. manipulating. #Bookofwhy

7.5.19 @10:00pm - (Replying to @PWGTennant @EpiEllie and 2 others) God forbid. What I can or cannot obtain has nothing to do with the meaning of an arrow coming out of BP, which simply says: someone is listening and responding to BP. I have not heard Ellie&Miguel say "yes", perhaps b/c the notion of "Well-defined" is not well defined? #Bookofwhy

7.5.19 @9:51pm - (Replying to @PWGTennant @EpiEllie and 2 others) Disagree. I believe the inconsistencies and hesitations of the "Not-Well-Defined" school have created more confusion than usefulness. All in the name of mimicking RCT's. #Bookofwhy

7.5.19 @9:44pm - (Replying to @PWGTennant @EpiEllie and 2 others) Identification should not mar questions of meaning. And would stay away for obesity because it is also ill-measured. So its better to focus on BP, which is well-measured and still debatable (at least in some circles). #Bookofwhy

7.5.19 @9:40pm - (Replying to @PWGTennant @EpiEllie and 2 others) I wish I could, like you, agree with both sides. But I still do not know what one side says about an arrow out of BP. When is it legit?

7.5.19 @9:33pm - (Replying to @melb4886 @EpiEllie and 2 others) By saying "because alcohol raises BP" you just admitted the presence of an arrow out of BP. Did you get a license from the guardians of Well-Definedness.? They are still debating if/when such an arrow violates scientific right and wrong. #Bookofwhy

7.5.19 @9:28pm - (Replying to @vijayant_k @Canadian_JACD and 2 others) But if BP is merely an indicator for some other risk, we should not see an arrow going out of BP. But in all DAGs presented here we did see this arrow. Can you check with the experts? Still, the fact that we ask this question means that drawing such arrow make a difference.

7.5.19 @9:10pm - (Replying to @vijayant_k) Last I heard from ML folks was they are all working on "combining CI and ML". I heard if even from people who have no idea what CI means (names withheld). So why would anyone object to be working on "advance ML"? #Bookofwhy

7.5.19 @9:03pm - (Replying to @s_monterohdz) "Advance" does not rule out "super-advance" in the future. I believe I did define the scope of SCM in my writing in "real phenomena", Each DAG you draw represent millions of "real phenomena" that fit the structure. #Bookofwhy

7.5.19 @8:56pm - (Replying to @ADAlthousePhD @EpiEllie) Thanks, I'll try to reply, instead of retweet.

7.5.19 @1:26pm - (Replying to @kareem_carr) Good point. I have been trying hard to have statisticians embrace CI and call it "advanced statistics" watch where they are. As to econometrics, intervention was the goal of even "elementary econometrics,", only the tools that are lacking. #Bookofwhy

7.5.19 @1:10pm - (Replying to @f2harrell) Let's hear ONE disagreement. We settled the previous ones, we can do same with this ONE.

7.5.19 @7:21am - Answer to What are the differences between econometrics, statistics, and machine learning? by Judea Pearl

7.5.19 @2:17am - (Replying to @EpiEllie) We are almost getting there. Quoting: "If.....then we can ask causal questions involving THAT INTERVENTION". But you did not say: "...causal questions involving Blood Pressure". Was it on purpose or by oversight? Moreover, would you then allow an arrow from BD? @Bookofwhy

7.5.19 @1:03am - (1/2) (Replying to @autoregress @EpiEllie) To interpret what @MariaGlymour wrote I would add: g-methods become valid "methods" only after you have a DAG to help you decide what variables to condition on, as in the back-door condition. Additionally, having or not having an IV is also a task dedidable by a DAG. Further,
7.5.19 @1:16am - (2/2) (Replying to @yudapearl @autoregress and 2 others) deciding whether a valid "g-method" exists (namely if identification can be done by regression" takes one glance over a DAG, it takes forever using ignorability assumptions as those used by Rubin, Imbens and Angrist. Lastly, I would not dismiss do-calculus #Bookofwhy

7.4.19 @11:07pm - (Replying to @EpiEllie) So, the answer is YES. We can include BP in the DAG, with an arrow going out of BP, since we do not require that BP be directly manipulable. It is enough that there are Well-defined interventions someplace in the model. Right? Can I tell this to my students? #Bookofwhy

7.4.19 @10:49pm - (Replying to @yudapearl @Canadian_JACD and 3 others) So, by all means, please post more and more DAGs with arrows emanating from BP. The more you post the less credible would their claim become, of the "new philosophy", that they are doing "experimental science" whereas you are doing "metaphysics". #Bookofwhy

7.4.19 @10:34pm - (Replying to @Canadian_JACD @EpiEllie @BL4PublicHealth) The "new philosophy" is the new school of "no causation w/o manipulation" that is brewing at Harvard (perhaps other places?) according to which you cannot talk about X causes Y unless X is manipulable. See @_MiguelHernan papers and for full view #Bookofwhy

7.4.19 @9:30pm - No No, please, continue. You are giving us straight answers: "blood pressure can be an important cause of things like the thickening of arteries". Please continue, because the very notion of "BP can be a cause" is under the danger of extinction, in the new philosophy #Bookofwhy

7.4.19 @9:18pm - I was waiting for this answer "the role of blood pressure in the causal question". It is the kind of answer @EpiEllie could not give, because it violates the dictates of RCT imitation. BP has no well-defined "role", only interventions have "roles". See

7.4.19 @9:07pm - (1/2) Not so easy. So the answer is YES. After thinking, weighting, assuming, understanding, specifying, etc. there might come a moment where you would add BP to the DAG. Now innocent me asks: why? Who needs this ill-defined entity there? Or, what can we do after adding it that
7.4.19 @9:07pm - (2/2) we couldn't do without, with all the thinking, weighting, assuming, understanding, specifying, etc. that gave us the license to add it? After all, if our research question is well-defined: "the effect of some drug", why do we need all this BP nuisance? Who cares? #Bookofwhy

7.4.19 @8:34pm - Not a simple question? OK, I will make it MUCH simpler: Is there any circumstance, after you do the thinking, thinking and more thinking, that you would add "blood-pressure" to your DAG? #Bookofwhy

7.4.19 @7:44pm - (Replying to @JadePinkSameera) It is a good question. And a very simple one too. And that's what makes it so hard to answer.

7.4.19 @7:22pm - "Are economists smarter than epidemiologists?" Our recent Twitter posts made me re-read the blog discussion we had in 2014: ... which I still find to be illuminating of the interplay between cultural and technical forces in scientific progress #Bookofwhy

7.4.19 @7:02pm - So, the simple answer to my innocent question is: NO. @EpiEllie and @causalinferenc and @_MiguelHernan would NOT include a variable "blood pressure" in a DAG before deciding how to weigh all the well-defined interventions on diet, drugs, exercise, etc, etc Am I right? #Bookofwhy

7.4.19 @6:29pm - (1/2) (Replying to @kareem_carr @jenniferdoleac and 2 others) @kareem_carr , you have infinite memory. You just confirmed my last Tweet. Many readers will resist my theory that an innocent curiosity of just two individuals can account for such profound differences between two disciplines. Many will seek differences in substance or
7.4.19 @6:39pm - (2/2) (Replying to @yudapearl @kareem_carr and 3 others) philosophy, or type of data etc etc. No way. It is as simple as your Tweet. Escaping from the echo-chambers of social bubbles is the strongest force that drives scientific progress. And it is becoming harder and harder in the age of internet and social media. #Bookofwhy

7.4.19 @6:15pm - (1/2) On the difference between Econ and Epi, I also said there are no substantive differences whatsoever between the two: ... The current differences in practice emerge from one fluke of history: Robins and Greenland were epidemiologists, not economists. Thus,
7.4.19 @6:15pm - (2/2) they followed curiosity and asked: what can DAG do for us? The rest is history. Why didn't economists follow their curiosity? Let others answer it because I do not want to spoil it for the many econs who ask this very question today. @causalinf , @jenniferdoleac #Bookofwhy

7.4.19 @5:31pm - All these complications need consideration, I agree. But let's ask an innocent question that comes before complications strike: "Can we assume that the variable "blood pressure" may appear in one of your DAGs and, when it appears, there is a arrow going out of it?"#Bookofwhy

7.4.19 @5:18pm - (Replying to @_gbmari) This self-fulfilling effect should be balanced against the spite-him effect of employees that are driven by the challenge to "prove him wrong". We agree I hope that both motivations demand causal models and cannot be accounted for by statistical considerations alone. #Bookofwhy

7.4.19 @4:07pm - Am I right to assume that the variable "blood pressure" may appear in one of your DAGs and, when it appears, there is a arrow going out of it, into other variables. Can we assume that much? #Bookofwhy

7.4.19 @2:02am - Phelps aside, do you think economists today are equipped to define+manage issues of "fairness" with their current tool set? #Bookofwhy

7.4.19 @1:23am - Your paper is an eye opener, glad you posted it. If it was not 1am I would have continued reading it, for it is written in a CI language and a well-structured style. The fact that it won @SIGMOD2019 best paper award may signify a new age for fairness folks. #fairness #Bookofwhy

7.4.19 @12:20am - I do not doubt that adequate causal defs of discrimination and fairness can be constructed by CI folks; they know causal models and counterfactual logic. I questioned the readiness of mainstream ML folks who are lacking those tools. Same goes for economists. #Bookofwhy

7.3.19 @10:35pm - (1/2) Yes, structural counterfactuals escape the torture chambers of the RCT's imitation-game. I think epidemiologists too are on their way to escape those chambers; how else can we interpret the arrows emanating from "blood-pressure" -- a well-measured yet
7.3.19 @10:35pm - (2/2) ill-manipulated variable that certainly has "effects" and often finds itself in DAGs drawn by epidemiologists? @EpiEllie @_MiguelHernan @Lester_Domes @MariaGlymour A puzzle. #Bookofwhy

7.3.19 @7:12pm - Thanks for re-posting, Brooke, and let our mantra for July be: "Make Zionophobia the ugliest word in town"

7.3.19 @7:03pm - For the next step, after #Bookofwhy, I will continue to recommend the PRIMER , , as long as it finds no match in clarity, examples and philosophy. Totally liberated from RCT's and other hangups, and essentially free. #Bookofwhy

7.3.19 @6:54pm - (1/3) Economists' "statistical discrimination," as it turns out, is both (1) the use of statistical associations to discriminate and (2) an attempt to define discrimination using statistical vocabulary alone. According to Phelps (1972), you discriminate whether you hire people by
7.3.19 @6:54pm - (2/3) education, race or zip code, as long as you base decisions on PREDICTED performance, rather than performance itself. So, all learning is "discriminatory" for it is based on past experience which PREDICTS, yet is not equal, the situation at hand. Conclusion: I was right
7.3.19 @6:54pm - (3/3) to suspect criteria entitled "statistical discrimination" and their ability to capture notions such as "fairness," in which causal relations play a major role. @JaapAbbring @steventberry #Bookofwhy

7.3.19 @6:24pm - (Replying to @swadhin_pradhan @geomblog) I dont recall giving such tutorial. Do you have a source?

7.3.19 @5:44pm - The counterfactual framework with which I am familiar (eg ) needs no interventions to have "effects". The paper you cite represents the "potential outcome" framework, a relic of an age when causation was forced-married to RCT (or imitations of). #Bookofwhy

7.3.19 @4:20pm - Statistics itself came from causal motivation for, surely, good predictions are essential for good decision (eg carry an umbrella). The missing ingredient is getting confidence to let these motivations out of the closet and articulate them mathematically #Bookofwhy

7.3.19 @3:31pm - I was under the impression that the issue of "manipulativity" was settled by explicating the non-manipulative aspects of "effects". Examples are: , . #Bookofwhy

7.3.19 @3:02pm - Footnote (1) in this paper ( ) confirms your point and explains why causal vocabulary is not at the "Frontiers" of Fairness research; it takes a generation to undo the statistical thinking that rules ML textbooks, classrooms and research practices.#Bookofwhy

7.3.19 @2:51pm - I cannot see how causal inference could "miss" injustices that statistical inference discovers, when the former represents reality and the latter a silhouette of reality, as projected in data. Example?? #Bookofwhy

7.3.19 @2:43pm - (Replying to @Adelaee) You made my day, Adelaee. The thought that your teenagers are reading my words compels me to write more, to make Zionophobia the ugliest word in town.

7.3.19 @1:48am - (Replying to @_onionesque @geomblog) I join you in this optimism. I have even blessed the pre-scientific hype for channeling good people to study causal and counterfactual models, then turn hype into science. It is happening, agree, because the tools for defining "fairness" correctly are available.#Bookofwhy

7.3.19 @12:48am - Thanks for correcting me on what economists mean by "statistical discrimination". I was not critical of the argument, but of the title, which alerted me to a possible new oxymoron, along with "probabilistic causality", "statistical confounding" "statistical mediation"..#Bookofwhy

7.3.19 @12:37am - (Replying to @NeuroStats) The first paper you cite is indeed the first I have seen on the right track. And I wish ML-folks will take notice. The second is an IV ACE-estimator, a rung-2 exercise, which cannot capture the counterfactual nature of "discrimination" (rung-3). #Bookofwhy

7.3.19 @12:19am - (1/3) If by "statistical discrimination" we mean the use of statistical associations in the decision, then it is perfectly harmonious with the causal definition of "discrimination" and I see nothing wrong with it. What makes me suspicious are attempts to DEFINE discrimination by
7.3.19 @12:19am - (2/3) statistical criteria. Note that in your example, causal considerations are unavoidable, for if the observed characteristics causally produce/prevent necessary skills, our notion of discrimination would change. This sensitivity to causal relations is absent from the fairness
7.3.19 @12:19am - (3/3) literature I have sampled thus far, and which seems to dominate recent discussions in ML. Note also that we now have the tools to define and manage criterion based on combined statistical+causal relations. BTW, the link you gave us is blocked. #Bookofwhy

7.2.19 @11:30pm - I've never dismissed an argument in which statistics and causality appear together. I am suspicious however of arguments that attempt to define inherently causal notions (eg discrimination) in terms of statistical vocabulary ALONE, void of causal relations. Wouldn't U? #Bookofwhy

7.2.19 @7:44pm - The title "statistical discrimination" worries me, because "discrimination" is a causal, not statistical notion. The words "from economics" makes me doubly worried, for reasons explained in #Bookofwhy. But I would love to hear how statisticians define "fairness". Truly curious.

7.2.19 @7:34pm - (Replying to @ahmaurya @alexdamour and 8 others) So how can we explain how FATE-motivated folks can study, write and speak FATE without applying some of the tools that CI is offering the ML community? For example I cant find the tools of Attribution Analysis and Causal Mediation invoked, Or am I missing them? #Bookofwhy

7.2.19 @5:12pm - (Replying to @eschisterman1 @AmJEpi) Congratulations!! Enrique.

7.2.19 @4:00 - (1/2) @marypcbuk @katecrawford @eliasbareinboim @benedictevans All definitions and examples of "fairness" that I have seen revolve around causal and counterfactual considerations, since they concern EFFECTs of policies on different segments of society. Thus, I venture to predict
7.2.19 @4:00 - (2/2) today's "fairness" hype will channel good people to study causal and counterfactual models, and help turn hype into scientific disciplines that define and algorithmitize the kind of fairness we wish AI systems to exhibit. There is virtue in pre-scientific hype. #Bookofwhy

7.2.19 @3:07pm - Agree. With two points of caution:(1) We need to study carefully the theoretical impediments to automated generation, to avoid falling for fake-gold, and (2) We need to study carefully what we can and cannot do with a causal model once we have it. The rest is just math #Bookofwhy

7.2.19 @6:38am - Thank you. I begin to understand. "Bias" is used here in the social sense, like discrimination. I am relieved. In our corner of the wood "bias" has technical meaning, standing for systematic deviation from expectation. eg confounding bias or "selection bias". Relieved #Bookofwhy

7.2.19 @6:26am - Interesting indeed, because I normally find people complaining of deep learning being opaque, here is one complaint: . Optical illusions are known to be mysterious. #Bookofwhy

7.2.19 @6:20am - I am not disputing the article, just expressing surprise at a term that I have not heard before. Can I conclude then that people who talk "AI Biases" mean the same thing as "limitations and biases of correlational ML." like those attributed to Rung-1 of the Ladder in #Bookofwhy?

7.2.19 @6:11am - Replying to @marypcbuk I couldn't pass the pay wall, but your myths are ML myths, which I can understand. I am still to understand what "AI bias" is which is not a myth. Are any of your 9 myths remediable by any of the 7 tools here:

7.2.19 @5:20am - (1/2) I did not realize that people call these problems "AI-bias". Thanks for bringing it to my attention and to the attention of my Twitter followers. Honestly, I have been swimming in the AI pool since the 1970's and never heard the term "AI bias" used before, especially not to
7.2.19 @5:20am - (2/2) describe problems that AI folks are about to solve in the near future. But, as they say in the Talmud: "Never too late to learn". Now, suppose I want to define the term on Twitter. Is every difficulty encountered by an AI program an "AI-bias"?? Which ones are not? #Bookofwhy

7.2.19 @4:36am - (Replying to @pentagoniac @benedictevans) I would love to see an example of a causal problem (say Simpson's paradox) that "*can* be corrected for without causal models. #Bookofwhy

7.2.19 @4:33am - (Replying to @benedictevans) But by saying "AI-bias" you give people the impression that such biases are inherit to AI, namely, permanently irredeemable by any AI program. Do we want to give general audience this impression?

7.2.19 @2:45am - I read @benedictevans on "AI-bias" and I still do not know what he means by "AI-bias", why not call it "ML-bias" or "curve-fitting bias" and how those biases can be avoided w/o attending to causal models as outlined here or here #Bookofwhy? A ML puzzle.

7.2.19 @2:22am - (Replying to @zacharylipton) I dont exactly understand what the time-management problem is. My students only worked on things that other people said are impossible, so mentoring was part of speaking to colleagues. I am not finished though, some colleagues still say they can do astronomy without telescopes.

7.2.19 @1:59am - (1/2) Nice and concise summary of #Bookofwhy. If I were forced to make a critical comment, it would be the way it tries to stimulates readers interest by pointing to existing interest. What if no ML conference had any session on causation, would that make the ideas of #Bookofwhy
7.2.19 @1:59am - (2/2) less compelling for enlightened ML folks trying to build intelligent systems? #Bookofwhy was written to change, not to follow habits. I hope it does.

7.2.19 @12:05am - (Replying to @mimblewabe) I considered changing "do" to avoid criticism like "You cant do this" or "some 'do's change everything". But I decided against it b/c there was no sub. We, Sapiens, do not distinguish between reality and models of reality. Even Abraham asked "what if there were 50...". #Bookofwhy

7.1.19 @11:56pm - Thanks for adding your smiles to this memorable event.

7.1.19 @11:52pm - For something more mathy, I will continue to recommend the PRIMER until something better shows up in the jewelry store. Note that it is now (essentially) open-accessed. #Bookofwhy

7.1.19 @11:22pm - (Replying to @causalinf @dlmillimet and 2 others) Astronomy will never be the same.

7.1.19 @11:19pm - (Replying to @IndexLlc @IARPAnews) I saw the abstracts and pre-BAA material, and wrote to the program manager for more information on the "counterfactual prediction" project. Awaiting his reply.

7.1.19 @11:12pm - DAGs are like optical lenses. First you use them as spectacles, to see things you already know, then you try them as telescopes and microscopes, to see things you never knew existed. Good luck my fellow econs; astronomy will never be the same. #Bookofwhy.

7.1.19 @12:22pm - (1/2) Mehdi Hasan's embarrassment will not end here. Once the conversation shifts to discuss the right of Jews to a homeland, he will be facing a dilemma: (1) To be honest and declare (like barghouti) that Jews are not a people or (2) To say that Jews, if well behaved, are entitled
7.1.19 @12:22pm - (2/2) to some semblance of sovereignty in the land. Theoretically, option (2) would get him off the hook but, unfortunately, he can't lose his support base of Arab rejectionists for whom such an admission amounts to a betrayal of 120 years of bloody wars and uncompromising denial.

7.1.19 @6:22am - The do-operator is not limited to physically "doing". It is an operation on your model of reality and it informs others about your model and about physical interventions that are feasible. On the interpretation of do(x), see: #Bookofwhy

7.1.19 @2:40am - (Replying to @LARichwine @NimaCNN @LAPressClub) Thanks @LARichwine for immortalizing this moment of grace. And thank you @NimaCNN for being part of a most memorable evening, and for honoring our son Daniel with your courage and integrity.

7.1.19 @2:32am - This quote, which sounds obvious in today's standards, was much debated in the 1980's, when AI folks labored to find proper formalisms to represent uncertainty in expert systems. This paper uses data insufficiency to expose a weakness in belief functions.

7.1.19 @1:48am - (1/2) We should all live to see Mehdi Hasan's face the first time he is called "Zionophobe" in front of an audience. On the one hand he is proud of being anti-Zionist, so he cannot deny the charges. On the other hand he will resist the analogy with Islamophobia which would force
7.1.19 @1:48am - (2/2) to philosophize on what "identity" is, be it religious, national or historical which, again, will steer the conversation to where we can win hands down: the moral imperatives of Zionism and the racist deformities of Zionophobia - the ugliest word in town.

7.1.19 @1:27am - (Replying to @EWilf @intelligence2) @EWilf , The title should have warned you of what the organizers tried to achieve: clearance from charges of AS. I stopped participating in such debates unless they change the title to: "Is Zionophobia racism?" or "Zionophobia on trial". ... "AS-- NO MORE"

7.1.19 @12:30am - (Replying to @zacharylipton) My goodness!! I never thought of becoming an icon of longevity, no way, playfulness yes, but longevity? Not in my worst dreams. Wait a minute! Perhaps playfulness is the secret to longevity? But isn't the latter simply a means to getting the former in greater quantity?#Bookofwhy

6.30.19 @2:58pm - (Replying to @IndexLlc @IARPAnews) Thanks for posting. I was not aware of this program and it is hard to tell from the BAA text whether its authors are aware of the fact that counterfactuals have been tamed, domesticated and algorithmitized, as in here . We need to find out. #Bookofwhy

6.30.19 @2:47pm - (Replying to @FJnyc @questionsin2014 @AshagerAraro) I share your suspicions. Zionophobic Jews-by-birth deny Jews rights that they grant to any other collective -- the right to define themselves. Does this make them racists? Painfully so!

6.30.19 @2:35pm - (Replying to @bnaibrithcanada) Thank you, @bnaibrithcanada for retweeting my speech to your followers. I am confident that, with your help, Zionophobia will become the Ugliest Word in Town. ps. Is there a Bnai Brith US? I wish they join us, in the trenches.

6.30.19 @2:34am - A transcript of a graduation speech I gave at UCLA last week is now posted on line: ... It explains what Jewish students are rallying for, and the kind of changes your campus will hopefully see next academic year.

6.29.19 @11:09pm - (Replying to @ShalitUri) Gee, I did not know about the Miao etal paper, perhaps because I was already immersed in #Bookofwhy. God bless Twitter for keeping us updated.

6.29.19 @3:39pm - (1/3) This paper identifies causal effects by proxies, a task shown feasible here and here . Blocking back doors is a sufficient condition for identification, not necessary; it turns out that, under certain circumstances, a proxy
6.29.19 @3:39pm - (2/3) can replace a blocker Z. Going from ATE to ITE ("individual" treatment effect) is not really hard (theoretically) if by "individual" we mean "c-specific effects" where c is a set of characteristics marking the individual, namely X=x. The nice thing about the paper you cited,
6.29.19 @3:39pm - (3/3) is that everything is spelled out in a language CI folks can understand, so that mysteries can be de-mystified. The authors should be commended. Effect-restoration was a breathtaking mystery to me in 2010, and it shows in the writing; I called it "far from obvious" #Bookofwhy

6.29.19 @12:42am - (1/2) To illuminate our discussion on PS, I am providing free access to Causality, Section 11.3.5 "Understanding Propensity Scores " . Note, in particular, Rubin's referring to PS matching "as if they had been randomized," and my closing remarks:
6.29.19 @12:42am - (2/2) "it is not enough to warn people against dangers they cannot recognize; to protect them from perilous adventures, we must also give them eyeglasses to spot the threats, and a meaningful language to reason about them." Readers of this Twitter understand it. #Bookofwhy

6.28.19 @12:54am - (Replying to @rlmcelreath) I've found that barriers among disciplines are lowered when we stress the questions to which CI seeks answers: (1) Effects of pending interventions, (2) Effects of undoing past events. PS. I couldn't read how the match-oxygen problem is related to attributable fraction #Bookofwhy

6.28.19 @12:16am - (Replying to @nghushe) Composure! Mom, Listen! This is the most complimentary message I received since my Bar-Mitzva. Three days after a prominent statistician accuses me of "calumny, caricature & confusion." I checked the dictionary, yes, the word exists; it just was not in my vocabulary. Thanks.

6.27.19 @11:21pm - (Replying to @edwardhkennedy) Fine. But I asked for a published paper that states so explicitly, to warn readers against assuming that PS has anything to do with asymptotic bias.

6.27.19 @11:17pm - (Replying to @jasonhartford @causalinf) HHMMM! you weren't talk about the real PRIMER ( ) ?? This too is a huge return on investment. For $20 and a few simple examples you get a glimpse at what causal inference can do for us, which is quite a lot. #Bookofwhy

6.27.19 @11:00pm - (1/2) I have no doubt that proving the asymptotic equivalence is immediate. Indeed, you can find such a proof in Causality ch 11 Eq. 11.10. What I asked however was whether anyone knows of a paper that actually states it EXPLICITLY, WITHOUT the assumption of non-confoundedness. The
6.27.19 @11:00pm - (2/2) paper you cite mentions Nonconfoundedness 17 times, thus contributing to the marketing myth that PS matching somehow contributes to bias reduction. See footnote 9, Causality p.349. #Bookofwhy

6.27.19 @5:14pm - I have not seen the connection, nor shown one. But I am a slow learner. The connection to complexity fascinated me in my youth: "On the Connection Between the Complexity and Credibility of Inferred Models," ... But I haven't tuned in since. #Bookofwhy

6.27.19 @3:19pm - Statistics as a discipline that helps us go from samples to distributions should be embraced and promoted. Statistics as an intellectual blinder that prevents one from seeing beyond distributions should be abandoned and shunned. #Bookofwhy

6.27.19 @7:11am - (Replying to @ildiazm) I never object to inclusion, especially inclusion of basic building block like stats. I object to exclusion. Like math teachers who exclude multiplication from arithmetic because they can always add a number to itself n times, i.e., the "classical way". #Bookofwhy

6.27.19 @2:46am - (Replying to @JDHaltigan) I agree with your take, and our challenge is to make 1>>2 in a climate where 2 control academia and demand submission from 1. #Bookofwhy

6.27.19 @2:39am - (Replying to @stephensenn) Exactly the way Lord describes what the 2nd statistician's does: Compare W_F of Diet-1 to that of Diet-2 for students of equal W_I. But Stephen, it is your turn now to teach me what I do not know about adjustment, I am listening carefully to new ideas, Just listening. #Bookofwhy

6.27.19 @2:21am - Thanks for posting, Jennie, now everyone can see what I mean by "causal inference" and why it needs a new logic, different from "the classical approach" we have been discussing here. #Bookofwhy

6.27.19 @12:41am - (1/3) I am curious, do you know of any paper that states explicitly the asymptotic equivalence of PS and the "adjustment formula"?? I have only seen it in Causality, ch 11, not elsewhere, Why? Perhaps someone is interested in marketing it as a magic wand? I am also curious to know
6.27.19 @12:41am - (2/3) how many readers hear this equivalence for the first time. It says that regardless of what covariates you use, as the number of samples increases, the bias of the PS estimator converges to the bias of the adjustment estimator. If many hear about it for the first time, it will
6.27.19 @12:41am - (3/3) serve as an example of information stifling that CI needs to liberate itself from. Anyone knows what the new CI textbooks say about PS? I would use it as a litmus test for authors understanding of modern CI. #causalinference #Bookofwhy @causalinf

6.26.19 @11:44pm - By all means. I hereby advocate, as always, that data-science rests on two equally important pillars: causal inference and statistical estimation. However, a glaring asymmetry can be seen today in academia: researchers in the CI pillars are thirsty for new
6.26.19 @11:44pm - (2/3) tools from the estimation pillar but not the other way. By and large, leaders of the stat pillars have zero interest in advances emerging from the CI pillar. The great majority of them truly believe in "do it the classical way" and "a causal model is a special case of a
6.26.19 @11:44pm - (3/3) predictive model". Moreover, in academia, the stat pillar dominates its CI partner by 100:1 ratio, and now insists on total dominion in the name of "its just a special case". Thus, to achieve equilibrium, I think CI needs academic autonomy, at least for a while. #Bookofwhy

6.26.19 @11:08pm - (1/2) Why the anger? Was I wrong in pointing to Gelman's blog as a "stronghold of statistical thinking"? (He has 21K followers, more than any other statistics-minded blog that I know.) Did I misquote it as advocating "do it the classical way"? or in stating that a "causal model is
6.26.19 @11:08pm - (2/2) a special case of a predictive model" or as advocating adjustment for all "pre-treatment differences among groups". I am always willing to learn more about what "adjusting" means. What is it? Teach us #Bookofwhy

6.26.19 @7:13pm - (Replying to @oacarah @PWGTennant) I think these papers miss the point. Agree? They treat PS as another method of identification, rather than an efficient estimator of the adjustment formula. #Bookofwhy

6.26.19 @7:01pm - (1/2) There was a time when statisticians were the guardians of prudence and caution, and causal folks the unruly children of unprincipled adventures. Today, if you peek at the strongholds of statistics (eg, Gelman's blog) you will find the opposite. CI folks hold the leash
6.26.19 @7:01pm - (2/2) on scientific principles, while statisticians advocate "do it the classical way", namely, rush into the mine-field of causation w/o a metal detector, and tell students that curve fitting is "causal inference". Yes, The Times They Are A-Changin' #Bookofwhy

6.26.19 @3:39pm - To summarize: (1) Is it just more explicit? No it is less explicit. (2) Helps with checking 'balance'? Sometimes, but 'balance' is the wrong criterion for covariates. (3) Encourages better model building? On the contrary. The advantage lies in mapping high-dim to 0--1 interval.

6.26.19 @3:39pm - In case you skipped Section 11.3.5 in Causality, titled "Understanding Propensity Score" (p.348), I would highly recommend it, for it clears up some of the myths connected with PS. This one may also be illuminating #Bookofwhy #EpiTwitter #Causalinference

6.26.19 @2:43pm - (Replying to @elizpingree @Jamie_Woodward_) Agree. Aside from its artistic qualities, the significance of the Lion-man in human development cannot be underestimated. Inspired by Harari's "Sapiens" , I featured it as the seed of counterfactual reasoning in the #Bookofwhy . Thanks for posting.

6.25.19 @11:57pm - (Replying to @lisabodnar) How powerful and truthful your uplifting words sing. I wish I could see them each time an academic troll unveils his/her pain. #Bookofwhy

6.25.19 @10:45pm - Thanks for your kind words. As to Aesthetics, our publisher insisted that general-purpose books should look different from technical books. Who are we to resist? #Bookofwhy

6.25.19 @10:40pm - (Replying to @adamdedwards) Good question. Thus far, I know of only two US philosophy dpts elevated to the age of causation: CMU and CalTech. Anyone knows of more? #Bookofwhy

6.25.19 @8:14pm - Agree, Epidemiologists are 98% there. All it takes is a final snip of the umbilical cord to mother-stat with a firm commitment to listen to what causal models tell us. #Bookofwhy

6.25.19 @1:13pm - (Replying to @DanielNevo) Same question can be asked about statistics. But it so happened that, in order to attract top scholars to the field, statistics insisted on academic autonomy, and dominion over data analysis, rather than being unappreciated minority in each data-using department #Bookofwhy

6.25.19 @3:06am - (Replying to @vkehayas) We have two additional sources of information: (1) playful manipulations (often called "interventions" or "experimentation") and (2) hearsay (often called "education"). #Bookofwhy

6.25.19 @3:01am - (Replying to @phi_nate) I twitted my take about two months ago, saying essentially: Can anyone translate what he is doing to our language so that we can prove that the claims do not violate any of the impossibility theorems we derived mathematically? Waiting. #Bookofwhy

6.25.19 @2:56am - (Replying to @zaffama) Agree, the culprit is in the interpretation. But notice an interesting phenomenon: the interpretation controversy did not rise when we accepted the axioms of probability, it has arisen only when we derived a consequence of those axioms, and interpreted it. What took us so long?

6.25.19 @1:50am - (Replying to @mariotelfig) We are talking about a one-line proof, not one-line statement of the theorem.

6.25.19 @1:48am - Eventually, I am sure, there will be more Causal Inference PhD programs than statistics PhD programs, possibly under the title "data science - causal inference" The question is which departments will launch it first, statistics or computer science?

6.25.19 @1:33am - Name another one-line theorem that has remained controversial for 250 years, and books like "the theorem that never dies" are written about it, and people are hired and fired in its name, etc etc. My take: it is not merely a theorem; it is a statement. #Bookofwhy

6.25.19 @12:26am - (1/2) Apropos Bayes. Does anyone thinks Bayes' Theorem is really a theorem? If it is, then it is the most trivial theorem in the cosmos, with a one-line proof. Can someone, even a Reverend, become immortal with a one-line proof? If it is more than just a "theorem", whence comes its
6.25.19 @12:26am - (2/2) added value? The #Bookofwhy answers this question (p. 102) from computational perspective, since I could not find any discussion of it in the statistical literature (@learnfromerror, @f2harrell, @stephensenn ) I believe it has more to do with psychology than with statistics.

6.24.19 @11:14pm - Agree. Bayesians first reacted to frequentists zeal, who persecuted them for contaminating statistics with "subjective knowledge". In time, they defined club membership by "priors on parameters", regardless if those priors conveyed knowledge or habits. I go Bayes 1763 #Bookofwhy

6.24.19 @8:58pm - (Replying to @kareem_carr @EpiEllie) Causal inference has been unified 10 years ago, see . True, the "schools" are still singing different anthems, laden with egos and tribalism, but you, as a rebellious champion of commonsense should look at the content, not the label. #Bookofwhy

6.24.19 @8:36pm - (1/2) My, My, Carlos, this paper was published 10 years ago and, so untypical of me, I am still behind every assertion. Remarkably, 10 years have passed, and statisticians are still resisting the distinction between causal and statistical notions (Section 2.1). Just this week,
6.24.19 @8:36pm - (2/2) Andrew Gelman @StatModeling wrote: "So I think it's a mistake to think of causal and predictive inference as being two different things." Your posting this review makes me both sad (0 inches - 10 years) and hopeful -- I detected sparks of awakening on Gelman's blog #Bookofwhy

6.24.19 @8:03pm - (Replying to @kareem_carr @EpiEllie) I wish you were right. My brief (83) encounter with statistics textbooks reveals a slightly different picture -- a complete prohibition on all assumptions, philosophical, causal or otherwise, except statistical assumptions that are but a tiny part of decision making. #Bookofwhy

6.24.19 @7:54pm - (Replying to @lisabodnar @EpiEllie) Congratulations, Lisa! Judea

6.24.19 @7:47pm - (Replying to @kareem_carr @EpiEllie) The tough question is: Is this a "well-defined" intervention?

6.24.19 @3:14pm - I think this discussion would benefit from a glimpse into the Bayesian vs anti_Bayesian controversy in AI, in the 1980's, about how to model epistemological uncertainties in expert systems. My recollections are here Biased, but honest #Bookofwhy

6.24.19 @4:28am - (1/2) When statisticians learned how to spray priors on parameters (of distribution functions) they formed an exclusive club called "Bayesian Statistics" and decided that he who won't spray priors on parameters is "not a Bayesian". My definition if Bayesianism is more broad, it
6.24.19 @4:28am - (2/n) follows Bayes paper of 1763, and it has to do with his interpretation of the phrase "given that we know X=x" and a license to invoke prior information. When I coined the name "Bayesian Network" (1995) I justified it on these grounds and I added Bayes' fascination with "cause"
6.24.19 @4:28am - (3/3) as another reason. Today I am only half Bayesian for reasons explained here , mainly because the bulk of our knowledge is causal, which cannot be captured by priors over parameters, definitely not parameters of a distribution of observations #Bookofwhy

6.24.19 @2:25am - True, I am very fond of my 1988 book "Probabilistic Reasoning", but I left Baeysian analysis in favor of causal inference, partly because most of medical reasoning is causal, not probabilistic. Its a whole new paradigm , and much fun #Bookofwhy.

6.24.19 @12:47am - (1/2) On a different occasion we will debate what DAGs cannot do, promise. Right now, the question is "Can a mechanism-seeking researcher answer causal questions from partial information alone, given in the form of "who listens to whom". Here are some examples:
6.24.19 @12:47am - (2/2) (1) The prisoner is dead. What if Rifleman-A refrained from shooting? (2) More people died from inoculation than from smallpox. Should we ban inoculation? (3) Is there a drug that is good for men, good for women, and bad for a typical person? #Bookofwhy

6.23.19 @11:35pm - (Replying to @yourbirlfriend @EpiEllie) @yourbirlfriend Fascinating! But cellular automata are driven by local forces and "minimum thread" is a global feature. How can you accomplish it? @yudapearl

6.23.19 @11:25pm - (Replying to @NeuroStats @shravanvasishth and 4 others) For computer scientists, the helplessness of Bayes analysis in causal reasoning comes glaring already in the notation. BDA invokes only one conditioning symbol, the vertical bar |X=x). Causal reasoning requires |do(X=x) Bingo! Done! You cant get "do" from "see" #Bookofwhy

6.23.19 @4:01am - (1/3) This article is worthy of our attention: ... 47 journal editors are offering guidelines to authors on ways to report results of causal inference studies. It is refreshing to see 47 editors reach consensus on a topic that only a decade ago was a sure ticket
6.23.19 @4:01am - (2/3) to discord. I believe the availability of DAGs as a communication language helped the process. I have trouble with some of the terminology (e.g., "causal association") but, overall, I welcome the timely rejection of "traditional" approaches of wholesale adjustment for
6.23.19 @4:01am - (3/3) everything one can conveniently measure. See @StatModeling for a lively discussion of opposing viewpoints, especially my explanation of why blindness to DAGs is an invitation to bias amplification . #Bookofwhy

6.23.19 @2:35am - (1/3) When you have a chance, please explain to readers on this Twitter how Bayesian Data Analysis (BDA) can help one think about causality. I have heard it from many statisticians and data analysts but I have never been able to understand what they find to be helpful and why.
6.23.19 @2:35am - (2/3) Is it the "model selection" part offered by BDA? Or the idea that you are properly combining prior knowledge with data? In my opinion, BDA is a siren song that lure people away from properly "thinking" about causation, as I argue here and in many
6.23.19 @2:35am - (3/3) other forums. I am appealing to you because, as an accomplished reader of #Bookofwhy I think you would be able to pin point to us where precisely BDA enthusiasts see the connection to Causal Inference, and why I am missing this connection. As always, a toy example is the KEY.

6.22.19 @7:59pm - (Replying to @RMartinChavez) I could not resist clicking on "like" to such a compliment, but the main thing is it is true: "At last, a science of causation!" And as much as I would try to minimize it the facts will scream at my face: a science of causation. #Bookofwhy

6.22.19 @5:58pm - (1/n) My remarks about being interested in seeing a Multi-level problem solved, and about traditionalists fearing toy-problems were in the context of the discussions on Gelman blog. As to #Bookofwhy, you are asking to be shown the "how" first and the debates second. This is exactly
6.22.19 @5:58pm - (2/n) what the book does. Chapter 1 tells you about the inference engine and what kind of questions it solves. What you see as a "fight" among competing approaches does not exist, because the history of debates about causation was not among "approaches" but among "ideologies"
6.22.19 @5:58pm - (3/n) 3/n that do not qualify for the title "approach". An "approach" should be armed with definition of what problems the approach attempts to solve and theoretical guarantees of reaching adequate solutions under certain conditions. Thus, the reason readers get the impression that
6.22.19 @5:58pm - (4/n) there is only one "approach", (ie SCM,) is that there is really only one approach armed with the needed theoretical guarantees, the alternative themes where merely "themes". Currently, I know of one competing "approach", ie, Rubin's potential outcome , which is logically
6.22.19 @5:58pm - (4/n) equivalent to SCM under the same assumptions but makes it hard for researcher to represent assumptions. It is described humbly and respectfully on pages 272-280. (I just made a similar point on Gelman's blog.) Its tough to tell the naked truth to readers who expect
6.22.19 @5:58pm - (5/n) something else. Some even view my claim about a transformative revolution to be nothing but "hype". I hope now, that you have finished the #Bookofwhy and had a chance to seek alternatives, you see that I had no choice but risk being called "hype" - "Si Muove" it really moves!`

6.22.19 @6:23am - (Replying to @djvanness) The calculus cannot help you with this choice, since you have not specified the state of knowledge upon which such choices are made. Now, since the calculus kicks in only after you made the choice, it can be considered an "assumption". #Bookofwhy

6.22.19 @5:42am - (Replying to @sakrejda @shravanvasishth and 2 others) You are being unfair by assuming apriori that what you see as "pettiness" is not genuine attempts to learn from each other. I, for one, would very much like to hear more about what we can learn from the "broader search" that goes beyond "causal inference". Care to explicate?

6.22.19 @5:31am - In causal modeling we do not call it a "choice" but an "assumption" (item (2)). If you think it is not plausible, go ahead and tell the calculus that Hall affects weight independently of Diet. (Perhaps one Hall has a dancing floor??). But we need a story to proceed #Bookofwhy

6.18.19 @5:35am - (Replying to @gjcampitelli @Lester_Domes and 4 others) I would stick to my SCM religion: Solutions must start by articulating the questions. What do we mean by "handling"? What are we expected to do after we differentiate between FE and RE in the model construction? Or, put more bluntly, why would anyone care? #Bookofwhy

6.22.19 @5:23am - It was great fun talking at your conference and assuming (wrongly) that audience never heard about causal inference. Thanks for posting this very slide "What is Causal Inference?" which is so timely in view of conversations we are having on Twitter and on @StatModeling #Bookofwhy

6.22.19 @5:09am - (Replying to @stephensenn) In Fig. 4 "Dining Hall " appears because it is taken from Wainer etal where it is another name for "Diet". Would you prefer we proceed with Fig. 4 and compute P(weight gain| do(Dining Hall)) ??? Give me a story. #Bookofwhy

6.22.19 @4:27am - (Replying to @stephensenn) Are you happy with the story of Fig. 6.9(b) ??. If not, please modify, if yes, we will go to: (3) data available. #Bookofwhy

6.22.19 @4:22am - (Replying to @stephensenn) These questions are vividly answered in Fig. 6.9 b. "Hall" is not in the graph, which means it is irrelevant to any research question. It shows W_I ----> D, which means "Diet may attract slim/heavy students", and so on, all assumptions are vividly displayed. #Bookofwhy

6.22.19 @4:04am - (Replying to @stephensenn) We need to analyze each version and see. So far we do not have a story -- what's the role of the "Hall". Is it just a place where Diet is served or a place that also affects weight or a place that attracts slim students, or a place that attracts students craving for a given diet?

6.22.19 @3:50am - (Replying to @stephensenn) Causal calculus does not "handle" things; it answers (1) research questions, given (2) assumptions and (3) data (experimental or observational). So far we got the question: (1) Find the causal effect of diet on weight gain, or P(gain|do(diet)). Now we need (2) and (3). #Bookofwhy

6.22.19 @2:23am - (1/2) I thought we agreed that in these versions Lord's paradox disappears. If you are still interested in "handling" these versions, independently of Lord, by all means, but you need to specify them as we specify any "toy problem", namely, (1) what do you wish to estimate?
6.22.19 @2:23am - (2/2) (2) What (causal) assumptions are you willing to make? and (3) what data is available to you? Let's start with ONE version, not TWO, to avoid going back and forth. #Bookofwhy

6.22.19 @2:01am - Unfortunately, I had to remove the Babylonian vs Greek analogy from the final version of the paper ; a reviewer insisted that this would offend ML folks. #Bookofwhy

6.22.19 @1:28am - Why do you assume that I am trying to "advocate' one approach or put down others? Why not assume that I am genuinely craving for ANY approach that enriches causal inference, and I can't satisfy this craving (my weakness) without seeing toy examples solved in other approaches.Why?

6.21.19 @10:48pm - I have the feeling that @StatModeling folks did try toy problems and the reason they dread them so consistently is that they realize they havn't got the tools. I would be very interested in seeing a simple Multilevel problem well posed and toy-like solved.

6.21.19 @9:55am - I like the term "susceptible to persuasion" (or "gullible"), because Ang Li has just finished a paper that formalizes this notion and gives it algorithmic teeth #bookofwhy

6.21.19 @12:37am - I just finished an adventurous visit to Gelman's blog ... and I am sharing here the last sentence: "if Causal Inference is 'Statistical Inference given causal assumptions' [as some claim] then Car making is car-painting given an engine and a body."#Bookofwhy

6.20.19 @10:30pm - Yes, Just in case you are at UCLA tomorrow, and have not gotten tired yet from listening to my Songs of the Revolution, join me tomorrow 3:30 pm where I will be singing one Aria to a slightly new libretto in front of social scientists and other mostly harmless folks. #Bookofwhy

6.20.19 @1:32am - I was truly honored to speak at the Algemeiner Gala, to be introduced by Sharon Stone and to warn against "Zionophobic Thuggery," which I hope will be recognized as the ugliest form of hostility on US campuses.

6.19.19 @4:52pm - @JewishPub. The more I read it, the more strongly I feel about including it in the list of 5 books that an "educated Jew" could not live without.

6.19.19 @3:49pm - (1/2) Many thanks. This is a good place to start chewing the eco. literature. One hurdle is the definition of "identified" which in eco. usually means identify the functions or their parameters, and in CI means identify a query Q which depends on the functions. I have just
6.19.19 @3:49pm - (2/2) noticed that the example you brought up is the usual IV setting, as in here , So it is already in our arsenal of "Powerful tools." Will see if more nuggets can be excavated. #Bookofwhy

6.19.19 @1:44am - (1/2) Great!! Here are some powerful results that hold for the linear extreme of the spectrum: . Can you or one of your students pin point to me which of them is extendable in some way to nonlinear systems? For example, Eq. (20) shows us how to estimate ANY
6.19.19 @1:44am - (2/2) counterfactual expression in ANY linear system. Can this capability be extended in some way to binary monotonic nonlinear systems? Just a hint, or eq.(#) would do, but please do not send me without guidance to a vast unchartered literature. Thanks #Bookofwhy

6.19.19 @4:30am - (Replying to @DanielOberski) There is a habit, especially on Twitter, by people who have not read things with sufficient depth to quote authors names, or provide links to cryptic papers and say: "It is treated here...". Done. "Without links" means: give me the basic idea, dont send me elsewhere. #Bookofwhy

6.19.19 @4:21am - (Replying to @WhitneyEpi @AdanZBecerra1 and 12 others) One should note that these ""classical" rules from the RCT context were merely "rules of thumb", lacking theoretical underpinning. It turns out that some post-intervention factors are actually safe, while some pre-intervention factors are unsafe. See Causality p. 339 #Bookofwhy

6.19.19 @3:14am - (1/3) I am partially familiar with Matzkins works, and I know that many economists refer to them. However, I have not been able to supplement the arsenal of powerful results now available for linear and NP extremes of the spectrum, with similar results applicable to mid-spectrum
6.19.19 @3:14am - (2/3) I examplified what I mean by "powerful results" here: ... Perhaps you or some other expert economist can fill me in on what is known about the mid-spectrum. Or, what is the one most important identification opportunity one would miss by not
6.19.19 @3:14am - (3/3) this literature as thoroughly as it deserves. For example, is the counterfactual estimability of linear models extendable in some way to monotonic models? or Binary models? #Bookofwhy

6.18.19 @10:43pm - I have posted a comment on Gelman's blog, explaining why causal and predictive inference are not the same thing, and why it is beneficial to solve each task in its own distinct vocabulary. #Bookofwhy

6.18.19 @3:03pm - (1/2) "Powerful results" are those that one remembers. For example, "Every causal effect in every NPSEM is either identified or quickly proven to be non-identifiable". "Every linear SEM that contains no bow-arc is identifiable" "Every counterfactual in linear SEM is estimable" etc.
6.18.19 @3:03pm - (2/2) Are there similar results for the spectrum between linear and nonparametric? For example, what do we know about the loglinear variety? What identification results from linear SEM are preserved? To be "powerful" they need to be articulable verbally w/o links . #Bookofwhy

6.18.19 @12:30pm - My thoughts? Hilarious!!! Here is one gem: "So I think it's a mistake to think of causal and predictive inference as being two different things." Here is mine: "So I think it's a mistake to think of any two things as being two different things - it's all arithmetic" #Bookofwhy

6.18.19 @5:35am - (Replying to @gjcampitelli @Lester_Domes and 4 others) I would stick to my SCM religion: Solutions must start by articulating the questions. What do we mean by "handling"? What are we expected to do after we differentiate between FE and RE in the model construction? Or, put more bluntly, why would anyone care? #Bookofwhy

6.18.19 @4:49am - (Replying to @fr_amodeo) If by "handling" you mean "provide distinct specification" then for the standard SCM the answer is NO. But why not add a marker "L" to any family that combines linearly, and leave the others unmarked. The beauty would be to decide if an arbitrary causal effect is fixed #Bookofwhy

6.18.19 @2:12am - (Replying to @juli_schuess @PHuenermund) Paul, why the wonder? This is the classical IV setting and, as @juli_schuess notes, all parameters are identifiable using either Wright rules or the method of moments. For generalizations, I would use #Bookofwhy

6.17.19 @8:50pm - (1/2) Got it. Thanks. And I am retweeing because many CI folks are likely to stumble on this jargon. Writing y=bx+u gives fix-effect, because P(y|do(x+1))-P(y|do(x)) is independent of u. But writing y=f(x,u) gives random-effect. So, nonparametric SCM assumes random effects, and the
6.17.19 @8:50pm - (2/2) distinction does not show in the DAG, only in one's declaration "Assume a linear model". An entirely different question is: "Is there a test that detects the presence of RE in a given population?" Yes, there is, and beautiful one too: #Bookofwhy

6.17.19 @7:59pm - (Replying to @Corey_Yanofsky @Lester_Domes and 2 others) A partial meeting of minds was achieved here, on Twitter, when my mind realized that what Senn objects to is the perfect alignment of the two ellipses, which would rarely occur in real life. This I consider to be orthogonal to Lord's question: "Who is right?" #Bookofwhy

6.17.19 @7:52pm - (Replying to @Lester_Domes @Corey_Yanofsky and 2 others) I will try, if only I knew what RE and FE are. (With examples please). Are the the causal effects estimated in an ideal RCT RE or FE ????

6.17.19 @6:59pm - (Replying to @PWGTennant @lisabodnar @ProfMattFox) Sorry to be missing your dazzling party. Now I have two reasons to attend SER-2020. Cheers!

6.17.19 @1:15pm - So, according to your perspective, modern SEM maintains the same input-output relation as SCM. Namely, INPUT = data + qualitative causal assumptions. OUTPUT= causal effect sizes. And no one is uncomfortable about the word "causal" there, even in published papers. #Bookofwhy

6.17.19 @12:59pm - (Replying to @andres_fandino) There is no such notion as "appropriate". If your understanding of the problem demands N exogeneous and M endogeneous variables, so be it. Reality comes first, estimability second. #Bookofwhy

6.17.19 @12:43pm - (1/2) Yours is a very encouraging perspective of modern SEM, namely, SEM = "SCM that accommodates parametric assumptions." Its badly needed. I am familiar with the linear and nonparametric ends of the spectrum; are there any powerful results in between? Say binary, or monotonic?.
6.17.19 @12:43pm - (2/2) Also, just out of curiosity, what percentage of your colleague/students can answer "Tell me if the partial correlation R_{XY.Z} is zero", or "Tell me which parameter is estimable by OLS". I have not communicated with them for several years. #Bookofwhy

6.17.19 @4:50am - (Replying to @y2silence) Beautifully put. And I wish more people learn to appreciate this miracle, which is truly unique. In fact, what else can one ask of a modeling methodology? #Bookofwhy

6.17.19 @3:28am - (1/3) In the last conversation I had with Peter Bentler he defined SEM as a "compact and meaningful representation of a covariance matrix". We know that "meaningful" is a round-about way of saying "causal". Fine. But since the purpose is declared as fitting a covariance matrix,
6.17.19 @3:28am - (2/3) one should ask: To what use can one put the conclusions of an SEM study (aside from getting your dissertation published in Journal of SEM), which usually comprises a huge list of path coefficients with their confidence intervals. I wish someone can explain why not list the
6.17.19 @3:28am - (3/3) estimated covariance coefficients themselves; a much simpler task. #Bookofwhy

6.16.19 @11:58pm - It is hard to discuss SEM since its practitioners are still not sure what it is. I tried to describe their confusion here and sway them to use SCM. Not successfully, and Psychomentika-type papers are my witnesses. Interesting social phenomenon #Bookofwhy

6.16.19 @11:29pm - (Replying to @ulusdd) True, but I dont know anyone who starts with a cyclic system and expects DAGs to handle it. DAG techniques are expected to compute properties of the input model, namely directed-acyclic systems. Besides, d-separatioin holds in linear cyclic systems. #Bookofwhy

6.16.19 @11:22pm - My perspective today: SEM is a community of researchers using SCM who refuse to commit to the causal reading of their models, argue endlessly about what that reading is, and refuse to benefit from the comp. power of DAGs. Too harsh? Would welcome opposing evidence #Bookofwhy

6.16.19 @11:09pm - (Replying to @omaclaren) I am not familiar with "rate parameters" "ODE models" and 'nonlinear least square". Can you explicate in terms of the input model, which is a DAG, be it linear or nonparametric. Of course a DAG cannot compute how much money you have in bank account, or halting problems etc.

6.16.19 @8:57pm - (1/2) "tools with which to think about models" is only one usage of DAGs. Another usage is computational; DAGs permit us to answer questions which otherwise are intractable. E.g.,"Tell me if the partial correlation R_{XY.Z} is zero", or "Tell me which parameter is estimable by OLS"
6.16.19 @8:57pm - (2/2) I am surprised that the computational capacity of DAG is under-appreciated by most researchers. Am I exaggerating? Is there another computational tool with which the questions above can be answered? Is there a question you wish to ask which DAGs do NOT answer? #Bookofway?

6.15.19 @6:28pm - (Replying to @robanhk) I love those "intellectuals" and have written a poem in their honor

6.15.19 @6:26pm - (Replying to @robanhk) I was not sure if you were sarcastic or honest. Now that you are equating "not loyal to Israel" with "denying Jews the right to a homeland" you made it clear. I do not know anyone who denies a people's right to a homeland who is also considered an organic part of that people.

6.15.19 @4:21pm - (Replying to @robanhk) Yes. I thought that the fringe group of Jewish-born intellectuals who deny Jews the right to a homeland have not felt targeted by Abdulhadi's rants; they run for safety by showing loyalty to her Zionophobic movement and do not need my encouragement.

6.15.19 @2:12pm - Excited towards keynote addressing the UCLA Jewish Graduation Ceremony on Sunday. Hoping to inspire the very students who were recently labeled "White supremacists" by a super BDS ideologist (still at large, still not condemned).

6.15.19 @1:28pm - (1/2) If instead of criticizing the way SEM's ARE used we take the trouble to explicate how they SHOULD be used, we will, I believe, end up with SCM's (Structural Causal Models) like those described in #Bookofwhy or . But,
6.15.19 @1:28pm - (2/2) now that we are discussing classical SEM, I am curious to know whether there is still a large community of SEM users who have not switched to SCM ??? If there is, can someone act as its spokesperson and explain what, aside from habit, prevents them from making the transition?

6.15.19 @2:35am - Sewall Wright had much harder time communicating with statisticians. He did not have the RCT metaphor to open the first window to their hearts, and he did not have the logic of counterfactuals to defend the structural assumptions behind his diagrams. A true hero. #Bookofwhy

6.14.19 @3:39pm - This is truly a gem -- thanks for posting. I have not read it for many years. And now that I re-read it, it covers so many topics that we have discussed here on Twitter. It even discusses modern hangups with "mimic RCT or else your causal effect is "not well defined". #Bookofwhy

6.14.19 @11:44am - The correct link to "Causal Foundations of SEM" is this There was a missing space. #Bookofwhy

6.14.19 @3:43am - Regarding the causal foundations of SEM, I have found another article that addresses this issue head on: . It also explains why teachers of SEM are so helplessly confused about a problem so simple. #Bookofwhy

6.14.19 @2:12am - Thank you, Rabbi Dunner, for immortalizing our meeting at the Algemeiner Gala. I felt truly fortunate to be able to contribute to our three common ideals "Truth, humanity and Jewish peoplehood" - a spec of sanity in the age of BDS madness.

6.14.19 @1:37am - Many have had the same SEM experience. The answer is that no magic dust is needed; the coefficients were always causal, and the association just help us to quantify those causal relationships. Bollen and I discuss it in "Eight Myths of SEM" #Bookofwhy

6.11.19 @8:15pm - Thanks, Lisa, for letting us know that you will MC tomorrow's Gala. Looking forward to seeing you again. And may the angels bless you and the other warriors in the trenches.

6.11.19 @6:25pm - @causalinf Have you looked into the way Peter Steiner represents RDD in DAGs? I have a link to his 2017 paper: ... #Bookofwhy @PHuenermund , @eliasbareinboim , @yudapearl , @Jabaluck @EpiEllie , @paulgp

6.11.19 @6:11am - (Replying to @TuomasPernu) I feel the same way, except that I found solace in the metaphysical assumption that reality acts as a "society of listening agents". So far, I have not found a more satisfactory metaphysical theory in the philosophical literature. Would be curious to examine alternatives.

6.10.19 @3:49pm - (1/2) "difference-making" or "counterfactual dependence" are various names philosophers used, but they fell short of operationalizing this relation through the simple mathematical concept called: "a function". I dont know what book you have on shelf, but if it is #Bookofwhy or
6.10.19 @3:49pm - (2/2) or or Causality, then you will satisfy you quest to see coherence among the assumptions and the conclusions. The philosophical literature, unfortunately, is still hung up on "what do we REALLY, REALLY... mean by CAUSE". Hopeless

6.10.19 @2:40pm - (1/2) One gift modern causal inference has given to us is a clear answer to your question: "what the assumptions and stages are, and how they relate to each other." See eg #Bookofwhy. Another gift is seeing all those assumptions expressed in terms of one primitive relation:
6.10.19 @2:40pm - (2/2) "One variable listening to others". One may argue that this primitive is just a causal relation in disguise. Perhaps, but it is still a gift to see ALL assumptions and ALL conclusions emerging from ONE easily grasped idea. #Bookofwhy

6.10.19 @1:55pm - Not entirely unfair. It lends support to the idea that perhaps "listens to" is an irreducible primitive in our mind which will stay irreducible until we can map the neural paths that are activated when we ask "why".

6.10.19 @1:53pm - (Replying to @PHuenermund) Not entirely unfair. It lends support to the idea that perhaps "listens to" is an irreducible primitive in our mind which will stay irreducible until we can map the neural paths that are activated when we ask "why".

6.10.19 @5:21am - (Replying to @fpgil) And I understand the Portugal is celebrating today its independence day -- the oldest independence day in Europe. Monday, 10 June: Dia de Portugal. I just heard it on Israeli TV, so, from Israel with love: Two small countries with rich rich history.

6.10.19 @4:09am - (1/2) I an not sure whether Woodward is satisfied with the structural account, see , but the philosophers that I read do (eg Spohn, hitchcock,Glymour). What is the point of asking for "the nature" of causation withouot telling us what type of answer will be
6.10.19 @4:09am - (2/2) accepted? The structural account, based merely on the notion of "listening to" is as satisfactory to me as any that I have seen in philosophy. And, in addition to its metaphysical satisfaction it can also solve the Simpson's puzzle, how can you ask for more? #Bookofwhy

6.10.19 @3:30am - Portuguese!! My goodness!! Now I can start arguing with all my Brazilian students on equal footing. Thanks for posting. #Bookofwhy

6.10.19 @3:21am - (1/2) The #Bookofwhy solution makes NO ASSUMPTION beyond what is specified by Lord himself. If you prefer to fantasize a complex multi-hall version of the problem, and fail to collapse it into Lord's distribution, do not blame Lord, nor #Bookofwhy . The key point is that, for
6.10.19 @3:21am - (2/2) the simple two-hall problem, each serving ONE diet, the solution provided in #Bookofwhy p. 217 is valid, and resolves a paradox that still baffles many good minds, even today, including many statisticians. Again, no extra assumption beyond those given by Lord and Fig. 6.9b.

6.10.19 @2:53am - (Replying to @stephensenn @analisereal) The distribution is given to us by the two ellipses; This is Lord's construction, not ours. NOTHING ELSE is assumed about variances and covariances. Mixed models were invented to help in the construction of distributions, not for handling a fully specified distribution.#Bookofwhy

6.10.19 @12:14am - The less background the better. Except, if you have not taken any stat-101 or econ-101 classes you will miss the fun of asking: "How come my professor never told us that causal inference is easy?" #Bookofwhy

6.10.19 @12:02am - (1/3) The #Bookofwhy is not about "what causal calculus cannot do" (eg, play chess, translate languages etc) but about the many miracles it CAN do. Among them resolving the simple version of Lord Paradox, with two dining halls, each serving one diet, and a very large sample. So,
6.10.19 @12:02am - (2/ ) believe it or not, but this simple version is still paradoxical to most mortals, and has been paradoxical for half a century. It is now resolved by causal calculus. Your multi-hall version may be of interest in a certain context, but I cant understand why you are insisting
6.10.19 @12:02am - (3/3) that this idiosyncratic version is essential for resolving the simple version, and that he who does not attend to your version is guilty of neglecting the foundations of statistics or worse. I do not buy it. Lets focus on the simple version -- are you happy with the solution?

6.9.19 @11:16pm - Good news for fans of #Bookofwhy. The book is now available in paperback ($14) from amazon: and, more importantly, friends tell me they got it delivered next day and it is much easier to read, handle and carry on the train. Tell your grandchildren.

6.9.19 @8:22pm - (1/2) I find VanderWeele's decomposition somewhat artificial. The two clinically meaningful components of mediation are: (1) The extent to which observed effects would be PREVENTED by disabling the mediating path and (2) The extend to which observed effects would be SUSTAINED
6.9.19 @8:22pm - (2/2) with the direct effect disabled. These two components (necessary vs. sufficient) collapse in linear systems, but are distinct when interactions are present. Each can be estimated using the mediation formula, , and , #Bookofwhy

6.9.19 @5:19pm - Congratulations, again, and I am re-tweeting your vision statement here, because it coincides so perfectly with my conception of where causality-land is heading. There may be some hidden cause in action here, but reality prevails regardless of models. #Bookofwhy

6.9.19 @5:00pm - (1/2) (Replying to @jwbelmon @jon_y_huang and 3 others) Some people use an "arrow on arrow" notation to indicate effect modification, but I find it unnecessary and confusion, because, if A modifies the effect of B on C then, from basic logic, B must modify the effect of A on C. In linear systems we can just use a product term
6.9.19 @5:04pm - (2/2) (Replying to @yudapearl @jwbelmon and 4 others) and in non parametric system, every parent is by default assumed to modify the effect of all other parents. I you need to find the degree of effect modification, you can use counterfactual logic, as VanderWheele is doing, or as we do in causal mediation #Bookofwhy

6.9.19 @3:26pm - (Replying to @stephensenn @analisereal) "The sample size is only two" My goodness!! we must be thinking about a different universe. I do not deny the existence of your universe but, given that I naively assumed that the two ellipses show a very large sample size, please describe your universe slowly slowly #Bookofwhy

6.9.19 @2:37pm - (1/n) Now that we a research question, we go to the next step: What data we have available. The answer is given by the two ellipse, showing (observational) measurements taken on 3 variables W_I W_F and D (Diet), no separate data on Hall. It is quite conceivable that each Hall #Bookofwhy
6.9.19 @2:37pm - (2/2) serves several diets, but the data does not provide separate measurement on H vs. D, so we assume Hall determines Diet unambiguously. Can we continue from this assumption? Or you prefer to introduce distinction between H and D.

6.9.19 @2:24pm - I second it! It is a great summary of the field, and makes the debate about "in statistics" or "extra-statistics" hollow and irrelevant. #Bookofwhy

6.9.19 @7:46am - (Replying to @stephensenn) Ring around the Rosie .... and we still do not what your research question is. A hint, perhaps? or a clue?

6.9.19 @7:20am - Perhaps Fisher/Nelder/Bailey/Speed/@PhilDawid can articulate what your research questioin is? I would gladly give them full credit for the ideas of symmetry/exchangeability if they can use them to solve your problem w/o knowing what you aim to estimate. #Bookofwhy

6.9.19 @7:09am - I cannot understand the issue without knowing three essential elements: (1) What you wish to estimate, (2) what assumptions (if any) you make and (3) what data you have available. Once specified, all "issues" are resolved mathematically. #Bookofwhy

6.9.19 @6:23am - (1/2) DV Lindley was a devout Bayesian, but he was also the first to understood that if (in Simpson's paradox) the same data leads to two different conclusions depending on the story, then Bayes analysis is helpless, because Bayesian methods are propelled by fitness to data. I have
6.9.19 @6:23am - (2/2) written about this in "Why I am only a Half-Bayesian" . It's a good paper because many Bayesians harbor the illusion that, if you only spray priors on models and take sufficient data, the posterior will peak around the correct model. #Bookofwhy

6.9.19 @4:54am - (Replying to @stephensenn) If you aim to analyze experimental studies from two different studies, on two different group, lets do it correctly, along the theory of data fusion. eg . Again, we need to articulate question, assumptions and type of data available. Ready? #Bookofwhy

6.9.19 @2:26am - (1/2) I am saying that the causal story behind the data determines which statistician is correct, and that the story should be articulated in the language of diagrams (not as Rubin and Holland told it). Thus, if we believe that diagram 3 is what Wainer etal meant, we go to diag 3.
6.9.19 @2:27am - (2/2) if we believe that Diet is different from Hall, and that another story describes how data were generated, we can draw the corresponding diagram and interrogate it regarding the correct analysis. Lets do it. Which diagram do you like? What is your research question? #Bookofwhy

6.9.19 @1:31am - I do not understand expressions such as "be all and end all"; I have not used any, nor implied any. As to your question, the diagram on p217 already contains "Hall"; it coincides with "Diet", exactly as presented in Wainer etal. Nothing changed. Just causal modeling. #Bookofwhy

6.9.19 @1:01am - (1/2) It is no longer secret. The lecture that @eliasbareinboim gave at Columbia last month ... was actually a job-talk and I wish to congratulate Elias on accepting a faculty position at Columbia, starting July 2019. I wish also to
6.9.19 @1:01am - (2/2) congratulate my colleagues at Columbia for strengthening their ranks with a top innovator of causal-inference research. May this marriage lead to major breakthroughs and smarter machines. Amen! #Bookofwhy

6.9.19 @12:26am - (1/4) I believe this set of slides reinforces what I tweeted earlier: ... -- its hard to cut the embilical chord to Mother-Stat. It occurred to me that this urgency to stay in Stat-womb was also the motivation behind the potential-outcome framework. The benefits
6.9.19 @12:26am - (2/4) were obvious, nothing is new, Y_1 and Y_0 are ordinary variables, with some missing values, so what? Everything else is ordinary statistics. The price, of course, was (1) Everything was tied to experimental "treatments", not to "events" or absence of events, and (2) we need
6.9.19 @12:26am - (3/4) to express knowledge in the language of {Y_1, Y_0} , namely, in the formidable language of "conditional ignorability". Some would argue: What's wrong in letting statisticians broaden the scope of statistics and then believe that "it is all statistics"? I believe PO is a good
6.9.19 @12:26am - (4/4) of what could go wrong. Distinctions play a role in science. Treating the Ladder of Causation as one chunk, in the name of "It is all statistics", while ignoring theoretical barriers between the 3 levels creates more confusion than the stat-womb warmth can sooth. #Bookofwhy

6.8.19 @9:08pm - (1/1) You seems to oscillate between your desire to validate the model and your desire to answer some causal question, eg. "is there a non-zero effect of Z on Y". Which of the two you wish to start? As @EpiEllie explained, conditioning on the mediator may lower or amplify
6.8.19 @9:08pm - (2/2) the association between X and Y, so this test is out. (Using your notation) If you can put down two competing models, one can tell you right away if there is a test that can distinguish them apart. The name of the game is to confess and let the mathematics work #Bookofwhy

6.8.19 @8:15pm - (1/2) Whether a problem area is causal or not depends not on the diagrams we can draw or intervention we can perform. It depends only on the research question we aim to answer. In my l last encounter with fMRI researchers I was disappointed by their lack of understanding of causal
6.8.19 @8:15pm - (2/2) inference and blind allegiance to potential outcome vocabulary. See . Are these authors representative of the field? If not, please describe what Dynamic Causal Models in fMRI are aiming to estimate, and one simple example of such a model #Bookofwhy

6.8.19 @7:57pm - (Replying to @jwbelmon @sjalexander @EpiEllie) @jwbelmon. I will try to help, but I do not understand what you want to do. If you can specify: your research question + your scientific assumptions, then whether it can be done or not can be answered mathematically regardless of what the field has been doing. #Bookofwhy

6.8.19 @7:49pm - @EpiEllie has done a good job describing the assumptions behind IV methods and their wide applications. I commented recently ... that epidemiologists are now proficient in describing IV methods in DAG's language. The next step is to REPAIR bad IV's #Bookofwhy

6.8.19 @2:50pm - (Replying to @y2silence @SALubanski) Most importantly, do not miss Carlos's final comment, which puts things in the proper perspective.

6.8.19 @5:29am - (1/n) (Replying to @jd_wilko) It is not really "provocative", just a gentle way of luring statisticians to cut the umbilical chord from Mother Stat. Instead of surgical do-operator, you condition on a variable F (force) which does the surgery for you. I used it in 1993 when
6.8.19 @5:43am - (2/n) (Replying to @yudapearl @jd_wilko) I thought that statistician are not prepared for a surgery. Other researchers, too, labor to create the illusion of remaining in the stat-womb. E.g., Heckman etal created a fix-operator to enforce this illusion . The folks at @harvardEpi still teach CI
6.8.19 @5:56am - (3/3) (Replying to @yudapearl @jd_wilko @HarvardEpi) 3 by "imitating RCT's". It is a tough umbilical chord to cut. Denis Lindley was the only statistician I met who said: CUT! In all these schemes we still need to import information from outside the data, which is the key to realizing that we are out of the stat-womb. #Bookofwhy

6.8.19 @4:18pm - Good luck and Happy landing! I am not exactly sure what @AVORA does or how, but when I see people taking causal modeling seriously, I know that #Datascience has a future beyond curve fitting. I will be looking forward to seeing example applications. #Bookofwhy

6.8.19 @3:51am - (Replying to @dlmillimet @PHuenermund) Great blog to teach us how "real world econonomists" think and work. Perhaps it is a good way of slowly introducing econ. readers to advanced methods, using graphs. E.g, the "IV with Endogenous Control" is solved here Others in

6.8.19 @2:49am - (Replying to @mimblewabe @viktorklang) Doubly agree. And this calls for a theory of explainability, namely, what is it about the CONTEXT of the conversation that makes us interpret "why" one way or another, say, "what for" vs. "by what means" #Bookofwhy

6.7.19 @7:09am - The closest would be the talk I gave at USC a few months ago: ... Some poetry, but mainly somber causal wisdom. #Bookofwhy @thinkmariya

6.6.19 @11:46pm - @_MiguelHernan A different take on Cochran's contributions to #observational studies is available here: , which emphasizes his "explanation of the mechanism by which the effect is produced" rather than "imitating" RCT's, which #Bookofwhy advises against.

6.6.19 @9:40pm - (Replying to @rakotesh) The flyer says: CIO UCLA Health Sciences IT TOWN HALL
Friday, June 7, 2019 10:00 AM to 12:00 PM
Pauley Pavilion, 301 Westwood Plaza
Arena Floor / East Bleachers
Enter through the NORTH EAST doors.
I will ask them about streaming, but you may have more clout.

6.6.19 @7:11pm - @EPIDEMIOLOGYY @mendel_random Now that epidemiologists are proficient using DAGs to describe selection bias: ... they will be happy to know that methods of repairing such bias are also available at the nearest DAG's brewery: , #Bookofwhy

6.6.19 @2:33pm - (Replying to @Abraham_RMI @_MiguelHernan @davidlederer) Not reporting would have been an option a few decades ago. Not these days, when we can explicate the assumptions behind our ES estimates and submit them as part of the report: If you buy assumptions XYZ, you must buy conclusions ABC. Conditional conclusions have value #Bookofwhy

6.6.19 @2:13am - If you happen to be at UCLA on Friday, 11 am, you are invited to attend a talk of mine in which I will try to summarize #Bookofwhy to health scientists, in 20 minutes of slides and poetry. See you there.

6.6.19 @1:33am - (1/2) Glad you plan to include precise conditions in your paper and, hopefully, relate them to identification theory. For example, gives examples of when surrogate experiments can decide Q and when they can't. It would be nice to simulate your method

6.6.19 @2:33pm - (Replying to @Abraham_RMI @_MiguelHernan @davidlederer) Not reporting would have been an option a few decades ago. Not these days, when we can explicate the assumptions behind our ES estimates and submit them as part of the report: If you buy assumptions XYZ, you must buy conclusions ABC. Conditional conclusions have value #Bookofwhy

6.6.19 @2:13am - If you happen to be at UCLA on Friday, 11 am, you are invited to attend a talk of mine in which I will try to summarize #Bookofwhy to health scientists, in 20 minutes of slides and poetry. See you there.

6.6.19 @1:33am - (1/2) Glad you plan to include precise conditions in your paper and, hopefully, relate them to identification theory. For example, gives examples of when surrogate experiments can decide Q and when they can't. It would be nice to simulate your method
6.6.19 @1:33am - (2/2) on such examples and see if the existence of h coincides with identifiability of Q (using the corresponding interventional data). Another interesting comparison would be to see how your h works when Q is identified from the (set of) graphs inferrable from the data #Bookofwhy

6.5.19 @11:36pam - (Replying to @mekarim7) Thanks, this was fixed already in our errata sheet.

6.5.19 @11:07pam - Suppose our target quantity is Q=P(y|do(x)), how would we get it from h(x)? Moreover, suppose our datasets are P_i(y|x1, for many many i's, x's and z's, Is Q necessarily identifiable from the available datasets? What happens if it is not? #Bookofwhy

6.5.19 @6:01pm - (Replying to @wgeary) The most inhumane atrocities in the history of mankind were committed by those who saw humanity on one side, and one side only. Now watch BDS puppets chanting: "human rights", "human rights", they can's see humans on the other side. Puppets can't see. They can only chant.

6.5.19 @4:15am - You are not alone, Petros. It is really the best introduction to causal inference that I know. And even the fear of sounding self-promoting will not deter me from saying so. One must be honest when so many communities are still under the boot of traditional education. #Bookofwhy

6.5.19 @2:04am - (1/2) Colleagues keep asking for my opinion on Bottou's method of revealing how the world works: ... I refer them to my May 12 Tweet, where I pleaded with followers to explain how he extracts invariances using training, but to explain it in the language
6.5.19 @2:04am - (2/2) of conditional, or interventional probabilities, P(y|x,do(z)...) , because, all we can get from training are such probabilities. So far, I have not received an explanation, so I remain speechless and mighty curious. Does anyone understand it? #Bookofwhy

6.5.19 @12:45am - @jjz1600 Hi James, Excuse my innocence, but what exactly is wrong with the NYT AD? I was there, in 1948, and I can verify every sentence in the AD as factual and well documented. I would welcome the opportunity to tell you personally what Tlaib's ancestors did to mine. @DrMikeH49

6.4.19 @9:02pm - (1/2) My puzzle is: How many of those econ seminars are aware of the fact that, in addition to "touching" on issues, a large chunk of them can actually be "solved"? I am not asking this question to prove anyone wrong. I REALLY DO NOT KNOW. It is already a year after #Bookofwhy, so
6.4.19 @9:02pm - (2/2) I am sure some econ. students are doing more than "touching", perhaps even with encouragement of their professors. But I really do not know the extent of this enlightenment. I don't see it in the universities that I watch, and I have no way of judging, unless you help me.

6.4.19 @4:34am - I got a review of Ian Stewarts forthcoming book Of chaos, storms and forking paths: How does statistics help us to understand the world? by Andrew Gelman ... My natural reaction: statistics and "understanding the world" are two different things #Bookofwhy

6.4.19 @3:15am - @jo_mendelson Glad you posted this quote from ... It is so easy for university administrators to hide behind "academic freedom" and theological definitions of anti-Semitism, and forget that thousands of creative academics have been criminalized on their campus

6.4.19 @2:50am - Remember Ang Li paper on how to select customers (or patients, or voters) that are most likely to respond to your request/action, though "responsiveness" is an unobserved, counterfactual notion? A revised and improved version is now posted here: #Bookofwhy

6.4.19 @12:27am - (Replying to @yudapearl @dbweissman) 2/2 namely, in econometric, but are afraid to do so even when needed? Or, to put it more mildly, what economics department is likely to prepare students who KNOW how to solve such problems? Does #Bookofwhy underestimate what these students actually know? Genuinely curious.

6.4.19 @12:14am - (Replying to @dbweissman) 1/2 OK, I am willing to reconcile with the existence of the gap between what people know and what they must do by their "fields". Fine. Can we conclude then that in the top economics departments in the US, students DO KNOW how to solve elementary problems in causal analysis

6.3.19 @5:00pm - (1/3) #Bookofwhy from the viewpoint of an enlightened economist. As you can probably guess, I am particularly interested in your comment: "His [Pearl] grasp of what economists, for example, understand ..
6.3.19 @5:00pm - (2/3) and don't understand about causal relationships is incomplete,". What is it that economists DO understand and that I assumed they DON'T? Has this part of "econ. understanding" been expressed formally in the econ. literature since Haavelmo and Cowell's Com.? Can economists
6.3.19 @5:00pm - (3/3) solve the toy problems posed to them here: ? I am genuinely trying to understand what they know that they are laboring to hide from us. E.g., Do they know which parameters can be identified by OLS? Which models have testable implications? etc. etc,...

6.3.19 @12:51am - (Replying to @BD_Zumbo @AlinaVdav) I salute your courage in joining commonsense, it is a dangerous road, as you must have felt already, but your students will thank you forever for saving them from outdatedness. #Bookofwhy

6.3.19 @12:30am - (Replying to @chandra1250) I really do not know, I am just a scribe. Perhaps the publisher knows.

6.3.19 @12:28am - (Replying to @KumailWasif) Thanks for you kind words. I will keep writing, but I need your help in expanding its potentials beyond my limited horizon. Thanks #Bookofwhy

6.3.19 @12:22am - (Replying to @NandoDF @dlowd and 5 others) I have not read Pinker's new book, but I thoroughly enjoyed his "Enlightenment Now". Perhaps because I am a born optimist, or because it brought to my attention data and insight I was not previously exposed to. #Bookofwhy

6.3.19 @12:17am - Your recommender must have been an extremely keen book-reviewer. I happened to have read this book already and, strangely, I enjoy reading it again and again. No dog in the fight, just the pleasure of a child listening to Aesop's fables for the tenth time. #Bookofwhy

6.3.19 @12:00am - (Replying to @riazshahzain1) Thanks for writing. And may that it inspires you to write an even better one.

6.2.19 @11:57pm - If I have encouraged even one professor/student at NYU to voice his/her knowledge and convictions about Israel, or to speak against the way Zionism is maligned on some BDS-occupied campuses, I would consider this Award the greatest honor of my life. Thanks!

6.2.19 @1:02pm - (Replying to @RevDocGabriel) Yielding to friends' wisdom. Will move to "incomplete". Thanks

6.2.19 @2:09am - (Replying to @DanielWhibley @EpiEllie @HarvardChanSPH) Note the difference. In #Bookofwhy you think: "who listens to whom?" In @HarvardChanSPH you think: "What if I intervene?". Luckily the difference does not affect practice until you get a variable like "Blood-pressure" in your model. Now begins the fun:

6.2.19 @12:44am - (Replying to @VDimitrakas) Willing to change adjective. "primitive"? "incomplete"? "limited" ? "insufficient"? "weak"?. Open to suggestions.

6.2.19 @12:28am - Unfortunately, what I said about statistics was misconstrued to mean "bashing" or "contempt". It was not. It was an objective assessment of what can be achieved with the impoverished language that statisticians had to work with. I hope #Bookofwhy is taken this way.

6.1.19 @10:13pm - I've never seen a book so cheap -- 2 pounds on Kindle -- a real bargain. I assumed readers lost interest, but it is still #1 in "probability and statistics" (after everything I said about statistics!) and "AI and Semantics". No explanation! Causal or otherwise #Bookofwhy,

6.1.19 @3:43am - Another supplement to Didelez etal paper on Mendelian randomization can be found in this re-posted blog discussion: ... It took place in 2014, when Bryant and Elias attempted to show Imbens how DAGs can help him live a fuller IV-life @mendel_random #Bookofwhy

5.31.19 @11:48pm - Some readers of Didelez etal paper are surprised by the absence of "d-separation" from the discussion. No need to worry; it is substituted by a surrogate called "Moralized graph", the way it was first proven in Europe by Lauritzen etal (1990)

5.31.19 @11:11pm - It all started when Nancy Pelosi elevated the bigotry of Ilahn Omar to the pulpit of the US Congress and no Democrat dared ask her to consider the consequences. I tried ... Left unchallenged hate is infectious and my fellow Democrats are now stunned helpless.

5.31.19 @12:56am - (Replying to @Basucally @edwardhkennedy) Not merely "to define", but also to derive properties that follow immediately from the definition. For example, consistency, ignorability, testability etc, See #Bookofwhy

5.31.19 @12:49am - This paper by Didelez etal is indeed a VERY good start. It is written in DAG's language and contains extensions to non-linear systems. It need be supplemented with new results on "conditional IV" and "Instrumental Sets", plus specific tasks where MR must listen to DAGs #Bookofwhy

5.30.19 @5:31am - (Replying to @mekarim7) Thanks for catching this typo. #Bookofwhy

5.30.19 @5:27am - It turned out to be more educational (for me) than I thought, and I will soon look into whether Mendelian randomization can be the next beneficiary of DAG's power. Bullet proof vests are back in the closet, awaiting a "mostly harmless" economist. #Bookofwhy

5.29.19 @4:24am - (Replying to @maximananyev @rodrikdani) I believe our conclusions are almost identical. Though I have hard time assessing the over all benefit of the "credibility revolution," as I explain here: . #Bookofwhy

5.28.19 @6:37am - (Replying to @bnielson01) The memory and traces of alternative options are missing from this definition. Does a thermostat have free-will? It seems to satisfy your "error correction" definition.

5.28.19 @4:27am - Finding yourself in this stellar company makes you really humble. What? Who? Me? You must be kidding!! #Bookofwhy

5.28.19 @2:14am - (Replying to @strangecosmos) Agree, yet moralists jump out of their wits if you ask them to accept that free will is just an illusion. This is why the ancients insisted on saying: Its not an illusion, its actually there, 'freedom of choice is always granted'. It seems to clash with determinism -- no more.

5.27.19 @11:37pm - (Replying to @zakkohane @segal_eran and 3 others) Feeling jealous being unable to be with you at the Weizmann Institute (I spent one sabbatical there) and discuss new ideas on AI/Big data with colleagues and students. Have a great meeting.

5.27.19 @11:08pm - (Replying to @strangecosmos) No difference, assuming that the organism that does not have free will has the computational facility to generate the illusion of having free will.

5.27.19 @7:56pm - It is program synthesis with an additional ingredient: the synthesizer module leaves traces in short term memory, so that it can go back to where it was before the incident, re-simulate itself, and perform the repair in accordance with the instruction. #Bookofwhy

5.27.19 @4:28pm - This is how "You should have slowed down" should be interpreted. The key point is, "I, the teacher, do not know your software, so I cant tell you exactly what to tweak, whatever it is, make sure that after tweaking, the resulting action should compute to "slow down". #Bookofwhy

5.27.19 @4:27pm - (Replying to @bnielson01) This is how "You should have slowed down" should be interpreted. The key point is, "I, the teacher, do not know your software, so I cant tell you exactly what to tweak, whatever it is, make sure that after tweaking, the resulting action should compute to "slow down". #Bookofwhy

5.27.19 @1:22pm - Fixing the program would fix the criminal robot, but will not tell others what was wrong with the offensive program. When we advertise widely "for bad behavior of type-I" an entire community of robots undergoes fixing; namely, reassignment of priorities to software #Bookofwhy

5.27.19 @1:09pm - It is the illusion of free will that allows us to supplement punishments with specific repair instructions, eg. "You should have slowed down" , as if the robot had an option to act differently. It's an effective communication trick [ Chapter 10, #Bookofwhy]

5.27.19 @2:33am - Readers remind me that this is precisely the motto quoted below the title of chapter 10 of #Bookofwhy. It reads: "All is pre-determined, yet permission is always granted." -Maimonides (Moshe Ben Maimon) (1138-1204). Note that he does not pose it as a dilemma, but as a fact.

5.27.19 1:28am - (Replying to @mendel_random @UCLA) Barring quantum noise, I am planning to be there. And if you have a list of discussion items, it will give me a chance to prepare the right bullet-proof vest. In gratitude, I will hand you a signed copy of #Bookofwhy and a personal invitation to join the "causal revolution."

5.26.19 @7:52pm - (Replying to @hlprmnky @TheAnnaGat) Godel is right! Our robot cannot have a detailed wiring diagram of its software, but this does not preclude having a blue-print of its software, which may be sufficient for conducting free-will flavored conversations, as we do with the rough blue-print of our software.#Bookofwhy

5.26.19 @6:25pm - (Replying to @TheAnnaGat) Once we understand the neural wiring behind the FW illusion there will be no FW problem. The sentence "I have an option" will be interpreted in terms of the wiring, an interpretation that precludes having an option. #Bookofwhy

5.26.19 @6:03pm - In the old days the FW dilemma was: "God can predict what you are going to do, how can he punish you for what you did?" Modernity replaced God with the equations of physics, messed with quantum uncertainties. Our problem is a robot which surely "has no FW", should we punish HIM?

5.26.19 @4:43pm - (Replying to @jrwill9) Compatibilists (like me) believe causation IS compatible with free will. The question is, does it make sense to tell a robot (driven by deterministic algorithms): "You SOB, you ought to have known better" same way as we talk a child who, presumably, does have free will.#Bookofwhy

5.26.19 @7:50am - (Replying to @chandra1250) I would not be surprised if it has. But I would love to see a reference. I did not call in "emergent property" because of interest to us is the CLASH between two levels of description, not the appearance or disappearance of a property. #Bookofwhy

5.26.19 @6:55am - (Replying to @DebyNavarroR) Zionophobia is an obsession, not a mode of thinking. The only logic it recognizes is the mechanics of slander. They are incapable of realizing that the fate of real human beings may be affected by their barking.

5.26.19 @6:32am - (Replying to @KerrShip) I dont take those attacks to be "nasty" or personal. Zionophobes truly believe they have a monopoly on truth, human rights and social justice. They are deeply shocked to find a thinking organism challenging their bubble. It was important to expose their moral blinders.

5.26.19 @6:19am - An interesting take on free-will. ... They are distorting somewhat what I say, but no harm. I say "there is no free will but it is beneficial to have this illusion and to act as through it is real". They say: "Pearl says there is free will" Not much different

5.26.19 @5:25am - (Replying to @Spinozasrose) One good thing about Zionophobic posts, they sometime link to gems of articles. I have found this gem of Amos Oz ... referenced in an article by a Zionophobic supremacist writer named Litvin. Its worth reading to empower the armies of sense and co-existence .

5.26.19 @4:35am - (Replying to @joshua_saxe @DebyNavarroR) An important clarification is needed: by "Jewish" we mean Jewish in the peoplehood sense, not religion. Otherwise the Zionophobes will jump at you with: "Aha, you want a theocratic state! Like Iran or Egypt!" Luckily the majority of Israelis are secular, bonded by heritage.

5.25.19 @10:45pm - (Replying to @droverbytrade) Are the actual slides available?

5.25.19 @9:14pm - Murray Gell-Mann (Nobel 1969) died yesterday at 89. ... I happened to meet Murray twice and, inspired by the IC algorithm, he showed keen interest in causal discovery. He was fun to be with, always curious and always offering an opinion, often blunt #Bookofwhy

5.25.19 @7:05pm - (1/3) As I tweeted earlier, I must terminate this strange discussion with enemies of co-existence who are genuinely SHOCKED to hear that some people hold different views on Zionism. This is typical of people in the self-righteous bubble of the far-far left. They see themselves as
5.25.19 @7:05pm - (2/3) God's anointed priests of human rights and anti-racism. So, when someone reminds them that other people have rights too, and that denying those right is RACISM, they undergo a traumatic mental SHOCK, "We? Racist?" they ask, "Unheard of!" Someone must tell them:
5.25.19 @7:05pm - (3/3) "Yes, look yourself in the mirror!" I hope the mirror convinces them. In the meantime, I going back to science, but not before offering readers another glimpse at "Who is indigenous?" here: ... and liberal definitions here: ...

5.25.19 @4:27pm - Seeing that you are starting to put ugly words in my mouth, I have to cut off this conversation. For the record, what I have articulated is that Zionism is a home-coming endeavor - a restoration of human rights to indigenous people: ... ...

5.25.19 @4:22pm - (Replying to @SamerAbdelnour @Vieroe @jvplive) Seeing that you are starting to put ugly words in my mouth, I have to cut off this conversation. For the record, what I have articulated is that Zionism is a home-coming endeavor - a restoration of human rights to indigenous people; ...

5.25.19 @3:59pm - (Replying to @SamerAbdelnour @Vieroe @jvplive) I dont think you read carefully what Zionism is. It is POLITE to let people who uphold an ideology define what it is, not those who labor to defile it. An ideology is defined by the way it is taught in kindergartens, not the way diplomats vote, then change their votes, then.....

5.25.19 @3:42pm - (Replying to @KerrShip) The settlement are built on wheels -- some will be uprooted, and some will remain as tolerated Jewish minority in a Palestinian State, once an agreement is reached. The obstacle is the agreement, not the settlements.

5.25.19 @3:29pm - (Replying to @jmourabarbosa) Sorry to disappoint you. But being a scientist means being honest, worship truth, and constantly tune to new evidence. It may disappoint people who have formed an opinion by reading hateful literature about Israel, but I was there, I am well informed and craving for co-existence

5.25.19 @3:12pm - (Replying to @SamerAbdelnour @Vieroe @jvplive) You are right! BDS is not anti-Semitic, it is Zionophobic, a more dangerous form of racism: See why: ... and also: Note: No BDS supporter has ever said: "I am not anti-Zionist" so, the label "Zionophobic' should be a badge of honor

5.25.19 @7:49am - (Replying to @Vieroe @SamerAbdelnour) You forgot that I am also a student of counterfactuals, so lets do a little exercise: Assume the occupation is lifted today, do we have any evidence that Palestinian's violence will subside rather than increase, given their improved position and what they promise their children?

5.25.19 @7:39am - (Replying to @SamerAbdelnour @mszargar @AsraNomani) They tell me that German's fatalities in WW-II were even more lop-sided. What does it prove? That the maximizer fails to accomplish its aims? Oh, I forgot to ask: can you now say: "equally indigenous"? In a hundred years from now? Two hundreds?

5.25.19 @7:28am - (Replying to @wgeary) That Israel tries to minimize fatalities is clear even to her enemies, otherwise they would not use civilians as human shield. After all, even if Israelis are inhuman monsters, fatalities are bad for public relation. Lets stop this nonsense, I know Israel, and you know it too.

5.25.19 @7:15am - (Replying to @SamerAbdelnour @mszargar @AsraNomani) When moral grounds are shifting we invoke "international law". Surely, no one is obligated to accept anyone's neighbor. But those who cry "oppressed" could strengthen their case by showing some commitment to a permanent peace, once oppression is ended, by agreement.

5.25.19 @6:58am - (Replying to @SamerAbdelnour @mszargar @AsraNomani) If I were to smell racism it would be in the glaring asymmetry between: "equally indigenous" on one side and "Me, Me, Me" on the other, followed again by empty slogans from BDS dictionary.

5.25.19 @6:46am - (Replying to @wgeary) Body-count says nothing about the conflict, especially when one side tries to minimize fatalities and the other brags on maximizing them. In statistics we call this a typical case of "selection bias". If you seek peace, support the minimizer, not the maximizer.

5.25.19 @6:32am - (Replying to @belial42) Never in human history was a nation threatened with extinction called an "oppressor". India never questioned the legitimacy of Great Britain, nor did Algeria threaten the sovereignty of France. Eliminationism has its price.

5.25.19 @6:21am - (Replying to @SamerAbdelnour @mszargar @AsraNomani) Calling people "settler colonialists" does not make them less indigenous than your family. Those "colonialists", remember, recite poetry written in that land, in the language spoken in that land, by heroes ruling that land. By dismissing their claims you diminish yours.

5.25.19 @6:05am - (Replying to @belial42) If I were under the boot and I knew that one word would set me free, I would say the word, even if I dont mean it. Trouble is, its hard to cheat, because if you really believe that Israel is temporary, you can't tell your child "Israell is permanent." Its tough to be eliminator.

5.25.19 @5:51am - (Replying to @Vieroe @SamerAbdelnour) Nice theory, except it does not match with facts. The besiegement started in 1936 (no occupation) when Haj Amin al Husseini declared a boycott on Jewish products, and a genocide by entrapment on European Jews seeking refuge from Hitler (My grandparents were among the entrapped).

5.25.19 @5:38am - (Replying to @KerrShip) How about one nation globally? How about just US and Mexico, share the land with equal rights to all, and happiness ever after? HMM! Now we begin to see some difficulties. Well, multiply those difficulties by 100 and we end up with "two states for two peoples". i.e., Zionism-101

5.25.19 @5:01am - (Replying to @Vieroe @SamerAbdelnour) So let us work together to allay this perception of threat. I will do it by telling my Israeli friends "It is not 150,000 rockets, it is only 130,000" and you do it by saying aloud: "equally legitimate and equally indigenous ". Do we have a deal?

5.25.19 @4:52am - (Replying to @SamerAbdelnour) All people, oppressed and not-oppressed, must accept each other right to freedom and dignity. Those who deny freedom to their neighbors must accept some responsibility for the consequences. What I hear on Twitter however is continuous one-side denial, "Me, Me, Me," not one "Us".

5.25.19 @4:35am - (Replying to @SamerAbdelnour @Vieroe) Truth can be shocking, agree. But anyone who aspires to co-existence needs to examine the state of mind of both sides, including Israelis, 95% of whom believe they are under siege by, first, rejecting neighbors, second, 150,000 Hezbolla rockets, third, Iranian proxies, more?

5.25.19 @4:22am - (Replying to @belial42) Sorry, but if you use populist slogans such as "apartheid, shooting children, bulldozing," it is hard for me to believe that you are genuinely "interested" in a comparison. It sounds like you really think Israel "shoots children" for pleasure. This is BDS thinking, tell us more.

5.25.19 @4:05am - (Replying to @Vieroe) I will describe it by, first , avoiding populist slogans (eg oppression, violation, illegal) which connote sadistic intents and, second, a temporary and unwanted predicament imposed on Israel by neighbors who openly promise her demise if she withdraws.

5.25.19 @3:50am - (Replying to @belial42) It is not a matter of "being nice". It is a matter of "as soon as they do not wish us dead, and openly say so". Some difference!

5.25.19 @3:46am - (Replying to @Vieroe) The Israeli government is a product of 71 years of beseigement under existential threats. "Give us one year of normalcy," say my Israeli colleagues "and we will show you what our government can do.

5.25.19 @3:35am - (Replying to @mszargar @NYTimesCohen) I have met literally hundreds of BDS supporters. None of them can utter the words "equally legitimate and equally indigenous". It is against their DNA. I even challenged some of them to

5.25.19 @3:28am - statehood once they grant this right to their neighbors. My friends in Israel are tuned to Palestinian schools and mosques for an INKLING of acceptance -- none thus far, just plans of elimination.

5.25.19 @3:06am - (Replying to @SamerAbdelnour @mszargar @AsraNomani) This is what BDS calls for: ...

5.25.19 @3:52am - (Replying to @SamerAbdelnour @mszargar @AsraNomani) I was right. You cannot bring yourself to say the words "equally indigenous" and you admit that SJP and the entire Palestinian leadership still deny the right of my people to a homeland in any borders. This is what I meant by "victim's cover." A cover for a plan of elimination.

5.25.19 @2:01am - (Replying to @SamerAbdelnour @mszargar @AsraNomani) The principles are simple: "Two states for two peoples, equally legitimate and equally indigenous". Your saintly "marginalized, non-violent movement" (SJP) will not be opposed if it adopts these principles. But it can't, because it was created to fight them under victim's cover.

5.25.19 @1:17am - What has BDS to do with Zionism? Everything! See ... What is Zionism? The right of a people to a homeland -- in some borders. A right denied to my people by every BDS supporter. A genocidal denial that started the conflict 72 years ago, and remains its core.

5.24.19 @7:06pm - (Replying to @causalinf @PHuenermund) I am speechless. What is the research question? assumptions? data?

5.23.19 @10:42pm - (Replying to @robertwplatt) Hard to implement. A review may be 3-5 pages, and all you want to tell the world is that seasoned reviewers in the 21st century still think that "causation is just a species of correlation" or that "confounding is a statistical concept" etc. History needs to know that. #Bookofwhy

5.23.19 @10:28pm - If Israel is so evil, why shouldn't SJP be supported? NYU campus needs to hear that President Hamilton's objections to BDS and SJP are based on moral principles and shared liberal values, not on donors or alumni pressure. Details are here: graduation-speech-quite-objectionable/ @AsraNomani

5.23.19 @9:27pm - Every time someone posts this speech I feel an urge to share it with others, finding myself re-identifying with these ideas, and with the realization that, unless re-enforced, they tend to fade into the background. I worked a lot on this commencement speech - my first. #Bookofwhy

5.23.19 @4:16pm - I propose to re-run our poll, now that we have discussed the pro and cons of publishing juicy reviews of submitted papers.
-- Is it ethical to publish selected extracts from anonymous reviews of your submitted paper?
72% yes
28% no 407 votes. 5 hours left

5.23.19 @9:03am - (1/1) This is amazing!! Thanks for finding it. I can't believe that econometric students would sit through this class and not laugh how hard the instructors sweat to avoid commonsense. I Imagine one of these students meeting an epi friend in the library, or reading #Bookofwhy .
5.23.19 @9:03am - (2/2) or hearing about d-separation, and asking: "Wow! And all this sweat just to prove that economists can do things independently and differently, home- grown style? It is still great progress for eco. education compared with PO w/o dags. See

5.23.19 @7:08am - (Replying to @JicMic) Truth and justice trump slogans and herd mentality.

5.23.19 @6:53am - (1/2) Thanks for posting this confession. I have felt the same when I had to renounce my Award. I have also come to realize that NYU administrators do not know how to handle the monster created on their watch, and no one tells them how. They dont realize that all they have to do is
5.23.19 @6:53am - (2/2) to tell the campus the reasons why Zionist students and faculty are welcome on campus, the distinct contributions they make to the cultural tapestry of NYU and the inspirational power of Israel's miracle on other minorities. They just need to be honest and tell the truth.

5.23.19 @4:09am - Sorry, typo! It is a Rung-1 task on the Ladder of Causation (see #Bookofwhy). Diagnosis goes from evidence to beliefs not caring whether the beliefs are about causes (of the evidence) or about consequences thereof.

5.23.19 @3:52am - An amazing investment in health science and data science: precision-medicine-research-590176 . I am mighty proud to see how the super-education I received in Israel is now spawning pioneering projects of such magnitude to benefit science and mankind. #Bookofwhy

5.23.19 @3:36am - The paper aims at finding minimal Boolean functions that separate positive from negative cases. A Rung-2 task. The function found explains why your diagnosis algorithm concluded that you have disease D. It does not explain why you have disease D or what would cure it. #Bookofwhy

5.23.19 @12:29am - (1/2) (Replying to @theblub @twainus and 10 others) Let me end with one more remark, to rule out any impression that I do not accept metaphysical questions as legitimate. Take the question: "what is a cause". It is legitimate if you allow me to translate it into "how can we explain the consensus among Sapiens about certain
5.23.19 @12:36am - (2/2) (Replying to @yudapearl @theblub and 11 others) cause-effect relationships. This now allows me to further translate the question into a computational one: Find a parsimonious mental representation of shared knowledge that explains how humans can access that representation and swiftly come up with same answer. #Bookofwhy

5.23.19 @12:13am - (Replying to @Kleee @DorotheaBaur) They played this role effectively until mid 20th century. They lost it when they rejected computational metaphors from entering the language of dialogue. My humble prediction: Searle and Dreyfus objections will become curious anecdotes in the history of AI and HAI. #Bookofwhy

5.22.19 @10:38pm - (1/2) (Replying to @theblub @twainus and 10 others) 1/2 I was not asking for a book or a paper, but for an idea or a method. Whether an idea is useful or not depends on what your question is. eg. Algebra is useful if my question is: what values of x satisfy certain conditions. The question "What is taste" on the other hand
5.22.19 @10:44pm - (2/2) (Replying to @yudapearl @theblub and 11 others) does not strike me as a legitimate question because I do not know what kind of answer would satisfy the questioner. Take: "taste is what unicorns swear by" or "taste is a combination of neural activation that result in a scream 'SPICY'", which would be more satisfying? Why?

5.22.19 @7:04pm - (Replying to @SiaKordestani @SFSU) No! Times are changing. It is not "anti-Semitism seeking to cloak itself as political discourse". Rather, it is Zionophobia seeking to cloak itself as normative anti-Semitism".

5.22.19 @1:30am - (Replying to @theblub @twainus and 10 others) The way I speak may betray years of frustration to understand this literature. But if you examine my work your will see that I have never turned away a useful idea because of a mind made up (at least I do not recall such mishap). What's the idea I cannot afford to miss?

5.22.19 @1:17am - (Replying to @Plinz) Everything could be part of higher education if so presented. But accusing you of child molestation in a class on human sexuality is not exactly conducive to higher education. Abdulhadi did precisely that, and we, normative Zionists, now request to be publically decriminalized.

5.22.19 @12:56am - (Replying to @theblub @twainus and 10 others) My language is: English, math, physics, programs and human cognition. Using this languages please tell me one idea that I should learn from Dreyfus , without which I am going to waste lots of time on dead ends. No ref to other philosophers please, no acronyms nor Heidegger

5.22.19 @12:17am - (Replying to @theblub @twainus and 10 others) I am dying to learn from them, but I have not found anyone who can translate them to a language I understand. In particular, to convey an idea from beginning to end without quoting five other phenomenological philosophers, so as to make you feel an outsider. Substance please.

5.22.19 @12:08am - (Replying to @Plinz) Naive it may be, but it need to be aspired to, and if no one says the word "racist" students may get the idea that Adulhadi is a normal person, perhaps even an "educator", so perhaps I and my Zionist colleagues are indeed "white supremacists" . Sorry. She is just a Zionophobe.

5.21.19 @11:47pm - (Replying to @Plinz) It is not "sacralization" but common decency. There is such a term in English called "racism", which is sometimes used improperly and sometimes properly. University administrators often condemn hate speeches that take place on their watch, to set the norms right. No sacralization

5.21.19 @11:28pm - (1/2) Readers ask if I was not too harsh calling Abdulhadi "racist". My answer: What would you call a guest lecturer who comes to your university saying "Muslims are terrorists, but I have nothing against Muslims who disavow Mohammad". Evidently, archaeology professors think it's
5.21.19 @11:28pm - (2/2) "educational". The fault is not really with the instructor, but with a university that does not internalize the equation: Zionophobia = Islamophobia, which means that religion does not have a monopoly on human identity, and all symbols of identity should be equally respected.

5.21.19 @10:56pm - The Dreyfus video reminds me of Marvin Minsky who once said: When I hear someone saying: "A computer cannot do Y", what he is really trying to say is: "I havn't got a clue what Y is". I will change my mind if I someone translates Dreyfus into English, math or program. #Bookofwhy

5.21.19 @6:19pm - Replying to @twainus @yhazony and 9 others Any idea when this video was taken? Dreyfus is repeating the same ancient arguments that I heard him state in the 1970's. Is he back in fashion? #Bookofwhy

5.21.19 @3:40am - This is amazing! So access to anonymous reviews IS available from some journals. Victory!! I feel less inhibited to start my juicy selection. #Bookofwhy

5.21.19 @12:08am - (1/3) Our poll shows a slight preference - 58/42 - to allowing publication of juicy quotes from anonymous reviews. I have hoped for a more decisive preference. Aside from the entertaining value of such a collection, and its encouraging effects on young researchers,
5.21.19 @12:08am - (2/3) I am concerned with its historical value. Written under the shield of power and impunity reviewers comments are the most honest and faithful reflections of the state of mind of a scientific community at any given period. Such historical treasure should not be allowed to rot
5.21.19 @12:08am - (3/3) in the archives of outdated journals. There should be at least some status of limitation before unveiling this information to the public. Anyone knows what happens to these archives? Can a historian request access to the reviews of Turing's paper of 1937 ?? #Bookofwhy

5.20.19 @6:57pm - For all who requested a copy of my contribution to the book "possible minds" (edited by J. Brockman, 2019), I now have a link to it: So let us continue our voyage, from Babylon to Athens. #Bookofwhy

5.20.19 @4:02pm - (Replying to @68kirk) Judea Pearl Retweeted Judea Pearl Causality is all about invariance, of course. But the relationships between these and machine learning should be articulated this way:

5.20.19 @1:47pm - My! My! Thanks for posting this picture --it made my day. I have enjoyed telling students my personal story on surviving 1948. Too bad I did not meet Miss Iraq, my wife is from Baghdad. Go on spreading the word about an inspiration called Israel. Happy Birthday!

5.20.19 @1:32pm - (Replying to @adviceonstock @IanCero) I have a strong issue with this statement. Machine Learning and Deep Learning models do NOTsupport casual inference, unless by "support" one means "can be used in", like "arithmetic supports classical mechanics." See #Bookofwhy

5.20.19 @3:00am - (Replying to @glarange72) CI is Causal Inference, and by "CI + ML hype" I mean the noise generated by the many who claim to be combining the two when they can barely do one. #Bookofwhy

5.19.19 @10:57pm - If you are inspired, as I am, by new perspectives on causal data-science, marinated in substance and free of CI+ML hype, you will enjoy this lecture by Elias Bareinboim which was delivered at Columbia last month Recommended to readers of #Bookofwhy

5.18.19 @8:25pm - (Replying to @maximananyev) Terrific PhD topic: Write a program that automatically generates such a news article. Input: voting data + polls + voters model. Output: a post-election article on "why the polls got it so wrong". Criterion for PhD: article accepted for publication in a top newspaper. #Bookofwhy

5.18.19 @10:45pm - (1/2)
For many years now I have been playing with the idea of publishing a selection of juicy reviews that my papers have received from anonymous, self-assured reviewers. Some colleagues advised against it, since it violates reviewers' trust in eternal concealment. Here is a poll
5.18.19 @10:45pm - (2/2)
1. As a reviewer, would you mind finding portions of your review in print (anonymously, of course)?
2. As an author, do you think this would improve quality and accountability of reviewers?
33%(1) yes; (2) yes.
09%(1) yes; (2) no.
41%(1) no; (2) yes.
17%(1) no; (2) no.

5.18.19 @6:47pm - (Replying to @aesopesque) I do not think risk-averseness would change any of the results; we can always assume that benefits are measured in "utiles" instead of "dollars". #Bookofwhy

5.18.19 @4:23pm - Good luck for the 2020 Eurovision, and happy singing. From Israel (2019), with love, to The Netherlands (2020).

5.18.19 @1:14am - How do you select customers (or patients, or voters) that are most likely to respond to your request/action when "responsiveness" is an unobserved, counterfactual notion? A new paper by Ang Li proposes a method. #Bookofwhy

5.15.19 @11:43pm - (Replying to @WLoosen @SZ) @Wloosen, Thanks for posting. Funny, I received the Dutch translation of #Bookofwhy in the mail today, and could not tell if they mean it. And saintly Mother Theresa on this article made me fall in love again. I love Dutch, the language of Huygens and Till Eullenspiegel

5.15.19 @10:10pm - When I first posted this birthday note, readers asked: Who are the 1948-deniers? Today we have a super-denier in Congress, Rashida Tlaib, who never heard Azam Pasha call for "a war of extermination and momentous massacre" against Israel; Sec. Gen of Arab League, Oct. 11, 1947

5.15.19 @7:46pm - Carlos etal have a new paper on sensitivity analysis which, to the best of my knowledge, is the first to recruit graphical models into this struggling enterprise. I expect to see soul-searching among traditionalists and awakening in the field. #Bookofwhy

5.15.19 @3:38pm - (Replying to @AndersHuitfeldt) How about this:

5.15.19 @2:02am - I was sent a new article on Evidence-Based policy. ... Perhaps it can enlighten readers to tell us where this enterprise stands in the microcosmos of causal inference. #Bookofwhy

5.15.19 @12:38am - (Replying to @DickeySingh) Sure! All we need now is to understand how you decide that the pressure causes the barometer reading, not the other way around. It may be possible, but to understand what's behind it, we need an explanation cast in conditional probabilities, not in training algorithms.

5.14.19 @11:27pm - Kudos to Maayan Harel @maayanvisuals for reminding us of the #Bookofwhy anniversary, and for inspiring us with her eye-opening illustrations. May 15 is also the general calendar birthday of Israel, Maayan's (and my) birthplace - We have much to celebrate tomorrow. Ko Lechai

5.14.19 @5:38pm - (Replying to @mrgunn @vgr @StatModeling) 1.8.19 @11:59pm -- Gelman's review of #Bookofwhy should be of interest because it represents an attitude that paralyzes wide circles of statistical researchers. My reaction is now posted on Related posts: and

5.13.19 @2:45am - (Replying to @HenningStrandin) Correct. But this does rule out "probabilities of counterfactuals" since those individuals may have different behaviors, with a distribution over the behavior types.

5.13.19 @2:15am - (Replying to @eadeli) Wait, wait, you did not get to the part where I encourage students to rebel against their textbooks. This will surely make you angry. #Bookofwhy

5.13.19 @1:58am - (1/ ) I was delighted to read your Causal Ladder blog-post, especially the way you explain the necessity of the counterfactual layer and the vivid examples you used. (I literally forgot the great party we had with Ann and Bob). A word about the exogenous variables U:
5.13.19 @1:58am - (2/2) These variables specify a "unit", be it an individual, an agricultural plot, time of day, etc, whatever refinement is needed to make all relationships deterministic. I hope this clarifies the dilemma posed in your last paragraph. #Bookofwhy

5.12.19 @3:39pm - I must fully endorse this recommendation . I did not realize how much I owe to my training in physics (eg., ...) until I had to face the model-blind thinking of modern statistics, and the difference between statistical and causal models. #Bookofwhy

5.12.19 @2:40pm - (1/ ) Communication between CI and ML folks will improve drastically if we can translate sentences such as: "Bottou trains his NN under conditions ABC" into sentences of the form: "Given the conditional probabilities P(y|x,do(z)...)". After all, what do we get from "training" if ..
5.12.19 @2:40pm - (2/ ) if not conditional probabilities, both observational and interventional. Another benefit for the translation: theorems of impossibility. CI has developed a theory that tells us if certain tasks can be accomplished given information in the form of probabilities P(y|x, do(z)...
5.12.19 @2:40pm - (3/3) We can use this theory to prevent disappointments from "training" schemes that lead to impossibilities. As far as I know, theories of what's possible or not possible were not developed (yet) for training schemes. Why not use what we have? eg, #Bookofwhy

5.11.19 @3:32pm - (Replying to @isli_amar) Your proposal is sincere and well-intended, but it crumbles against the logic of Algerian children who are asking: If Israel is a colonial endeavor (like French rule in Algeria) why should it be recognized in ANY border? Indeed, WHY? Are you prepared to tell them the truth?

5.11.19 @1:48pm - (Replying to @_fernando_rosas @KordingLab) Terrific question!! (which every student want, but is afraid to ask). Answer: We learn quantitative "effect size", P(y|do(x)), while before we had only qualitative information, eg. "X does not listen to Y", or P(x|do(y))=P(x). #Bookofwhy

5.11.19 @1:41pm - (Replying to @nlpnyc) "1948-deniers" are authors (eg Gelvin) who deny that the 5-army attack on Israel in 1948 was GENOCIDAL IN INTENT: "a war of extermination and momentous massacre which will be spoken of like the Mongolian massacre and the Crusades" (Azam Pasha, Sec.Gen. Arab League, Oct. 11, 1947)

5.11.19 @1:18pm - (Replying to @_fernando_rosas @shell_ki) The observed data is just "observed data" ie, a bunch of correlations among variables. It is generated by causal relationships, yes, but it does not tell us what those relationships are, unless we assume extra-distributional assumptions (eg graph) called "causal model" #Bookofwhy

5.11.19 @3:32am - (Replying to @isli_amar) Yes. I remember 1967. The world was waiting for a plausible peace process when, on August 29, the Khartoum Arab League Summit issued its famous: "Three No's"; No peace with Israel, no recognition of Israel, no negotiations with Israel. They can still reverse it-- we are waiting!

5.11.19 @12:26am - Had a great celebration yesterday of Israel's Independence Day - speaking to students as a "1948 eyewitness", feeling like an endangered species, and thinking gloomily: who will bear witness when my generation is gone and the professional 1948-deniers take over?

5.10.19 @11:46pm - When I was a student, speaking against #racism was an obligation, not "courage." It's not funny, the change happened on our watch!!

5.10.19 @2:28pm - (Replying to @sweichwald @atypical_me and 3 others) Your definition is a good one. But I would like to embrace partially specified models under the label "causal model" . For example: X----->Y + X<---U--->Y which does not identify P(Y|do(X)) is still a causal model; it tells us P(X|do(Y)) = P(X) #Bookofwhy

5.10.19 @1:40am - (Replying to @eigenhector @ericjang11 and 8 others) In CI, we call it "functional model" or "completely specified structural model", where the response of each listening variable is functionally defined. In such a model all counterfactual queries, conditional as well as marginals, are estimable. Why invoke quantum? #Bookofwhy

5.10.19 @12:57am - (1/ ) (Replying to @ericjang11 @eigenhector and 8 others) In CI we classify problems along 3 dimensions: 1. what we know, 2. what we wish to know, 3. what type of data sources we have. We then ask: Can we obtain (2) from (1)&(2)? We try to avoid giving agency to algorithms and acronyms. For example, "Model-based RL" is too vague for
5.10.19 @1:03am - (2/ ) (Replying to @yudapearl @ericjang11 and 9 others) evaluation, because its capabilities depends on the kind of "models" invoked, and because you can do everything with a highly refined "model". Whether RL and its varieties can accomplish one task or another depends on (1) and (2), see ... The best way to
5.10.19 @1:12am - (3/ ) (Replying to @yudapearl @ericjang11 and 9 others) communicate about capabilities is to use canonical examples. For example, how would acronym ACR handle the napkin problem? or Joe's "would be salary" problem? Is the output guaranteed to be consistent? In these canonical examples, (1) and (3) are formally specified,
5.10.19 @1:17am - (4/4) (Replying to @yudapearl @ericjang11 and 9 others) as in , leaving no room for ambiguity, thus enabling us to determine which query can be answered from a given combination of knowledge and data #Bookofwhy

5.9.19 @10:50am - (Replying to @theAlexLavin) Great way to think, agree. But what principle need we assume to show that y=f(x,e) is more "invariant" than say x=g(y,e').

5.9.19 @10:40am - Not so. Not when two sides agree to a win/win sharing partnership. Besides, we have not heard this equation on Cinco de Mayo, last week. Reminds me of a parody I wrote on editors who feel the urge to spoil birthdays with equations: Happy Birthday Israel!!

5.9.19 @2:45am - (Replying to @PHuenermund) My, My! I did not realize Israel shares a birthday with EU, both determined to prevent another genocide, and both succeeding thus far, to some extent. I wish however that Israel's peace would depend on something so tangible as coal trade.

5.9.19 @2:31am - (Replying to @BreyonWilliams) Congratulations, Breyon. I usually add: well deserved, after reading one's dissertation but, in your case, I am willing to bet it is. So, welcome to PhD-land, and I hope you revolutionize econometrics.

5.9.19 @1:23am - Today, May 9, is Israel's Independence Day. I invite all readers to join me in celebrating the 71st birthday of the country where I learned to speak, and thanking her for what she has contributed to mankind, and for redefining the meaning of "miracle." Happy Birthday Israel !!!!

5.9.19 @12:14am - (Replying to @arih1987 @Stanford and 2 others) The trick is not to dis-invite racist speakers but to politely explain to them why they should dis-invite themselves, as I tried to do with Cornell West ... The host, the audience and the public got the idea, and some say the speaker got it too.

5.8.19 @10:41pm - (Replying to @miquelporta @socestadistica and 5 others) @miquelprta, what brings you to mention the NYT review of #Bookofwhy? True, it is one of the best written reviews, but how is it related to p-values and to other topics discussed there, on Mt. Olympus?

5.8.19 @4:36pm - (Replying to @AngeloDalli @ylecun) A free (signed) copy of @Bookofwhy to the first person who extracts the principle from the equations.

5.8.19 @3:36pm - (Replying to @babeheim @MPI_EVA_Leipzig) I vow for all equations, Keep us informed when you dig into counterfactuals. Have fun!

5.8.19 @3:28pm - (Replying to @ach3d) good choice!!

5.8.19 @3:27pm - (Replying to @AngeloDalli) Did you get the principle? I can't get to the paper itself, can you? Invariance is always at the heart of causation,; do we have a new method of interrogating the invariance? Let me know if you dig it.

5.8.19 @3:15pm - (Replying to @NevinClimenhaga) Interesting abstract but it hides the principle/assumption till the thick of the paper. Can you summarize it for us in Tweeter length? As I tried to do here , Tool # 7: causal discovery. #Bookofwhy

5.8.19 @4:39am - (Replying to @JDHaltigan @juli_schuess @doinkboy) I sure do, but not as a stand-alone argument. Why? Because to model-blind researchers "confounding" is just "nonignorability" which is defined relative to a "treatment," and requires no notion of "cause". See how Imbens and Rubin labor to explain "unconfoundedness." #Bookofwhy

5.8.19 @2:21am - (1/ ) Depressing to see a once friendly campus (UCSD) consumed in slander. As I note here: ... -"in the grand opera of BDS's slander machine, it is not the libretto that matters but the stage and the megaphone. The charges may vary from season to season,
5.8.19 @2:27am - (2/ ) the authors may ' rotate, and it matters not whether a resolution passes or fails, nor whether it is condemned or hailed. The victory lies in having a stage, a microphone, and a finger pointing at Israel saying, "On trial!" It is only a matter of time before innocent students
5.8.19 @2:27am - (3/3) mostly the gullible and uninformed, will start chanting, "On trial!" It worked in Munich, and it has worked on some campuses. The effect will be felt among the next generation of policy makers.

5.7.19 @11:20pm - (Replying to @JDHaltigan @juli_schuess @doinkboy) I love this argument, but I dont think manipulationists like Hernan will buy it. They would say: OK,"U causes Y" colloquially, but thou shall never say "causal effect of U." Sounds a bit inconsistent? Agree. But we have not heard from them since #Bookofwhy

5.7.19 @6:51pm - Faculty of New York University: Oppose Academic Boycott of NYU Tel Aviv - Sign the Petition! via @Change

5.7.19 @6:33pm - (Replying to @djinnome @juli_schuess @sweichwald) Thanks for noting the break. We will fix it soon (I hope).

5.7.19 @1:47pm - (Replying to @martin_garcia_a @_MiguelHernan) Fine, but I am questioning the benefit of separating "description" from "prediction", skipping "diagnosis" and lumping together "intervening" and "retrospecting" under one opaque category "causal inference".

5.6.19 @8:56pm - (Replying to @thehuntinghouse @IsraelCampus and 6 others) So is the racist mentality of BDS cronies, an extremely interesting object of academic study. A newly evolved aberration.

5.6.19 @8:31pm - As I wrote to NYU president Sexton here: ... "When a group of self-appointed vigilantes empowers itself with a moral authority to incriminate the academic activities of their colleagues, we are seeing the end of academia..."

5.6.19 @8:17pm - I was happy to read today that a petition to restore commonsense to NYU has gathered over 3,000 faculty, students and alumni signatures. ... Refreshed by this response, I feel proud again being a NYU alum.

5.6.19 @1:47pm - (Replying to @ehud) Worrall lives in the pre-causal era (ie, probabilistic causality) and so, it does not satisfy my curiosity: What is "evidence-based-Med"? Is it a QUEST for principles, or a SET of principles? If the latter, is there a simple example where I NEED one such principle?.#Bookofwhy

5.6.19 @4:07am - (Replying to @XiXiDu @wtgowers) We have to take into account that some observers, even in academia, have not watched the "peaceful demonstrators" in action, and truly believe that Israelis shoot civilians for sport. We tend to underestimate the power of Hamas propaganda because it sounds so absurd, but it works

5.6.19 @3:38am - (Replying to @wtgowers) De-legitimization is a very good predictor of how one would act given the chance. The fact it has not diminished over the past 83 years means that one is very very serious about acting, given the chance.

5.6.19 @1:31am - (Replying to @ArashBroumand) You keep mentioning "politics" and "rhetorics" where I see none. I see humanity and I see human rights and I see a conflict that can be resolved when each side acknowledges the human rights of the other. This makes me an optimist, because I have found one such side already.

5.6.19 @12:58am - (Replying to @ArashBroumand) I wish I could share your optimism. Unfortunately, I know that peace can only prevail if both sides agree to the "equally indigenous" principle. I also know that the vast majority of my Israeli friends agree to it - curious what you hear from your Palestinian friends. Optimist?

5.5.19 @11:30pm - (1/ ) (Replying to @mohomran) "Being in a position" is an interpretation. The fact is that Palestinians have been denying Israel's right to exist RELENTLESSLY, for 71 years, 24/7. From school-teachers to TV anchors from journalists to imams from academicians to intellectuals. These are the people whom my
5.5.19 @11:48pm - (2/2) (Replying to @yudapearl @mohomran) friends in the Israeli peace camp are tuned to. None came forth with the word "coexistence", or "equally indigenous". Why 71? It is now 83 years of Arab "me, me, me!" rejectionism. Wait!, how about an intellectual like yourself, can you say the words "equally indigenous"?

5.5.19 @9:53pm - (Replying to @Doc_Yemen @aynumazi) We both know that Hamas/Rashida want more than just "basic necessities". We both read the Hamas Charter (Israel destruction), and we know that Rashida can never say the word "co-existence." We also know that Gazan can have all necessities were it not for what Hamas/Rashida want.

5.5.19 @8:34pm - (Replying to @Doc_Yemen @aynumazi) I was there, in that counterfactual world. Jerusalem was caged in 1947, no food no water and we were promised "momentous massacre" by the Arab League. Still, all we asked our neighbors (+ UN) was a legitimate co-existence for two indigenous peoples. Not denial of the other!!!

5.5.19 @7:08pm - (Replying to @aynumazi) I am VERY serious, and I know something about DAGs and about the logic of cause and effect. So you are invited to join me in examining the logic of Goliath/Rashida/Hamas as they deny their neighbor the very rights they demand for themselves -- freedom and self-determination.

5.5.19 @6:08pm - When will the world learn that Palestinians' conception of "freedom" is somewhat different than the ordinary. It entails freedom to de-legitimize the freedom of their neighbors.. A strange conception indeed, but some think it nevertheless deserves the Hall of US Congress.

5.5.19 @5:48pm - Goliath are those who deny their neighbors what they demand for themselves: right to self-determination.

5.5.19 @4:54pm - When will the world stop dehumanizing Goliath who just want to be free, and who just happened to forget that David and his shepherd brothers likewise, just want a day for freedom.

5.5.19 @2:06pm - (Replying to @nathansttt @ehud) Reading your wife's overview I get a strong suspicion that EBM is an emotional appeal for principles by which we can integrate epidemiological studies with informal opinions of individual physicians. Has it advanced beyond the appeal?

5.5.19 @1:53pm - The more I look at Hernan's classification of data-science: {description, prediction and #causalinference} the more I prefer the Ladder of Causation {Association, Intervention and Counterfactuals} as in #Bookofwhy. Perhaps b/c #causalinference is becoming "what everyone is doing"

5.5.19 @12:57am - (Replying to @ehud) I have never been able to figure out what "evidence based medicine" is all about. Perhaps an expert can explain? Is there a Medicine not based on evidence? Are there requirements on the methods? #Bookofwhy

5.4.19 @10:21pm - I totally agree with @_miguelHernan that the competition devised by DORIE et al. (2019) to compare "methods for causal inference" provides no information on "methods for causal infererence" and should have been titled "methods for estimating certain formulas" #Bookofwhy

5.4.19 @9:27pm - (Replying to @jim_adler @Toyota_AI_VC @PitchBook) I am trying to understand the task that the CAT is asked to perform. Anyone can explain to the uninitiated?

5.4.19 @9:23pm - (Replying to @joaoeira @gwern @Jabaluck) We tried a blog-based discussion on the same issues with Guido Imbens, see ... -- to no avail. I actually prefer Tweeting because it forces you to cut the baloney. #Bookofwhy

5.4.19 @7:58pm - (Replying to @MariaGlymour @danielwestreich @UCSF_Epibiostat) @MariaGlymour You say: "If effect modifiers are differentially distributed, trial effect estimates won't match the target population effects." Is this a new complication that is not captured by the standard conditions of or ?

5.4.19 @7:34pm - (Replying to @chophshiy) Anyone doing causation outside "conventionally qualifying institution" deserves a free copy of #Bookofwhy. If you can drop by my office at UCLA it will be waiting for you. [My sec. is on vacation this week, try next one]

5.3.19 @2:50pm - (Replying to @_julesh_) Not clear to me why the last item is "theory" instead of "counterfactuals". The first two items specify what we can do, not how we do it, shouldn't the 3rd also tell us what we can do with the "theory" that we cannot do with the other two.?? #Bookofwhy

5.2.19 @10:47pm - Confession: the juicy stuff in #Bookofwhy was partly inspired by the juicy stuff of modern academia and its submission to egos and clans. Funny, graphs are still prohibited in certain departments, and top PO researchers can't do the homework problems in

5.2.19 @7:14pm - (1/ ) (Replying to @IsraelCampus @nyuniversity and 5 others) This is a perfect time for President Hamilton of NYU to defend the Study Abroad program on moral grounds and expose the hypocrisy of the department of Social and Cultural Analysis by stating: "A country whose existence is under daily threats cannot be expected to invite in
5.2.19 @7:16pm - (2/2) (Replying to @yudapearl @IsraelCampus and 6 others) people who openly seek its destruction. I [Hamilton] advise members of the SCA department to spend their energy in support Israel's right to exist before criticizing her protective laws or policies."

5.2.19 @2:32am - (Replying to @omaclaren @bjh_ip) This distinct notation is useful, because it allows us to distinguish "; theta)" from "|X=x)". In the former, theta can be an arbitrary index; in the latter, X=x must be an event in our probability space. #Bookofwhy

5.2.19 @12:50am - (Replying to @bjh_ip) Relatedly, most modern Bayesians define their craft as that of assigning priors to parameters and computing their posteriors. Rarely do they examine Bayes' original dilemma: How do we express probabilities that we WISH to estimate in terms of those we CAN estimate. #Bookofwhy

5.1.19 @4:09pm - (Replying to @bjh_ip) Modern Bayesians would do well to take a second look at Bayes' original paper, this time from a causal inference perspective, and follow the evolution of the epistemological term: "Given that we know". I believe Chapter 3 of #Bookofwhy does a good job of presenting this idea.

5.1.19 @2:48am - (1/2) (Replying to @AlexMGeisler @clibassi) Such a course is urgently needed for data scientists as well. The reason I am calling attention to the foundations of counterfactuals is that I see even noted champions of @causalinference frequently abandoning those foundations. For example, the principles discussed here
5.1.19 @2:53am - (2/2) (Replying to @yudapearl @AlexMGeisler and 2 others) are hardly recognized by "experimental" economists, or decorated integrators of ML and CI, and I am still not sure about Harvard epidemiologists. Supreme Courts need clarity, cohesion and consensus. #Bookofwhy

4.30.19 @6:40pm - While debating Gorshuch on how causes and #counterfactuals DONT WORK, let's take a look at how they DO WORK, and what it takes to extract them from regression. Here is a gentle introduction, as harmless as they come: . #causalinference, #Bookofwhy

4.29.19 @6:53am - (Replying to @yoavrubin) Thanks, but don't forget the Primer , which stands between #Bookofwhy and Causality, and which I continue to recommend for people who want to DO #causalinf rather than talk ABOUT it.

4.28.19 @4:28pm - (Replying to @causalinf @agoodmanbacon) Mistakes like this can make history. The more I hear of such mistakes, the more I am tempted to forgo modesty and recommend this book honestly and in the strongest possible terms: Read it! For fun and insight! Here is Chapter 4 for a bait #Bookofwhy

4.28.19 @2:26am - (1/ ) I have decided to retweet my last reply because the distinction between the methodological "science of adjustment" and the substantive "science of diseases" seems to be vulnerable to ongoing confusion. Moreover, one cannot over-emphasize the miracle of the former. Not only are
4.28.19 @2:26am - (2/2) the assumptions qualitative, they are also meaningful and natural, namely, judgments about what variables determine the value of another are the easiest ones for a domain expert to articulate and communicate. Conclusions: adjustment is not an "art," it is a science.#Bookofwhy

4.28.19 @1:51am - (Replying to @EngineerDiet @TuckerGoodrich and 4 others) It is healthy to separate the "science of adjustment" from the "science of diseases". The former is valid whenever your assumptions about the former are valid. Moreover, and miraculously, the assumptions needed are qualitative: who affects whom. The rest is algebra. #Bookofwhy

4.27.19 @3:07am - (1/ ) (Replying to @sifogrante1) You are probably right about the effectiveness of MOOC format. However, we lack the administrative/institutional support needed to launch it. All we have is truth and commonsense, and we trust those who audit MOOCs to compare what they learn to modern ways of doing things,
4.27.19 @3:15am - (2/2) (Replying to @yudapearl @sifogrante1) and share their experience with us on Twitter. In particular, check if the tools taught are compatible with CI as defined in ... or, better yet, if they permit you to solve the toy problems of - my ultimate litmus test. #Bookofwhy

4.26.19 @4:10am - I was sent this illuminating talk by David Gross (Nobel, Physics) on the meaning of "truth" and the scientific method. ... The first part classifies scientific questions as in the Ladder of Causation, and should be recognized by readers of #Bookofwhy.

4.26.19 @4:02am - (Replying to @a40ruhr) Bidirected arrows do not create cycles, the simply state that the correlation is created by some hidden common cause. Still, I would turn this arrow into unidirectional and pose it as a homework for IV enthusiasts. #Bookofwhy

4.26.19 @3:45am - (Replying to @a40ruhr @USF_Economics and 2 others) I fail to detect cyclicity. Can you point to it? I think the structure is ideal to serve as a homework in modern econ. textbooks, asking students to identify the causal effect using IV method. This should lure "experimental economists" to catch up with modernity. #Bookofwhy

4.25.19 @11:17pm - I am intrigued by your DAG, can you flesh it out so we can see the labels? It seems to demonstrate that, contrary to Yale's economists, even simple IV cases need DAGs for identification. IOW, I dont see how Yale students would identify it using "mostly harmless" p.85 #Bookofwhy

4.25.19 @7:48pm - The article you posted realizes that data analysts need to: "understand the data from a wholly new perspective". Fortunately, this realization is taking place while the field is still young, so re-thinking will be less traumatic than, say, statistics or economics. Hope!#Bookofwhy

4.25.19 @1:24pm - (Replying to @Soroush_Saghaf @JoshuaSGoodman @oziadias) Eager to learn from your teaching experience. What, in you opinion, was the most useful tool, or concept, that econ. students learned from #causalinference, which they could not get from econ. textbooks, say Angrist or Wooldwridge. #Bookofwhy

4.25.19 @3:10am - (Replying to @HenningStrandin) It is definitely non-manipulationist, and that is why Rubin&Co +Harvard&Co refuse to use it, despite the fact that it gives meaning to counterfactuals, which they do use, but only when enslaved to a "TREATMENT" ; they are missing out on "undoing of past events. " #Bookofwhy

4.25.19 @2:56am - (Replying to @yskout) The beauty of this formalism is that, for certain questions (eg interventions), all you need is the graph structure; for others (eg retrospective counterfactuals) we need the functional form as well. Its a miracle, dont miss. See #Bookofwhy

4.25.19 @12:41am - Here is a more intuitive definition of Structural Causal Model (SCM). It is a society of LISTENING variables, and a specification of how each variable would respond to what it HEARS from the neighbors. Though somewhat informal, I find it more meaningful than the formal.#Bookofwhy

4.24.19 @5:19am - (1/ ) (Replying to @totteh) Agree with the wisdom, and agree with the pioneering value of this book. However, one weakness is that it does NOT give an example where POs and DAGs are used to describe the same causal structure. More severely, POs and DAGs are treated as two separate frameworks. It would be
4.24.19 @5:30am - (2/2) (Replying to @yudapearl @totteh) helpful to show readers that the two emerge from the SAME mathematical object, Structural Causal Model, as is shown for example in Section 4.2 of , especially "The Fundamental Law of Counterfactuals" Eq. (4.5). #Bookofwhy

4.24.19 @5:06am - (Replying to @nagpalchirag) If you show me one example where identifiability is easier to see in PO, I will convert to voodoo. #Bookofwhy.

4.23.19 @9:20pm - Sad news. Nils Nilsson passed away An AI pioneer, and a mentor to many of us since the 1970's. Always encouraging and always insisting on understanding new ideas, and how they fit together in the grand scheme. I will miss him immensely.

4.23.19 @8:58pm - (Replying to @nagpalchirag @DaphneKoller) First time I hear someone saying so. Could you share with us the first place where you found PO illuminating do-calculus. Eager to learn. #Bookofwhy

4.23.19 @4:53am - (Replying to @dlmillimet @PHuenermund @causalinf) Personal choices -- I will salute to that!! As long as we know what the options are. And, to accelerate progress, we can let our students too see what the options are. #Bookofwhy

4.23.19 @3:20am - (Replying to @DorotheaBaur @bobehayes) Do you think Kalev Leetaru (Contributor) read #Bookofwhy? Worth checking.

4.23.19 @3:09am - (Replying to @ingorohlfing @DToshkov) The do-calculus is "type"-level, and its relation to Woodward's interventions is discussed here:

4.23.19 @3:01am - (Replying to @DToshkov) "Intervention" is type level, eg. careless driving causes accidents. Counterfactuals are 'token' level. eg "this accident would not have occurred had you driven carefully." Both are handled harmoniously in the SCM framework, for philosophers to rejoice #Bookofwhy

4.23.19 @2:49am - (Replying to @smhall97 @alienelf @elanvb) This is a beautiful paradox, thanks for reminding me. In my former life I solved it using Bayesian analysis. Today I would approach it as a causal problem and model the process by which the envelops are stuffed; the prob. of doubling A is not 1/2 but varies with A. #Bookofwhy

4.23.19 @12:08am - (Replying to @PHuenermund @dlmillimet @causalinf) Cruel? Please compare to modern treatments of measurement error here: and here , in both clarity and generality. Cruel??? #Bookofwhy

4.22.19 @11:02pm - (Replying to @UusitaloLaura @KirsiNorros @teemu_roos) I have not seen #Bookofwhy in a picture frame before. Glad it was translated into Finnish. Happy voyage to you.

4.22.19 @9:51pm - (Replying to @jkrt168t) What makes experimental economists strangers in CI-land is the fact that they do not start with the available causal information and use it (with data) to answer causal questions. Instead, they start by applying a pre-canned procedure to the data and then ask: Does the estimate
4.22.19 @10:55pm - (2/2) (Replying to @yudapearl @jkrt168t) obtained have "causal interpretation?". Or, "Under what conditions can we interpret the estimate obtained as a valid answer to our question?" Those conditions are then articulated in opaque language (eg ignorability) far removed from ordinary scientific thinking. #Bookofwhy

4.22.19 @4:26am - (Replying to @RevDocGabriel) I am not familiar with multi-omics data.But I can refer you to integrating data from heterogenous sources involving, for example, diverse populations, diverse samples, experimental and observational data etc. See or #Bookofwhy chapter 10, pp 350-358.

4.21.19 @11:06pm - (1/2) For AI researchers still immersed in the debate between model-based vs. model-blind AI, I am retweeting a response of @wellingmax to Sutton's blog. Max agrees by and large with "The Seven Tools of CI", , albeit w/o noting the theoretical impediments.
4.21.19 @11:06pm - (2/2) Along a similar vein, I was asked to retweet the last line of my slides in the Why-19 symposium . Gladly; it reads: "Only by taking models seriously we can learn when they are not needed." And I still vow for it. #Bookofwhy

4.21.19 @8:14pm - (1/ ) In view of persistent ambiguities regarding the definition of "causal inference" (CI) I am sharing here the definition that has guided me successfully throughout my journeys. CI is a method that takes data from various sources, as well as extra-data information, and produces
4.21.19 @8:14pm - (2/ ) answers to questions of two types (1) the effects of pending interventions and (2) the effects of hypothetical undoing of past events. See Causality (2000) Chapter 1. A vivid and recurrent example of a non-causal question is any question that can be answered from the joint
4.21.19 @8:14pm - (3/ ) probability distribution of observed variables, eg, correlation, partial regression, Granger causality, weak and strong endogeneity (EHR 1983) etc. See . This definition excludes Pearson's (1911) and Fisher's (1925) descriptions of statistical tasks
4.21.19 @8:14pm - (4/4) and I would reserve judgment on how "experimental economists" fit into this definition. I believe that, in due time, "experimental economists" will manage to articulate formally what "extra-data information" they use, and thus become bonified members of CI. #Bookofwhy

4.21.19 @3:16pm - (Replying to @ahmaurya) The vast majority of economists that I know would be offended if labeled "regression analysts", in the same way that physicists would be offended if labeled "arithmeticians", though they use arithmetical operations every hour of the day. Regressionists live and die w/o causes.

4.21.19 @1:45pm - (1/ ) (Replying to @ahmaurya) You do not understand me correctly. The economentric literature is motivated by causal questions, and has pioneered modern causal inference. See . Traditional regression analysts, however, shun causation, which evoked my surprise at the paper discussed
4.21.19 @1:59pm - (2/ ) (Replying to @yudapearl @ahmaurya) which starts and ends with regressional questions and, surprisingly, invokes a causal diagram. "For what purpose?" I asked. But may I suggest that, instead of putting words in my mouth, please articulate the research question you claimed I have been avoiding. #Bookofwhy

4.21.19 @12:08am - (1/2) (Replying to @ahmaurya) Are you asking why I am surprised to find "regression analysts seeking the wisdom of causal diagrams when they are not asking causal questions"? Ans. Because I have not seen it done in the regression literature, not even in the econometric literature and its "tricks". But
4.21.19 @12:21am - (2/2) (Replying to @yudapearl @ahmaurya) I thought you have earlier accused me of avoiding a burning research question, the answer to which is revealed in the econometric literature. Glad we are no longer there. Or, if we are, what is that question that I avoid? #Bookofwhy

4.19.19 @11:20pm - (Replying to @ahmaurya) I wish I knew what I have done wrong to earn this Tweet. What questions are you asking whose answers I avoid? Try me, I would love to learn from you and the books you read.

4.19.19 @4:54am - (1/3) I have read this paper with great interest, trying to understand what makes regression analysts seek the wisdom of causal diagrams when they are not asking causal questions and labor merely to assess the magnitude of measurement errors. The answer seems to be two fold.
4.19.19 @4:54am - (2/3) (1) The diagram allows them to use Wright's Rules to compute correlations among latent variables (X,Y) in terms of correlations among observed proxies (x',Y'). This could be done, of course, w/o the diagram, but only at the cost of painful algebraic
4.19.19 @4:54am - (3/3) derivations, as in econ. (2) The problem is in fact causal in disguise. Why else would anyone be interested in cov(X,Y) as opposed to cov(X',Y') which is estimable from the data and is sufficient for all predictive tasks? Curious if other readers agree. #Bookofwhy

4.18.19 @2:09pm - (Replying to @y2silence) Amazing discovery!! Unveiling the origin of ideas. #Bookofwhy

4.18.19 @2:06pm - (Replying to @abesilbe) I just spoke with students from NYU Realize-Israel. Harrassment, threats and intimidation is unfortunately the modus operandi of the NYU SJP chapter. Please speak to them. Curious, what gives you the hope that they would be different?

4.18.19 @1:15pm - (Replying to @abesilbe) I fail to grasp the logic of this new theory of evidence. You attended a symposium that was not interrupted by SJP, from which we should infer that SJP does not resort to disruption tactics nation wide, and did not interrupt the May 17 "indigenous people" meeting at UCLA?

4.18.19 @4:33am - An update from NYU: ... It seems that NYU administrators have discovered a new and courageous way of handling disruptive student organizations: Give them awards and do not show up to the ceremony. Dont ask, dont tell.

4.17.19 @4:33am - (Replying to @arnoldroth @epavard) And a joyful Passover to you @arnoldroth, and to your family. We will never forget your beautiful daughter Malki. She and my son Daniel are the torches to the light of which this insane world may discover one day why the normalization of evil is twice as evil as evil itself.

4.17.19 @2:33am - (Replying to @HolgerSteinmetz) Strange, I feel the opposite. My impression is that, at least on this Tweeter forum, the number of such causality-dead people is shrinking by the hour. But this impression is infected of course by heavy (and hopeful) selection bias. #Bookofwhy

4.16.19 @4:19pm - (1/ ) This sad happening at NYU unveils the power of ignorance in the electronic age. I bet my esteem colleagues at NYU do not know that their university is awarding a "president service award" to SJP, a student organization that prides itself on crushing meetings of other student
4.16.19 @4:19pm - (2/2) organizations. How can you tell when your university administrators are embarrassed by their own words? When they start lecturing you on "free speech" -- the ultimate blanket for inaction or lack of courage. Any NYU alumni on this Tweeter?

4.15.19 @3:29am - (Replying to @AngeloDalli) Interesting. But we need to ask Miguel if this is what he meant by "description". His examples do not match this interpretation. #Bookofwhy

4.15.19 @12:39am - In the interest of many readers of Primer who requested to see a derivation of Eqs. 4.13 and 4.14, I am re-tweeting here an earlier post by Ang Li. It may look complicated, but it is a straightforward application of probability calculus. Thanks @Ang_UCLA

4.15.19 @11:51pm - (1/2) I believe it is a mistake to assume that business applications care only about interventions, not about counterfactuals. An astute businessman wishes to spend his advertisement budget on people who are "swayable potential customers" not on "captive buyers". The distinction
4.14.19 @11:51pm - (2/2) between these two groups is counterfactual in nature, and requires counterfactual logic for definition, analysis, identification and estimation. There is more to counterfactuals than meets the eye, and it is all here: , told and exemplified. #Bookofwhy

4.14.19 @3:33pm - (Replying to @DrJohnKang @cd_fuller @julian_hong) Precisely, except that we need to add: "...would behave under conditions different from those Joe's encountered in the the trial." See . Also, the word "generalizability" is reserved for extrapolations across diverse populations. #Bookofwhy.

4.14.19 @2:39pm - (1/ ) An important correction. I actually do NOT agree with Hernan's classification of #datascience. First, I do not see substantive difference between "description" and "prediction'. Second, the "counterfactual" layer should be split into two, intervention and counterfactuals, as
4.14.19 @2:39pm - (2/2) in the Ladder of Causation #Bookofwhy or . The reason is that these two layers of the ladder require different types of knowledge. You can never tell if "Joe's headache would have gone had he not taken aspirin" by conducting RCT on aspirin and headache.

4.14.19 @5:17am - (1/ ) I just finished Hernan's "Second Chance.." and I recommend it strongly to data scientists, statisticians and even eocnomists. If #Bookofwhy has not convinced you that curve-fitting is not sufficient for CI, Hernan's paper will. I mention economists b/c based on our Twitter
4.14.19 @5:17am - (2/ ) conversation, "experimental" economists will not buy Hernan's "the validity of CI depends on structural knowledge". Worse yet, economists conception of ML & Econ, [Millainathan, J. Econ Pers. cited by Hernan] confirms my suspicion that their resistance to structure is more
4.14.19 @5:17am - (3/3) than a passing fad. I differ from Hernan on the primacy of RCT as a conceptual model for CI, as well as on the taxonomy of causal tasks. I am sure nevertheless that this paper will sway more #Datascience enthusiasts to the #causalinference fold.

4.14.19 @3:21am - (Replying to @twainus @katecrawford and 17 others) I must be too late for this trolley, I can't see how it relates to Turing machine, probabilities and artificial intelligence. Suppose probabilities were invented to model us (the way we handle uncertainty), what then? are we closer to GAI? #Bookofwhy

4.13.19 @12:47pm - (Replying to @kareem_carr @EpiEllie) I believe that, deep in their heart, data scientists dream of an "automated data scientist". #Bookofwhy

4.13.19 @12:39pm - (Replying to @rlmcelreath) Most profound observation. Suggested remedy: (1) Post a "preliminary version" on your site, written for your students and friendly colleagues. (2) Cite the "preliminary version" in your final article. (3) We, readers, will ignore the "final" and read the "preliminary". #Bookofwhy

4.13.19 @12:29pm - Note that Hernan's trichotomy is different from the Ladder of Causation in , which raises 2 questions: 1) Where is "diagnosis" (or "abduction") situated? and 2) Where is the barrier between intervention and (unit-level) counterfactuals situated?#Bookofwhy

4.12.19 @10:58pm - Gee, thanks. I have almost forgotten about this overview and, now, as I read it 10 years later, I am amazed at how little I would change in it. Highly recommended!! Adding (1) Transportability and (2) Missing data, and would I still stand behind it today. Amazing. #Bookofwhy

4.12.19 @4:19pm - (Replying to @cqfdee @edwardlandesber @ericcolson) Good point. I called it "a gift of the Gods" in one of my slides. It is hard to believe that smart people like Heckman would go through all possible acrobatics to avoid it. Any theory for his motivation? #Bookofwhy

4.12.19 @3:44pm - (Replying to @elibressert @ericcolson) If you have a copy, can you share the Table of Content? Amazon does not allow us to peek inside. Strange.

4.12.19 @12:40am - (Replying to @paulfkrause @QuantaMagazine) I am an AI researcher who discovered, in the 1990's that economists of the 1940's had some good ideas about causation which were abandoned since. I have borrowed them for AI and tried to share with econ., but the forces of resistance turned out insurmountable. #Bookofwhy

4.11.19 @5:19pm - (Replying to @the_dismal_tide @PHuenermund and 4 others) The real world is always beyond us, all we can hope for is to capture faithfully our understanding of reality, which is depicted on the r.h.s DAG. And I never resonated to Box slogan: all models are wrong. It's empty unless you can tell us WHY some are more useful than others.

4.11.19 @12:44am - (Replying to @PHuenermund @WvanAmsterdam and 3 others) A beautiful depiction of the whole discussion of "experimentalists" vs. "structuralists". The left part is the template the experimentalist seeks to fit, the r.h.s is the mental DAG he/she possesses while fitting. We see clearly how the fitting process can be automated.#Bookofwhy

4.11.19 @12:18am - @ale_martinello called my attention to an excellent 2014 article on causal inference, history of ideas, in The New Atlantis: ... Rarely can we find such lucid writing combined with profound understanding of this difficult topic. Highly recommended. #Bookofwhy

4.11.19 @12:04am - (Replying to @MPenikas @ale_martinello and 5 others) Agree! The New Atlantis paper is excellent. Rarely can we find such lucid writing combined with profound understanding of this difficult topic. I dont know how I missed it thus far. Thanks for posting. #Bookofwhy

4.10.19 @10:18pm - (1/3) Commending you on so skillfully navigating the waters of DAGs and PO. That the two are compatible comes from the fact that both are derived from structural causal models (SCM). DAGs are used to encode what we know and PO what we wish to know. However, I find it hard to
4.10.19 @10:18pm - (2/3) understand why you say that "PO are most useful for estimation". Assuming that we have obtained an estimand using DAG-based identification, isn't the estimand itself sufficient for estimation? Do we really need to dress it in PO cloths before proceeding with the estimation
4.10.19 @10:18pm - (3/3) phase? This dressing habit, I believe, is a remnant of bygone age when, lacking DAGs, people attempted to identify queries of interest in the PO language. But why go through that tormented experience today, when we do have DAGs.??? #Bookofwhy

4.10.19 @4:36pm - Primer's Errata has been updated: . Def. of front-door condition corrected, to read "no backdoor path". Thank you @juli_schuess .Good catch. #Bookofwhy

4.10.19 @3:18pm - (Replying to @ale_martinello @edolaw76 and 4 others) Another Grazie !!!

4.10.19 @2:50am - (Replying to @juli_schuess @_MiguelHernan) The issue of "multiple versions" attains clarity when cast as "disjunctive action" e.g. P(Y=y| do(X=a or X=b)) [paint the wall green or purple]. It is illuminated in and , which hopefully resolve @IanLundberg dilemma. #Bookofwhy

4.10.19 @12:21am - (1/2) (Replying to @ycaseau) Thanks for posting this adventurous hunt for causation. The only comment I have is on "Granger causality". In 1991, I had a quiet dinner with Clive Granger in Uppsala, Sweden. Between the 2nd and 3rd glass of wine, he confessed to me that he feels embarrassed by the name:
4.10.19 @12:26am - (2/2) (Replying to @yudapearl @ycaseau) "Granger causality", since it has nothing to do with causality, but he can't stop people from using it; they need some way to express what is needed. I think we should honor him by echoing his distinction. #Bookofwhy

4.9.19 @12:12am - In all modesty, this is one of the most profound statement I read on Twitter. My heart goes out to students who struggle with PO w/o first solving a toy counterfactual problem like Joe's score in Fig. 4.1 here: . It borders on child abuse. #Bookofwhy

4.9.19 @12:12am - In all modesty, this is one of the most profound statement I read on Twitter. My heart goes out to students who struggle with PO w/o first solving a toy counterfactual problem like Joe's score in Fig. 4.1 here: . It borders on child abuse. #Bookofwhy

4.8.19 @8:25pm - (Replying to @annasdtc @causalinf and 2 others) Too much to read!! And not enough time!!! If you were my student at UCLA I would tell you: "One toy problem is worth ten books - Play!" #Bookofwhy

4.8.19 @9:49am - (Replying to @TheShubhanshu @fhuszar and 5 others) Great that you are curating this resource. Let us know how we can help you.

4.8.19 @1:31am - (Replying to @fuiud @omaclaren and 2 others) Yes, I have. Perhaps a later version. JP

4.7.19 @11:13pm - It smells like spring is rising in econometric education -- DAGs before regression. Keeping my fingers crossed, and wishing Econ. students a safe voyage and a bright future. #Bookofwhy

4.7.19 @10:36pm - (Replying to @nickchk) The idea of postponing discussion of regression is what econ. education has been waiting for. The confusion between structural and regression equations is THE major hindrance for econ. students and pushing the latter to the estimation phase should be a life saver. #Bookofwhy

4.7.19 @10:25pm - (Replying to @omaclaren @eliasbareinboim @analisereal) To help me navigate through the pages, can you tell me where I would fumble if I were to assume that all is fine and rosy in my naive conception of identifiability and estimation. What difficulties I would end up facing, etc.

4.7.19 @10:20pm - (Replying to @pedrohcgs @omaclaren and 2 others) Is it fair to say that the functional form of the estimand determines critical properties of the estimator, such a being regular vs irregular. ??? Or is it some other factors in the problem, say the data?

4.7.19 @5:00pm - A tiny comment on "repackaging of things we know". What exactly do we "KNOW" that #Bookofwhy has "repackage"? We KNOW the rules of chess. These rules dictate which board position is a WIN for white. Do we KNOW which board position is a WIN for white?" Care to explain 'repackage'?

4.7.19 @4:46pm - (Replying to @RuochengGuoASU) Agree. Now imagine a whole book on IV, named "harmless", with not a single DAG, and a whole movement in econometric named "credibility" cheering this book for replacing DAGs with ignorability judgments: "Z is conditionally independent of the potential outcome of Y, had X been....

4.7.19 @4:36pm - (Replying to @KevinDenny) Show me one who does not find it handy, and I'll show you one who has not tried to use conditional IV, or one who pretends to be able to judge if "Z is independent of the potential outcome of Y had X been x, given W=w". Which one did you have in mind? #Bookofwhy

4.7.19 @4:24pm - (Replying to @y2silence @HannesMalmberg1 and 4 others) Bingo!! I was going to recommend these two papers, and you beat me to it. Thanks @Bookofwhy

4.7.19 @4:15pm - (Replying to @HannesMalmberg1 @autoregress and 3 others) I must repeat and repeat the CREDIBILITY aspect: It is not merely to "encourage formalization" (which is also the aim of PO, etal ) but to encourage formalization at the level where knowledge RESIDES, as opposed to a level where knowledge is transformed to appease ID-strategists

4.7.19 @4:02pm - (1/3) The simplicity of IV validity quickly disappears with nuances. see ... But, a more important aspect of the "repackaging" is CREDIBILITY, namely, judgments are recruited from where they reside, not from where they are distorted to appease the identifier.
4.7.19 @4:02pm - (2/3) Consider the IV validity again, and ask yourself "what judgements were necessary to execute this exercise?". Mark them. Now compare to the judgments required in the PO framework, which are cast in ignorability language, (See Angrist etal). Finally, ask "What type of judgments
4.7.19 @4:02pm - (3/3) would be more CREDIBLE if I were the one to make them?" I am sure your assessment of the value of "repackaging" will become one of greater appreciation, perhaps even one of necessity. #Bookofwhy

4.7.19 @2:35pm - (Replying to @HannesMalmberg1 @autoregress and 3 others) Theory is easy as long as you have 4 variables in the canonical template. Add a nuance and simplicity is shaken. Add one "control variable" (ie conditional IV) and theory is helpless, experts are embarrassed and must resort to: "how often has it happened in practice". #Bookofwhy

4.7.19 @2:12pm - (Replying to @HannesMalmberg1 @Jabaluck and 4 others) Those who are less courageous salute you.

4.7.19 @2:10pm - (Replying to @HannesMalmberg1 @Jabaluck and 4 others) Start with teaching, then time will tell, and time will forever be indebted, till the and times are done. @Bookofwhy

4.7.19 @5:32am - Summarizing our discussion of IV's, here is how DAGs distinguish valid from invalid IV's: ... Prudent economists may wish to use these conditions in research, teaching and communication, then convert to a DAG-free language for publication. #Bookofwhy

4.7.19 @2:35am - (Replying to @yudapearl @nickchk and 2 others) 2/exemplified in @nickchk DAGs, but it goes further, to include cases with multiple unobserved variables, multiple paths between Z and Y, as well as auxiliary "control variables". A good source for generalized IV is while #Bookofwhy is a conceptual primer

4.7.19 @2:16am - (Replying to @nickchk @Jacobb_Douglas @Fhanksalot) Discussants on this thread might enjoy a glimpse at the general condition for Z to be a valid IV. The condition goes: 1) There is an unblocked path between Z and X, and 2) Every unblocked path between Z and Y contains an arrow into X. This condition confirms of course the cases

4.6.19 @4:17pm - (Replying to @EpiEllie @Jabaluck and 4 others) Judea Pearl Retweeted Nick HK And here is the accompanying thread as well @nickchk ... I am dying to see an experimental economist, versed in PO language, 2sls, and "credit expansion" (but no DAG) explain to his/her students what kind of proxies should qualify as "good IV's" . #Bookofwhy

4.6.19 @3:52pm - (Replying to @EpiEllie @Jabaluck and 4 others) Yes, that's the one. Thanks

4.6.19 @3:50pm - (Replying to @fuiud) Only DAG-averse researchers use this scare tactics, and talk about "an enormous DAG", larger than absolutely necessary for scrutinizing the assumptions upon which the conclusions rest. I am surprised people are still buying it. #Bookofwhy

4.6.19 @3:10pm - (Replying to @Jabaluck @autoregress and 4 others) The thread I tried to recover was much longer, just two days ago. But we are not going anywhere on this track. I wish you could say: I, Jason, can answer these questions correctly w/o DAG. Instead, we hear about @autoregress doing it, or "Go read about credit expansion". I'm out.

4.6.19 @2:30pm - (Replying to @Jabaluck @autoregress and 3 others) I have tried to recover this thread, but it disappeared under a hundred later "notifications". I believe @EpiEllie retweeted it, just two days ago. Any recollection, @EpiEllie ? We are searching for a Twitter discussion on "Must an IV be a cause of the treatment?" #Bookofwhy

4.6.19 @2:20pm - (Replying to @yudapearl @Jabaluck and 5 others) When I hear Mr. X saying: "A computer cannot do Y", what Mr. X is really trying to say is: "I havn't got a clue what Y is". I differ with Minsky on many issues, but not on this one. #Bookofwhy

4.6.19 @2:11pm - (Replying to @Jabaluck @robertwplatt and 4 others This is the first time I hear about a task that (1) humans do, (2) computers are on their way to do, and (3) a professor says: "It can't be done." I heard it being said about emotions, religion, free will, etc. not about a scientific task. Marvin Minsky once said:

4.6.19 @1:12pm - (Replying to @Jabaluck @robertwplatt and 4 others) Luckily, the "automated-Angrist" is on its way, where hand-waving talks about "intuitive identification strategy" become "principled identification strategies" #Bookofwhy

4.6.19 @1:04pm - (Replying to @Jabaluck @autoregress and 3 others) My blindness!! "discusses at length"!! And yet, in 2019, dozens of economists argue (on Twitter) "What if I can only measure a proxy of an IV ?".. (and other variants on the theme.) Should we send them to "any applied paper worth a salt"? I fail to name one. Help?. @Bookofwhy

4.6.19 @12:38pm - (Replying to @Jabaluck @HannesMalmberg1 and 4 others) Beg to differ. Barring cycles, DAGs are analogous to writing down mathematical equations for the first time! That part of the model which is not captured by its DAG's abstraction entails assumptions that makes modelers uncomfortable, eg, parametric forms... #Bookofwhy

4.6.19 @12:19pm - (Replying to @taragonmd @SF_DPH) Third time!!! We need to fix that!! What did we do wrong? It was supposed to be like a jingle, you hear it once and you ham it the rest of the day. Please brief us here, in Twitter's trenches, how the book club is going. #Bookofwhy

4.6.19 @3:41am - (1/ ) You keep sending us to decypher a subject-specific body of "applied papers" that are ostensibly "concerned" with certain questions, but where those questions are not asked explicity and where it is very hard tell whether/how those "concerns" are pacified to the satisfation of
4.6.19 @3:43am - (2/ ) (Replying to @yudapearl @Jabaluck and 4 others) their authors. Lets take the question: "Is X a good IV?" I have just seen a long Twitter discussion on whether a non-exogenous X can be a good IV. This is 2019, 90 year after the invention of IV, and I do not think the applied economists who are tormented by this
4.6.19 @3:50am - (3/ ) (Replying to @yudapearl @Jabaluck and 4 others) 3/ very applied question would find the answer in the "applied papers" you uphold as supreme oracles of economic wisdom. #Bookofwhy

4.6.19 @3:24am - (Replying to @HannesMalmberg1 @autoregress and 4 others) I did a similar experiment in dozens of Econ departments nation-wide. My mistake was to publish the results. Do not repeat, even if you have a tenure. Observe the ferocious efforts your colleagues are making to maintain the myth that "you can also do it in PO". #Bookofwhy

4.6.19 @2:22am - (Replying to @TenanATC @NateSilver538 and 2 others) Tracking trolling discussions may be a fun pass time but hardly a learning exercise. My mantra: Solving ONE toy problem in causal inference tells us more about science than 10 debates, no matter who the debaters are. #Bookofwhy

4.5.19 @8:32pm - (Replying to @EllieAsksWhy @aesopesque) The SCM framework actually saves economists a huge amount of effort, see The arrow-phobic resistance in economics is driven mostly by egos and insularity. It is changing now; students are beginning to see what the phobia has denied them of. #Bookofwhy

4.5.19 @8:11pm - (Replying to @Jabaluck @autoregress and 3 others) Judea Pearl Retweeted Judea Pearl They arise all the time. Unless practicing economists have ceased to ask themselves: (1) Is X exogeneous? (2) is X a good IV? (3) ... namely, questions that need to be answered from where knowledge resides i.e., DAGs, not trasformations of DAG. #Bookofwhy

4.5.19 @3:31pm - (Replying to @autoregress @eliasbareinboim and 3 others) I think what @eliasbareinboim meant to emphasize is that the two systems are not equivalent in terms of representational transparency, ie,. (1) deciding plausibility of assumptions, (2)their consistency (3) their completeness (4)deciding testability, etc. After all,

4.5.19 @2:54pm - (Replying to @PauSchae @PHuenermund) True, in econ models A=B and B=A are two different equations. However, the temptation to equate the two was so intense that generations of economists ended up with blunder after blunder (citation withheld to spare embarrassment). Notational distinction explicates the asymmetry.

4.5.19 @2:38pm - Good to watch all WHY-19 slides in one package. Thanks @eliasbareinboim and all speakers. #Bookofwhy

4.5.19 @2:09pm - (Replying to @DToshkov @eliasbareinboim and 3 others) Thanks to all responders. I now understand "agent-based models". Input: a fully specified SCM model for each bird, encoding its reaction to environmental conditions as well as to behavior of other birds. Output: Aggregate behavior of the flock. Simple. #Bookofwhy

4.5.19 @1:52pm - (Replying to @yoavrubin) Having survived a decade since authoring this footnote, I agree with your assessment: "the best footnote ever". I especially like the end: "citations withheld to spare embarrassments" , though it did not protect me from the wrath of the embarrassed. #Bookofwhy

4.5.19 @5:58am - (Replying to @robertwplatt @Jabaluck and 4 others) I thought you and Jay are epidemiologists, contaminated by DAGs since 1995, and differing culturally from mainstream statisticians. #Bookofwhy

4.5.19 @5:35am - (Replying to @robertwplatt @Jabaluck and 3 others) I judge by textbooks and history books (eg. Stigler). But my last comment was meant more to economics. Statistics, I agree, has spawned some pockets of acceptance, still not strong enough to penetrate education. As bemoaned in this interview #Bookofwhy

4.5.19 @5:22am - (Replying to @juli_schuess @autoregress and 3 others) It is interesting, indeed. I use this proof in many of my lectures, including why-19, since PO researchers are embarrassed to discuss the comparison with DAGs. #Bookofwhy

4.5.19 @5:12am - (Replying to @robertwplatt @Jabaluck and 3 others) An inertia that lasts for over 3 decades borders on a pathology, and is indicative of the power of those "pockets of resistance" you mentioned, which thrive in the absence of "pockets of acceptance". #Bookofwhy

4.5.19 @3:11am - (1/n) Thank you @PHuenermund for summarizing so vividly the Why-19 symposium. I agree with most of your observations and recommendations, especially those pertaining to causal inference in economics. Last week saw a huge interest on Twitter coming from economists, triggered
4.5.19 @3:11am - (2/n) possibly by the challenge to analyze a causal chain using PO. While it unveiled the obvious advantages of DAGs in compactness, transparency and inference complexity, some bystanders might still have gotten the impression that one can do
4.5.19 @3:11am - (3/ ) without them through a heavy investment in PO training. Only passive on-lookers could come to such conclusion, not one who actually tries to analyze the chain using the two languages side by side. I therefore continue to advise readers: Do not rely on on-lookers, try to
4.5.19 @3:11am - (4/ ) solve this problem yourself, from beginning to end, its not too hard, yet it reveals the essential differences between the two representations, one a direct mapping of your knowledge, the other a convoluted transformation of that knowledge. Next time an economist asks you:
4.5.19 @3:11am - (5/ ) "What do I get by using DAGs?" you will be able to assert first-handedly; you get the ability to answer certain questions that you would not be able answer otherwise, and these are questions that economists ask themselves 12 times a day: e.g., Is this variable exogeneous?
4.5.19 @3:11am - (6/ ) perhaps conditionally exogeneous? Is this parameter estimable using OLS? Does my model have testable implications? Are these two models statistically distinguishable, and more and more... I listed some in , Try them out, for fun and profit. They are not
4.5.19 @3:11am - (7/ ) meant to prove that economists do not know X or Y, but to entice them to enjoy the power of new tools, still absent from their textbooks. Conclusion: Do not rely on "On-lookers", listen to your own experience. Good luck, and Tweet if any questions. #Bookofwhy

4.4.19 @2:23pm - (Replying to @causalinf) Reminds me of that math teacher who admitted: Yes, multiplication can be useful, once you learn it, but we can do everything with addition. And all bystanders, who did not try to actually solve a problem with and w/o multiplication concluding: HMM, its just a matter of training

4.4.19 @1:12pm - (Replying to @EpiEllie) Thanks @EpiEllie . But as an incurable engineer I can only absorb new information when it comes in the format of: Input --> Output. ie, What information is needed to start the analysis (or simulation) and what conclusions come out of it. #Bookofwhy

4.4.19 @3:08am - Illustrated slides for causal inference instructors. Marco Zaffalo was kind enough to share with us slides that he made for a master course, based on the Primer book . They can be downloaded here . Thank you Marco. #Bookofwhy

4.4.19 @12:56am - (Replying to @EpiEllie) Can you describe in Twitter length what agent-based models are? I heard this term used occasionally but never explained. How about it? #Bookofwhy

4.4.19 @12:16am - (Replying to @KordingLab @robertwplatt and 3 others) I am suspect of having a dog in the fight, so my credibility can't match that of the "credibility movement" which, by definition, is both home-grown and credible.#Bookofwhy

4.3.19 @11:27pm - (1/2) (Replying to @robertwplatt @KordingLab and 3 others) Agree, but let's assume that an entire literature of a certain field, by some mysterious force, adamantly resists using tools that neighboring fields have found useful, for 3 decades. Isn't it reasonable for an insider to call out the entire field: Look what WE are missing!"
4.3.19 @11:35pm - (2/2) (Replying to @yudapearl @robertwplatt and 4 others) Now assume that instead of WE, the insider names the field, and explicates in details what students and researchers in that field cannot do, so as to call attention to those mysterious forces that hold back progress. Isn't he/she justified calling out an entire field?

4.3.19 @10:52pm - (Replying to @paulgp @autoregress @Jabaluck) Right, the effect of X on Y would no longer be identifiable. Though the effect of Z1 on Y will be. It's all in the diagram, I am not making it up. #Bookofwhy

4.3.19 @4:07pm - (Replying to @autoregress @Jabaluck) This is even better that Smoking and Tar. Because the decisions at Z and Y are "man made", ie, controlled by humans who follow a scrutinizable protocol, much like the canonical IV examples, which dependent on man-made lotteries. Great step forward, #Bookofwhy

4.3.19 @3:58pm - (Replying to @a_strezh @autoregress @Jabaluck) This is indeed how it was first introduced to economists in the 1990's, see , it was named an "instrument", and for a good reason; it is a variable that is instrumental in the identification. #Bookofwhy

4.3.19 @3:28pm - (Replying to @robertwplatt @Jabaluck and 2 others) The apple pie is "good science" and "most of the econ seminars I've been to". So everything is fine and dandy in the cause of "good science". Shouldnt we also identify by name the supernatural forces that resist progress? I am forbidden from doing so, how about others? #Bookofwhy

4.3.19 @2:10pm - (Replying to @yudapearl @robertwplatt and 3 others) Such a "Hey Look" alert invariably evokes the question: "What will we be missing if we ignore the Hey Look?" And then the answer: "You wont be able to do X Y Z" creates professional indignation: We can't do X Y Z? Conclusion, do X Y Z before going back to Apple Pie. #Bookofwhy.

4.3.19 @1:59pm - (Replying to @robertwplatt @Jabaluck and 2 others) How can anyone object to "Mother's Apple Pie" and "Lets continue what we are doing"? Not sure T. Kuhn will agree. There comes a point when totally new ways of "clarifying assumptions" are born, and someone has to say: Hey Look! At the risk of irritating others. #Bookofwhy

4.3.19 @7:39am - (Replying to @robertwplatt @PHuenermund and 2 others) I do not see anything missing. More and more people get the idea that you use DAGs to encode what you know and PO to express what you wish to know, and that we now have the logic of going from the former to the latter. The rest is up to subject matter experts. #Bookofwhy

4.3.19 @6:14am - (Replying to @robertwplatt @PHuenermund and 2 others) DAG/SCM does not commit to any identification strategy, it accommodate them all. Likewise it is not GLOBAL, but accommodates incremental construction of the model, The idea that SCM starts with a huge DAG was invented by people who could not find another reason for not using DAGs

4.2.19 @11:58pm - (Replying to @PHuenermund @Jabaluck) I love quasi-experiments because, as I argued in , they are neat exercises in structural economics, awaiting to unveil more power if assisted by a "structure" . Glorifying them in their current structure-less state seems premature to me. #Bookofwhy

4.2.19 @11:15pm - Good news for European passengers. Our UK publisher (Penguin) is informing us that a paperback edition of #Bookofwhy is now out. See ... (Unfortunately they are not allowed to ship to US.) Another step towards seeing students climb the Ladder of Causation.

4.2.19 @8:38pm - (Replying to @analisereal @autoregress @causalinf) There is another aspect to consider. The most reliable question people can answer is "does X affect Y?" DAGs are built on this primitive question, suggesting that causal knowledge is organized in DAG structure. Encoding it differently results in the mess we see on the left side.

4.2.19 @7:34pm - (1/2) This discussion seems to be pitting PO vs. DAGs as two competitors. They are in fact complementary. DAGs are used to encode what we know and PO what we wish to know. SCM is a bi-lingual framework, , tying the two through structural equation semantics
4.2.19 @7:34pm - (2/2) The reason some researchers still attempt to encode what they know in PO is purely cultural -- they were deprived of DAG education. Glancing at how a simple chain is encoded in the two languages makes you wonder why any bi-lingual would choose PO over x-->y-->z #Bookofwhy

4.2.19 @4:54pm - Link repaired!!!) Readers have alerted me to a broken link to my interview on 3:am. It is repaired now, and can be clicked here: ... #Bookofwhy

4.2.19 @3:20pm - (Replying to @aricaroline @causalinf @Jabaluck) Very reasonable, and I hope @causalinf finds the time and inspiration to do it. However, my golden rule is: One example "of it" is worth ten debates "about it". I say so because I have seen many flip overnight as soon as they solved one toy example hands on using DAGs. #Bookofwhy

4.2.19 @2:41pm - @y2silence Thanks for remembering this blog page, I forgot all about it and I find it so timely (in view of latest debates) that I am retweeting it to all readers. Please review carefully and feel free to use whenever someone claims that everything can be done in PO #Bookofwhy

4.2.19 @1:07am - (1/3) In the interest of keeping this Twitter conversation as a platform for genuine learning, and saluting our Golden Rule: "One example outweighs ten debates", I strongly recommend that readers try to work out this toy example: It calls for analyzing a causal chain X-->Y-->Z
4.2.19 @1:07am - (2/n) in two frameworks: 1. DAGs, 2. Potential outcomes. It has two stages: (a) specify the model assumptions in both languages, and (b) decide if those assumptions have testable implications. The example is extremely important for understanding the often-heard claim:
4.2.19 @1:07am - (3/3) "The two frameworks are "provenly equivalent"" and its counter-claim: "logical equivalence ain't computational equivalence." It is a great opportunity to engage in a fun example that most debaters have tried to avoid. Good luck. #Bookofwhy

4.1.19 @11:00pm - (1/2) You are making heroic effort to establish the superiority of IV over MIV (mediator IV); it's commendable. But no effort can escape the need to establish no-confounding in BOTH cases. The adjectives "desparate", "fruitless" "not understood" "horrible" are equally applicable to
4.1.19 @11:00pm - (2/2) BOTH cases. No judgment of "good natural experiment" can be made without ruling out billions of "unmeasured confounders". Likewise, an MIV is no less "natural " than IV, each may have billions of determiners hence requiring a "no confounding" judgment. #Bookofwhy

4.1.19 @7:46pm - (Replying to @analisereal @autoregress @causalinf) Agree. Only he who has not tried it would say "it is a matter of taste". I urge all passengers to try it before speaking "about it". Computer science students learn that good "representation" is more than just taste, its the difference between doable and undoable. #Bookofwhy

4.1.19 @6:49pm - (1/n) Tired of caricatures? Note that we never construct a dag by listing 150,ooo variables. We start by asking: can you think of a variable affecting both X and Y? Is it measured? If not, is it significant? If yes, lump it together with all other such variables and mark it U,
4.1.19 @6:49pm - (2/n) "unobserved confounders", ONE node. Next you ask: Can you think of a variables that is either (1) on the X-Y path and shielded from U, or (2) affects X and is sheilded from U and not affecting Y (except..)? The former is front-door the latter is IV. And so on and on. At each
4.1.19 @6:49pm - (3/n) stage the question arises: What is "shielded"? and the answer is given, again, in term of: "Can you think of a variable that resides here or there...and has a property that can easily be verified in the "mind's DAG" which is expert in answering only one primitive question:
4.1.19 @6:49pm - (4/5) "Who is listening to whom?". Caricatures are not helpful. Note that you need the "mind's DAG" to certify any candidate IV. Note also that you need the theory of identification to interrogate your mind's DAG toward identification templates, eg. backdoor, frontdoor`, IV, ..
4.1.19 @6:49pm - (5/5) conditional IV, etc. etc. Finally, once you construct the DAG in this incremental way, guided by hoped-for identification templates you may end up with 4-10 variables, and become uncertain of identification -- go to do calculus, the ultimate arbiter (for NP). #Bookofwhy

4.1.19 @5:54pm - (Replying to @autoregress) If you have a DAG, you can see many things immediately. The problem is with people who are DAG-averse, and claim that they can do everything in the "provenly equivalent" PO language, including model specification. It is fun to try #Bookofwhy @causalinf

4.1.19 @2:40pm - (1/2) Not really. Consider the causal chain X--->Y--->Z. Students of pictures can immediately conclude that X and Z are independant given Y. I do not know ANY student of Greek symbols who can easily come to same conclusion from a symbolic representation of the chain, say using PO
4.1.19 @2:40pm - (2/2) (potential outcomes). It is doable, of course, but it would take you a good 5-30 minutes of derivations. You must try it yourself to appreciate the difference and, if you fail, you might wish to take a look at the solution:
4.1.19 @2:40pm - (3/3) or give it to a PO expert, for fun. #Bookofwhy

3.31.19 @9:33pm - (Replying to @Jabaluck @MariaGlymour) Aha!!! Now we are getting someplace. Listen carefully: "searching for... to fit" what is that space in which we search? For me, it is a mental representation of knowledge which shares many DAG's features. For others it is a nebulous cloud, not to be depicted. What about you?

3.31.19 @9:21pm - (Replying to @Jabaluck) I keep on begging you to help me do so by describing (conceptually, w/o acronyms) what lessons need be imported. All I hear is: read his papers, go to Sudan, or "add Angrist type techniques" (which I find already in SCM.). Give me ONE to add, b/c I am dying to automate it.

3.31.19 @9:07pm - (Replying to @Jabaluck) I am tired of hearing that I need to study income inequality in Sudan before I can appreciate the methodological lessons that someone learned. Just tell us what those lessons are and we will incorporate them in SCM, dont send us back to Sudan please. One m. lesson, and formal.

3.31.19 @8:37pm - (Replying to @MariaGlymour @Jabaluck) I see "looking for good quasi-experiments" to be within structural economics, namely, economics that explicates scientific knowledge in SE, or in DAG, so that we can search for "good" quasi-experiment, namely one that our knowledge certifies as meeting the IV conditions.

3.31.19 @7:07pm - (Replying to @metrics52) Josh, I think you would like your doppelganger's performance. He goes by the name "Least Harmful" and can execute all kind of acrobatics; front-door, back-door, IV,... you name it. A true partner in credibility. How about helping us shaping his motion? #Bookofwhy

3.31.19 @6:18pm - (1/4) I think the Garbage Theory is fundamentally flawed. Since "credible inference" is subsumed by "structural economics", garbage generation is a logical impossibility. This is expressed clearly in , Section on "Experimentalists". Quoting:
3.31.19 @6:18pm - (2/4) Quoting: "to the extent that the "experimental‚" approach is valid, it is a routine exercise in structural economics. However, the philosophical basis of the "experimentalist" approach, as it is currently marketed, is both flawed and error prone." (The Refs are illuminating)
3.31.19 @6:18pm - (3/4) Thus, Good news to all sailors and passenger on this unassailable ship: "The garbage attack is over." Moreover, the more you hear about things "you dont even know" (eg. "what data you need") the closer we get to an automated-Angrist, because, if this is what we need to know,
3.31.19 @6:18pm - (4/4) we are already there. See on how smart structural economists can select data sources for identification. What else do they ("credibility" claimants) think "you don't even know"? The more the better. #Bookofwhy

3.31.19 @9:10am - (1/4) Just woke up, to the sound of garbage flying. Wow!. Must pacify some deadlines, but not before stating: The aim of causal inference is to automate the process of generating id-strategies, starting with mental models of the domain. I do not see any theoretical impediment to
3.31.19 @9:10am - (2/4) automate the process by which Angrist&Comp are generating their identification templates from their conceptual knowledge of the world, since the knowledge contained in the former is derivable from the former. This is what the Inference Engine is all about in #Bookofwhy p.11
3.31.19 @9:10am - (3/ ) An "automated Angrist" is not a far-fetched dream, it is partially implemented already in Elias software, which searches a model for nuggets such as frontdoor, backdoor, IV, napkins and, if none is found, goes to do-calculus. Thus, freeing economists to engage in things
3.31.19 @9:10am - (4/4) they do best, accumulate knowledge (through empirical studies), refining knowledge and encoding it in a natural and transparent way, without worrying about id-strategies, which computers can perhaps do better. It is doable.

3.31.19 @8:39am - (Replying to @MartinSGaynor @Jabaluck and 7 others) qq

3.30.19 @8:42pm - We were fortunate to hear Yoshua's lecture at while he was under a strict embargo, to conceal the news he received a week before. After the lunch he had to make an important phone call... The rest is in NYT.

3.30.19 @6:43pm - (Replying to @steventberry @Jabaluck) It is refreshing to read Angrist again, especially after our discussions; I recommend it to all Tweeters. What he does not get is that "identification" need not be "transparent" if the model is. Knowledge comes from our mental model, not from our manufactured id-strategies.

3.30.19 @6:14pm - (Replying to @a40ruhr @snavarrol and 4 others) Not surprised. The "giants" refuse to accept that even straightforward judgments, such as exogeneity or exclusion", come from a mental model that, once explicated, regardless how sketchily, would make judgments so much more reliable. No, "more credible" says the flag. #Bookofwhy

3.30.19 @2:52pm - (Replying to @steventberry @Jabaluck) Conditional independence was anointed "Principle-2" in . Do you also use "Principle-1"? i.e., "the law of structural counterfactuals"? This would certainly turn Angrist against you. Where are your discussions with Angrist's camp aired? #Bookofwhy

3.30.19 @2:12pm - (Replying to @Jabaluck @snavarrol and 3 others) Heckman and Pinto critics should be read in the context of which defines non-parametric identification, explains its rationale, and outlines how LATE IV can be incorporated within the SCM framework. It also bemoans graph-avoiding derivations. #Bookofwhy

3.30.19 @1:21pm - (Replying to @snavarrol @PossebomVitor and 3 others) Yet at the conceptual level the Nonparametric IV problem is well defined: We want to identify the effect of X on Y but we can only randomize Z. When is it doable? A complete (ie, necessary and sufficient) condition can be found here: . #Bookofwhy

3.30.19 @7:06am - (Replying to @PossebomVitor) The difference between the IV and CF condition will show up if you add an arrow from U to Z. Under such a model Z is no longer a valid IV, but it becomes a valid IV upon conditioning on U. See "Generalized IV" (A good paper for economists) #Bookofwhy

3.30.19 @6:36am - (Replying to @Jabaluck @ho_ben and 2 others) I surely agree that those assumptions have substantive content that requires PO notation in addition to DAGs, but I would not know what that content is, or how to assess its plausibility if it were not for the SCM semantics that tells me what I am assessing. #Bookofwhy

3.30.19 @6:21am - (Replying to @Jabaluck @ho_ben and 2 others) No need to reassure me that you are not DAG-averse, I can see that, and I consider it an asset to econometric education. I wish more scholars like you would approach the arrow-phobic cult of econometric and say: Come on, it shows, and it is embarrassing all of us. #Bookofwhy

3.30.19 @6:05am - (Replying to @Jabaluck @ho_ben and 2 others) These "parametric" assumptions (eg monotonicity) are sometimes "more plausible" than exclusion assumptions, not necessarily "weaker". They can easily be incorporated into the DAG framework, using the same PO notation, which derives its logic from SCM semantics, #Bookofwhy p.276

3.30.19 @5:44am - (Replying to @yudapearl @Jabaluck and 3 others) not of identification. Glad you concur. IOW: this series of methodological papers all assume away the identification problem, invoking some template ignorability assumption, and, for a fixed given estimand they provide new estimation strategies. Agree? #Bookofwhy

3.30.19 @5:28am - (Replying to @Jabaluck @ho_ben and 2 others) The issue was " identification templates" and the contested sentence was: "Athey and imbens have been providing a series of methodological papers that push the boundaries of those templates. " My understanding too was that the boundaries pushed were boundaries of estimation

3.30.19 @3:26am - If you are in NYC this coming Monday, please note that @eliasbareinboim will be speaking at Columbia on "Causal Data Science", to the tune of . I wouldn't miss it!
For details: ... @ColumbiaCompSci @DSI_Columbia @ColumbiaMed @CUSEAS

3.30.19 @3:54am - (Replying to @ho_ben @causalinf and 2 others) Strange. I have heard dozens of people speak about the "Use of basic ideas from machine learning" for identifying treatment effects, yet none was able to tell me what those basic ideas were. Perhaps our Twitter discussants can weigh in and illuminate the uninitiated? #Bookofwhy

3.30.19 @2:29am - 3 remarks on DGs (Directed Graphs) and simultaneity:
1. The 3-steps of computing counterfactuals are valid.
2. d-separation is valid in linear systems
3. The 3-steps of identifying counterfactuals are valid in linear systems.
REF: p. 96. #Bookofwhy

3.30.19 @2:28am - (Replying to @steventberry @Jabaluck @causalinf)
3 remarks on DGs (Directed Graphs) and simultaneity
1. The 3-steps of computing counterfactuals are valid.
2. d-separation is valid in linear systems
3. The 3-steps of identifying counterfactuals are valid in linear systems.
REF: p. 96. #Bookofwhy

3.30.19 @1:27am - (Replying to @yudapearl @steventberry @Jabaluck) Or was it perhaps my explanation of "What kept the Cowles commission at bay?". in that you object to? I hope you agree with my claim that they lacked "syntactic machinery for reading counterfactuals from a model or solving simple id-problems. #Bookofwhy

3.30.19 @12:44am - (Replying to @steventberry @Jabaluck) I am trying to find which "summaries of "econ thought"" could have triggered your objection. Do you not agree that "The causal lens is badly obscured in econometric history-writing."? Has any historian narrated Haavelmo's three key insights described here

3.30.19 @12:23am - (Replying to @steventberry @Jabaluck) Well taken, and internalized. But curious: Is "Modern Structural Modeling" different from DAGs? I usually call what I do "Structural Causal Models" (eg #Bookofwhy p.276). Of which DAGs and PO are 2 abstractions. Am I stepping on someone's toes? What does MSM say about control?

3.29.19 @7:23pm - (Replying to @steventberry @Jabaluck) Do you mean "causal" or "casual"? If casual, I wonder why "You could make the same mistake with DAGs," I can't recall seeing even one student making "control" mistake with DAGs. Have you? And btw, did the ancient religion ever discuss the issue of adding "controls"? #Bookofwhy

3.29.19 @4:44pm - (Replying to @causalinf @Jabaluck and 2 others) @caisalinf, I beg to disagree on one tiny issue-PO. No amount of command over PO notation can make up for the (usually ignored) fact that PO is a derivative of structural models, as in . Have you met a PO expert who can estimate Joe's would-be salary p.94 ?

3.29.19 @3:46pm - (Replying to @causalinf @rdahis @Jabaluck) I am glad @causlinf is undertaking the task of summarizing this conversation. If I writes it, I would be accused of trying to sell "MY approach". Plus I am still a firm believer in the Axiom of Choice: "Solving one problem in DAGs is worth 100 debates ABOUT dags." #Bookofwhy

3.29.19 @3:31pm - (Replying to @ho_ben @causalinf and 2 others) Miscommunication. When I asked you to share "new templates pushed by Athey and Imbens" I did not mean links to their papers, but sharing ONE template that appealed to you personally, whose conceptual basis can be communicated on Twitter #Bookofwhy

3.29.19 @3:15pm - (1/3) @Undercoverhist , First, Thanks for doing your tweetstorm on Eco-history, which I have found to be illuminating. Second, I do not believe @causalinf was joking. Eco-historians (including Morgan, Qin, Epstein, Hoover) failed to distinguish "theory-data tension" from
3.29.19 @3:15pm - (2/3) "Causalily-data tension" which, in my opinion, accounts for most hurdles in eco. For the school of Hendry (and Sargan), for example, "economic theory" meant the statistics that governs data, not Haavelom's theory which he viewed as a collection of experiments conducted by
3.29.19 @3:15pm - (3/3) mother nature. I hope you make this distinction in your upcoming book. I tried to make it in my paper on Haavelmo but I do not think it penetrated mainstream eco thinking. The causal lens is badly obscured in econometric history-writing. #Bookofwhy

3.29.19 @1:42pm - One can find good use of PO&DAG combination in every paper on mediation (eg ) because PO and DAGs are both derivatives of Structural Causal Models (SCM), DAGs encodes what you know, and PO what you wish to know - no conflict. #Bookofwhy

3.29.19 @12:59pm - (1/2) My paper contains a section: "What held the Cowles commission at bay?". It may sound heretic to most economists, bc it attributes the decline to a stalemate in notation and computation, yet I still believe that, lacking the ability to SEE (ie compute)
3.29.19 @12:59pm - (2/2) simple features of one's model, makes it impossible for an economist to go beyond Marschak and F Fisher. Eco-Historians minimized the role of Haavelmo and Wold, bc the causes-data tension was not their priority. It takes a chemist to write the history of alchemy #Bookofwhy

3.29.19 @11:14am - (Replying to @ho_ben @causalinf and 2 others) What conclusion should I draw? Sit down and relax? Or assist the progressives in their efforts to educate the remaining islands of resistance? BTW, I am not familiar with the new templates pushed by Athey and Imbens. Can you share one? #Bookofwhy

3.29.19 @4:00am - (Replying to @ho_ben) Not exactly. By studying relativity seriously we know the limitations of Newtonian physics. By NOT studying DAGs economists are still unaware of the limitations of their template-driven methods. Did you say "many of us are working on same"? Please share, eager to hear #Bookofwhy

3.29.19 @2:47am - (1/n) I see a spark of agreement looming from this conversation. It is based on (I hope) everyone's agreeing that "we need a DAG for inference, bc it carries the info we need for id." Another spark is the fact that everyone (I hope) is talking about at least TWO DAGs, one residing
3.29.19 @2:47am - (2/n) in the mind and tacitly stores your understanding of the relevant domain, and one (called Full DAG) is what you eventually explicate when you decide to draw it on paper for full analysis. Call the former "mental DAG" (or m-DAG) and the latter ex-DAG (for explicit).
3.29.19 @2:47am - (3/n) Scott also introduced a project-specific DAG, or a premade DAG defined by the id-strategy one wishes to use. Call it t-DAG (for template). Barring two repairable exaggerations, I generally agree with Scott's depiction of "practical economists" and CI-theorists.
3.29.19 @2:47am - (4/n) But I need to add another brush stroke to facilitate full agreement: The mental DAG and the Full-DAG are the same, while the t-DAG is a fragment of the former, selectively extracted to match a specific id-strategy.
With these points of agreement, we see that the Full-DAG is
3.29.19 @2:47am - (5/n) used in two different roles. First, mentioned by Scott, to alert us to new id-strategies (eg front-door) which our ancestors have missed lacking complete id-logic. Second, and perhaps more important, to validate matching between the postulated template-DAG and our mental DAG
3.29.19 @2:47am - (6/n) - the ultimate arbiter of plausibility. For example, if we selected an IV strategy, our t-DAG would be the canonical IV-DAG, and matching involves checking whether the exogeneity and exclusion properties assumed in that canonical t-DAG hold in our mental-DAG, which is waiting
3.29.19 @2:47am - (7/n) passively to be interrogated. This is what we normally call "judging for plausibility". Why then do we insist that even practicing economists learn to read DAGs before engaging in heavy empirical work? Because the task of matching requires reading your m-DAG. What do we mean
3.29.19 @2:47am - (8/n) by READING a DAG? We mean taking an arbitrary 4-variable DAG and checking if properties such as exogeneity, exclusion or conditional exclusion hold in it. It is a matter of checking the plausibility of one's assumptions, not of discovering new id-strategies.
3.29.19 @2:47am - (9/9) This is why we get suspicious when leaders of "credibility movements" tell us that they do not need to read DAGs, since they deal with "real life" problems." IOW: "We hate to show you how poorly we do when things are explicit, trust us to do better in real-life, #Bookofwhy

3.29.19 @1:57am - If one believes that Economists SHOULD learn DAGs, how can one trust leaders of a so-called "credibility movement" who proclaim they do NOT need to learn DAGs because they deal with "real-life" problems. I would be a bit suspicious of their un-aided judgment. #Bookofwhy

3.28.19 @5:35pm - (1/n) This is new and interesting observation: You are 1 and we are many. I am a believer in crowd wisdom, because it integrates many perspectives, each taken from a different angle. Can we say that about econometrics? Are there many angles if NBER postings are by "members only"?
3.28.19 @5:35pm - (2/n) Moreover, the claim "causal inference is all we THINK about" is true for children, homosapiens, scientists and alchemists, yet #Bookofwhy distinguishes "thinking" from committing our thinking to some mathematics. And here, all my economist friends bemoan econ textbooks, see
3.28.19 @5:35pm - (3/n) . Another moreover, my history book tells me that Alchemists too had norms and practices, and they were quite effective in developing metals and alloys. The only thing they failed on were "toy problems"; eg,they could not explain why substances often
3.28.19 @5:35pm - (4/n) combine in integer proportions. They dismissed the "toy problems" with great indignation, probably saying "we are solving important practical empirical and large scale problems, show us what we did wrong!". It was by taking those toy problems seriously that we have science.

3.28.19 @4:42pm - (Replying to @robertwplatt @Jabaluck @PHuenermund) I remember coming to the conclusion that the word "design" could stand some formal clarification, otherwise it sounds like "if you do the right thing nothing else matters". I love it. Every student (and robot) should remember this important principle -- a life saver.#Bookofwhy

3.28.19 @2:38am - (Replying to @juli_schuess @PHuenermund @jim_savage_) @jim_savage, I am curious if you understand how "credibility" can be attached this way. I'm truly curious, and would value your insight, because I have the feeling that you did try to solve a problem with dags, so you know where PO's come from and why it is crucial. #Bookofwhy

3.27.19 @8:33pm - (Replying to @causalinf @JeffDenning and 4 others) I have just received Mostlyharmless from Carlos, three days ago, and I read it with all the curiosity that our tweeter discussions have evoked. Let's compare its treatment of PO with that of Primer: ttps:// then say "credibility" to our students, with straight face

3.27.19 @5:29pm - (Replying to @JeffDenning @causalinf and 4 others) I got a copy of Mostlyharmless and, if this book is representative of the "credibility revolution", I find it hard to understand how anyone who tried to solve a problem with dags can attached the word "credibiity" to a methodology that ignores our models of reality. #Bookofwhy

3.27.19 @9:49am - (Replying to @juli_schuess @Jabaluck @yskout) Well put. Glad you took the time to read mostlyharmless and, like me, tried to make sense of how they justify adjustments and IV using PO. Only those who have not used DAGs to justify things would prolong the myth that such acrobatic constitutes a "justification". #Bookofwhy

3.27.19 @9:20am - Exciting news to all sailors and passengers in AI and ML: 2019 Turing Prize awarded to 3 ML pioneers. ... Our Why-19 symposium was fortunate yesterday to host Bengio's brilliant lecture just before hearing of the good news. #Bookofwhy

3.27.19 @8:31am - (Replying to @juli_schuess @Jabaluck @yskout) It is rather implausible that our mental representation of what's going on in the world is tailored to any "identification strategy" that 20th-century scientists have devised. Such representation simply tells us what affects what, prior to engaging is "identification" exercises.

3.27.19 @2:06am - (Replying to @yskout @Jabaluck) Agree. But it is the DAG that tells you if an IV is "GOOD". For simple cases we do not need a DAG, but in cases where exogeneity and exclusion are questioned, or when they need to be created, eg we need DAGs.

3.27.19 @1:24am - The identification problem is NOT about determining the logical consequences of assumptions Rather, it is about determining if the estimability of a desired causal effect follows logically from the assumptions and, if so, how it is to be accomplished. #Bookofwhy

3.27.19 @1:10am - Replying to @peter_mourfield @PooyanJamshidi Thanks for posting.

3.25.19 @1:52am - Taking a leave from Twitter for two days, on pilgrimage to Stanford, speaking Monday, March 25, 2:00 pm, at Jordan Hall, 420-40, . Title: "The Foundations of Causal Inference, with Reflections on ML and AI". Admission free for Stanford students.See you.

3.25.19 @1:08am - (Replying to @MartinRavallion @SteveRo48195125 @Jabaluck) Conceptually, I agree. #Bookofwhy for example claims that the RCT is valid because it mimics backdoor, not the other way around. Same goes for the validity of natural experiments.

3.25.19 @12:00am - I must add that, in SCM, talks about embracing and unifying are not just words. Examine again this chapter and judge for yourself how naturally counterfactuals (hence PO) emerge from a structure, how they are used to solve practical problems and, and, and

3.24.19 @10:39pm - (Replying to @robertwplatt @Jabaluck and 2 others) The symmetry does not hold here. Why? Because Structural Causal Modeling (SCM) embraces PO and IV and DAG (eg mediation and attribution are all done with PO). May I take it that you recommending SCM? If so, hat off, welcome to the causal revolution #Bookofwhy

3.24.19 @7:19pm - (Replying to @CarnaticPrior) I do not doubt that de Finetti theorem is central to some statistical tasks. But that does not make it relevant to causation and, certainly, I do not see how even comes close to deciding whether one should condition on M to get unbiased effect estimates. #Bookofwhy

3.24.19 @6:34pm - (Replying to @HannesMalmberg1 @SteveRo48195125 and 2 others) Well put! I could not say it better. What IS quasi-exp? The insiders say: its some setup that mimics RCT. The ousiders say: No, it is a model of reality which, after doing mental checking, is judged to satisfy exogeneity. Now come DAGs and say: Lets replace mental with formal

3.24.19 @6:22pm - (Replying to @smilleralert @autoregress and 3 others) I have never been a fan of the expression "can help think through things" but after seeing how it can be abused I pledged NEVER to use it as long as I live. DAGs actually compute what economists are struggling to THINK through. Analogy: algebra computes solutions, not help think.

3.24.19 @5:15pm - (Replying to @SteveRo48195125 @MartinRavallion @Jabaluck) My understanding of backdoor is that we do not ADD variables. THEY ARE THERE in nature, we just acknowledge their existence in our mental picture of reality and check if they are needed. We ADD them to our identification strategy, if appropriate, not to the model. #Bookofwhy

3.24.19 @4:16pm - (1/2) First time I hear this: "PO can be helpful for thinking about what to control for. " I am retweeting this statement as an example of how dangerous the expression "helpful for thinking" can be, and why students of causality should work out simple problems in both languages, as
3.24.19 @4:16pm - (2/2) as I have done in my book, and in my slides. In my adult life, I have not seen a SINGLE case where "PO can be helpful for thinking about what to control for", but since I am biased, test it on PO-experts, eg or #Bookofwhy

3.24.19 @3:08pm - Thanks Scott for another candid narrative on the role DAGs play in your research and teaching. I wish there were more econ like you who say, lets first TEACH it, then see what new practical opportunities open up. This is how Epi started - I dont think they are sorry #Bookofwhy

3.24.19 @2:52pm - (Replying to @Jabaluck @analisereal and 4 others) If I were to sniff the econ literature for cases of bad control I would be accused of negative attitude, if not "bashing econ." Instead, I am presenting "seat-belt usage" as a case of bad control authorized by super statisticians . Still not convinced?

3.24.19 @2:37pm - (Replying to @analisereal @Jabaluck and 4 others) Glad you emphasized "disciplining current ID strategies" which has been almost forgotten in the discussion, as if PO+IV+RC do not require formal assistance in visualizing, elaborating and strengthening the assumptions behind them. They do. #Bookofwhy

3.24.19 @1:47pm - (Replying to @RealKevinThmpsn) I appreciate your candid narrative of change. I am trying to stimulate in economists both the feeling of "I am missing something" and "I could do so much more". A very thin difference between the two, the latter being more productive. #Bookofwhy

3.24.19 @2:10am - (Replying to @diomavro @Jabaluck) proof of equivalence in Causality chapter 7. But this only says that if we start with same assumptions we end up with same conclusions, it does not guarantee that encoding the assumptions in one system will be comprehensible or compact. See and slide.

3.24.19 @12:36am - My favorite M-bias example is "seat-belt usage" as described in Section 2 of this paper . What's interesting about it is that "Seat-belt usage" was actually controlled for in Rubin's study. #Bookofwhy

3.23.19 @9:57pm - (Replying to @yudapearl @Jabaluck and 2 others) one of us can whisper on Josh's ear that the "difficult" problem was solved and made easy in 1993. It can't be me, because Josh will suspect that I want to SELL him DAGs, and econometric will go through another decade without a solution. #Bookofwhy

3.23.19 @9:46pm - (Replying to @Jabaluck @EpiEllie @autoregress) This is an illuminating blog from "mostlyharmless", not so much for discovering M-bias, but for confessing that the issue is "difficult" and may depend on whether variables are "exogenous" or "mediating", namely, the economist is advised to interrogate the mind's DAG. Perhaps

3.23.19 @9:21pm - (Replying to @Jabaluck @mathtick) Jason, I respectfully ask to be excused from answering your Tweets, especially those that interpret or assess Pearl's "claims" or "quarrels"; assume I have none. If you think econometrics has areas of potential improvement, please Tweet; I would be curious to read and

3.23.19 @9:24pm - (Replying to @yudapearl @Jabaluck @mathtick) occasionally add my two cents. But I would rather not engage in theories on how "Pearl" can be a better salesman. I am interested in econometric methodology, not in salesmanship. @StuartBuck1 #Bookofwhy

3.23.19 @7:14pm - (1/2) I actually believe economists have positive values and honest scientific attitude. Unfortunately, they inherit dogmatic education and rigid cultural allegiance that stifle innovation. One hope is their inquisitive students. Here, our tweeting readers can help. How about
3.23.19 @7:14pm - (2/2) petitioning the editors of the five top econometric journals to invite an introductory review article on "causal inference among our neighbors", just to wet their students appetite. If I were an Editor, I would consider it my duty to thus counter allegations of insularity.

3.23.19 @5:17pm - (Replying to @stewarthu) Compare econ literature to social science a la "Morgan and Winship" (2007). The latter embrace diagrams, while the former bans them at all cost, denying students even the rudimentary ability to read independencies in their own econ models. Not exactly "mathy", #Bookofwhy

3.23.19 @5:03pm - (Replying to @stewarthu) Econ used to be "mathy", searching for the right math to solve eco problems. Then, invaded by statisticians, they started searching for problems to fit the math. It is indeed hard to stomach the idea that a graph is a mathematical object, as honorable as an equation.#Bookofwhy

3.23.19 @3:45pm - (Replying to @yudapearl @stewarthu) Resistance to math can be justified when the math seems totally unrelated to your research (eg abst. set theory to an economist). But resisting a math that helps you solve problems that you yourself declared important and hard, that really takes cultish thinking. #Bookofwhy

3.23.19 @3:30pm - (Replying to @stewarthu) What happened to Epi can be described as "bloodless" revolution, namely, a dramatic change of perspective and tools. Economics is different. There, professional pride and insular structure entail stiff resistance for every change. #Bookofwhy

3.23.19 @3:08pm - (Replying to @IgnacioPerezR @jreileyclark @_MiguelHernaa) It also has a nice title: "Draw Your Assumptions Before Your Conclusions" which encapsulates the difference between CI and empirical economists. The latter go by the principle: "First Decide on Identifying Strategy, then search your mind for assumptions that support it"#Bookofwhy

3.23.19 @4:15am - (Replying to @DoctorActivist) I never took it "personal", I take it as tiring attempts to justify the status quo, instead of helping to change it. My writing is indeed a "call out": econ. students need to know how far behind their textbooks are, else the status quo will change incredibly slow. #Bookofwhy

3.23.19 @3:55am - Your question drove me to re-read - the chapter on counterfactuals in Primer. My gosh! So many debates and misconceptions could be avoided through this chapter. I recommend it again to all readers and followers -dont miss! I have not found an alternative

3.23.19 @2:41am - (Replying to @diomavro) For more technical details, I recommend Primer e.g.,

3.23.19 @2:35am - (Replying to @diomavro) I try not to say "directed graph" when talking to economists; this totally shut their mind off. I start with a STORY, eg what would Joe's salary be had he stayed one more year in college, given that currently his salary is 30K and that he won two prices for technical innovation.

3.23.19 @1:56am - (1/3) To be honest, Jason's repeated lectures on what I need to do to be taken seriously by economists began to get on my nerves. I wrote #Bookofwhy having given up hopes of ever seeing old-guard economists adopting of causal inference tools, so Jason's lectures
3.23.19 @1:56am - (2/3) sounded hollow to me, and I tried not to reply to him. I haven't given up on econometric students though. So, if readers wish to help econometrics catchup with advanced tools, alert their students to what they are missing and, as a pre-requisit, teach them how to solve
3.23.19 @1:56am - (3/3) a 4-variable problem before sending them to Kenya to study income inequality. Physics students learn to solve equations of motion before they touch a telescope or visit an observatory.

3.23.19 @1:09am - (Replying to @Rodrigo20980033) The issue is not the citation but the narrative, which conflicts with that of #Bookofwhy (p.334-5) on decomposition and identification. Check and

3.22.19 @3:24am - (1/3) Thanks for posting this article on missing data and I am glad you found graphical models to be beneficial for both visualizing and organizing the results obtained. I am not sure, however, if you are familiar with recent works on missing data, as summarized for example in
3.22.19 @3:24am - (2/3) this review article and in which general graphical criteria are derived for recovery and testability . I hope you find it useful for
3.22.19 @3:24am - (3/3) would yield biased results. Your graphs seem to imply such criterion but I could not derive one since MI is model-blind. Congratulations on an illuminating article in a hard field that was once thought to be the sole province of statistical analysis. #Bookofwhy @IJEeditorial

3.21.19 @11:40pm - (1/2) Thanks for the link. The dippel etal paper may well be the first applied econometrics paper on mediation, and it even has a DAG in the appendix: courageous, and structural equations with their relations to counterfactuals. I would am tempted to forgive the authors
3.21.19 @11:40pm - (2/2) for not citing where the mediation formula first appears (see #Bookofwhy p. 334-5 for the story). This is progress, but note, contrary to ruling paradigm, they put down the structural model on paper before searching for "identifying assumption". Bold strengthening your results. One issue which was not clear to me is whether you have a criterion for detecting when MI

3.21.19 @11:15pm - (Replying to @autoregress @btshapir) I wish I could say, yes, I hope so too. But I am stubborn believer in the Golden Axiom: He who has not solved a 4-variable problem, will not find areas where more complex formalization can be "fruitfully applied. But your students will. It is the Law of lampposts. #Bookofwhy

3.21.19 @10:03pm - (Replying to @RahelJhirad) Yes, it is free, according to AAAI headquarters, and a personal mesg I received from Carol Hamilton, our Tzar. @eliasbareinboim is closer to the ruling party, so he should be able to confirm it. I hope it is as free as I was given to understand #Bookofwhy

3.21.19 @8:40pm - (Replying to @quantadan @autoregress @btshapir) Love your analogy. Though I no longer attempt to lead the horse to water. I am just hoping that econ. students are smart enough to notice that, if they want water, they should and can get it on their own. And they will, for it is becoming more and more accessible. #Bookofwhy

3.21.19 @7:00pm - (Replying to @ChrisSeveren @metrics52) I have "Master Metrics" (2015) on my Kindle. Is that representative of the culture? or is "Mostly Harmless" more advanced? I am not surprise that the language in which you "grow up" comes naturally. The question is how flexible the culture is to language enrichment #Bookofwhy

3.21.19 @3:59pm - (Replying to @ChrisSeveren) I would really appreciate an example, so that I will understand what they mean by "stating" the assumption, and 'providing" supporting evidence. Not that I doubt whether they do that, I trust you, but as a computer scientist, I am interested in the language they use. #Bookofwhy

3.21.19 @2:05pm - (Replying to @PHuenermund @Jabaluck @autoregress And once you agree that the exclusion restriction is not plausible, what then? Do you try to see if it can be remedied? eg by conditioning on an intermediate variable. I have seen it in the literature, but I may have missed it.

3.21.19 @1:08pm - (Replying to @_scott_fleming_ @DavidnLang and 5 others) I understand that you just show up and show your Stanford affiliation card.

3.21.19 @8:48am - (Replying to @autoregress @btshapir) I dont know how this thread turned me (again) into an enemy of the people. I stopped trying to convince empirical economists to use DAGs - they wont. I'm still eager to show their students how equivalent representations differ, which should be useful once they build a model.

3.21.19 @8:02am - (Replying to @autoregress) OOps, a little snug here. Normally one assumption is insufficient to estimate any effect. Even in simple problems we need 10-20 assumptions (see slide) the combination of which may be sufficient. Shouldnt we wait to validate 10 assumptions first, one at a time? then proceed?

3.21.19 @7:49am - (1/3) I would never use such language. Economists are not supposed to know about DAGs, and I never expect them to view things through DAGs. I only expect them to commit to paper what THEY believe is plausible, in a way that THEY think faithfully represents their beliefs, then ...
3.21.19 @7:49am - (2/3) continue to estimate what THEY think need be estimated. What I find missing in this natural sequence of steps is "committing to paper". Indeed, this step is missing in the writings of empirical economists, who prefer to jump directly from mental thoughts to estimation.
3.21.19 @7:49am - (3/3) This explains why they fail to appreciate the derivations of do-calculus, and always go back to ask: "we need to think hard if this assumption is plausible". Without committing to paper they never get to saying: assume for a moment my assumptions are valid, just for a moment.

3.21.19 @7:16am - (Replying to @autoregress) ok, we thought and we thought and we argued and argued and we came to the conclusion that U has no effect on Z. Now everyone is in agreement. Next step =? Do we keep the agreed story in our mind? or commit it to paper? Recall, we have 10 other things we agreed on. Please advise.

3.21.19 @6:51am - (Replying to @autoregress) I am not clear what we disagree about. Perhaps you can state your position in the form of "I think that enough thinking will enable us to......" or "We still need more thinking to do before....". Sorry.

3.21.19 @6:38am - (Replying to @f2harrell) Not sure of the context. Mediation? Generlization? 3-languages?

3.21.19 @6:34am - (Replying to @autoregress) Agree, the two steps should be taken in sequel. The slide deals with choices available in the second step: encoding our understanding of the world. Once we encode, there is no more thinking necessary, we can sit back, relax, and let the do-calculus take over. #Bookofwhy

3.21.19 @5:52am - (Replying to @autoregress) Once we come to agreement on how to think about the world, we need to decide how to represent our thinking formally, so as to combine it with data and estimate what we wish estimated. Thinking alone is insufficient for deciding how to estimate things. #Bookofwhy

3.21.19 @3:31am - (1/2) With all the talks about "logical equivalence", "transparency" "defensibility" "testability" and "empirical economics", I presume that only few students have had a chance to see an actual comparison of a simple example articulated side-by-side in 3 different languages:
3.21.19 @3:31am - (2/2) 1.English, 2.Potential Outcomes and 3. Structural Models. Here is one for your enjoyment: a slide that I will be discussing at Stanford. The research question is of course: Estimate the effect of Smoking (X) on Cancer (Y), given samples from P(X,Y,Z).

3.20.19 @9:05pm - This blunder permeates the literature, and leads to the mediation fallacy #Bookofwhy p.315 which econ. students should learn to avoid. Students interested in understanding mediation should go straight to the model-based literature, eg

3.20.19 @7:40pm - I can identify the skinnerian level as a subset of Rung-1 (which embraces retrodiction.) and Popperian level as Rung-2 (love the emphasis on "models"). The other levels are not described in term of capabilities -- hard to place. And I dont find counterfactuals, strange.#Bookofwhy

3.20.19 @5:22pm - (Replying to @depistemology) I really haven't thought about the relations between the two classifications. Where can one find the best accessible window to Dennett's levels? Something as concise and exemplified as #Bookofwhy ?

3.20.19 @4:33pm - Confirming my pilgrimage to Stanford, I will be speaking Monday, March 25, 2:00 pm, at Jordan Hall, see . My title: "The Foundations of Causal Inference, with Reflections on ML and AI". Admission is free and no topic left behind.

3.20.19 @2:28pm - (Replying to @akelleh @eprosenthal @EvanBianco) Great name "Causal Data Science" and great blog too. How do you survive Columbia under such name? Last time I spoke at Columbia they almost crucified me at the podium for suggesting they should name their Institute "Reality Science" instead of "Data Science" #Bookofwhy

3.20.19 @12:27am - (Replying to @raymondshpeley @Jabaluck and 3 others) I would replace 2 and 3 with 2'. encode that knowledge in a defensible model M 3a. check if your needed effect is estimable from M (and how) 3b. If so, estimate it, if not, elaborate M. [I would avoid "identifying assumption" like the mother of all miscommunication.] #Bookofwhy

3.19.19 @2:15pm - Another role played by the orientation of the ellipse can be seen in Lord's Paradox, #Bookofwhy page 213, elaborated in . But this takes us from regression to causality-land. Great for class demonstration.

3.19.19 @12:40pm - (Replying to @TiernanRayTech) Thanks for posting. I would never have imagine that Pharma folks read #Bookofwhy.

3.19.19 @4:13am - Trygve Haavelmo (1944, p. 14) drew an analogy between two sorts of experiments: "those we should like to make" and "the stream of experiments that nature is steadily turning out from her own enormous laboratory, and which we merely watch as passive observers." Angrist-Krueger 2001

3.18.19 @6:34pm - (Replying to @gelbach @Susan_Athey) You must be wondering what makes me work so hard on trying to understand where our communication failed. The story is a long one but, right now, I am trying to understand if economists have a way of distinguishing "exactly-identifying assumptions" from other kind of assumptions

3.18.19 @5:27pm - (Replying to @gelbach @Susan_Athey) After walking together through 20 Tweets, is it too much to ask for help in refreshing by failing memory and granting me one last word: Is the task: trivial? important? solved in the literature? taught in econ. classes? irrelevant? Outdated? Your word? #Bookofwhy

3.18.19 @5:12pm - (Replying to @raymondshpeley @eliasbareinboim and 2 others) Agree. But sometimes the slow bites are the hardest ones of all. I truly do not know what "good causal empirical work" is, anyone knows?. Note the hardship we go through to have economists opine if identification is important for policy analysis. All small bites are giants

3.18.19 @4:43pm - (Replying to @gelbach @Susan_Athey) OK, no words. What about the TASK of "examining all the assumptions conveyed by your econ. model and deciding whether they imply that a given variable X is conditionally exogenous." Is that task trivial? important? solved in the literature? taught in econ. classes? Opinion?

3.18.19 @4:05pm - (Replying to @gelbach @Susan_Athey) But what word should I use in the future?

3.18.19 @4:01pm - (Replying to @gelbach @Susan_Athey) Good. Lets not used the word "ascertain" any more. What word do you use for the act of examining all you model assumptions and computing/determining whether a variable X is conditionally exogeneous (given subset Z in the model).?? I promise to use your term only.

3.18.19 @3:16pm - (Replying to @ToreEllingsen1 @CarterPaddy) GMM is a totally different beast; it aims to estimate parameters in statistical models instead of causal effects in nonparametric econ models. The best way to see the difference is to try it on a simple problem (eg frontdoor #Bookofwhy p.224) Its essentially undoable

3.18.19 @2:44pm - (Replying to @gelbach @Susan_Athey) I take then it that you agree that ascertaining "conditional exogeneity" is of primary importance for "policy evaluation", for it determines whether or not we need to resort to bounds or other semi-identification methods economists have developed. I hope I read you correctly?

3.18.19 @1:31pm - (Replying to @f2harrell @AndersHuitfeldt) I missed your reply, thanks. So you are emphasizing the "reliability" of a parameter estimate, given a finite sample. As compared to our notion of generalizability which emphasizes alternative models under infinite sample. Did I get it right?

3.18.19 @1:25pm - (Replying to @gelbach @Susan_Athey) You may be right. I would be eager to hear what kind of engagement can lead to better communication, avoiding "shade-throwing"? What is your opinion? Can policy analysis be conducted without one's ability to ascertain "conditional exogeneity"?

3.18.19 @1:02pm - (Replying to @f2harrell @AndersHuitfeldt) What about the generalizability problem I posed: "Find if the OLS estimator of parameter beta in Model 1 estimates parameter gamma in Model 2 without bias." Is this part of the generalizability problem addressed in your book? Should it be? #Bookofwhy

3.18.19 @12:59pm - (Replying to @f2harrell @AndersHuitfeldt) I get a double kick , because I still do not know "where prob. models come from" and I am beginning to suspect people do not have such models in their minds. The math models we have are attempts to patch up the fragments we do have in our minds. #Bookofwhy

3.18.19 @12:50pm - (Replying to @NicholasStrayer) Thanks for highlighting this important distinction

3.18.19 @12:49pm - (Replying to @elliottmcollins @CarterPaddy) You can always skip multiplication and replacing it with addition -- just add a number to itself N times. The miracle of multiplication is that with a meager investment of memorizing one multiplication table you can do science that would be left undone without that investment.

3.18.19 @3:29am - (1/2) Readers of #BookofWhy may wish to attend the Inauguration of Stanford Institute for Human-Centered AI (HAI), Monday, March 18, which will include a symposium and remarks by Bill Gates. A livestream will begin 9:15am Pacific.
3.18.19 @3:29am - (2/2) I will not be able to attend, unfortunately, but as a Distinguished Fellow of HAI, I wish the co-director Fei-Fei Li @drfeifei all the success that a Center for such critical and timely mission deserves.

3.18.19 @2:10am - (Replying to @yskout @eliasbareinboim and 2 others) A worthwhile endeavor! But be careful to choose a paper that, in case your result differ from the author's, you will be able to convince readers that yours are right. Perhaps by questioning one of the assumptions. So, choose an author who is explicit about assumptions - Rare!

3.18.19 @1:50am - (Replying to @nyarlathotepesq @CarterPaddy) Totally irresponsible, I agree, even delinquent. However, if someone demonstrates to us that lacking certain tools we cannot solve simple problems that we OURSELVES declared to be important and fundamental to our field. Shouldn't we then try to expand our tool set? #Bookofwhy

3.18.19 @1:30am - (Replying to @sarahmrose) The reason it is hard (if not impossible) to unveils the flaws in "real examples" is that we do not have ground truth against which to gauge glaws in case we get different result than our predecessor. Also, "toy examples" are quite "real" (drug, age,cancer) but are manageable.

3.18.19 @12:56am - (Replying to @sarahmrose) We would ask the same questions, but on larger problems, problems that we either discarded as hopeless w/o DAGs, or that we treated by ad-hoc methods, not having the tools to choose methods to fit the characteristics of our models.

3.18.19 @12:24am - (Replying to @maximananyev @Susan_Athey @joshgans) Good idea. Although it is hard for me to envision an economist convinced by missed opportunities when he/she has no language to gauge the amount missed. Plus, market efficiency disappears when trade embargoes are imposed. #Bookofwhy

3.17.19 @11:49pm - (Replying to @maximananyev @Susan_Athey @joshgans) Thanks for posting this discussion which illustrates nicely the M-bias which Rubin has denied since 2009 . Another thing I learned from it is that the word "design" is used for "data generating model" (ie., SEM). Good to remember.

3.17.19 @10:25pm - (Replying to @f2harrell @AndersHuitfeldt) Here is a toy generalizability problem in parametric models (Dunan, 1975): "Find if the OLS estimator of parameter beta in Model 1 estimates parameter gamma in Model 2 without bias." :// Is this the kind of generalizability problems addressed in your book?

3.17.19 @9:55pm - (Replying to @maximananyev @Susan_Athey @joshgans) I have a different theory on why epi. embrace DAGs and economists ban them; it has nothing to do with the research problems in the two fields (the structures are identical.) In the 1990's Epi was led by Greenland and Robins, and Econ. by disciples of Don Rubin. Simple.#Bookofwhy

3.17.19 @8:33pm - (Replying to @Ideal_Health18 @Susan_Athey @joshgans) Thanks for reaching out. Tell me your research problem and I will try to find a nice introduction for you. Such introductions do exist, all we need to know is the language in which you feel comfortable and ops we go. #Bookofwhy

3.17.19 @5:57pm - (Replying to @Susan_Athey @joshgans) How does one judge "better quality" in "empirical applications" where assumptions are debatable, hardly articulable, and where we don't have ground truth against which to gauge improvement. The Golden Axiom says: He who cannot do it on a 4-variable problem, cannot do it period.

3.17.19 @5:18pm - (Replying to @paulnovosad @Jabaluck and 2 others) I think you meant "econs are good at backdoor criterion" INFORMALLY, because I know only N<10 economists who know it formally, or who can use it beyond 4-variable problems. If you know of N=11, there is hope.

3.17.19 @5:11pm - (Replying to @analisereal @Jabaluck and 2 others) Plus, it is always helpful to quote the golden axiom: He who cannot do it on a 4-variable problem cannot do it on REAL WORLD EXAMPLES, where we have no way of testing claims.

3.17.19 @4:47pm - (Replying to @elliottmcollins @Jabaluck @CarterPaddy) Ufortunately, First-graders solutions would not improve with multiplication; they were perfectly sufficient and deeply sophisticated for the exam given, which was selected by the teacher to cover addition, not problems requiring multiplication.

3.17.19 @3:04pm - (Replying to @Susan_Athey) What about the "treatment effect" literature, also going by "policy evaluation" etc. Is CE of primary concern there? Does the DAG-avoiding description of that literature make sense? or is the "avoiding" unintentional, or scientifically informed, or non-existing? #Bookofwhy

3.17.19 @2:55pm - (1/2) Nowhere has anyone presupposed that understanding "our" work is "as foundational for doing causal inference as understanding multiplication is to doing mathematics." What IS foundational (as multiplication) is deciding conditional ignorability in a scientific model
3.17.19 @2:55pm - (2/2) (ie, a model describing "how the world works") by ANY method. Plus the axiom that he who cannot do it on a 4-variable problem cannot do it on REAL WORLD EXAMPLES where we have no way of testing claims. But, hating to further insult the innocent, I'll quit this thread.

3.17.19 @2:26pm - (Replying to @f2harrell @AndersHuitfeldt) I am all for parametric models. But I have not seen the problem of external validity, in the parametric context, expressed in layman terms, i.e., What's given, what's assumed, what's needed. Does it take more than 2 Tweets? #Bookofwhy

3.17.19 @2:18pm - (Replying to @Susan_Athey) Would you consider inability to decide if a variable is conditionally exogeneous (given an econ. model) to be a hindrance to identification of counterfactual inference? Do you think that DAG-avoiding economists do not suffer from this inability? #Bookofwhy

3.17.19 @1:20pm - (Replying to @CarterPaddy @fuiud @Jabaluck) This is indeed my stance. See , where I expressed surprise at the heirs of Haavelmo, Wold and Maschak for not pursuing these teachers, and not jumping enthusiastically at the opportunity to overcome the obstacles that held the Cowell's Commission at bay.

3.17.19 @1:03pm - (Replying to @fuiud @CarterPaddy @Jabaluck) I am listening attentively to you (not to "our field") Please describe to me your attitude to external validity. ie. what are wet trying to estimate, what information is available to us, what assumptions we are making etc. I am eager to talk with econ who says "I can" #Bookofwhy

3.17.19 @12:45pm - (Replying to @fuiud @CarterPaddy @Jabaluck) I dont buy the "reality of applied work" as an excuse for not trying synthetic problems, whose correct answers can be obtained either analytically or by simulation. I would love to learn new attitudes to external validity issues, explained from scratch. I'm listening. #Bookofwhy

3.17.19 @12:27pm - (Replying to @Jabaluck @btshapir @CarterPaddy) In my humble field of computer science, if an editor smells that new tools are being developed in neighboring fields, possibly applicable to another, he/she sees it as an editorial duty to invite an introductory article to educate readers. Even stat journals do so, not econ.

3.17.19 @12:08pm - (Replying to @Jabaluck @CarterPaddy) A passing word on those revered "top journals". The editors of three such journals advised me not to submit papers because "they cannot find qualified reviewers". I hope econ. readers advise those editors to retire if they want econ. to enter the age of modernity. #Bookofwhy

3.17.19 @11:56am - (Replying to @CarterPaddy @Jabaluck) There would definitely be a sweeping changes even in published papers (ie, using addition). E.G. "conditional ignorability" would not be assumed apriori, but be justified on scientific grounds, namely, by appeal to "how the world works" instead of "everyone makes this assumption"

3.17.19 @11:47am - (Replying to @Jabaluck @CarterPaddy) True, if is not just "save time" but solving correctly problems that were discarded by economists (multiplication) for being beyond the pale of their tools (addition) hence they never appear in the "top journals".

3.17.19 @11:42am - (Replying to @CarterPaddy @Jabaluck) Not only "easier", but "possible". eg. Deciding to treat or not to treat in Simpson's paradox is intractable to the unaided human mind. Same with choosing the right set Z for "control" of confounding. Same with deciding if an RCT result is generalizable. #Bookofwhy

3.17.19 @11:33am - (Replying to @Jabaluck @CarterPaddy) Why would it be irresponsible for an observer to say that first-grade students do not really understand arithmetic, if all the problems they solved thus far are problems in addition? One can say so without showing that the solution would improve with multiplication. #Bookofwhy

3.17.19 @6:33am - (Replying to @AndersHuitfeldt @f2harrell) Right. We are dealing with non-parametric models which, by nature, allow unrestricted heterogeneity. I therefore presume that if "interaction" plays aany role, it must be that one is working in some parametric model, which I missed. Can we clarify afresh what the problem is?

3.17.19 @4:12am - (Replying to @CarterPaddy) The way we can handle feedback is to unfold the loop into a sequential, time varying DAG and, unless we can assume linearity, this becomes quite messy. Luckily, when linearity cab be assumed, d-separation holds, and some identification results can be obtained w/o unfolding.

3.17.19 @3:22am - (1/5) Reading this justification of X||Y_x|Z, I was ready to plead ignorance of "cost sharing" "copay" "actuaries" and "utilization trends" and quit before it gets too domain-specific. But out of respect to your genuine attempt to capture the meaning of this statement,
3.17.19 @3:22am - (2/5) I offer my version, in generic terms. (1) The cryptic statement X||Y_x|Z, also named "conditional ignorability" (CI) by PO folks, is a feature of the population under study and, when valid, provides a license to estimate the ATE using regression, simply "controlling for Z"
3.17.19 @3:22am - (3/5) CI is the key assumption behind all works in PO. (2) Being a feature of the population, it can be validated from our model of the world, without thinking about what we do or wish to do. It depends only on how Z is related to X and Y in the presence of other variables if any.
3.17.19 @3:22am - (4/5) (3) If our model comes in the form of a DAG (a depiction of economic structural equation model) we can validate CI by simply checking if all paths between X and Y are "backdoor-blocked by Z". (4) The notion of "backdoor-blocked by Z" is a fun, game-like criterion on DAGs
3.17.19 @3:22am - (5/5) that can be mastered in 5-12 minutes by any economist who is serious about finding out if ATE is estimable by regression. See #Bookofwhy or PRIMER . Shunning DAGs, PO folks must assume CI apriori, unjustified, and some economists follow them blindly.

3.17.19 @2:14am - Very well put!! And there is an additional aspect to it. DAGs are not merely visualization devices to help economists spot opportunities, they are also computational engines that spot those opportunities and exploit them, while you relax and think only: "how the world works"

3.16.19 @4:49pm - Demanding "same treatment effects" is basically same as "genralizability". The next question is: Given what we know about how the target and study populations DIFFER, decide if the latter still "represents" the former. Examples:

3.16.19 @4:30pm - (Replying to @JWSBayes @f2harrell and 2 others) I know of two publications that tell it quite nicely, perhaps different from @f2harrell and @stephensenn, but addressing a clear question: When can you generalize and when you cant. Here they are: and

3.16.19 @3:22pm - (Replying to @Jabaluck) Forget DAGs. "scientific grounds" is anything that resembles the way scientists communicate, or the way scientific knowledge is stored in the mind of a researcher who is asked to judge if a given condition is plausible or not. You are now to justify X||Y_x|Z. You choose X,Y,Z

3.16.19 @2:50pm - (Replying to @Jabaluck) Great, if only PO researchers could justify those "sufficient conditions" on scientific grounds. But, again, we should stop hypothesizing what other researchers can or cannot do. It is time for us to hear one researcher saying: I CAN DO. Willing to volunteer? #Bookofwhy

3.16.19 @2:29pm - (Replying to @Jabaluck) Proving you wrong: It is time we hear "I am an applied researcher, and I understand this and this, e.g., I can detect conditional exogenetity when it exists in my model, and other features that are necessary for CI " instead of "I think so and so already understand... #Bookofwhy

3.16.19 @2:01am - (1/3) The criterion you propose (could econ. do better with DAGs) suffers from selection bias. Could first-grade students do better with multiplication? No! The exam was only about addition. Lacking CI tools, economists had to work on simple problems, where their lamppost methods
3.16.19 @2:01am - (2/3) (LATE, dif-i-diff, reg-disc.) seemed applicable. When I say "slow progress" I look at areas such as: identification, mediation, transportability, sensitivity, which flourished in soc. sc. (eg. Morgan &Winship, Imai etal) in the past 20 years, with almost no progress in econ
3.16.19 @2:01am - (3/3) And when we compare progress, let us discard hearsay eg., "so and so did it" and count only those who say: "I can do it". Eg. "I can detect conditional exogeneity" or "I can repair violations of exclusion". We hear very few "I can" in the econ. literature. We need more.

3.15.19 @12:56am - (Replying to @Jabaluck @analisereal) I wish I could believe it. But watching econometric research progress over the past 25 years, it is hard to believe that laziness alone made them slip behind other CI disciplines, eg Epi or Soc Sci., especially in identification, mediation, transportability, selection, more ..

3.14.19 @12:59am - (Replying to @Jabaluck @analisereal) Curious: Do you believe economists' systematic avoidance of DAGs reflects informed scientific judgement or cultish dogma? #Bookofwhy

3.14.19 @6:45am - (Replying to @Jabaluck @analisereal) This paper is an example of how assumptions are made not because they are defensible on scientific grounds, but because they allow certain mathematical derivations to go through. If this is "more commonly used analysis in econ" I would be really worried. Arn't you?. #Bookofwhy

3.14.19 @5:49am - (Replying to @thosjleeper @analisereal @causalinf) I fail to see why we learn more "empirically" when we tacitly assume exogeneity and exclusion for IV, as opposed to representing these same assumptions explicitly in a DAG?

3.14.19 @5:16am - (Replying to @thosjleeper @analisereal @causalinf) DAGs are perfect for representing absence of prior information, just fill them with arrows and unobserved confounders. #Bookofwhy

3.14.19 @3:47am - (Replying to @analisereal @thosjleeper @causalinf) I would phrase it a bit differently. Instead of building models that represent their problems, they impose assumptions that permit identification. The former are defensible on scientific grounds, the latter are fabricated for convenience, hence titled "identifying assumptions."

3.13.19 @4:58am - (Replying to @TPA_Debray) So suppose I am dying to estimate the probability the Unicorns have purple horns. It is quantitative, because I asked for "probability" and it is surely "uncertain". Is it a statistical question?

3.13.19 @4:12am - (1/3) Imagine a modern chemist traveling back in time to a 16th Century alchemist laboratory. This is how I felt upon reading this paper on "suppression" [ie. where adding a new variable to a regression model, totally unrelated to the outcome,
3.13.19 @4:12am - (2/2) increases the predictive power of the model]. Being an incurable whiggish historian, I have found it fascinating to watch how a phenomenon that has baffled social scientists since the 1940's is unfolding gracefully through the lens of causal analysis. #Bookofwhy

3.13.19 @12:39am - (1/3) Jeff has a point. If we start with "colliders" and "confounders" economists can say: "yes that makes sense but I didn't need a DAG to understand that." For this reason I usually start with "look at your own model, the one you authored. Can you tell me a few things about it?"
3.13.19 @12:39am - (2/3) For example: "which parameter can be identified if we assume correlated disturbances between Z and W?" Now, the economist feels embarrassed not to answer simple questions about his OWN model, and, if he is curious, he would say: "can you do it with DAGs?". If he is not, ....
3.13.19 @12:39am - (3/3) he/she would get defensive, and say "How can you belittle the whole field of econometrics...? " [ familiar?]. The trick is to educate the former without offending the latter. Tough but doable. There is hope to Eco. #Bookofwhy BTW the paper by West is pure correlation, unfit.

3.12.19 @11:36pm - (Replying to @boback @EpiEllie) What is Kardashian ? A treatment? or an alternative universe?

3.12.19 @11:33pm - Many readers have questioned whether "Causal Revolution" is an appropriate name for what we are currently seeing in @causalinference . Here is a short note touching on Kuhn's "Structure of Scientific Revolution", which puts things in historical perspective

3.12.19 @11:12pm - (Replying to @l__ds) Pearson and Yule observed the disappearance of correlation upon aggregation, not reversal. I checked their papers years ago. See Pearson's quote in #Bookofwhy

3.12.19 @11:04pm - (Replying to @deaneckles @metrics52) This is great, because it sounds exactly like Pearl. Did @metrics52 join the revolution? Or is he going to clash with Pearl in the next chapter, where students need to learn how to actually compute counterfactuals? Lets wait and see. #Bookofwhy

3.12.19 @10:22pm - (Replying to @ghoshd @kareem_carr @EpiEllie) It means that, unlike "obesity", people agree on how to measure it. If wealth is thus measured then, indeed, the query is just "probability of necessity" as defined in #Bookofwhy and even better in PRIMER

3.12.19 @9:08pm - (Replying to @EpiEllie) Yesterday I did not know if Kylie Janner is a person or a tarantula. So, I would not use this example to learn about SCM. I dont see what makes this example different than asking for the "probability that factor F is necessary for outcome Y", which is in Primer Chp 4 & #Bookofwhy

3.12.19 @8:50pm - (Replying to @chrisalbon) If Causality flashbacks are too painful try #Bookofwhy, you will forever swim in "flashforwards"

3.12.19 @8:13pm - (1/2) A few notes on the history of Simpson's Paradox. (1) The first reversal was noted by Cohen and Nagel (1934, p.449). (2)Simpson was indeed the first to note that what we judge as "sensible" depends on the STORY, not on the data [did not say "causal"] (3) In Causality (p. 177)
3.12.19 @8:13pm - (2/2) I attributed this discovery to Lindley and Novick (1981), but Hernan etal (2011) corrected me; Simpson already noted it in 1951. (4) Not all examples of the paradox invoked confounding. Lindley etal used mediation as in Fig. 1(b) of and #Bookofwhy

3.12.19 @3:09pm - (1/2) I did not realize that Ed Simpson's was still alive, and I feel really sad with his departure. He will be remembered as an intellectual family member of all students of causality. The paper you cite is a timely citation in light of recent discussions
3.12.19 @3:09pm - (2/2) on whether causality is IN or OUT of statistics. I dont believe anyone would claim IN, in good faith, after reading the history of Simpson's Paradox and noting the persistent, century-long effort by leading statisticians to avoid its causal dimension. Amazing! #Bookofwhy

3.12.19 @5:37am - (Replying to @EpiEllie @kareem_carr) By causal model I mean qualitative model specifying who listens to whom. Surely, if we have good measurements on all relevant variables and data on how successful people are in Jenner category, then the question translates to prob. of necessity #Bookofwhy, which can be bounded

3.12.19 @12:16am - (Replying to @kareem_carr @EpiEllie) We can do the moon without data because the theory of gravity gives us a fully specified dag, with functions attached to the arrows. Lacking functional specification, we need data. #Bookofwhy

3.12.19 @12:09am - (Replying to @dccozine @ale_martinello @eliasbareinboim) I could not agree more. But if they were only visualization-communication devices DAGs would end up where SEMs are today, ie, nice diagrams all over the articles, yet when it comes to identification and testable bad mouthing..... see Journal of SEM. #Bookofwhy

3.11.19 @11:48pm - (Replying to @the_dismal_tide @kareem_carr) I would join one of the dozens Data Science Centers now erected across the country and keep on bugging my mentors: "Hey, and where is the Science in our building?" . Simultaneously I would take one of the obstacles of ML, say transfer learning and overcome it using "science".

3.11.19 @10:08pm - (1/2) In the wake of on/off discussions of external validity vs generalisability vs transportability, I am glad to note that Boston will soon get a glimpse at aunifying science of "data fusion", as in . Elias Bareinboim will be speaking at Harvard on
3.11.19 @10:08pm - (2/2) March 21 and will illuminate the topic with conceptual, mathematical and algorithmic results.@HarvardHealth @Kennedy_School @HarvardChanSPH @_MiguelHernan @harvard_data @HarvardEpi

3.11.19 @9:16pm - (Replying to @kareem_carr) I mean "data science' as defined in one of my ancient papers: The use of data to interpret the world. Not as it is used today, which is "the use of data to summarize data" #Bookofwhy

3.11.19 @8:49pm - (Replying to @mathtick) I would not use the word "fraud" because it connotes intention to deceive. Econ leaders are just lazy to buy (or even try) a new pair of eyeglasses. No fraud committed. But who would volunteer to tell their students? #Bookofwhy

3.11.19 @8:37pm - (Replying to @RussellSPierce @balexanderstats) Natural experiments are Rung-1, i.e., observational studies in which the experimenter makes an assumption of exogeneity on one or more variables. There is no need for continuum, because the the presence/absence of physical intervention is usually crisp. #Bookofwhy

3.11.19 @8:30pm - (Replying to @kareem_carr @EpiEllie) You suspect wrongly about UCLA. Reading carefully, we find that to estimate P(Y|do(X)), X needs (1) to be unambiguously measured (e.g. blood pressure, or sex) and (2) population data and (3) a causal model must be available. #Bookofwhy

3.11.19 @8:13pm - (Replying to @kareem_carr) Great question!! It will eventually be housed in Data-Science, embracing both Data (now housed in statistics) and Science (now scattered among various scientific disciplines, the common denominator of which will come under the second arm of Data-Science). #Bookofwhy

3.11.19 @8:07pm - Newly announced: Admission to the Why-19 Symposium will be FREE for @Stanford community. I will be talking Tuesday, 3/26 2-4 pm on "What is Causal Inference" and will be delighted to see you there. @StanfordMed @StanfordEng @StanfordAILab @drfeifei

3.11.19 @2:57pm - (1/2) How can anyone say that economists "never cared about such questions"?? characterizes Haavelmo as the first to define causal effects in economic models. Still, one can expose the current stalemate of econometrics and its insular structure without
3.11.19 @2:57pm - (2/2) being accused of "belittling others" How else can one convince Econ students that they are being taught outdated methods, if the champions of those methods refuse to tell us how they solve a toy problem (eg. your DAG)? Shouldn't students take this refusal as outdatedness?

3.11.19 @2:11pm - (Replying to @ale_martinello @_limbs_) @_limbs_ is asking a humble question: "shouldn't we question the usefulness of traditional vs modern methods?" Why is he being accused of "belittling/denigrating others"? When/how can we compare methods and display their faults and merits without being thus accused? #Bookofwhy

3.11.19 @12:58pm - (Replying to @80Data @kareem_carr) @learnfromerror Interesting. And now (2019) do you still believe Granger causality has anything to do with "causality"? #Bookofwhy

3.11.19 @12:55pm - (1/2) On the other extreme, one can takes the position that statistic is EVERYTHING, subsuming all scientific endeavors because, whenever you ask a scientific question the answer is "a missing data problem" - Bingo. To avoid such extremes I have found it useful (Causality p.38)
3.11.19 @12:55pm - (2/2) to delineate the boundaries of "statistics" accordance to its practices: "the study of relationships governed by distributions of observed data", which also coincides with Fisher's definition of "reduction of data", and makes causal relations "extra-statistical." #Bookofwhy

3.11.19 @11:57pm - (Replying to @RussellSPierce @balexanderstats) This discussion revolves around observational studies. Interventions take us to Rung-2 of the ladder, and takes a different logic. For example, "Would an intervention be sufficient for answering our research question?" is derivable in that logic. See

3.11.19 @4:47pm - (Replying to @balexanderstats) I am going to use statistics, but only where it is useful, namely in the estimation phase of the causal exercise. See the structure of the Inference Engine in and also in #Bookofwhy

3.11.19 @4:43pm - (Replying to @Hl60464759) Some examples of extra statistical information are: Mud does not cause rain, symptoms do not cause diseases , the drug does not change patients gender. In short, all the missing arrows in the examples of #Bookofwhy

3.11.19 @4:13am - (1/5r) No offense, and I appeciate your sharing impressions with other readers. I am even more grateful for mentioning @StatModeling which should give readers a glimpse at how some 2019 statisticians think. I quote: "I find it baffling that Pearl and his colleagues keep taking
3.11.19 @4:13am - (2/5) statistical problems and, to my mind, complicating them by wrapping them in a causal structure." This quote from Gelman's blog should enter the archives of scientific revolutions as proof that my depiction of the inertial forces paralyzing statistics is not made up; and my
3.11.19 @4:13am - (3/5) description of causal inference as a "revolution" is not a fantasy. The resistance to accepting needed assumptions as "extra statistical" is alive even in 2019. Moreover, readers of this quote take it at face value that problems solved in #Bookofwhy can also be solved by
3.11.19 @4:13am - (4/5) Gelman's students w/o "wrapping them in a causal structure". This is the power of blogs; no one asks you how: eg, "Can you show us how 'traditional statistical methods" would solve problems?" People assume they somehow do. For otherwise, the all powerful science of statistics
3.11.19 @4:13am - (5/5) would be deficient, which is inconceivable, because someone would have noticed. Well, the truth is "traditional statistics" IS deficient and her obedient students cannot solve those problems without "wrapping them in a causal structure". And challenging them is no "bashing".

3.11.19 @2:30am - Glad you like #Bookofwhy. But where is "bashing of statisticians"? This is what the book says about statisticians: "...they declared those questions off limits and turned to developing a thriving, causality-free enterprise called statistics." This is not "bashing", it is history.

3.11.19 @2:11am - (1/2) Your DAG above can help us understand how classical economists (i.e. your mentors) used to approach the problem of identification before graphs. From your recollection, did they use matrices? method of moments? LISREL? PO? Imagine one of them addressing a class today.
3.11.19 @2:11am - (2/2) How would he/she convince the class that all the betas are identified even when IQ_0 and C are unobserved. Today we can do it no hands but, as a part-time historian I am curious how the world looked like before windshield wipers. Do you remember? Does anyone? #Bookofwhy

3.10.19 @7:27pm - (1/2) Well phrased! And I am glad you mention our paper on Econ texts . Intriguingly, all my econ. colleagues agree: "our testbooks are an embarrssment, they dont represent our research, but who has time to write good texts". Moreover, some of the authors
3.10.19 @7:27pm - (2/2) of the reviewed textbooks wrote to me (in secret) for advice on how to repair things in the next edition. So, no reason to despair, these new editions will be out in this and next year, econometrics can be reclaimed and its current gurus bypassed. I am convinced #Bookofwhy

3.10.19 @3:49pm - (1/2) I have corrected you twice before, first on what I said about economists and, second, on what we can believe about the research that some economists produce. I do not understand your need to distort what I say. I have also explained (above Tweet) why economists conservatism
3.10.19 @3:49pm - (2/2) is a natural phenomenon, somewhat more glaring than in Computer Science, but still forgiven. "No hard feeling if they dont, gold medals if they do." But, Hey, you owe us an answer on N>4. "No hard feeling if you dont, gold...." #bookofwhy

3.10.19 @3:26pm - (1/4) Well put. Referring to a scientific paradigm committed to certain tools by the name of the professions that have adopted that paradigm IS NOT AD HOMINEN. But, to pacify the offended, let me declare publically that by saying "economists cant solve problem X" I mean:
3.10.19 @3:26pm - (2/4) "Researchers who have acquired ONLY the tools taught to them by economics textbooks and economics professors adhering to those textbooks cannot solve problem X, unless they enrich those tools by their own initiative". No hard feeling if they dont, gold medals if they do.(cont
3.10.19 @3:26pm - (3/4) We cannot ignore the fact that scientific research and education are organized into tightly shielded paradigms (or cults) laboring to keep members from defection. In some extreme cases the tools that an individual uses in research are perceived as proof of allegiance to
3.10.19 @3:26pm - (4/4) the cult or its gurus, an allegiance that entails significant academic benefits. The sooner we admit this cultish structure the better we can correct for it. In summary, it takes courage to defect: No hard feelings if you dont, Kudos and better science if you do. #Bookofwhy

3.10.19 @7:52am - (1/2) You and I make such claims all the time, and no one gets offended when we assert that certain tasks cannot be performed absent certain tools. It does not reflect on the professional honor or creativity of those lacking the tools. But I am still curious about your own personal
3.10.19 @7:52am - (2/2) experience among researchers who are offended by my observation that certain tasks require certain tools. Have you really met anyone who can judge the plausibility of ignorability assumptions in problems of size N>4? Please share your experience with us. #Bookofwhy

3.10.19 @7:12am - (Replying to @pophealth3) The analogy seems valid to me: Even with 3-dimensional objects, we still need to calculate volumes using 2-dimensional surfaces.

3.10.19 @5:47am - (Replying to @ale_martinello) I am truly curious to know how economists did it w/o DAGs, even in linear systems, in the presence of unmeasured confounders. What was the prevailing technique?

3.10.19 @4:57am - I am careful NOT to pose questions in terms of DAGs, that would be cheating. The questions I asked are generic, e.g., which parameter in YOUR model can be estimated using OLS? I dont care if they use DAGs or not to answer such important questions - I am surprised they dont care.

3.10.19 @4:27am - (1/2) In some simple cases, people who make ignorability assumptions can translate them (mentally) into scientific assumptions and judge their plausibility. In most cases they can't and, then, the conclusions are as opaque as the assumptions. Interestingly, last I checked they
3.10.19 @4:27am - (2/2) have refused to take a test to evaluate their ability to judge plausibility of their own ignorability assumptions. But I am eager to learn from your experience. Have you met anyone who can judge the plausibility of ignorability assumptions in problems of size N>4? #Bookofwhy

3.10.19 @3:50am - This is beautiful quote, thanks. First, note that I specifically refer to "causal question posed in ". Second, how many economists can you name who can answer those fundamental questions? Seriously, don't leave us in suspense., more than 10? #Bookofwhy

3.10.19 @12:40am - (Replying to @thosjleeper) Again, I am from a small village and cannot make sweeping statements like the above. In this paper I was very specific about the kind of questions most economists cannot answer. If you can answer any, tweet, and I will gladly note the exception #Bookofwhy

3.10.19 @12:25am - (1/2) @petemohanty brought up Casella and Berger for a good reason, they are "pure" statisticians who never claimed explicitly to deal with causal questions. Still, some statisticians today believe that, if you dig deeply into the foundations of statistics in C&B book, you will be
3.10.19 @12:25am - (2/2) able to answer such questions. The people you cited (Angrist, Imbens, Rubin, Gelman/Hill, Athey, Freedman, etc.) are not "purists", since they do talk explicitly about CI, and some even admit to be doing so with extra-statistical assumptions (eg. "ignorability") #Bookofwhy

3.10.19 @12:03am - (Replying to @thosjleeper) I come from a small village. I do not speak of "the vast majority of empirical research" nor of "all current statistical evidence". I deal with one claim at at time. Which claim would you like me to evaluate? #Bookofwhy

3.9.19 @11:45pm - (1/2) I have invited you to take a leisurely journey into a book that you chose, search for one toy example that you deem to be "causal" and decide if the book solves it to your satisfaction. So, why am I unfair? I have extended this invitation to every colleague who tells me:
3.9.19 @11:45pm - (2/2) "Statisticians have done it decades ago." Thus far w/o success. So, where have I been unfair to "decades of work..."? Fisher, Lord, Kruskal, and Lindley acknowledged that certain questions reside outside the province of statistics; what would it take to convince you?

3.9.19 @10:59pm - (Replying to @BDataScientist) It reminds me of the finding (in the 1970's) that men earn a higher salary than equally qualified women, and simultaneously, men are more qualified than women doing equally paying job. It is described and treated in under "reverse regression" #Bookofwhy

3.9.19 @7:30pm - (1/2) The disconnect between statistics and causality is unfortunately not a matter of "imperfection" and no amount of effort "to improve existing work" could bridge this disconnect. My analogy: it is like asking a 2-dimensional creature what "volume" is. To prove my point,..
3.9.19 @7:30pm - (2/2) let's take an exploratory journey through the toy examples of Casella and Berger and examine if ANY of them answers a causal question. If we find one, I will take back the 2-dimensional-volume analogy and go study statistics. #Bookofwhy

3.9.19 @2:16am - Confessing negligence. I have been so fascinated by how much millage we get from the primitive relation "is a function of" that I have neglected to explore other types of semantic networks. With the exception of "Is part of", in 1986, see #Boodofwhy

3.8.19 @6:46pm - (1/3) There isn't really great need to differentiate external validity vs generalisability vs transportability, since we now we have a unified framework to handle them all, as in . The most important distinction one needs to make is about the disparities
3.8.19 @6:46pm - (2/3) between the study and target populations, i.e., whether such disparities are "man-made" (as in recruiting subjects) or "nature made" (eg age differences). The interplay between the two is described in . Still, however we taxonomize these subproblems,
3.8.19 @6:46pm - (3/3) I would be very weary of any theory that does not provide you with playful solutions to at least some toy problems, for example, the three toy problems in Fig. 3 of . #Bookofwhy @BrownUniversity #kolokotrones @HarvardEpi @harvard_data

3.8.19 @5:59pm - (Replying to @djvanness) I am familiar with Manki's writings, and I am curious to hear from his other readers HOW he represents that "extra statistical information", and whether he does that very well. My litmus test for "very well" will be (as usual) the abilitiy to solve a toy problem #Bookofwhy

3.8.19 @3:46pm - (Replying to @malmyros) Causal diagrams in #Bookofwhy are a special (primitive) case of semantic networks described in Minsky's book. The nodes are all "variables" (not "objects") and the arrows represent ONE relation: "is a function of". Next generation AI will extend CI to full semantic networks.

3.8.19 @3:30pm - (1/2) If statisticians understand that "you cannot learn much of anything from a statistic without `extra statistical' information," then one would expect statistics textbooks to spend at least a few pages on how that "extra statistical information" can be represented ...
3.8.19 @3:30pm - (2/2) mathematically, and how it can be combined with data in order to obtain what statistics alone cannot provide. I have not found such a textbook yet. But would be eager to learn. #Bookofwhy

3.8.19 @3:12pm - (Replying to @jwalkrunski) The statistical and causal notions of identification are compared in footnote 7 of . The unique character of Q(M) is discussed in the other sources sited.

3.7.19 @6:02pm - (Replying to @omaclaren @eliasbareinboim) Any link to Stark?

3.7.19 @4:42pm - (Replying to @jwalkrunski) This concept of Q(M), and the identifiability definition that it entails are the foundations of CI and appear in many works: from 1995 to 2014 to 2018 #Bookofwhy. I even called it "Principle-1" in

3.7.19 @5:36am - (Replying to @ad_pickering) I havn't checked SEMNET lately. Are they still using regression? Havn't they heard the bells of the revolution? Avanti Popolo .... The yoke of tyranny is broken... How do we bring them into the fold? Do they still read The Journal of SEM? #Bookofwhy

3.7.19 @2:25am - (Replying to @iamconscious2 @RebekahBaglini) I wish I could be there. But, seriously, is anyone doubting that we are having a causal revolution? I have an untested theory in mind: Those who doubt do not know how to solve Simpson paradox (or any toy problem), and those who do not doubt can solve it. Worth a test. #Bookofwhy

3.7.19 @1:41am - (1/1) The tools of Transportability analysis are almost unknown to researchers most badly needing them, namely traditional trialists, probably because these tools are often buried in opaque ignorability jargon. (see Students attending the
3.7.19 @1:41am - (2/2) transportability symposium at @BrownUniversity would undoubtedly be introduced to modern methods, invoking graphs, which should turn every trialist to an expert in transporting study results (see #Bookofwhy p. 354, and Tool 5) . Merry Transporting!

3.6.19 @10:45pm - (Replying to @omaclaren @analisereal) I understood your argument to be that allowing "all functions" behind the nonparametric model is too permissive, some functions are pathological and would prevent estimation. I proposed excluding such pathologies the same way we exclude non-positive distributions. Why not?

3.6.19 @10:05pm - (Replying to @omaclaren) You make it sound like I disagree with your conclusions. I dont. I disagree however to taking syntactic similarity of two definitions and calling it "exact equivalence". In one, the model is defined by the thetas; in the second, Q is a function computable FROM the model.

3.6.19 @8:33pm - (1/2) Surely for every model M there is one distribution P(M). But this is where the 1-1 "index set" analogy ends. The key feature of causal models, not share by statistical models, is that for every M there is one Q(M), which the answer to our causal question Q. Identifiability
3.6.19 @8:33pm - (2/2) does not require inverse mapping M(P). Instead, it requires that whenever two models agree on Q they also agree on P. Allowing "arbitrarily indexed family of distributions" is like allowing arithmetical operations; thanks, but it misses the essence of causation. #Bookofwhy

3.6.19 @4:01pm - (Replying to @omaclaren) We normally attached the assumption of "positivity" to non-parametric analysis. Would the sun shines brighter if we augment "positivity" with "non-pathological"? Can we protect ourselves against pathological cases? #Bookofwhy

3.6.19 @3:38pm - (Replying to @cmirzayi @staci_risman) I agree. Authors are often surprised by who finds their books useful or entertaining. I get letters such as: "I dont understand any of the math, but I admire Sewall Wright and Barbara Burk" or "Now I understand what they tried to teach me in Econ. 101". ..etc. #Bookofwhy

3.6.19 @3:03pm - (Replying to @omaclaren) Questions. Suppose there is a class PD of pathologial distributions. Can one recognize them from M? from the estimand? from the data? Can you demonstrate one such PD using 3 variables, 4 ? Suppose PD exists, does this minimize the importance of identification? #Bookofwhy

3.6.19 @2:53pm - (Replying to @LMcHugh_Russell @UnlearningEcon and 2 others) It seems to me that #Bookofwhy should provide what you are looking for, especially when coupled with PRIMER for fun and homework.

3.6.19 @7:24am - (Replying to @statwonk) I thought someone would rise to defend the honor of statistics. But we are dealing here with "identifiability" not with causal intuition, nor with RCT. Take the simplest problem of selecting covariates for adjustment, how many statistics professor can do it TODAY? #Bookofwhy

3.6.19 @7:03am - Correcting a link. The relation between the statistical and causal notions of identifiability is explicated in footnote 7 of this paper .

3.6.19 @1:33am - I second @eliasbareinboim . The WHY-19 Symposium will give me, personally, the pleasure of meeting readers and followers and profess/confess what I believe Causal Inference is all about, at least from my humble, totally unbiased perspective. #Bookofwhy.

3.6.19 @1:11am - (1/3) Now it is my turn to say: No way! The "identifiability analysis" conceptualized as a "property of a family of arbitrarily indexed probability measures," is good for statistical identifiability but is totally useless for #causalinference. This conception is precisely what
3.6.19 @1:11am - (2/3) #Bookofwhy asks readers to abandon, and it may explain perhaps why statistics has not made ANY progress in identifying causal effects for over 150 years. To do that we need extra-statistical information. Footnote 7 in explicates the relation between
3.6.19 @1:11am - (3/3) the two notions of identifiability. Some economists tried to import this "indexed probability" conception to econometrics with known results. Readers of #Bookofwhy: Watch out for this trap; a causal model is NOT a set of distributions, one for each choice of parameters.

3.5.19 @6:24pm - (Replying to @omaclaren @edwardhkennedy and 2 others) Once a causal parameter is identifiable you forget about it being "causal". Dress it up as a "statistical" parameter, and ask "Is estimation possible?" as if it came from a respectable statistical textbook. Causality plays no role in making the estimation hard or easy. #Bookofwhy

3.5.19 @3:09pm - (Replying to @edwardhkennedy @omaclaren and 2 others) Pearl spends more time on identifiablility, because he believes that the other task (estimation) can be handled by traditional, 150 years old statistical methodology, and that it is our new understanding of identifiability that will make #causalinference a reality. #Bookofwhy

3.5.19 @2:43pm - (Replying to @kareem_carr @generativist and 5 others) Not Typical??? Perhaps this was the reason I thought: Eurika! I found someone at Harvard who can explain to me what is, and what is NOT a g-method. Now I am back at the mercy of my inability to get it the official way. Hard! #Bookofwhy

3.5.19 @5:17am - (1/2) (Replying to @kareem_carr @generativist and 2 others) Your bite is an invaluable resource for the outside world to understand the workings of Harvard culture. I will add a couple of comments to those made by @analisereal. 1. SCM is more than DAGS+do. It contains all seven wisdoms of . #Bookofwhy
3.5.19 @5:17am - (2/2) (Replying to @yudapearl @kareem_carr and 4 others 3. If g-methods are taught at Harvard as "effects which epidemiologists care about." then what is NOT a g-method? i.e., what kind of effects epidemiologists do NOT care about? Front-door perhaps? Napkin? Or mediated effects? Or is everything a g-method? Why not? #Bookofwhy

3.4.19 @2:31pm - (Replying to @kareem_carr @generativist and 2 others) There is no room for embarrassment in trying to taxonomize a new science, where every new trick gets a fancy new name. As the Mishna (220 AD says: lo habayshan lamed, v'lo ha kapdan l'lamed. -- A bashful person cannot learn, nor can an impatient person teach. #Bookofwhy

3.4.19 @8:40am - (Replying to @kareem_carr @generativist and 2 others) Three comments:
1. I do not find SCM (Structural Causal Models) on your list. See eg #Bookofwhy Does SCM unify all the rest? (my school)
2. Are these schools independent of each other? Or nested within each other?
3. What are G-methods? Or, what is not a G-method?

3.2.19 @8:36am - (Replying to @wgrosso @shakir_za and 2 others) The point is that the limitations articulated in are not merely "interesting" but challenge the foundations of Bayesian epistemology. Bayesian practitioners, however, continue to speak as if those limitations do not exist: This strikes me as very bizzare.

3.2.19 @8:33am - Agree. And those who think that CI is a missing data prob. should read and compare the transparency of the assumptions needed, the theoretical guarantees obtained, and the tools produced by each paradigm. It is a streetlight vs. flashlight comparison.

3.2.19 @3:10am - (Replying to @wgrosso @shakir_za and 2 others) Thanks for noting that the expression "waiting for their posteriors to peak" may be given grotesque interpretation, with "posteriors" being the rear parts of one's body, rather than "posterior probabilities continuously updated by data." Begging Bayesians for forgiveness.

3.1.19 @1:57pm - (Replying to @shakir_za @blei_lab @fhuszar) I am glad you mentioned my "half Bayesian" paper which makes me more convinced in "halfness" each time I read it. I wish I could get serious feedback from the "full Bayesians", but they are busy waiting for their posteriors to peak on the truth.#Bookofwhy

3.1.19 @6:34am - Paul, I agree with your observation. However the issue is not parametric vs. nonparametric as much as it is causal vs. associational. ML vocabulary (at least in 2019) is purely associational or predictive, while econometric parameters are causal, as you noted, residing out of ML

3.1.19 @6:24am - (1/2) @KoelleM While admiring your attempt to bridge #MachineLearning and #economics I must add a note of skepticism. Economic parameters are CAUSAL, while the vocabulary of ML is statistical. The two do not mix. E.g., Beta is defined causally as in , Eq. (3)
3.1.19 @6:24am - (2/2) True, not many econometric books define beta this way. But those that do, enjoy clarity and consistency, and escape the embarrassment of those that dont. See for example [The authors confessed to embarrassment] #Bookofwhy

3.1.19 @2:16am - (Replying to @fhuszar) To be honest, I meant the students in the audience but, now that you mentioned the panelists, I would appeal to model-blind extrapolation. #Bookofwhy

3.1.19 @1:40am - I came across this interesting article about Causality in Machine Learning ... which also links to a recent panel on causality ... It is interesting to see how young ML researchers react to the new hype. Very encouraging #Bookofwhy.

2.28.19 @5:44pm - (1/1) The theory of transportability is not only a "nested version" of Transfer Learning (TL), but the SOLUTION to the set of problems coined TL. It is listed as Tool-5 in and proven complete in . To the best of my knowledge
2.28.19 @5:44pm - (2/2) the enterprise of TL has yet to pass Pearl's Litmus test, of solving one TOY PROBLEM, as those solved in and . Reason: TL is a probabilistic enterprise while "transfer"requires causal assumptions. #Bookofwhy.

2.28.19 @12:38am - (Replying to @t_matam_t @audibleDE) Glad you found one. Can you share the link with other readers -- I can use one myself. Thanks.

2.28.19 @1:30am - Kudos! And an overdue closure! I am extremely happy to see @DAGophile bridging modern mediation analysis with the long, tormented and hopeless struggle that philosophers have had trying to capture intuitions about direct and indirect effects using inadequate languages. #Bookofwhy

2.27.19 @11:28am - (Replying to @ConiByera @stephensenn @deaneckles)

2.27.19 @11:28am - (Replying to @ConiByera @stephensenn @deaneckles) You hit it on the nail. Lord Paradox is paradoxical in finite as well as infinite populations. It was introduced in the latter form, and has challenged statisticians in that form. I do not see what insights can be gained by introducing finite sample complications. #Bookofwhy

2.27.19 @2:14am - (Replying to @HenningStrandin @kareem_carr) I would take X=x to mean "the gravitational pull at the molecule location is a vector x." And I would not even think about changing the moon location. I would compare rushing under X=x to rushing under X = x' and that would give me the causal effect of X on Y. #Bookofwhy

2.27.19 @2:05am - (Replying to @stephensenn @deaneckles) True, I did not know Nelder. But I looked up the paper you cited and I did not see any potential outcomes to define the question, nor DAG or PO to define the assumptions. Perhaps he invented his own causal language. If so, can you translate?

2.27.19 @1:43am - (Replying to @stephensenn @deaneckles) If Nelder defines his research problem and encodes his assumptions we can compare his solution to that of #Bookofwhy and decide if it is identical or elegant. But I doubt he can do that. Why? Because both problem and assumptions are causal, and Nelder did not speak causation.

2.27.19 @1:15am - (Replying to @kareem_carr) You seem to have forgotten an alternative C : Forget all this torment about defining interventions. Imagine water molecules rushing to create a tide. Do they worry about intervening with the moon's location? They just LISTEN to the gravitational pull and rush. Effect=listen+rush

2.26.19 @11:21pm - (Replying to @stephensenn @deaneckles) #Bookofwhy provides a solution to Lord's Paradox given any DAG, including ones where initial weight has an effect on "Hall". See examples here: . The trick is to 1. Define you question, 2. Encode your assumptions (DAG); the rest follows (+fun), why resist?

2.26.19 @10:26pm - (Replying to @deaneckles) There is no longer any disagreement. We agree to talk about the same DAG, authored by Miguel, with whatever humility and verisimilitude Miguel attributes to it. We see that DAG carrying a variable named "gender" and we agree that it has causal effects. No trepidations. #Bookofwhy

2.26.19 @3:22pm - (Replying to @thosjleeper) To resolve disagreements, we must address them one at a time. Disagreements about the structure of the DAGs are one kind, and those about whether a variable in the DAG has a causal effect are of different kind. We are dealing with the second kind. #Bookofwhy #causalinference

2.26.19 @2:42pm - We must stick to the rules. We do not talk about "your DAG"; only about "Miguel-authored DAG". If Miguel ever includes a variable "gender" or "blood pressure" in a DAG, then the causal effect of that variable is quantified and interpreted with no trepidation. #causalinference

2.26.19 @2:29pm - (Replying to @omaclaren @ildiazm and 2 others) That's the whole point. Because I know that you don't really believe "mud causes rain" I expect you to rebel against a formal theory that does not enable you to articulate this simple belief. And, as a statistician, guess what this deficient theory is #Bookofwhy.

2.26.19 @2:20pm - Good point.,Thanks. I meant to write MOOT point. But, as a result, manipulability should also become MUTE, namely, stricken from scientific discourse.

2.26.19 @2:09pm - (Replying to @kareem_carr @cshalizi) @kareem_carr, where is this quote from? It is full of things I cannot agree with. But it is all a MUTE question because (I hope) Miguel and I have just reached a consensus: If "gender" appears in your DAG, then it has a causal effect . see

2.26.19 @1:54pm - (1/2) Miguel, we have never been closer to agreement than this very moment --just replace "your DAG" with "Miguel's DAG". Do you agree that the causal effect of any variable in any "Miguel-authored DAG" can be quantified (if identified)?? If I hear a YES, we can jointly pronounce
2.26.19 @1:54pm - (2/2) the problem of manipulability a MUTE problem, to be stricken from the scientific literature, past and future. Agree? And note that explains why we need to resolve "blood pressure" first, before we go to "Obesity". So good to be in agreement #Bookofwhy

2.26.19 @1:27pm - (Replying to @ildiazm @_MiguelHernan @deaneckles) Ivan, thanks for defending commonsense. If it were not for you, no one would dare ask Miguel: Come on! You do not really believe Pearl would put Unicorns/Cinderellas in a DAG, do you? So what's the point of evading the issue as articulated in #Bookofwhy

2.26.19 @3:30am - (Replying to @GuillermoBurr) I am tempted to agree with your "three body" constellation, except that I consider theories to be part of the solution, instead of the problem. #Bookofwhy

2.24.19 @2:22pm - Readers might enjoy the way "The Seven Tools of Causal Inference" appears in the March issue of CACM ... and the accompanying video . Too bad the Ladder of Causation was not illustrated in artistic colors #Bookofwhy

2.22.19 @11:50pm - (1/3) Thanks for defending my reputation. I have been thinking about why I sound provocative to statistics-trained researchers doing #causalinference. I believe the reason is that I take the difference between the two domains to be a "clash of civilizations", not of cultures,
2.22.19 @11:50pm - (2/3) and definitely not of a meeting of two "approaches". It is like trying to explain the notion of "volume" to a two-dimensional creature. That creature will never forgive your provocative insistence on adding another dimension to a comfortable 2-dimensional world. #Bookofwhy
2.22.19 @11:50pm - (3/3) Even in 2019, Gelman "finds it baffling" that we take comfortable statistical problems and "wrap them in a causal structure" -- It is unforgivablly provocative. Yet you will never understand "volume" unless you internalize the provocative idea of escaping from a 2-dim. world

2.22.19 @3:24pm - I would be very weary of anyone who argues pro or con any approach who refuses to solve any toy problem in that approach. I spoke about the limitations of PO here: only after solving problems, side by side, in three different approaches. #Bookofwhy

2.22.19 @2:23pm - (1/2) Since #Bookofwhy is aimed not at the establishment but at free-thinking economists, what prevents these economists from ignoring the establishment and communicating the indispensibility of SCM to other enlightened economists? Fear of retribution? Lack of conviction?
2.22.19 @2:23pm - (2/2) As a student of Heckman, do you believe Jim now teaches his students some of the indispensible elements of #causalinference, eg (1) structural counterfactuals (2) ways of deciding NP-identification of causal parameters, (3) ways of deciding testability of causal assumptions ?

2.22.19 @12:32pm - (1/2) (Replying to @ghoshd) I fail to see why saying that "until the 1980s. The rest of statistics, ... remained in the Prohibition era," would "alienate" any group. If I were a statistician, I would rejoice hearing that something great happened to my field in the 1980's. As to Gelman's post
2.22.19 @12:32pm - (2/2) (Replying to @yudapearl @ghoshd) it is tainted with years of tormented attempts on my side to have Gelman solve ONE toy example before speaking about @causalinference in statistics. And I would appeal to your honest assessment: How many PhD statisticians do you think can solve such an example? #Bookofwhy

2.22.19 @12:11pm - (Replying to @hangingnoodles) I am wondering, are accusations of "alienating people" typical to scientific paradigm shifts? Thomas Kuhn does not elaborate on this aspect of shifts. Does anyone have more pointed references to historical parallels? #Bookofwhy

2.22.19 @11:51am - (Replying to @learnfromerror) I have hard time relating "sample & parameter" to the issue of knowledge organization. Curious; what do Bayesian statisticians say about the way humans organize probabilistic knowledge in their heads? #Bookofwhy

2.22.19 @11:40am - (Replying to @cmirzayi) I believe I substantiated these characterizations with hard facts. What do we call someone who excludes Israeli women from participating in a women's movement? Isn't "outright bigot" a mild characterization? There is a sharp difference between "disagreement" and moral deformity.

2.22.19 @11:25am - Very keen observation. The reason you have not seen it is because Bayesian statisticians do not feel comfortable making assertions about how people store probabilistic information; they leave it to psychologists. They forget that priors come from people not from tables.#Bookofwhy

2.22.19 @10:34am - (Replying to @_limbs_ @loudquack @shravanvasishth) Well put! Here I did not imply deliberately avoiding causality for some udlterior motives. I attributed the missed opportunity to (1) their justified excitement over the discovery of correlation and (2) lacking a language to go beyond it. Deliberate avoidance was practiced later.

2.21.19 @4:49am - (1/2) Two comments on this thread. First, Pearl does not "go after people". He goes after outdated ideas and, occasionally, people invested in those ideas feel threatened. Second, while it is true that the threatened tend to cling to their street lamps and totally ignore advances
2.21.19 @4:49am - (2/2) in #causalinference, I think you are underestimating the curiosity and critical thinking of their students, and their ability to read the forbidden literature and to understand the limitations of what they are being taught. I count on them. #Bookofwhy.

2.20.19 @10:17pm - (Replying to @Andrew___Baker) No problem. But if not too much trouble I would still like to read at least one (accessible) article where a legal decision hangs on which of the two causal type we weigh more heavily. It should be illuminating for #Bookofwhy 2nd Ed.

2.20.19 @8:39pm - (Replying to @jim_savage_ @Andrew___Baker) I dont get it. (Born naive). Can you be more specific? What is it about "nobody has had a thought before".? #Bookofwhy

2.20.19 @7:29pm - (Replying to @Andrew___Baker) Blatantly wrong? #Bookofwhy it does not claim philosophers did not distinguish necessary from sufficient causes, it claims that "what weight" to assign to each component was left open. Am I wrong? Any counter-citation? Can we discuss weights w/o having quantities to weigh?

2.20.19 @10:33am - (Replying to @albertocairo @mendel_random) Agree with @albertocairo because, to me (as a computer scientist), the litmus test of "taking them seriously" is commitment to NOTATION. As #Bookofwhy argues, it was only in 1920 that slogans were broken with new notation for causal relations, escaping the dryness of tables.

2.20.19 @1:43am - (Replying to @MHC_UNC @lsarsour and 2 others) It is sad for me to see an esteemed university and a reputable School of Public Health fall for the charms of an outright bigot. ... @UCCpublichealth

2.20.19 @12:35am - (Replying to @maqartan @aecoppock and 4 others) But is it true that at the end of the analysis all we can get is the ATE in a sub-population that is both unidentified and uninterpretable. I am not minimizing the importance of the analysis, I am just checking if I understood its limitations. #Bookofwhy. #causalinference

2.19.19 @12:27pm - (Replying to @aecoppock @SonjaASwanson and 4 others) Can't we just ignore the presence of U1 ? According to the classical definition (eg Causality fig. 7.8(b)) Z remains a perfect IV with or without U1. #Bookofwhy

2.19.19 @2:12am - Just received a Kindle version of . Eager to see if my esteemed colleagues contributed earth-shaking ideas to the discussions we have had here on Twitter. #Bookofwhy

2.18.19 @7:15pm - PRIMER is full of DAG puzzles designed carefully to build skill and intuition (supported by Daggity.) See sample solution manual: . (Available upon request.). My biased opinion: PRIMER is THE BEST introduction to CI. #Bookofwhy

2.18.19 @4:28pm - Statisticians themselves recognized that they can't be true to their own slogans and the "no opinion" exclusion was officially struck out from the Royal Statistical Society manifesto in 1858. I wish we had a record of the debate leading to that decision. #Bookofwhy

2.17.19 @8:40pm - Readers and passengers who happen to be in LA on Thursday 2/21 are invited to attend this lecture by Stuart Russell, entitled: Human-compatible Artificial intelligence #Bookofwhy

2.17.19 @8:03pm - (Replying to @mayfer) We saw a microsoft resource tweeted around, but usually the direction goes the other way: neural network modules are assisting a causal inference engine. #Bookofwhy

2.17.19 @7:47pm - (Replying to @zarzuelazen @Grady_Booch) I wish I could resonate with your architecture, but seeing "Information Theory" next to "Action" my antenna rises: Information theory is purely probabilistic, how can it govern Action ? It does not sit. Note: #Bookofwhy does not mention entropy, or Shannon's mutual information.

2.17.19 @2:42pm - (1/2) Toy examples are educational devices absent of which a researcher cannot assess the limitations of any given design, be it observational or experimental. It is necessary therefore for understanding the limitations of RCT, whether they can be cured or replaced by obs. studies.
2.17.19 @2:42pm - (2/2) Here is an example of how researchers who have not mastered toy examples missed an adequate assessment of how RCT limitations can be cured: . I am speaking of limitations such as selection bias and changing target populations. #Bookofwhy

2.17.19 @2:25pm - (Replying to @albertocairo) Glad they confirm the observation of #Bookofwhy, and I love "The dryer the better" even more than "exclude all opinion" (ie. no priors) which I found in the inauguration manifesto of the Stat. Royal Society.

2.17.19 @6:27am - (Replying to @thosjleeper) Agree!! But I see the simplicity of economists models a symptom of their inability to think about more complicated ones. Solving ONE toy problem will open their eyes to what they can do, especially in generalization and selection bias tasks, where they are so painfully behind.

2.17.19 @5:40am - (Replying to @thosjleeper) Disagree!! "It is only by taking models seriously that we learn when they are not needed or not useful." (Must be Aristotle, 384-322 BC)

2.17.19 @5:29am - (Replying to @sidinusofaiii @intensivemargin and 4 others) speaks 3 languages. Also, some folks are working on a Glossary of terms, to improve the translation. But they need your help. Which concept, tool, or assumption from #Bookofwhy do you find needing a translation to your vocabulary of choice? Please help.

2.17.19 @5:00am - (1/2) (Replying to @thosjleeper) I am proposing toy examples as educational, not persuasive device. Folks who need tables of numbers to be persuaded about methods can easily find them in the epidemiological literature and abstract back to economic problems -- the methodology is identical. Moreover
2.17.19 @5:10am - (2/2) (Replying to @yudapearl @thosjleeper) Those who have difficulties setting up the DAG for their problems will not benefit from seeing how a problem was formalized in a totally different application. Toy problems can help those folks by highlighting the commonality of structure. No substitute for toys. #Bookofwhy

2.17.19 @1:47pm - (1/2) @PHuenermund was kind enough to remind us of his "not-invented-here" article ...where he also provides the quote from "Mostly Harmless Econometrics" which you cited. As you can see, a model with 4 variable requires @metrics52 to spend pages of informal
2.17.19 @1:47pm - (2/2) arguments. Imagine what an economist would go through in models of 10-20 variables. (eg. fig.2 of ). It is cognitively intractable and should be deemed impossible. This, I know, will not convince the "not-invented-here" folks, but outdatedness will.

2.16.19 @10:10pm - (1/3) You ask if there is a "shorter" #Bookofwhy, and I assume you want to get the technical meat w/o reading the stories. Yes, there is. If you take a look at Section 2 of , you will find the whole book summarized in 3 pages. But it must be supplemented
2.16.19 @10:10pm - (2/3) with the toy problems of Section 3. No matter how many books one reads ABOUT economics, shying away from solving toy problems would leave one where econometrics is today -- two decades behind the time. Plus, it is fun to see important methodological problems escaping their
2.16.19 @10:10pm - (3/3) textbook handcuffs and rejoicing game-like solutions. I therefore recommend: do not skip the toy problems in and their solutions. Try one - its better than reading a whole book. Among the easy ones: Can your research question be answered using OLS?

2.16.19 @8:07pm - (Replying to @jehosafet @CarterPaddy) As #Bookofwhy narrates, epidemiology was fortunate to have the enlightened leadership of Greenland and Robins (see ). Econometrics is still begging for such leadership, and the gap gets wide in time. It's strange how much leadership counts.

2.16.19 @7:32pm - (Replying to @sidinusofaiii @intensivemargin and 4 others) ALPHA is any parameter in your economic model. Most economists have trouble swallowing that elementary questions in Econ 101 (eg. about ALPHA) require tools developed elsewhere; they see it as an insult to the profession, so they hope @metrics52 has it. Does he? Please check.

2.16.19 @6:10pm - (Replying to @mkessler_DC @CarterPaddy @t0nyyates) I think it should not take longer that a 10-item glossary to do the translation. Let me start: Exogeneity = Nonconfoundedness, iv = iv, Control for = condition on, Error term = Unobserved factor..Care to continue?

2.16.19 @5:24pm - (Replying to @intensivemargin @CarterPaddy and 2 others) That question of estimating ALPHA is identical to the issue of "bad control" being addressed in MHE. If you can distinguish "bad control" from "good control" you also know whether ALPHA is estimable by OLS. Thanks for bringing up the relation between these two issues. #Bookofwhy

2.16.19 @5:02pm - (Replying to @intensivemargin @CarterPaddy and 2 others) I don't have a ready copy of "Mostly Harmless Econometrics", but I am nevertheless willing to bet that the authors cannot answer the question: "Given a general economic model, can parameter ALPHA be estimated by OLS?" How come? Because it is super-human to answer it w/o graphs.

2.16.19 @4:32pm - (1/3) Your question: "why I wasn't taught the graphical approach" was raised by many economists on this Twitter, and I have partially answered it in and . Without going too deep into Psychology, the answer is "Not home grown!". (cont.
2.16.19 @4:32pm - (2/3) Your second question: "Would I be [taught it] today" is tricky. @Susan_Athey says there is no need, because economists already have the answers. (see ...According to others (eg @marcfbellemare, @PHuenermund @causalinf), economists are beginning to rebel
2.16.19 @4:32pm - (3/3) against the tyranny of outdatedness. This workshop: will provide an opportunity for both rebels and conformists to present their cases before Clio, the Muse of history. #Bookofwhy.

2.16.19 @1:13am - (1/2) The DAG can only give us a partial preference-order on the functionals, as discussed in . In however you can see a separate analysis, which declares estimator # 2 superior to the other two. Interestingly, although functional #3
2.16.19 @1:13am - (2/2) is identical to # 1 (for all distributions) I am not sure the corresponding estimators are equally powerful. Can the theory of influence functions shed light on these questions? @autoregress @LauraBBalzer @edwardkennedy #Bookofwhy

2.15.19 @11:56pm - (Replying to @autoregress)
I corrected it. It should read:
1. E[Y|x]
2. SUM_z E[Y|z] P(z|x)
3. SUM_z E[Y|x,z] P(z|x)
And the answer is: ??????? See

2.15.19 @1:23am - Correction: the three estimands are:
1. E[Y|x]
2. SUM_z E[Y|z] P(z|x)
3. SUM_z E[Y|x,z] P(z|x)
Which is most powerful?

2.15.19 @12:24am - (1/2) Here is a simple puzzle for estimation experts: Consider the chain model X--->Z-->Y. Below are three valid estimands of E[Y|do(x)]:
1. E[Y|x]
2. SUM_z E[Y|x,z] P(z)
3. SUM_z E[Y|x,z] P(z|x)
Each estimand defines a consistent estimator if we decide to take the MLE
2.15.19 @12:24am - (2/2) of each factor and combine them by the formula. Question: Which estimator is the most powerful? (1), (2) or (3) ?? Once we agree on the correct answer, we will ask whether TMLE can help us decide correctly, and how. #Bookofwhy

2.15.19 @12:12am - (1/2) I am glad to see this paper getting closer to publication. It was written in response to a flood of Epi papers on generalizing study results, all from the potential outcome perspective, suffering of course from the basic limitations of that perspecive (see page 1).
2.15.19 @12:12am - (2/2) I hope this paper opens new vistas for Epi researchers seeking to generalize study results. #Bookofwhy

2.14.19 @1:32am - The sentence that gave me the chuckle is "Economists were responsible for asking and answering all of these [causal] questions." It is 2019, and I can hardly name a handful (<6) of economists who can answer even one causal question posed in #Bookofwhy

2.14.19 @4:12pm - (Replying to @charleswangb) It is because A B E F and G are the parents of X, so they block all backdoor paths from X to Y. #Bookofwhy

2.14.19 @5:48am - (Replying to @ildiazm @CalebMiles16) Agree. But where/when is the decision made that one estimand is better than the other? Does TMLE examine all estimands? Before looking at the data, or after? Is there an algorithm to extract optimal estimands from a given DAG? #Bookofwhy

2.14.19 @12:48am - (Replying to @CalebMiles16 @ildiazm) I do not believe the information provided by the mediator Z is a convincing explanation as to why the product of the estimators is better than the one-shot estimator. If we look at the saturated case (Appendix II ), observing Z does not help. #Bookofwhy

2.14.19 @12:16am - (Replying to @LauraBBalzer) I thought that once the TMLE obtains an estimand, she never looks back at the DAG, she simply takes the estimand and finds the best estimator for it, given the data. If so, what would tell it that the product of the estimates is better than the estimate of the product? #Bookofwhy

2.13.19 @2:06pm - (1/2) It was an inspiration to hear Laura speak at UCLA, and see TMLE dissected to its logical roots. But nights bring second thoughts: Is it always true that optimizing overall estimand is better than optimizing its individual components? Consider the chain X-->Z-->Y, where
2.13.19 @2:06pm - (2/2) the regression coefficients obey: R(YX) = R(YZ)*R(ZX). Guess which estimator is better? The MLE of R(YX) or the product of the MLE of R(YZ) times the MLE of R(ZX)? Surprise! It is the latter! In blunt defiance of what TMLE would suggest. See #Bookofwhy

2.12.19 @11:52pm - I plan to stick around for another 2-3 decades, so just tweet if you want to dissolve any disagreement. Lucky for us, modern #causalinference leaves no room for lingering disagreements, they can now be examined under the microscope and breed poetry. #Bookofwhy #epitwitter

2.12.19 @9:32pm - (Replying to @abrahamnunes @GunnarBlohm and 4 others) I have not read "anticipatory systems", but if you can summarize "what a model is" in Twitter's length I will try to comment on whether it matches my understanding of what a model is.

2.12.19 @8:48pm - (Replying to @jamessseattle @DavidDeutschOxf) I am not familiar with Constructor Theory, and the little that I read about tells me that it does not provide explanation that we, mortals, would call "explanations".

2.12.19 @6:05am - (1/3) Dear readers and followers, I noticed that our Twitter audience has swelled to over 15K, which prompts me to thank you for energizing me with your comments and questions and for giving me the illusion of being somewhat useful in this game of cause and effect.
2.12.19 @6:05am - (2/3) I have tried to keep my Tweeting focused on the science, minimizing foods, cute pets and funny cartoons, yet entertaining enough to have you try causation's new tools that do what we used to think we can't. I hope I can continue this experiment and make
2.12.19 @6:05am - (3/3) and make Tweeting exchange a valuable learning experience. As they say in the Mishna (220 AD, anticipating Twitter): "Speak briefly and act fully" [In Hebrew: "Emor Meat Veaseh Harbe"] #Bookofwhy

2.11.19 @11:23pm - I am surprised that 51% thinks RCT has anything to do with the DEFINITION of causal effect. Chapter 4 of #Bookofwhy labors to convince reader that it is just a means for "interrogating nature", that is, nature has the answer before the interrogation. Plenty of work for educators.

2.11.19 @10:47pm - Stefan Conrady, Managing Partner of Bayesia, was kind enough to send us an interesting selfie he took with the Lion Man that is featured in Chapter 1 of #Bookofwhy . Here it is

2.11.19 @7:46pm - (Replying to @Jabaluck @PHuenermund) I am also speaking about combining observational methods and experiments in the same population, yet I find it hard to read Angrist's paper, because I cant quickly see the principle: What is taken from each study? and how is it combined? Principles suffer when models are avoided.

2.11.19 @6:05pm - (Replying to @Jabaluck @PHuenermund) If you are working on a fusion paper then I am sure you will find the do-calculus to be helpful if not indispensible, as in and ... . This is something economists cannot use, given their model-avoiding mindset. #Bookofwhy

2.11.19 @1:17pm - (Replying to @Jabaluck @PHuenermund) This is indeed the key question. And the answer depends on whether the RCT was conducted on 12K or 12 subjects, whether they were diabetic with likely allergy to the drug, etc. etc. Some of these issues are discussed here: #Bookofwhy

2.11.19 @1:07pm - (Replying to @charleswangb) No relation between the three layers. The assumptions could be interventional or counterfactuals. Same with the Query. #Bookofwhy

2.11.19 @5:50am - (Replying to @NikHarmon @PHuenermund) On the other hand, there are (infinitely) many unobserved Zs that can spoil the exogeneity of your IV, and infinitely many unobserved Ws that can spoil the exclusion of your IV. Thus, we need to see the model before deciding on identification strategy #Bookofwhy

2.11.19 @2:46am - While I was bugged down in debates on "obesity" and "design", my faithful Google Alert noticed that the field of #causalinference has advanced forward by this comprehensive paper on generalized identification: ...It's worthy of close attention #Bookofwhy

2.11.19 @12:40am - (Replying to @TheLeanAcademic) It is the first time I hear about it. Who would be a typical beneficiary of the list?

2.10.19 @5:52pm - (Replying to @KordingLab @djinnome and 2 others) Model blind is not impossible if you assume auxiliary conditions (eg ignorability) that turn your statistical routines into causal estimators. If pressed hard, one can label these assumptions "a model", but then the causal story is gone, with all the fun of "seeing" it working.

2.10.19 @4:45pm - (Replying to @KordingLab @djinnome and 2 others) #Bookofwhy is about causal MODELING. The course, as I read it, is about MODEL-BLIND methods. There is some overlap there, once we extract from the model all the information it can provide (e.g. estimability, testability) and we no longer need it. What remains is statistics.

2.10.19 @4:15pm - (Replying to @robertwplatt) If you "get it" then you do not need my help. The distinction between "analysis' and "design" follows logically from their respective characters. But I am open to learn what language a person who "gets it" should choose to describe mathematical derivation of causal quantities.

2.10.19 @1:23pm - (Replying to @KordingLab @djinnome and 2 others) @KordingLab, I am unable to recommend this course, except to seasoned #Bookofwhy students who wish to see with their own eyes what an "alternative approach" looks like.

2.10.19 @12:50pm - (1/2) Agree. Look how careful and accommodating I am speaking to statisticians or publishing in http://Stat.Sc . But there are other folks there in #causalinference, e.g., ML folks, philosophers, and students yet unharmed by traditional education.
2.10.19 @12:50pm - (2/2) These folks will take it as pathology if I defy logical standards and name a simple "mathematical derivation" as "design". But, by all means, if you have conduits to traditional "designers", use their language to show them that "design" is no longer "thinking hard".#Bookofwhy

2.10.19 @6:14am - (Replying to @thosjleeper) As an enlightened reader of both Causality and #Bookofwhy, would you respect me if I play diplomacy and call formal mathematical derivations "DESIGN", just because many people were educated in an era when these derivations had to be done by informal "thinking hard". ???

2.10.19 @6:00am - (1/3) Causally speaking, it is the other way around. Those who "get it" cannot bring themselves to call mathematical analysis "design" or "thinking hard". Moreover, speaking of "implementation of these methods" sounds like there are "alternative methods". I do not see any; (cont.
2.10.19 @6:00am - (2/3) The choice is between doing thing systematically or retreating to the days of "thinking hard", namely, between survival and staying behind. At the same time, if you and others think that mild language would create more survivors, by all means, you have my blessing. But
2.10.19 @6:00am - (3/3) please do not ask me to call formal mathematical derivations "DESIGN", just because many people were educated in an era when these derivations had to be done by informal "thinking hard". Readers of #Bookofwhy will not forgive me, because they do "get it".

2.9.19 @10:36pm - (Replying to @thosjleeper) I just answered Robert Platt on this very question: No, The only researchers who truly get it are those who are shaken by shifting vocabulary to realize that "the times they are a-changin" and, to survive, they must re-think old molds. Diplomacy dupes them them into slumber .

2.9.19 @10:25pm - (1/2) The distinction between "statistical analysis" and "causal analysis" is made visibly and consistently in all the #causalinference literature. The little confusion that remains dwarfs in comparison to the benefit of shaking the old "designers" to understand that much of their
2.9.19 @10:25pm - (2/2) torment and "hard thinking" can be accomplished today by "analytical methods" The only researchers who truly get it are those who are shaken by shifting vocabulary to realize that "the times they are a-chanin" and, to survive, they must re-think old molds. It's #Thebookofwhy

2.9.19 @7:52pm - (1/2) In the old days, estimation was the only task that folks could submit to formal treatment, so they called it "analysis". All the rest was entrusted to "hard thinking", so they called it "design". Today, many of the latter tasks can be treated analytically and should (cont.
2.9.19 @7:52pm - (2/2) therefore be called "analysis". But traditional "designers" refuse to hear the bells of change and prefer to retain their posture and terminology. Awaken them, "for the times they are a-changin'" #Bookofwhy

2.9.19 @7:26pm - (Replying to @MariaGlymour) I am yielding to your epi experience if you feel that way. But before quitting I would just leave you with an open question: Are you sure the clarity you obtain from thinking RCT is not a placebo effect induced by our statistically dominated education? #Bookofwhy

2.9.19 @6:01pm - (Replying to @MariaGlymour) I get clarity and specificity by thinking about the actual policy that is about to be implemented Monday, by so and so, who is about to issue an instruction to so and so. Why should commonsense thinking need the help of hypothetical trials, distorted by randomization? #Bookofwhy

2.9.19 @4:26pm - (Replying to @MariaGlymour) I dont see the relation between specificity of policy and RCT. There is no "randomization" in my thinking about minimum wage policy, not even "trial", there is only implementation details and some knowledge of the forces involved. RCT is one way of testing, not a way of thinking.

2.9.19 @3:53pm - (Replying to @MariaGlymour) But what if I do not have any RCT in mind, not even hypothetical. All I have is a burning desire to predict how a 1$ increase in minimum wage would affect unemloyment, if enacted tomorrow. Must we think RCT when we can think what variables would change, how and why? #Bookofwhy

2.9.19 @3:07pm - (1/n) Of course I am not using "analysis" the way applied social scientists do. Recall, the boundaries between "design" and "analysis" were drawn before people realized their research question can be written down, submitted to analysis, and receive an answer. (cont.)
2.9.19 @3:07pm - (2/n) My definition of "analysis" is simple: Everything that the inference engine can conclude without human input is "analysis". Any input needed from the researcher is "design". See #Bookofwhy . True, most applied social scientists are not aware that inference engines exists.
2.9.19 @3:07pm - (3/n) Some even see the engine as a threat, namely, not in their textbooks. But living in an exciting period of paradigm shifts entails redrawing boundaries, redefining outdated concepts and re-formulated the fundamental question "what is the study goal?" I am excited! #Bookofwhy

2.9.19 @11:49am - (Replying to @pophealth3) The "graphs are bad" argument (and culture) explains why Rubin had no model, and had to resort instead to a "think hard" recommendation. Unaware of this misguidance, some people still think that "mostly harmless econometrics" is not harmful. #Bookofwhy".

2.9.19 @11:22am - Analysis is what tells you what confounders are relevant. Lacking tools for this analysis ("graphs are bad") Rubin recommended "think hard". Analysis means "Stop thinking so hard", let the mathematics think for you. U "think hard" only once, when you commit to a model. #Bookofwhy

2.9.19 @4:18am - (Replying to @thosjleeper) By "imitation" I mean e.g. watching for placebo effects, or selection bias. Why do we take those precautions? To create conditions which are similar to those created by the do-operator -- the definer of "causal effects". See #Bookofwhy pp. 143-9

2.9.19 @2:44am - (1/4) Glad we agree on content. But if I recall correctly you said (on Target Trials) "this is a design issue, not design" So what in your opinion is the dividing line between design and analysis? My dividing line: whatever the inference engine can infer mechanically is "analysis"
2.9.19 @2:44am - (2/4) And whatever requires human judgment (on top of the inputs to the engine) is "design". I would be curious to know if you agree. This means that when AI succeeds in producing an "automatic scientist", everything will be "analysis". An interesting anecdote, I believe Don Rubin
2.9.19 @2:44am - (3/4) was first to pit "design" against "analysis" in a 2008 paper "FOR OBJECTIVE CAUSAL INFERENCE, DESIGN TRUMPS ANALYSIS." But the number of mistakes in that paper does not speak much for "design". That paper also sounds like the scientific manifesto of "Target Trials" #Bookofwhy
2.9.19 @2:44am - (4/4) I think the idea that an observational study should imitate some RCT is totally misguided. It is RCT that should imitate things, not an observational study conducted in the natural habitat of a population, where causal effects are derived properties of the model. #Bookofwhy

2.9.19 @1:53am - (Replying to @TunnelOfFire) If all variables are observed then the causal effect (of X) is identified by the adjustment formula, with the parents of X as the "controlled for" covariates. See #Bookofwhy p.220, or Causality, Eq. (3.13)

2.9.19 @12:00am - (Replying to @austinvhuang) I am not after winning debates. I am here to show researchers that "they need not distinguish manipulable from non-manipulable variables; both types are equally eligible to receive the do-operator and both benefit us by evaluating its effects." #Bookofwhy

2.8.19 @10:10pm - (1/3) Yes, I have done the "science thing" and have shown here that E[Y|do(X=x)] is empirically testable even when X is non-manipulable, which was one of your requirement to be "part of science". This still far from computing the effect of Obesity, but
2.8.19 @10:10pm - (2/3) as we know, science progresses incrementally. Obesity presents TWO obstacles, non-manipulability and lack of agreeable measurement. Progress demands that we tackled one obstacle at a time, so I took on blood-pressure which is well-measured but non-manipulable. I am satisfied
2.8.19 @10:10pm - (3/3) with the way the non-manipulability obstacle was removed -- are you? If not, then I would really like to know how your team treats such variables when they show up in the dag. No trap here, just innocent curiosity. #Bookofwhy

2.8.19 @8:40pm - The arguments on "Well-definedness" , pro and con, are laid out here and here . The latter contains a subsection on "off policy" actions (in RL) and how they can be evaluated using "in policy" actions, once we have a causal model.

2.8.19 @4:35am - (Replying to @BL4PublicHealth @mathtick) Let's take the question "who r u going 2 recruit?". How do you make this decision? Dont you take into account what question you need to answer, whether the character of the recruitees impedes or assists your ability to answer that question? etc., This is what "analysis" is about.

2.8.19 @4:22am - (Replying to @robertwplatt) I suggest that having this notation, and having the machinery to infer the answer to our research question from this notation is "analysis". Moreover, design options that matters (eg what to measure) can be decided mechanically using this notation, and knowing the question.

2.8.19 @2:29am - (Replying to @thosjleeper) Please examine the "design choices" you have and ask if you can make them by whims, or you need to think twice before making them. If the latter, who is guiding you in "thinking twice"? What knowledge do you consult in this thinking? Can't this thinking be replaced by analysis?

2.8.19 @1:56am - (Replying to @mathtick) It all becomes clear when we focus on a concrete example. Lets say we have a "design choice" what measurement to take: Z or W, or both. The choice is not arbitrary, it takes analysis to tell us if we can get by with Z only, W only or we need both to answer our research question.

2.8.19 @1:36am - (Replying to @stephensenn) Its OK to tell your students that such a notation was provided by Adam and Eve, as long as you can also teach them how to combine observational and experimental studies to get results that none can provide in isolation. See or #Bookofwhy page 357

2.8.19 @1:26am - Moreover, I cannot think of any "design choice" that does not require a close guidance of analysis. I question, therefore, whether the word "design" still conveys respectability or, rather, escape from responsibility. #Bookofwhy

2.8.19 @1:05am - (Replying to @Profinit_EU) I love this robot. Amazing how asking "why" can make you look smart. #Bookofwhy

2.8.19 @10:48pm - Today we have mathematical notation to describe how data are generated, how treatments are assigned and how subjects are recruited. See . Armed with this notation, do we still need to talk about "design" as competitor to "analysis"? #Bookofwhy

2.7.19 @9:52pm - (Replying to @iwashyna @eliasbareinboim and 6 others) The English language offers us many ways of expressing Y_x informally, so as to engage informal audience. For example, using counterfactuals: "what if X were x," or simply "how X affects Y". These linguistic utterances precede the RCT metaphor by 2000 years, see #Bookofwhy

2.7.19 @9:43pm - (Replying to @eddericu @eliasbareinboim and 6 others) Speaking of "well-definedness", I am waiting to hear from @_MiguelHernan how they handle a measurement of blood-pressure. Do they delete it from the model? or leave it in the DAG and mark it as "dangerously non-manipulable"? If so, what do they do with the markings? #Bookofwhy

2.7.19 @6:43pm - Related quotation: "If our conception of causal effects had anything to do with randomized experiments, the latter would have been invented 500 years before Fisher." --- #Bookofwhy (page 135)

2.7.19 @9:36pm - (Replying to @eddericu @eliasbareinboim and 6 others) No way!. It is "covariate-specific effect" because the authors keep on talking about the intended population (or a person) for which the study result will eventually apply. "Well definedness" is a red herring as shown here and here

2.7.19 @6:11pm - (1/2) The danger of introducing Latin terminology without warning, is that there are dozens of students anxious to tackle new problems with causality's new tools. They are wasting hundreds of hours only to discover that "Target Trials" is just Latin for "causal effect". #Bookofwhy
2.7.19 @6:11pm - I would also question the premise that researchers speak RCT more fluently than they speak causal effects. Most mortals are born with causal intuition, we then indoctrinate them to speak only RCT and, now, we justify Latin as if they were native RCT speakers. Back to #Bookofwhy

2.7.19 @4:40pm - (Replying to @EpiEllie @eliasbareinboim and 6 others) I am not familiar with any causal effect or estimand of causal effect that is not "ideal" in the same sense, namely, it is what RCT is trying to emulate. So, I do not understand the point of introducing a new adjective into the crowded glossary of causal vocabulary. #Bookofwhy

2.7.19 @3:45pm - (Replying to @robertwplatt @EpiEllie and 4 others) Suppose I never heard about the word "design" and, naive me, all I know is how to estimate the causal effects of treatment X on a given subpopulation from a combination of experimental and observational data on other subpopulations. What do I gain by singing "design"? #Bookofwhy

2.7.19 @2:54pm - (Replying to @eddericu @iwashyna and 5 others) Delving more and more does not tell us whether the research question answered by "Target Trials" is the same as the one answered by "covariate-specific effect". Beside, why should we care about what RCT emulates if we can answer directly the research question at hand? #Bookofwhy

2.7.19 @2:33pm - (Replying to @robertwplatt @EpiEllie and 4 others) The word "design" is sometimes used to avoid analysis, I hope this is not the aim of #TargetTrials. For example, the question: "What is it appropriate to compare that person to?" is answerable by analysis, once you explicate what the research question is. What is it? #Bookofwhy

2.7.19 @2:50am - (Replying to @HenningStrandin) But my dissatisfaction with the causation-avoiding statistics is aimed primarily at the inadequacy of their language (ie probability), hence their methods. What is it about trad. empiricism that you dislike? #Bookofwhy

2.7.19 @2:44am - I second @analisereal idea. We have proposed a summer school in causal inference to IPAM (at UCLA). So, the more evidence of interest we get from readers the greater our chance to have it approved by their (rather conservative) board. #Bookofwhy

2.6.19 @11:30pm - (Replying to @edwardhkennedy @EpiEllie) I wonder if this quote should still holds in the era of causation. Is it the case that statistics has studied the way scientists think about and process scientific data? Something to ponder about, but I can feel the doubts crawling all over me. #Bookofwhy

2.6.19 @11:22pm - (Replying to @iwashyna @EpiEllie and 3 others) Fascinated by the term "#TargetTrials I asked if it can be translated into a problem familiar to my students. I believe the answer is: Yes,it is the problem of evaluating "covariate-specific effect" as in , p. 70 (w identification conditions)No? #Bookofwhy

2.6.19 @2:49pm - (Replying to @PWGTennant @babylonhealth) That's an interesting scenario. And while the #ML and #DL and #RL folks are busy assuring us that: "Yes, we are all in favor of #causalinference", the business community will outdate them by just doing it. Interesting. #Bookofwhy

2.6.19 @8:15am - (Replying to @PHuenermund) Well put. The puzzle is not so much "Why do economists not trust ..." but "Why do economists authors avoid the question you raised"? Embarrassment? Lack of answer?Fear of exposure? Love of darkness? #Bookofwhy

2.6.19 @6:24am - (Replying to @KordingLab @RivasElenaRivas @manuelbaltieri) If you organize such a workshop in LA, I will be happy to participate.

2.6.19 @4:51am - (Replying to @manuelbaltieri @KordingLab) There are several advertised. But I would watch for impostors. My litmus test: Can the instructor solve a toy problem in causal inference? For example, the problems illustrated in #Bookofwhy

2.5.19 @6:10pm - For ML folks who asked: "what's in it for me?', the paper on non-manipulable variables, , is now updated with a sub-section on untried actions in reinforcement learning (RL) #Bookofwhy #causalinference

2.5.19 @2:35am - (1/2) Thanks for a beautiful summary of the two perspectives. It re-ignites an ancient question I once asked: Do the Harvard Epi people really distinguish manipulable from non-manipulable variables in practice? How? What do the do with a variable like blood-pressure that happens
2.5.19 @2:35am - (2/2) to appear (as obs. covariate) in their DAG? Do they delete it entirely? Leave it in the DAG and mark it as "dangerous"? What do they do with the markings? Do they play with the incoming and outgoing arrows? Anyone knows? #Bookofwhy #causalinference #Epitweeter @epiEllie

2.4.19 @3:09am - (Replying to @NeuroStats) I am not aware of a different word, though "leveraging prior knowledge" is implicitly assumed in all philosophical accounts of causation. In Lewis account of counterfactuals, for example, one has to specify prior knowledge about which worlds are closest to ours. #Bookofwhy

2.3.19 @7:47pm - Seriously, @learnfromerror , can you share ONE ingredient that you think my characterization of Bayesianism is lacking. Hopefully it can be described in Tweeter grammar, same as I tried to do, without going through volumes of intricate literature. #Bookofwhy

2.3.19 @7:20pm - (Replying to @cubic_logic @learnfromerror) It is always desirable to run more stringent test to sharpen our knowledge. But what do we do if we need to bet before sharpening? Do we have a coherent methodology to do so? Bayes said: YES. Cast you suspicions in priors and treat those priors as if they were proportions.

2.3.19 @4:18pm - (Replying to @FamedCelebrity) If Bayes formula "isn't per se needed" can you mention an alternative? In my experience of the 1980's all those who offered alternatives (eg. Dempster-Shaffer, Fuzzy logicists) were careful (and proud) to proclaim themselves "non-Bayesians". @Bookofwhy

2.3.19 @4:10pm - Good question, but "use prior information" is not an afterthought; it is an empirically testable statement: Would our frequentist bet differently knowing that the coin was obtained from the neighborhood grocery as opposed to a shady gambler? See #Bookofwhy page 90

2.3.19 @3:45pm - I beg to differ. "Must be assigned a probability" is vacuous unless it is a commitment to obey the laws of probability after the assignment. Bayes' innovation was to translate "given that we know X" into "conditioned on X" as well as all the laws of conditioning #Bookofwhy p. 102

2.3.19 @3:31pm - This is seriously all that I meant, with all the epistemological commitments implied by the statement "assign prior probability" (#Bookofwhy bottom of p.102). But from your question I gather that you require additional commitments from a red-blooded Bayesian. Can you share ONE?

2.3.19 @12:35pm - (Replying to @amine_ouazad @shell_ki) Right! "Ignorability" is not a statistical concept despite the fact that it is dressed in probability outfits and sings the songs of conditional independence. The causal character of this concept is betrayed by the counterfactual variable Y(1), which is not an observed variable.

2.3.19 @11:57am - (Replying to @learnfromerror) Leveraging prior knowledge is the essence of Bayesianism. When appropriate, that knowledge can be encoded through conditional probability, or priors on parameters, and when it is not appropriate it should be encoded differently. I'm eager: What is your definition of Bayesianism?

2.3.19 @11:44am - (Replying to @alexpghayes @f2harrell and 3 others) Causality without probability is formulated nicely in structural equations, prior to assigning probabilities to the U terms, also called "error terms" or "units". See the firing squad example in #Bookofwhy (chapter 1) or

2.3.19 @11:03am - Nothing makes me happier than to hear a leading statistician say: "convincing". Nowadays I hear it only from PhD students and post docs, which is OK for the future of statistics but leaves me with the feeling that I could have explained things better. Perhaps I could. Thanks.

2.2.19 @5:20pm - (Replying to @learnfromerror @elementary_peng @richardtomsett) At leaset people who eat "hot dogs" agree on what "dogs" are. Do Bayesians agree on "who is Bayesian"? I gave a humorous definition: "assigning priors to parameters." Do we have a more technical definition that would satisfy the high priests? #Bookofwhy

2.2.19 @1:29pm - A friend sent me a mini-review of #Bookofwhy which appeared on The Verge's "AI Reading List" e-reading-list-books-scifi . Written by Rumman Chaudhury, it is a most perceptive and accurate description of what the book is trying to say, in the context of current discussions of AI.

2.2.19 @12:49pm - (Replying to @AdrianBauman @JamesSteeleII and 5 others) Having seen painful problems induced by vague terms such as "interrelatedness" I am now more cautious, to the point that my mind demands to know what makes direct relatedness different from indirect relatedness. Surely your algorithm treats them differently, should it?

2.2.19 @12:30pm - (Replying to @HenningStrandin @_MiguelHernan) If you include me among "people in causal inference" then I would differ. My claims do concern objective properties of causation, as encoded in the "listening" metaphor. Causation is the algorithm by which Nature assigns values to variables.

2.2.19 @11:36am - (Replying to @JamesSteeleII @SamueleMarcora and 5 others) What I would truly prefer is an explication of what the arrows (or edges) mean. The caption says "drive activity", which sounds causal and directional. More important, what do we want to convey when we connect X to Y but NOT to Z, even though Y is connected to Z. #Bookofwhy

2.2.19 @12:54am - Sorry for the wrong url. My Toronto speech is linked here:

2.2.19 @12:03am - My college route was from Rutgers to Brooklyn Poly, but I only took it three times: 1.Written exam, 2.Oral exam. 3.Thesis defense. Classes were boring. I publically confessed, and apologized for this unruly behavior in a commencement speech at Toronto:

2.1.19 @10:50pm - (1/2) This passage demonstrates the monumental linguistic hurdle that statisticians had (and still have) in dealing with causation. Lindley was the only one to recognize the need for statistics to lift itself from the two-dimensional plane to a 3-dimensional view of the world....
2.1.19 @10:50pm - (2/2) Even today, few statisticians are able to acknowledge this barrier. Gelman, for example, writes: "I find it baffling that Pearl and his colleagues keep taking statistical problems and, to my mind, complicating them by wrapping them in a causal structure." And this is 2019!

2.1.19 @10:16pm - Thank you, Gerry, for remembering Daniel.

2.1.19 @12:21pm - (Replying to @SFMagus @schmarzo and 2 others) I was quite involved in DST, late 1980's, and even wrote two summary papers: and . Readers of #Bookofwhy would have fun trying DST on the Monty Hall problem and seeing it fail. Lesson: always try your theory on a toy example.

2.1.19 @11:57am - (Replying to @SFMagus @schmarzo and 2 others) I am not familiar with DST-type algorithms, and have chronic forgetfulness to acronyms in general. Would be glad to look into it if you can provide a link to a concise description, hopefully in input-output terms. #Bookofwhy

2.1.19 @4:32am - (Replying to @f2harrell @tvladeck) Just downloaded? This is the fate of papers submitted to "handbooks" or "antologies" -- no one reads them except librarians. But is truly worth reading. I could not have said it better, except perhaps expanding on the potential outcome framework.

2.1.19 @4:19am - Causality (2009) is the mathematical expansion of #Bookofwhy. See book page: . Sitting between them, in smiles and playfulness, is Primer . All three are dealing with human intuition; each amplifying that intuition in its own way.

2.1.19 @4:07am - To the best of my knowledge a Kindle version of #Bookofwhy is available from Amazon (I have one), and so is an e-audible version.

2.1.19 @3:56am - One pleasure that people in my age and my feeble memory are blessed with is the joy of re-reading and get inspired by their own writings, as if they see them for the first time: My! This is so true! Please post it again. "The siren song of curve-fitting." #Bookofwhy #AI #ML

1.30.19 @8:20pm - (Replying to @tvladeck @f2harrell) The causal revolution and bayesianism are not at odds. They are orthogonal to each other. What is odd is for Bayesians to claim ownership on prior knowledge and insist on expressing that knowledge in only one way: assign priors to parameters, else you ain't a "Bayesian".

1.30.19 @3:38pm - (Replying to @f2harrell @davidmanheim) Before asking about "putting a prior", which is an outdated stat habit, lets ask what "probability of causation" is, how to define it formally, and what we want the answer to tell us. The #Bookofwhy does it through PS (prob. of sufficiency) and PN(necessity). No prior needed.

1.30.19 @1:50pm - (Replying to @The_RickMc @dccozine) What you are saying is that "the vast majority of work in political science and economics" is stone-aged and in crying need for re-tooling. I fully agree. Indeed, see the Three Bullets" - easy to remember, and easy to apply. #Bookofwhy #causalinference

1.30.19 @1:03pm - One can spray priors on any set of mathematical objects, turn the Bayesian engine on and wait, as do most Bayesian mechanics, for the posteriors to peak on something meaningful. In most cases they won't, and tell us why. Let's think 'why' #Bookofwhy

1.30.19 @12:22pm - Jim, you have a good eye for good papers. I just re-read and I find it truly informative and convincing (no dog in fight). Yes, here is the link to all our papers, new and old . Some of the least popular are my favorites. @Bookofwhy

1.30.19 @11:44am - (Replying to @F_Vaggi) If I have a full causal diagram, why would I rush to spray priors around, unless I was brainwashed by a professor who feels uneasy unless he/she sprays priors. And this holds even when I am unsure about the diagram; spraying priors does not repair lack of causal knowledge.

1.30.19 @11:34am - (Replying to @The_RickMcs) Correction: Modern social scientists do use Pearl's framework, eg Morgan&Winship. And while mud and rain is not their main concern, they do need to somehow encode the knowledge that the price of beans in China does not effect the election in LA. #Bookofwhy

1.30.19 @10:56am - This paper explains why priors on parameters is not a good way of encoding the BULK of human knowledge, which is causal, not statistical. For a quick example, let us try to express our prior belief that mud does not cause rain. My prior is high #Bookofwhy

1.30.19 @2:09am - I side with you. This paper explains why @_MiguelHernan definition of "well-definedness" is overly restrictive. Naturally, when one starts with narrow definitions, every scientific question appears "ill-defined", and our job is to broaden them. #Bookofwhy

1.30.19 @12:42am - @ruchowdh My!My!"Humble writing style" is the most flattering compliment I heard from a reader. The statisticians protesting "We knew it all along" should hear it. And you may be right, perhaps the writing style IS humble, considering all the misconceptions #Bookofwhy had to undo

1.30.19 @12:21am - (1/2) Bayesian Networks are Bayesian in the sense described in #Bookofwhy (Chap 3) and further defined in . It is also the sense that makes some territory-minded statisticians have adopted ...
1.30.19 @12:21am - (2/2) a narrower view of Bayesianism: "if you don't assign priors to parameters you ain't a Bayesian." I dont buy it, do you? Priors on parameters is just one way of encoding prior knowledge, and not a very good one.

1.28.19 @8:35pm - (1/2) I thank readers for congratulating me on this award, which carries special meaning to every Israeli-born and to every history-minded observer. For context, the Hebrew University in Jerusalem was inaugurated in 1925 by Albert Einstein and Sigmund Freud among other
1.28.19 @8:35pm - (2/2) scientific giants, and came to symbolize both the historical revival of Jewish peoplehood and its age-old commitment to the pursuit of learning. I feel truly honored to be given the opportunity to express admiration and show support for Israeli scientists and academicians.

1.25.19 @2:26am - (Replying to @Miles_Brundage) Very interesting article, thanks for pointing. The fact that they are using the Ladder of Causation already indicates that the authors are on the right track. Now I need to understand if the connection to RL is substantive or just in name. #Bookofwhy

1.25.19 @12:11am - And for readers who enjoy artistic science, here is a non-numeric counterexample:
This yields Y = X and P(M|do(X))=1/2 and P(Y|do(M)) =1/2
So, the composition formula gives zero effect while the true effect is 1. #Bookofwhy

1.24.19 @12:17pm - (Replying to @omaclaren @JohannesTextor) We cant end this beautiful discussion on a negative note. It sure DOES guide us on what to do. First, completeness tells us that there exists no np solution if Y---M are confounded -- great help!!!. Second, it suggests creative remedies. E.g., measure an M2 mediating M-->Y, etc.

1.24.19 @3:14am - (Replying to @4lertus @nsaphra @KaiLashArul) Sorry, I usually refer to the book Causality (2000, 2009) as "causality". #Bookofwhy

1.24.19 @2:10am - (1/3) Your composition formula is right ONLY if we assume no M--Y confounding. But in general, even if M is the only mediator between X and Y, the formula does not hold. This is in fact a beautiful example how modern #causalinference deconstructs strongly held intuitions #Bookofwhy
1.24.19 @2:10am - (2/3) Intuition says: We have two RCT studies, one gives P(Y|do(M)), the second gives P(M|do(X)), So, if we have a new study with X--->M--->Y the ONLY directed path from X to Y, we should forget confounding and chain the two causal effects together to get P(Y|do(X)). WRONG! We do
1.24.19 @2:10am - (3/3) need to worry about confounding, but only Y--M, not X--M. It is also beautiful because it is not easy to show that the intuition is wrong, so I bet most traditional researchers would go wrong here. Funny, some researchers still refuse to believe in the revolution #Bookofwhy

1.24.19 @12:38am - (Replying to @omaclaren) The way to answer your question is to try and prove your formula through do-calculus in the same way that front-door was proven (Causality p.87-88) and see if you get stuck. If you get stuck (and you will) then you know the formula is wrong, because do-calculus is complete.

1.24.19 @12:14am - Contrary to intuition, the answer to your question is NO. When Y-M are confounded, the chain-rule does not hold. We can get:
P(y|do(x)) = SUM_M P(M|do(x)) * P[Y|do(x), M]
but we cannot write P[Y|do(x), M) = P[Y|do(x),do(M)] because Rule 2 of do-calculus is not satisfied.

1.23.19 @12:56am - (Replying to @PauSchae) Because "main effects" are properties of "race" itself, and "side effects" are properties of those "nature intervention" (eg. cosmic radiation, genetic ancestry, ...) that determine why one person is black and the other white. #Bookofwhy

1.23.19 @12:42am - SUTVA excludes "race" because it insists that treatments obey the "consistency rule", namely, those "nature's interventions" which determine "race" should be free of side effects. We can't even conceive of those race-determining interventions, can we speak of their side effects?

1.22.19 @1:51pm - (Replying to @AnalyticExec) I can imagine that "why" would be of great importance to the insurance industry, but I am not aware of a related technical literature. Have they attempted a formalization of "why" comparable to "but for" and "NESS" that we find in the legal literature? #Bookofwhy

1.22.19 @4:39am - (1/3) I am surprised and delighted that economists found interest in my paper on non-manipulable variables. Why delighted? Because economists deal with such variables all the time. Why surprised? Because they do not have an expression comparable to E[Y|do(X=x)].(cont.) #Bookofwhy
1.22.19 @4:39am - (2/3) Even economists who drifted to Rubin's camp are stuck. True, they can write Q=E[Y(x)] and pretend that it is the same as E[Y|do(x)]. But is it? Y(x) comes from Rubin's PO framework where X=x is a name of a manipulable treatment that has sworn allegiance to SUTVA. (cont.)
1.22.19 @4:39am - (3/3) and SUTVA forbids race, gender and even "minimum wage" to enter the privileged company of "well-defined treatments". So what do economists do? Same as other stuck people. They speak SUTVA in writing and use commonsense in practice. I hope my paper helps them speak straight.

1.20.19 @5:13pm - (Replying to @NeuroStats @analisereal @paulportesi) Thanks @NeuroStats. Great window into the mind of a 1952 statistician who had the courage to admit: "I did not know anything about path analysis...Why?" And this is 30 years after Sewall Wright introduced path analysis and was left to defend it alone. See #Bookofwhy chapter 2.

1.20.19 @2:25pm - (Replying to @mesuturkiye @Cambridge_Uni @CambridgeUP) Sorry, can't travel out of California these days. Wish Twitter could carry my warm signature across the Atlantic. Warm e-signatures - the next challenge for AI.

1.20.19 @1:53pm - (Replying to @MariaGlymour @PHuenermund and 3 others) Should the water molecules wait for "legitimacy" before they "perceive" the gravitational pull of the moon and rush to join the tide? I wrote this paper to allay the hesitation of those poor molecules, as well as other natural beings (eg humans) #Bookofwhy

1.19.19 @6:17pm - (Replying to @IFindAnomalies) Anxious to ask: Why Y ?

1.18.19 @11:18pm - (Replying to @statwonk @bajwa_jamal @nntaleb) I don't think even Gelman would classify #causalinference methods as "qualitative". Yes, the input knowledge is qualitative (ie, "who depends on whom") but the output is an "estimand", namely, a RECIPE for processing data to get a QUANTITATIVE answer to our question. #Bookofwhy

1.18.19 @10:51pm - (Replying to @sarahkmels @stephensenn) So the hard question is: If two half-Bayesians co-author a paper, will it be frequentist or Bayesian? The #Bookofwhy says: it all depends on their half-priors.

1.18.19 @7:21pm - (Replying to @TheLeanAcademic @StatModeling) What seems to be the problem. Please share with us the voice of practical virtue.

1.18.19 @12:01pm - (Replying to @eddericu) Good catch. It will be corrected soon.

1.18.19 @11:59am - But someone needs to explain to stat students why their professors speak P when the world speaks P+M. Someone needs to explain to them that it is temporary, that their professors will catch up. Now, when you say that, the enlightened professors get offended: "I'm P+M!" #Bookofwhy

1.18.19 @12:46am - (Replying to @MaartenvSmeden) You must be lucky to be in an enlightened stat environment. But look up the last ten presidential addresses on the RSS website. Not one touches on M. Talk with all chairs of stat departments within a 300 Km radius from you - not one would know an object outside P. Try it, I did.

1.17.19 @11:51pm - Very interesting! But can't we really account for path-dependence through side-effects. Surely my headache would depend on whether I take aspirin from my cabinet or go buy it from a North Korean pharmacy. But the difference should show up in M, if the latter option is serious.

1.17.19 @11:14pm - What seems to be the problem? Please share. I used to be a Bayesian myself, see why I half-quit . So my heart can't allow a Bayesian to remain frustrated, especially with counterfactuals, the gems of scientific thinking. #Bookofwhy

1.17.19 @11:06pm - (Replying to @MaartenvSmeden) There is no "blame" here. Missed opportunity? Yes. But no "blame". Statisticians chose to focus on P, not M, (90% of them still do) and so they have not encountered the need to define "estimand" with the extra provision of being "a property of P". No blame, just facts #Bookofwhy

1.17.19 @10:54pm - (1/3 Indeed #Bookofwhy gives Neyman an honorary mention as a pioneer of the revolution. However, his counterfactual subscripts are not events but indices of treatments. Thus they cannot be related to nonexperimental data. (See Rubin's Fisher Lecture) The first formal connection
1.17.19 @10:54pm - (2/3) to data was made by Sewal Wright in 1920, and that is why a whole chapter is devoted to him. Rubin and Rosenbaum also made this connection in 1983, through the consistency rule Y=x Y_1+ (1-x)Y_0 and, of course, the PO framework is part of the revolution. But, let us examine
1.17.19 @10:54pm - (3/3) ....examine what replaces P in the PO framework, so we can articulate and answer causal questions. The revolution is defined by the result of this examination. Why resist the title "revolution"?. Can't your students do today ten times more than anyone could two decades ago?

1.17.19 @8:23pm - (1/2) The primary reason some authors are reluctant to use the do-operator in #causalanalysis is the oddity of saying do(X=x) when X is nonmanipulable (e.g ., sex or blood-pressure). To diffuse this reluctance I am posting a new paper that explains why EVERY
1.17.19 @8:23pm - (2/2) EVERY variable should be given equal right to invite the do-operator. The paper explicates the scientific meaning of this invitation as well as its benefits and testable implications. Conclusions: Manipulativity prohibitions - a relic of the past. #Bookofwhy

1.17.19 @3:44pm - (Replying to @MaartenvSmeden) I see no "bashing" in explaining to people why they cannot find the item "property of the distribution" in their dictionary definition of "estimand". I see no "bashing" in explaining that, prior to the revolution, EVERYTHING was a property of P. Isn't it true? So, why "bashing"?

1.17.19 @2:30pm - (1/4) Since the word "estimand" seems to be giving readers problems, let me share my definition. First, it is indeed that which needs to be estimated, ..However, in the era of #causalinference, one needs to add something, to avoid confusion. #Bookofwhy
1.17.19 @2:30pm - (2/4) One needs to add that an "estimand" is a property of the distribution that governs the observed data,e.g. E[X], var[Y], E[Y|X=x],. We need it because causal quantities e.g., E[Y|do(x)] or E[Y_x|Z=z] may also need to be estimated, but they are not expressed as properties of P.
1.17.19 @2:30pm - (3/4) hence they are not "estimands" unless they are "identified", i.e., reduced to properties of P, and when they do, they can be regarded as "recipes" of what to do with our data (see #Bookofwhy) to get a valid answer to our causal question. This important distinction is not in
1.17.19 @2:30pm - (4/4) favorite text, or encyclopedia, or handbook. Why? Because those were written by statisticians prior to the causal era. For them, everything was properties of P, including the research question. The were not interested in causal questions which are not defined by P but by M.

1.16.19 @7:08pm - (Replying to @Grady_Booch @IntuitMachine and 4 others) We call this facility "causal modeling" and, miraculously, the model gives you both "what if I see x" and "what if I do x" from the same compact representation. (Not to mention "what if did x' instead of x"). It is really a miracle. #Bookofwhy

1.16.19 @6:00pm - (Replying to @IntuitMachine @tdietterich and 4 others) From my humble lens, Deep RL is semi-interventional, for it cannot infer consequences of untried actions (eg ban cigarettes, raise minimum wage). I would also hesitate making a big fuss about abduction; it has become a technical term, meaning different things to different folks.

1.16.19 @5:44pm - (Replying to @ceptional) Well summarized.

1.16.19 @5:43pm - (Replying to @tmorris_mrc @kareem_carr) Tim, If you understand how statisticians think, please tell us how 20K followers of Gelman's blog do not rise up to demand an explanation for the logic of doing estimation before identification. Is it out of politeness, lack of curiosity? or what? #Bookofwhy

1.15.19 @11:44pm - Yielding to readers' interest in my exchange with Andrew Gelman, I have posted a portion of it on my blog and am offering it here for Twitter inspection: . I have learned a lot from it, but still not quite sure how many statisticians think. #Bookofwhy

1.15.19 @11:22pm - (Replying to @sksq96) Wishing you smooth sailing and a poetic sense of accomplishment.

1.15.19 @11:16pm - (Replying to @IntuitMachine @tdietterich and 3 others) Beautiful lecture by Hassabis. I could not resonate his the interpolation-extrapolation-inventiveness ladder, perhaps because curve-fitting gives both interpolation and extrapolation. But I was struck by his repeated calls for counterfactual thinking. He would like #Bookofwhy

1.15.19 @5:39pm - (Replying to @austinvhuang @omaclaren) Agree. In this case, instead of yes/no refutation, we have almost-yes/almost-no refutations. This is always the case in noisy worlds.

1.15.19 @5:04pm - (Replying to @austinvhuang @omaclaren) You do not need to guess the functions that drives your model. You can falsify the structure without knowing the functions. #Bookofwhy

1.15.19 @4:24pm - (Replying to @omaclaren) No, you do not need to guess the function. You can falsify w/o guessing. Overall I cant figure out what you are after. If you want to convince me that there are other tasks in physics and engineering that use physical models, and do not need identification, I am convinced.

1.15.19 @3:43pm - I assume many readers are wondering about the origin of those funny "doors". They come from an ancient (1993) paper "Graphical Models, Causality, and Intervention" and, since no one complained, I kept using them and even #Bookofwhy

1.15.19 @3:29pm - (Replying to @omaclaren) Labels aside, no one is excluded. If you know the functions, you are welcome to use them but, then, you will not benefit from the magic of "nonparmetric identification", which miraculously estimates your query of interest WITHOUT knowledge of those functions. #Bookofwhy

1.15.19 @1:53pm - (Replying to @omaclaren) Oliver, the Pearl I know has not done "parameter estimation" since WWII. On the other hand, "statistical falsification" is something he is very proud of: It's Tool #1 here: . For more about his work, try #Bookofwhy and (dont miss!)

1.15.19 @6:52am - If you only heard #Harvard side, Primer will poison you irreversibly. For the first time you will meet free counterfactuals, liberated from their leashes to "treatments" and RCT's, smiling in their natural habitat, ready to be estimated and help others get estimated. Irreversible

1.15.19 @4:31am - (Replying to @LydiaManiatis) Good catch. It should read: "Glad I had the time, in 2012, to prepare lectures."

1.14.19 @11:09pm - (1/2) My favorite toy problem is Simpson's paradox: Would you or would't you recommend the drug to a randomly chosen person. Just asking yourself: "What do I need to know to decide?" already elevates you two levels above most data analysts. Next, I would go to #Bookofwhy
1.14.19 @11:09pm - (2/2) and to especially , with dozens of worked out examples. Leave the philosophers to argue what causality REALLY is, and enjoy solving causal problems they can't. I had a professor who said: "From doing comes the understanding".

1.14.19 @6:25pm - I do not understand this hunger for Twitter fights, when solving one toy problem can teach us so much more, both conceptually and operationally. If you watch any such debates you can tell immediately who solved a problem and who just talks "about" things from a smelling distance.

1.14.19 @3:01pm - (Replying to @quantadan @NeuroStats) Quanta, you are so so right! And I say it at the frightening risk of sounding self-serving: "There is no better introduction to causal inference." I dream of a world in which data scientists communicate informedly on models, actions & counterfactuals see

1.14.19 @1:52pm - Worth following, agree. In the past two days, the discussion culminated in a conditional-convergence based on a division of labor, and on new evidence that Pearl is not as thoughtless as Gelman's review portrays him to be. #Bookofwhy

1.14.19 @1:32pm - (Replying to @adibzaman @eliasbareinboim) If we were to ask this question on every ML paper published we would have had some AI in the output. #Bookofwhy

1.14.19 @1:22pm - I like this lecture, thanks for re-posting. It goes quickly over the fundamentals (including the two basic laws of causal inference), and delves into novel applications. Glad I had the time, in 2002, to prepare lectures. #Bookofwhy

1.14.19 @8:27am - (Replying to @DimDrandakis @ipfconline1 and 4 others) Not really. Folks doing prediction and object-recognition are not really trying, and those doing model-blind deep-learning are trying the impossible. My explanation: . #Bookofwhy

1.14.19 @7:58am - (Replying to @JohannesTextor @mendel_random and 2 others) Thanks for pointing to, and correcting the ije summary, its really misconceived, almost as bad as the "pluralists" whining. My take:

1.12.19 @11:04pm - (Replying to @omaclaren @eliasbareinboim and 2 others) Gelman's statement: "Statisticians use models all the time, statistics textbooks are full of models" is plain wrong --I just examined all stat textbooks on my shelf. We could perhaps excuse engineering/physics books that estimate parameters in systems of equations of known form.

1.12.19 @7:17pm - (Replying to @omaclaren @PHuenermund) DAGs are to regression as fire is to water. But this is secondary. I mentioned the problem of "deciding which covariates to adjust for" because it remains an elementary problem in @causalinference that 90% of statisticians cant do. Can BHM folks do it? Can anyone that you know?

1.12.19 @6:39pm - (Replying to @omaclaren @PHuenermund) This is indeed a causal model that, in some deep sense, can compute the effect of rain on mud. Thus, presumably, one can use HCM to do every task in #causalinference, ie, decide which covariates to adjust for, decide if an ACE is identifiable, etc. Would you like to try on a toy?

1.12.19 @6:24pm - (Replying to @omaclaren @PHuenermund) But does it present any method of estimating ACE? I do not question the quality of the book, I just want to know if it is relevant to works which normally fall in the category of #causalinference. #Bookofwhy

1.12.19 @5:51pm - (Replying to @omaclaren @PHuenermund) I was craving for example of a causal problem, solved by standard statistical means, (or by hierarchical models,) not just any example. Does the book has ANY causal example? Say finding an average causal effect, or effect of treatment on treated, or counterfactual, etc... ?

1.11.19 @8:59am - (Replying to @pophealth3) Clinical trials became textbook material in 1935. This is one place where statistics allows for causal considerations (albeit informal), and #Bookofwhy acknowledges this allowance. What about observational studies, Haavelmo, Rubin and Robins? Were not the ten presidents aware?

1.11.19 @8:12am - (Replying to @pophealth3) (1/2) What ten presidents say over a period of 20 years tell us something about how the community as a whole viewed causation; a high priority challenge or a nuisance that should better be handled by others? It also reflects on Gelman's claim to have been doing CI for 30 years.
1.11.19 @8:15am - (Replying to @yudapearl @pophealth3) (2/2) Was he a rebel or a leader in the eyes of an establishment that failed to even mention CI as a worthy challenge? Was he really doing CI or just imagined so? #Bookofwhy

1.11.19 @1:00am - Sorry, the link to the 10 Presidential Addresses is here: covering the period 1999-2017, i.e., the heat of the "Causal Revolution", with not an echo from the statistical leadership. #Bookofwhy

1.11.19 @12:49am - An interesting document came to my attention, linking to 10 RSS presidential addresses . It illustrates the amazing disinterest of statistics in #causalinference, as noted in #Bookofwhy and protested/denied by leading statisticians (eg Gelman, Senn...)

1.10.19 @12:35am - It is only after tweeting "SUTVA is for PO folks who speak no structure" that I came to realize, indeed, how clearer #causalinference classes could become by skipping discussions of SUTVA and going straight to structure. Thanks for raising the SUTVA issue. #Bookofwhy

1.10.19 @10:32pm - (Replying to @cdsamii @eliasbareinboim and 3 others) (1/2) I confess to be among those who are trying to understand your new framework, and can't. Since you are assuming away ID through ignorability, there remains a new framework for estimation. Can we start from here, specifying what is needed and what is given? I will try ...
1.10.19 @10:45pm - (Replying to @yudapearl @cdsamii and 4 others) (2/2) (1)Needed: a measure of program effectiveness in the target population (2)Given: information coming from a source population. But since we assumed the two populations are essentially the same (Mod. ignorability), what makes this estimation problem unique? The estimand? How?

1.10.19 @3:01pm - (Replying to @SueMarquezR @jim_savage_ and 4 others) Thanks Sue, this is a thoughtful and precise paper on a deeper-level "structural economics" literature. However, it appears to me that anyone doing this deeper analysis should be able to do counterfactuals in structural equations. Yet the latter is oddly missing. #Bookofwhy

1.10.19 @2:39pm - (Replying to @cdsamii @PHuenermund and 3 others) Can you help us understand what "old stuff" you are using? I for one am not familiar with any "old stuff" that would solve even a simple transportability problem (unless the two environments are (conditinally) the same, ie ignorability). One "old stuff" would do. #Bookofwhy

1.10.19 @2:20pm - (Replying to @_limbs_) Agree. The arrows in hierarchical models are set-subset relationships, not cause-effect as in DAGs. In this sense, it is still fair to say that Gelman does not use DAGs in his writing. E.g., he cannot infer conditional independence in his model, not to mention identification.

1.10.19 @1:57pm - (Replying to @jim_savage_ @JimMinifie and 4 others) Appreciate the warning, good point. But I am not saying "must use my method". I am genuinely not aware of any definition of counterfactuals in structural equations, or any method of identifying counterfactual statements in the economics literature. Concrete pointer would help.

1.10.19 @1:41pm - (Replying to @cdsamii @PHuenermund and 3 others) Paul is right. Even if the issue is estimation, readers need to know what identification method is used, what assumptions are made, and perhaps what asymptotic results prevail. Your paper sounds like solving an identification problem. Why not help readers? #Bookofwhy

1.10.19 @7:17am - (Replying to @dailyzad) Good question. I know how. Give me a toy example, and I will not try to avoid it.

1.10.19 @12:57am - (Replying to @mendel_random @StatModeling Some folks think differently, e.g., They understand that solving toy problems is pre-requisite to understanding causation, and avoiding them cast doubts on the validity of "real world answers" #Bookofwhy

1.10.19 @12:35am - It is only after tweeting "SUTVA is for PO folks who speak no structure" that I came to realize, indeed, how clearer #causalinference classes could become by skipping discussions of SUTVA and going straight to structure. Thanks for raising the SUTVA issue. #Bookofwhy

1.10.19 @12:15am - (Replying to @JimMinifie @jim_savage_ and 4 others) @JimMinifie I am not sure who those stat/econometricians are. Heckman ? Angrist? Imbens and Rubin? If the latters, I have related their work to SCM here: , I did not repeat it in #Bookofwhy (only in passing) because it would have sounded too negative.

1.9.19 @11:08pm - (Replying to @_limbs_) Impossible. Gelman has never used DAGs in his work (no Rubin's student ever has) so his notion of "model" is different. For him postulating a parametric family of distributions (normal, Bernoulli, logistic etc) constitutes "a model". The idea of "causal model" is unknown there.

1.9.19 @9:46pm - (Replying to @jim_savage_ @Andrew___Baker and 2 others) @jim_savage I am willing to take the test and would like to find out how "policy counterfactuals" can be constructed without the help of DAGs. Can we start with a simple "policy counterfactual" problem with 3-4 variables, so that we can see our way through it? I am ready

1.9.19 @8:15pm - John is an economist who never heard about SUTVA. He simply learned to compute counterfactuals from structural equations, as in Primer: Will John ever go wrong in his scientific career? He won't! SUTVA is for PO folks who speak no structure. #Bookofwhy

1.9.19 @5:35pm - (Replying to @JadePinkSameera) It is hard to illustrate "reluctance", but take a look at Stigler's "The Seven Pillar of Statistical Wisdom", which is a thorough account of the achievements of 20-Century statistics, and judge for yourself how proud he is in the contributions of statistics to #causalinference.

1.9.19 @2:55pm - An ounce of advice to readers who comment on this "debate": Solving one toy problem in causal inference tells us more about statistics and science than ten debates, no matter who the debaters are. #Bookofwhy

1.9.19 @11:24am - (Replying to @PHuenermund @eliasbareinboim @heckmanequation) Heckman loved the do-operator so much that he renamed it "fix-operator" and went on to praise "fix" and criticize "do". He forgot that the operator comes with the benefits of graphical models, and left economists hungry for transparent causal inference.

1.9.19 @4:04am - (Replying to @rquintino) Simpson's is the litmus test of causal thinking. You want to know what a statistician knows about causation? ask him/her about Simpson's paradox. [And you will be mighty surprised at the majority of the answers. Try it.] #Bookofwhy.

1.9.19 @3:54am - (Replying to @mgaldino) Oh, No, Pearl does get it, for he has heard the "less convincing" argument used many times in the past. But today his request is aimed not at "convincing" Gelman but at establishing credibility and a common language of communication. One toy problem please, forget "convincing".

1.9.19 @1:09pm - (Replying to @f2harrell) Would love to, but the wealth of topics overwhelms me. I can only contributes to discussions where I know something about. Do you want to start with Gelman's views on causality? Or his concept of a "model"?

1.8.19 @11:59pm - Gelman's review of #Bookofwhy should be of interest because it represents an attitude that paralyzes wide circles of statistical researchers. My initial reaction is now posted on Related posts: and

1.8.19 @8:37am - (Replying to @PHuenermund) My point is that had he understood the #Bookofwhy he would have realized that the toned is justified. e.g. there is NOT a single model in any statistics text if "model" is taken in the sense defined by #Bookofwhy, namely, assumptions about the world external to the data.

1.8.19 @8:09am - The qualitative/quantitative distinction is forgivable. When we write down a model we are only expressing qualitative relationships. But the rest of Gelman's review entices me to pose a new discussion topic: Which is Gelman's main obstacle to understanding/accepting #Bookofwhy?

1.7.19 @7:07pm - In view of recent interest in front-door estimators and their relation to regression, it would be interesting to note that front-door actually hatched from regression analysis, as seen in this 1993 paper . #Bookofwhy

1.7.19 @6:14pm - (Replying to @fhuszar @paulwillemjvr and 3 others) Much relieved.

1.7.19 @2:47pm - (Replying to @paulwillemjvr @shakir_za and 3 others) Anti-causal in 2019? Can't be!

1.7.19 @5:39am - (Replying to @juli_schuess @SLMsociology and 2 others) It is really beautiful and short, but it takes an artist to appreciate beauty, and a philosopher to appreciate the benefit of unification. Glad beauty is still a virtue among readers of #Bookofwhy. Most PO enthusiasts get upset when you show them how Y_x is computed from a model.

1.6.19 @5:38pm - (Replying to @omaclaren) My copy says: . "Here B can precede X; yet we should still not condition on B, because that would violate the back-door criterion." [#Bookofwhy Readers: If in doubt about typos, please check "errata sheet" on . All typos were corrected on 4th printing]

1.6.19 @4:41pm - (Replying to @omaclaren) I checked my copy. I see no typo, nor reverse causation. Are you still having problems with the M-bias? #Bookofwhy

1.6.19 @4:26pm - (Replying to @seanmcarroll @littlebode) I am eager to read you expansion of this idea, which I mimi-leveraged to explain compatibilstic free will (page 364 of #Bookofwhy). More needs to be done there, especially to show that the thermodynamic direction of time remains in tact under determinism. Please keep us informed.

1.6.19 @12:34pm - (Replying to @littlebode) What does Sean Carroll say about causation? I hope good things, perhaps new ideas? new problems? new tools?

1.6.19 @12:29pm - I have hoped the #Bookofwhy would shake up stagnation by emboldening students, but you found a more effective way -- embolden writers. Brilliant! And may progress be your bounty.

1.6.19 @11:39am - (Replying to @yskout) The relationship is not clear to me. When I learned factor analysis, it was presented as a purely probabilistic problem (eigen values of covariance matrices), hence, totally divorce from causality. But perhaps my teachers failed to motivate it from a more causal angle.

1.6.19 @11:33am - (Replying to @thedmca @jnd1er and 3 others) Thanks for the kind mention.

1.6.19 @6:27am - (Replying to @ChrisAdamsEcon) I dont see why guidelines for economists should be vastly different from those issued in epidemiology. The mantra "No causation without clear communication" should rule both. #Bookofwhy

1.5.19 @10:29pm - This is an incredible vote of confidence in the methodology of #causalinference, especially in observational studies. The standards that regulate scientific writing shape scientific thoughts and practice. I wish editors of economics journals generate similar guidelines.#Bookofwhy

1.5.19 @8:41pm - (Replying to @SLMsociology @erikdbwestlund @stevepowell99) Why should anyone disagree? If readers can solve the toy problems on page 93-100 of Primer , kuddos to you and your book. If you feel they could stand some help, it is easily done. We have enough disagreements with the arrow-phobes. #Bookofwhy

1.5.19 @7:20pm - (Replying to @dccozine) Strange, and I thought no one would notice that in the insular bubble called "econometrics". Evidently, some folks read the details -- there is hope to Eco. #Bookofwhy.

1.5.19 @7:06pm - (Replying to @erikdbwestlund @stevepowell99 @SLMsociology) @SLMsociology and Winship book is the greatest thing that happened to social science since Duncan and Blalock. The only missing ingredient (reparable in one page) is to show how counterfactuals emerge organically from SEM, thus unifying the chapters. #Bookofwhy.

1.5.19 @6:43pm - (Replying to @erikdbwestlund @stevepowell99) From the last lecture I heard (by Imbens, NIPS 2017) it appears that many folks are still using "matching methods informed by Rubin" but, nowadays, someone who read #Bookofwhy (page 274) is likely to ask: "Is it valid?".

1.5.19 @5:18pm - Your summary amounts to a whole new book, or a valuable R-supplement to Primer . Why not publish it? One comment: ETT should be introduced as another research question, rather than an estimation technique. See p. 106 of #Bookofwhy.

1.5.19 @1:37pm - (Replying to @jon_y_huang) Thanks for the pointer. But though they explain how to get Stochastic IDE, I am unable to grasp what research question it answers and whether it can be used to approximate NIE, whose meaning I do grasp. Can anyone articulate it in Twitter language? BTW, is it a do-expression?

1.5.19 @1:24pm - Reading #Bookofwhy before Athey's lecture is a good way of getting perspective of the field. You will be puzzled though by the almost complete overlap of topics and zero overlap of methods, bringing us back to a cultural enigma called "econometrics"

1.5.19 @2:48pm - (Replying to @CaitLamberton @linzcpage) Yes, there has been some change in causality-land, and I am glad it came to the attention of folks in your field. I dared call it "causal revolution" in #Bookofwhy, to highlight the sudden outflow of results. Enjoy.

1.5.19 @1:56am - (Replying to @dccozine) Methodological allegiance, I have found, is the greatest impediment to scientific progress. Behold how robust and long-lasting it is. You can easily guess an author's PhD advisor decades after graduation. Scared to offend the boss and be charged with treason. #Bookofwhy

1.4.19 @9:56pm - (Replying to @MartinRavallion) The "evaluation" community was an honest revelation. I knew of course that every researcher should and is concerned with evaluating the impacts of programs, but I did not realize that the task of evaluation presents unique problems that justify a separate discipline. #Bookofwhy

1.4.19 @7:37pm - (1/2) Thanks for a positive and informative review. I like it because (1) I learned from it about a community called "evaluation", new to me, and how it has been dealing with causation. (2) The reviewer is not scared to say that this community can learn something from #Bookofwhy (a
1.4.19 @7:37pm - (2/2) (a hard thing for a statistician to say, and high treason for an economist). (3) I learned what I could have done better in #Bookofwhy to make the reading obstacle-free for evaluation researchers. If you have not read #Bookofwhy yet, I receommed reading this review first.

1.4.19 @11:07am - (Replying to @MarcoIppolito) Your conclusion is right (if I properly understood you question). Path-0 is not blocked by U and Path-1 is blocked by U. Conclusion: All backdoor paths from B to E are blocked by U.

1.3.19 @4:44pm - (Replying to @david_colquhoun @EpiEllie @stephensenn) Quite the opposite. DAGs are the perfect language to express the possible existence of unanticipated confounders, in case you really fear them. The double arrows <--> were invented to express this belief, why not use it? #Bookofwhy

1.3.19 @1:24am - (1/2) A reading group of both @_MiguelHernan's book and Primer would be of great value, especially to economists, and I would start with Primer. First, it's much shorter. Second, it is good to have a 3-dimensional view of the world before examining it through a microscope. Specif..
1.3.19 @1:24am - (2/2) Specifically, seeing how counterfactuals emerge naturally from an economic model before they are hooked to treatments and RCT's, and how the notational distinction works between interventions and counterfactuals (do(x) vs. Y_x) before it disappears in one big PO. #Bookofwhy

1.3.19 @1:04am - But the most gratifying news came yesterday, when #Bookofwhy became #1 on Amazon bestsellers list for BASEBALL: And economists are telling us there is deep wisdom in crowd.

1.2.19 @6:11pm - (Replying to @iamnemo8 @amychua @skdh) Thanks for the favorable mention #Bookofwhy

1.1.19 @11:44am - (Replying to @LaurentFranckx) Now that you brought it up, I see that it needs repair. It should read: "... using naive statistical methods". Thanks. #Bookofwhy

1.1.19 @5:03am - (Replying to @mathtick) I also had difficulties understanding Wang&Blei. They claim identification, but do not present an estimand. I could not tell if they leverage external information, as in here . A hard to decypher enigma. #Bookofwhy

1.1.19 @3:38am - (Replying to @mathtick) Please refresh my memory on the "blessing of multiple causes"

1.1.19 @2:24am - (1/n) Dear readers and followers. This is a brief end-of-year report that I would like to share as 2019 is about to enter our space. I am very gratified with the warm and wide acceptance of #Bookofwhy and its key messages: the inevitability of causal modeling and the logic ...
1.1.19 @2:24am - (2/n) ... the logic imposed by the Ladder of Causation. The need to marry ML and CI has become a popular banner even in communities that are strangers to CI. The introductions of free courses on graphical models at @HarvardEpi and Microsoft (DoWhy) [there may be others] promise to.
1.1.19 @2:24am - (3/n) ...promise to end decades of confusion and bewilderment that were perpetuated by sluggish educational and editorial establishments. Some disagreements do remain, among them the debate over manipulability and the need for a three-level hierarchy.
1.1.19 @2:24am - (4/n) But these differences are bound to disappear, I believe, as soon as the full power of structure-based tools become familiar to the community. To that end, I would like to provide free access to Chapter 4 of Primer ( ), which develops and exemplifies ...
1.1.19 @2:24am - (5/n) harmonious symbiosis of counterfactuals, structural models and the do-operator -- . Enjoy, and may 2019 make our collective understanding dwarf that of 2018. #Bookofwhy

12.31.18 @7:58pm - (Replying to @NeatWitTweet @chriskroiss) I recommend this course very strongly, especially to economists who are already two decades behind the time @HaravardEcon Do not wait for your professors to recommend it, they won't. Once you acquire these tools, the hardest problems in your Econ classes become a child play.

12.31.18 @7:30pm - (Replying to @AlvaradoMoutin1 @HarvardEcon) In such hopeless cases it is doubly important to know what simplifying assumptions are capable of moving you towards a contingent solution. Blindly applying textbook routines (eg IV) is not wiser than searching under the lamppost. #Bookofwhy

12.31.18 @6:49pm - (Replying to @theute_ @junpenglao @generativist) Ask the Stanford professor if he can teach any of the seven tools described here and, if not, what he would suggest to creative students who wish to acquire these tools. #Bookofwhy

12.31.18 @3:52am - (1/2) I am not familiar with the two courses you mentioned, so I am basing my opinion on the advertised content. Sobel's course is more dogmatic than Rubin's. It is likely to be harmful in its religious aversion to graphs, to structural models, and to modern way of thinking. (con.
12.31.18 @3:52am - (2/2) Roy's course stands between @HarvardEcon and #EpiTweeter. It makes limited use of graphs, but no unification with PO. Mostly harmless. #Bookofway

12.30.18 @11:53pm - (Replying to @jaiz30) Glad the #Bookofwhy made it to another smart collection.

12.30.18 @1:07am - I'm not sure what we can infer from your job market list; (1) that #HarvardEcon has chosen jobs over science, (2) that some applied works do not require science, or (3) that #HarvardEcon students acquire scientific tools by independent reading? I pray it's (3), not (1) #Bookofwhy

12.29.18 @1:45am - (Replying to @AlvaradoMoutin1 @HarvardEcon) I need more context to answer your question. Note however that an IV analysis is not regression-based but an exercise in causal analysis and, so, an economist is allowed to say whatever logically follows from the causal assumptions embedded in the IV structure. #Bookofwhy

12.29.18 @1:36am - A note of appreciation to David Auerbach @Auerbachkeller for including #Bookofwhy in such stellar company.

12.29.18 @12:09am - (Replying to @Awesomnomics) I would not go as far as "the first thing," but it should at least be brought to their attention. The current isolationist attitude of educators and editors is breeding an obsolete generation of econometric students, bound to perpetuate isolationism. #Bookofwhy @HaravardEcon

12.28.18 @11:47pm - (1/3) I am not searching for traces of one work or another. My question was whether you believe @HarvardEcon students are prepared for the challenges of modern #causalinfernce. e.g., Can they determine conditional exogeneity or testable implications in a given economic model? C....
12.28.18 @11:47pm - (2/3) Can they compute counterfactuals in a given systems of economic equations? More pointedly, do they have the tools to solve simple problems like those listed in Sec. 3.2 of , which Epi students learn as basic alpha-bet? Or will they stay behind? (con.
12.28.18 @11:47pm - (3/3) Many economists on this Twitter look up to @HarvardEcon for leadership in the reformation of econometric education into the age of modernity, among them seasoned educators who are expressing an urgent need for such reformation. It's time for a bold step forward. #Bookofwhy

12.28.18 @4:24am - (1/2) In 2014, the editors of Econometric Theory asked me to comment on the "experimentalists" vs. "Structuralists" debate in econometrics. I have hoped Sec. 4 of would convince economists that there is more to causation than IV. Isn't there? @HarvardEcon
12.28.18 @4:24am - (2/2) Speaking of "foundations of of causality". Do economists take IV analysis to be foundational? ie, sufficient for deriving other tools and results in #causalinference developed in the past two decades? (eg identification, mediation, generalizability) #Bookofwhy @HarvardEcon

12.26.18 @4:56am - (Replying to @totteh @EpiEllie) I agree, but we still need to explicate what makes the causal effects of "infeasible interventions" scientifically meaningful, namely, informative to policy makers who, like Harvard Epi students, care only about the impacts of "feasible interventions". #Bookofwhy #causalinference

12.26.18 @12:10am - It is encouraging that another "father of deep learning" acknowledges the crucial role of #causalinference in AI and, remarkably, not for managing actions and counterfactuals but for facilitating adaptation and generalization, as in Tool 5 of #Bookofwhy

12.25.18 @8:58pm - (Replying to @autoregress @PHuenermund and 2 others) Perfect question, closer to the core, once we clarify what is meant by "infer from data". Surely, nothing can be inferred from obs. data alone; But can we infer those elusive causal effects from data+ASSUMPTIONS,-- same assumptions we use in standard obs. studies? #Bookofwhy

12.25.18 @6:51pm - (Replying to @autoregress @PHuenermund and 2 others) Perfect question, closer to the core, once we clarify what is meant by "infer from data". True, nothing can be inferred from obs. data alone; But can we infer those elusive causal effects from data+ASSUMPTIONS,-- same assumptions we use in standard obs. studies? #Bookofwhy

12.25.18 @3:16pm - (Replying to @shell_ki @melb4886) Who are the gurus of social epi to dictate what one is supposed to use or not to use? The whole purpose of #Bookofwhy is to examine the wisdom of those gurus, unveil its pillars, and show how shaky those pillars are in the light of causal logic. #Bookofwhy #causalinference

12.25.18 @2:31pm - (1/2) The intenventionist debate demonstrates indeed why the Greeks invented logic, and why it does not surface in SCM, where things are "meaningful" if they are authorized by the logic of "listening". The debate rages in Rubin's PO, crafted to mimic RCTs, where "listening" is
12.25.18 @2:31pm - (2/2) meaningless, and non-manipulable variables, eg. "blood-count" or "gender" require special handling (still to be explicated). Is there a coherent logic that properly accounts for such variables as "effect modifiers" yet refrains from ascribing them "causal effects"? #Bookofwhy

12.25.18 @2:32am - (Replying to @igon_value) Much of the power of informal medicine stemmed from the #causalinference engine that physicians carried in their heads, and its ability to conclude: "no point fighting obesity if obesity is harmless", now prohibited by the interventionist school, since "harmless" is "ill-defined"

12.25.18 @2:18am - (Replying to @stevesphd) Not sure if you read #Bookofwhy. If you do, note that we, in modern #causalinference, do not "refuse" tools, nor remedies, nor methods, whether we understand them or not. We push towards understanding, yes, and we oppose those who "refuse" or "exclude" without the understanding.

12.25.18 @12:43am - This sounds like medical practice in the prescientific age. I hope some dissidents on the Harvard team are busy marking the limits of strict interventionalism. Even the Royal Navy rejoiced the late (1912) discovery of vitamine C. #Bookofwhy #Epitweeter

12.24.18 @8:23pm - (Replying to @Niels_Bremen) Santa was graciously creative to you. Structural equation models play key role in #Bookofwhy though always in conjuction with the do-operator, thus giving meaning to SEM parameters which your friends on SEMnet are laboring to estimate. Enjoy.

12.24.18 @4:33pm - My last verse of Greek poetry : "People of Athens, I can't turn my migraine-headache on and off, but it sure has well-defined causal effects, see why: " Even sworn interventionists (@EpiEllie) open their heart to poetry when logic compels. #Bookofwhy

12.24.18 @1:56am - (Replying to @AlfredoMorabia @IntJObesity and 3 others) Truth is in the survival of the legend. Why else would this story survive in our textbook if it were not for scientists admission that truth can be surprising, even very surprising, and that the scientific community is not always charitable to the surpriser. #Bookofwhy

12.23.18 @10:53pm - (Replying to @benmbrew @EpiEllie @kareem_carr) Agree! What #causalinference needs these days is 1 ounce of commonsense and 9 ounces of poetry.@kareem_carr proves them non-contradictory. Kudos! #Bookofwhy

12.23.18 @4:30pm - (1/2) (Replying to @edwardsjk) I agree. There are times when addressing a problem in some language is better than not addressing it at all. But my naive question deserves addressing too: Is the language of analysis adequate, or outdated? Wouldn't researchers be more likely to deal with measurement error..
12.23.18 @4:40pm - (2/2) (Replying to @yudapearl @edwardsjk) ... if given cognitively meaningful tools to manage it? More specifically, does the PO framework offer cognitively meaningful tools for managing such problems as: measurement error, missing data, generalization, and even confounding??? #Bookofwhy #causalinference #epitweeter

12.23.18 @6:41am - (Replying to @_eleanorina) By non-manipulable I mean we do not have a "Well-defined intervention", I , such that every member of HEPO will feel comfortable saying "the effect of blood-pressure on Y in none other but the effect of I on Y". #Bookofwhy

12.23.18 @5:22am - (1/2) I am writing an addendum to to further clarify the difference between SCM and HEPO (Harvard Epi school of Potential Outcomes). Can anyone familiar with HEPO practices describe how they treat a non-manipulable variable (like blood-pressure)
12.23.18 @5:22am - (2/2) (like blood-pressure) that happens to appear (as an observed covariate) in the DAG? Do they delete it from the model entirely? or leave it in the DAG and mark it as non-manipulable? or play with its incoming and outgoing arrows? #Bookofwhy #causalinference #Epitweeter ???

12.23.18 @12:32am - My perennial naive question: Why cast a problem in a lesser understood framework? When we cast "missing data" as a causal problem, we gain tools and insight: . Same with measurement error: . Why go back? #Bookofwhy #causalinference

12.22.18 @11:39pm - Very useful overview of ML latest. Thanks. I would be interested in a pointer to those who "argued that causality is somewhat of a theoretical distraction," and how they think we can achieve interpretability or explanations without invoking causation. #Bookofwhy

12.21.18 @5:58am - (Replying to @timssweeney) Any discussion of consciousness is in danger of ending up in metaphysical deadends. However, if we take the position that consciousness is a shallow blueprint of one's software, we can embed do(x) in a very shallow blueprint with one actor-observer and one environment.#Bookofwhy

12.21.18 @2:56am - (Replying to @Luis_de_Miranda @harari_yuval) Thanks for the pointer to Bergson's "fabulation". Is there a good summary of his ideas? #Bookofwhy

12.21.18 @1:27am - I didn't realize data-scientists would view #Bookofwhy from these angles, but perhaps it is good that I didn't, there is much to be learned from viewing the world through new windows. #causalinference

12.20.18 @7:17pm - I am thrilled #Bookofwhy is found useful by top experts in decision science and risk analysis. It reminds me that, bottom line, rational decision is the ultimate goal of data science. #causalinference

12.19.18 @8:58pm - (Replying to @BreskinEpi @EpiEllie and 2 others) I love it! It tops my current favorites: "I dont trust graphs so let's assume ignorability" or "I don't trust untestable assumptions so let's use propensity score", or "its dark where I lost it, so let's search under the lamppost". Great! #Bookofwhy #causalinference or

12.19.18 @5:55pm - Adding completeness to accuracy, EHPO has been examined recently here and compared to standard scientific thinking, in which exposures, blood-pressure, and earthquakes do have causal consequences, separate from their stimulants #Bookofwhy #causalinference

12.19.18 @12:54pm - More accurately, the PO framework is split between Rubin-PO and Epi-Harvard-PO (EHPO). The latter (correct me @EpiEllie) welcomes graphs but restricts all causes to be "well-defined interventions", eg, disallowing sex, earthquakes and blood-pressure.#Bookofwhy #causalinference

12.19.18 @2:16am - (Replying to @HenningStrandin @mendel_random) Beautifully put. And I also prefer it to other metaphors suggested by philosophers. After all, no matter how rigorous one wants to be, any "understanding" must end up with primitives that are universally meaningful. "Listening" is super-universal. #Bookofwhy #causalinference

12.19.18 @1:16am - (Replying to @mendel_random) Both ways. I use "causal character" for both "cause" and "effect". When X listens to Y then both X and Y earn "causal character", X for being the "cause" and Y for being the "effect". #Bookofwhy

12.18.18 @11:51pm - (Replying to @TJ_Kelleher @Scott_E_Page and 8 others) Its a miracle that you found time for the #Bookofwhy. Thanks.

12.18.18 @11:32pm - (Replying to @DaveBrady72 @Eule_Geheule and 2 others) The differences are encapsulated in "Three Bullets" , Easy to remember, and easy to apply. #Bookofwhy #causalinference

12.18.18 @11:26pm - (Replying to @Eule_Geheule @DaveBrady72 and 2 others) The notations are complementary. The "frameworks" are miles apart. SCM uses both graphs and counterfactuals, PO is religiously opposed to the former. It shows. #Bookofwhy #causalinference

12.18.18 @2:16pm - (Replying to @AlfredoMorabia @IntJObesity and 3 others) "Let's use all the tools ..." said the Greeks. Then, when Hippasus of Metapontum proved that there are irrational numbers, they threw him overboard; they could not stomach their own logical tools. Can we accept the conclusions of our own assumptions? #Bookofwhy #causalinference

12.18.18 @2:02pm - (Replying to @deaneckles @hoktay) A paper published 3-4 years ago becomes "ancient" because, at my age, I forget why for heaven sake I spent so much time on it. Reading the "introduction" reminds me, and makes me feel less guilty.

12.17.18 @10:44pm - (Replying to @hoktay @deaneckles) This ancient paper was really fun to do. The realization that if you only knew how bad your measurements are you could repair them threw me off completely. Thanks for the nostalgia. #Bookofwhy

12.17.18 @4:24pm - (Replying to @robertwplatt) We have all seen models in which we agree on the assumptions and in which we can formally identify P(y|do(x)) and still X is not controlled by any "well-defined intervention" (say sex or race), thus leading strict-empiricists to claim: I do not know what P(y|do(x)) means.

12.17.18 @2:33pm - (Replying to @robertwplatt) Sure: Drug--->BP--->Cardiac arrest , with confounders all over the place. Again, the objection is not to the model or its estimability, the objection is to the audacity of claiming that the causal effect of BP is "well-defined" even if the model permits identification.

12.17.18 @2:08pm - (Replying to @robertwplatt) This assumption is needed to separate the issue of measurement from that of control. Hernan's objections to the do-operator hold for blood-pressure too, which is well-measured but escapes direct control. To focus on one issue at the time, can we speak about the effect of BP???

12.17.18 @6:20am - True, but in our lucky case we have both: (1) a beautiful formalism and (2) it can be easily applied to real-world scientific problems. I can't name any real world scientific problem that (1) HAS a solution and (2) escapes the beautiful formalism. Note the cap HAS. #Bookofwhy

12.17.18 @5:27am - (Replying to @WC_Lab @johnsontoddr4 and 3 others) "Minimal surgery" means making the least perturbation to the model and still ensuring that X be equal to x. If your model is a DAG, it translates to removing the arrows entering X and setting X to x. It is illustrated in #Bookofwhy chapter 1 and other chapters as well.

12.17.18 @4:20am - (Replying to @WC_Lab @johnsontoddr4 and 3 others) The do-operator is a mathematical operation on a model. Of course it does not "exist" as intervention in the physical world. But it can be given an interpretation as if Nature performs a minimal surgery. If you are bothered by the poetry, replace "do-operator" by "Operation-23".

12.16.18 @11:55pm - (Replying to @OptimizingMind ???? unpack ????

12.16.18 @8:37pm - (Replying to @eddericu) They co-exist, because "consistency" is defined differently in these two frameworks. In PO it is defined relative to the actual intervention that made X=x. In SCM it is defined relative to an ideal intervention that establishes X=x without side effects. #Bookofwhy #epitweeter

12.16.18 @8:26pm - (1/2) (Replying to @talyarkoni) Let's assume an extreme case that all this gibberish about BP and concentrations is just pre-scientific jargon that says nothing about the world, only about our perception of the world. Still, as computer/cognitive scientists, isn't it worthy of formal analysis? #Bookofwhy
12.16.18 @8:30pm - (2/2) (Replying to @yudapearl @talyarkoni) What exactly does scientist 1 say to scientist 2 that enables them to swiftly reach this common illusion they call "understanding"? Should we program robots with this powerful jargon, just for the sake of effective communication, never mind lowering BP or saving lives.

12.16.18 @7:31pm - (Replying to @deaneckles @LuGram12) Is this connected to Aronow's book? Are you suggesting that pre-modern methodology (eg Angrist-Pischke) encapsulates wisdom that post-modern methods (eg Morgan-Winship) ignore? Or that social scientists are better off arguing without graphical models? Like PO folks? #Bookofwhy

12.16.18 @7:20pm - (Replying to @loudquack) Is #Bookofwhy bashing statisticians? See two blog posts on-to-the-book-of-why/ ...Is anyone bashing statisticians or are they making it up?

12.16.18 @5:22pm - (Replying to @loudquack) Why are you uneasy? If the criticism is justified it should be welcome, if not, it should be refuted. You seem to side with the former, so why feel uneasy? #Bookofwhy

12.16.18 @5:08pm - (Replying to @deaneckles @LuGram12) In revolutionary times the "standards setups" are changing daily. Is Angrist&Pischke the "standard" or Morgan&Winship? Is Little&Rubin the "standard" on missing data, or the modern ? If Aronow book stops at pre-modernity, it will be hard to recommend.

12.16.18 @4:07pm - (Replying to @johnsontoddr4 @aruncann and 2 others) You got it on the nail. I can only speculate on the cultural roots of Miguel's objections: Rubin's wedding vow to RCT. #Bookofwhy #causalinference #Epitweeter

12.16.18 @3:49pm - (Replying to @LuGram12 @deaneckles) I see a section devoted to causal inference, page 256, which is a welcome novelty for a statistics book. But I cant tell to what extent it covers the "causal revolution" ie, whether it stops at ignorability (1974-1983) or goes beyond. Perhaps @deaneckles can illuminate?#Bookofwhy

12.16.18 @3:21pm - (Replying to @johnsontoddr4 @aruncann and 2 others) I have tried my best to write a very clear exposition here: I will try to improve if you only tell us which point needs further clarification.

12.16.18 @2:15am - (Replying to @LuGram12) I believe the whole #Bookofwhy is about "contolling for unmeasurable confounders". Pseudo-randomized designs is just one way of achieving this "control". #Bookofwhy

12.16.18 @2:05am - (1/3) In the past two days readers had a chance to carefully examine (Appendix) and verify that health scientists communicate in terms of states of variables, as opposed to manipulations of variables. They talk about agents and substances being ``present''
12.16.18 @2:05am - (2/3) or ``absent'', being at high concentration or low concentration, smaller particles or larger particles; they talk about variables ``enabling,''``disabling,'' ``promoting,''``leading to'' ``contributing to,'' etc. Branding these causal relationships "hopelessly ill-defined",
12.16.18 @2:05am - (3/3) "extrascientific" or "White Magic" does not do justice to the bulk of scientific discourse. What for? Just because PO started with RCT perceived as the mother of all knowledge? #Bookofwhy #causalinference #EpiTweeter

12.16.18 @12:42am - (Replying to @BD_Zumbo) This is a brave move, which will usher clarity in your field. The behavioral sciences, from my reading, are still torn between the PO and the SEM perspectives. Unifying the two would be a healthy relief to many researchers. See "Eight Myths.." #Bookofwhy

12.15.18 @7:42pm - Many PO-educated folks draw comfort from excluding all but "empirical phenomenon." Still, once they learn to derive a PO from a model, they draw extra comfort from answering same "research questions on empirical phenomenon" (& same assumptions) using more powerful tools. Try it.

12.15.18 @6:59pm - (Replying to @_creepinatshirt @snavarrol) Thanks for noting. The second link is "Reflections of Heckman and Pinto", written after they sweated over 8 pages to prove that we dont really need graphs to derive the front-door formula. Economists will risk everything to defend their econ. textbooks.

12.15.18 @5:51pm - (Replying to @snavarrol) Unfortunately, Heckman is still reluctant to teach his students (and many followers) how to derive a counterfactual from a toy model. See . Care to guess why? See . And because econ students are not rebelling (yet!) #Bookofwhy

12.15.18 @4:05pm - Not entirely "empirical phenomena". You rely on ignorability assumptions in observational studies. Where do they come from? Ans. A model which resides in your mind, and which you suppress when you only allow it to imitate RCT's, instead of deriving PO's. #Bookofwhy #EpiTweeter

12.15.18 @2:54pm - (1/n) Your observation is super insightful. One of the major communication obstacles I have encountered with potential outcome (PO) folks is the notion that causal effects and counterfactuals, Y_x, are PROPERTIES OF OUR MODEL. They cannot swallow it because, in the PO framework
12.15.18 @2:54pm - (2/n) of Rubin (1974) there is no such a beast as a MODEL. All he had were conditional probabilities of potential outcomes {Y(0),Y(1)}. Subsequently, those who entered #causalinference through a PO-education suffer from the same deficiency -- no model. #Bookofwhy. (cont....)
12.15.18 @2:54pm - (3/n) Instead, the potential outcome Y(1) is defined in the context of a real-life RCT, not a model of how a population responds to RCT. Hard-core PO folks, including Rubin's disciples in economics continue to operate in this model-blind conception. DAG-using Epidemiologists
12.15.18 @2:54pm - (4/n) have advanced towards model-hood, but not all the way. Since Robins and Greenland works (1986) were rooted in Rubin's PO, many epidemiologists today still view DAGs as tools for serving the RCT conception of potential outcomes, not as a mathematical object (cont.
12.15.18 @2:54pm - (5/n) that DEFINES potential outcomes. This explains why @_MiguelHernan depicts it as black magic when I assert that an ideal intervention is defined as a property of one's model. This conceptual barrier continues to impede communication until ..(YES)... a metamorphosis occurs...
12.15.18 @2:54pm - (6/n) Once a PO-schooled researcher learns to derive a counterfactual from a simple model he/she is liberated for life, never to return to the darkness of model-void. So, derive ONE today, as in #Bookofwhy eg, estimate Mary's salary had she not quit school

12.14.18 @10:18pm - (2/n) the utility of defining ideal mathematical constructs, and the practical benefits of attributing causal qualities to non-manipulable variables, from blood-pressure and temperature to gender and obesity. #Bookofwhy #EpiTwitter

12.14.18 @10:10pm - (1/n) Your suggestion to start a conversation was taken seriously in (2018),which provides rigorous analysis of each of your 2016 arguments. In particular, it analyzes the logic of "consistency," the difficulties of defining "well-defined interventions,"
12.14.18 @10:10pm - (2/n) the utility of defining ideal mathematical constructs and ideal manipulations, the semantics of multi-version interventions, and the practical benefits of attributing causal qualities to non-manipulable variables, from blood-pressure and temperature to gender and obesity.
12.14.18 @10:10pm - (3/n) The Appendix further examines ordinary conversations among health scientists, points out the ubiquity of non-manipulable causes and their communicational benefits. I have received no objection to any of these arguments and assumed their power and transparency convinced you
12.14.18 @10:10pm - (4/n) and your followers in #causalinference. I still believe they deserve serious considerations. #Bookofwhy #EpiTwitter

12.14.18 @5:24am - (Replying to @_MiguelHernan @dingding_peng) It seems that both of us are looking forward to a serious conversation. I suggest we start by agreeing on what a model is, for I need to make sure you do not really believe that: "Pearl believes that any causal effect we can name must also exist." Do you? #Bookofwhy #Epitweeter

12.13.18 @11:23pm - (Replying to @dingding_peng) Now that I read this article again, I find it hard to believe that anyone could adhere to that ill-defined notion of "well-defined intervention". Science seems to be making progress, even in the 21st century, when students are kept on short cultural leashes #Bookofwhy #EpiTweeter

12.13.18 @4:02am - (Replying to @IntuitMachine @raulincze and 2 others) Honestly, I am not sure any of the DL researchers grasps today what I was talking about then. Why? They are busy reaping the fruits of success. To understand the ladder one needs to take the time and solve ONE toy problem from A to B, eg Mary's income had she not quit college.

12.13.18 @1:20am - Is #Bookofwhy bashing statisticians? Carlos posting on Cross-Validated has been updated with two of my posts on-to-the-book-of-why/ Is anyone bashing statisticians or are they making it up?

12.12.18 @4:06pm - My final Confucius proverb: What do we tell a statistician, an economist or a ML researcher who insists on doing model-blind learning? Confucius says: "It is only by understanding how models should be used when we have them that we can learn how to live without them!" #Bookofwhy

12.12.18 @1:49pm - (Replying to @raulincze @dennybritz @Waymo) Beautiful symbiosis between a model of the world (map) and a local machine learning imitation. #Bookofwhy

12.12.18 @4:50am - Readers ask: "Why lonesome, and why elite force." Loneliness will strike when you discover that your editor, reviewers and even respected colleagues know very little about CI. Force will emanate from #Bookofwhy tools that enable you to see their weaknesses and do what they can't.

12.11.18 @11:44pm - (Replying to @prof_goldberg @BrookingsInst) As you might have heard, the #Bookofwhy is advancing standards for proof of causation that are grounded in observational studies, and could serve as alternatives to RCT's. These were criticized as endangering public safety. Would you say that they should serious be considered?

12.11.18 @6:11pm - (Replying to @malmyros) "And he whispered to me try the do-calculus....and the problem got solved in 3 steps." #Bookofwhy. Imagine what students are missing whose professors deprive them of the do-operator, not to mention do-calculus. They do not even know when to complain. #EpiTwitter #causalinference

12.11.18 @5:13am - Welcome to the elite force of the army of commonsense and, at the risk of sounding like a "sage" I will add that commonsense is a very lonesome road. Still, it is also fun, because history is on your side; my high-school teachers told me so, see end of

12.11.18 @2:58am - Thanks, Frank, for posting @DrewLevy illuminating review of #Bookofwhy. Even those who read the book cover to cover will find it valuable to recapture the spirit of the book and its key messages. I especially enjoyed "Why do the heathen rage?" on the psychology of the resistance.

12.10.18 @4:52am - (Replying to @sytelus @YosephBarash) I've found it, thanks for asking; nostalgia is a healthy emotion, makes you feel young and useful. "How to Do with Probabilities What People Say You Can't," was first presented in 1985 and then published in 1988. Enjoy. #Bookofwhy

12.10.18 @3:18am - (Replying to @HenningStrandin) Very nice example. It highlights several dimensions of "action" which I need to organize in my mind before answering. Will be back!. #Bookofwhy

12.10.18 @1:38am - (Replying to @Physical_Prep @_MiguelHernan @ChristophMolnar) Whose article did you love? No, I have not read yet the book by Molnar but, from his description, it seems that he takes interpretability to be machine's ability to show that it did what it was told, not why it was necessary to do. The latter requires a causal model. #Bookofwhy

12.10.18 @1:12am - (Replying to @bensprecher @GaryMarcus and 2 others) Pleading a chronic illiteracy to acronyms. But if GAN can help us predict the effects of banning cigarettes I would be curious to know the PRINCIPLE by which it extracts the needed information from data. Once we understand the principle, you can call me a "proud well-trained GAN"

12.9.18 @7:00pm - (Replying to @tdietterich @GaryMarcus @ylecun) And vice versa: "Symbolic causal models, in principle and in fact, provide grounding for the semantics of sensory data that current deep learning methods lack." Conclusion: "current methods" on both sides need be informed by the "current methods" of the other side. #Bookofwhy

12.9.18 @3:41pm - (1/2) (Replying to @tdietterich @GaryMarcus @ylecun) In my God little acre of the physical world, we do perceive and act on symbols, not on sense-data. For example, I am tweeting this text after I imagine its impact on readers, which is an exercise in symbolic representation. Note, I have never tweeted this mesg before, so...
12.9.18 @3:49pm - (2/3) (Replying to @yudapearl @tdietterich and 2 others) ... it could not possibly be stored in the arsenal of DL functions. Another way of looking at it, in #causalinfeence we seek a grounding for data, not for symbols; we want to find a real-world interpretation of sense data, not a summarization of data which DL gives us.

12.9.18 @3:13pm - The debate about manipulability is summarized in and is orthogonal to my latest tweet, which commends your nips talk and informs #causalinference students that a more refined taxonomy of causal questions yields great benefits

12.9.18 @1:01pm - (Replying to @jfeldman_epi @_MiguelHernan) #causalinference and #Bookofwhy is all about the idea that "theory matters". I am surprised you found nothing substantive there. I for one find only substance there. And I would not use the word "priors" which connotes prior probabilities; causal theories require more.

12.9.18 @1:46am - (Replying to @pfau @dileeplearning and 4 others) My personal journey into Belief Propagation is narrated here , a chapter we had to discard from #Bookofwhy for space considerations. But many find it educational, especially the story about Bill Gates.

12.8.18 @11:17pm - (Replying to @GaryMarcus @tdietterich @ylecun) I wish I could join this discussion but I can't parse Tom's "DL provides grounding for the semantics of internal representations that current symbolic methods lack." In my simple world, it is symbolic representation that provides semantics, not DL. What am I missing? My personal journey into Belief Propagation is narrated here , a chapter we had to discard from #Bookofwhy for space considerations. But many find it educational, especially the story about Bill Gates.

12.8.18 @10:19pm - (Replying to @NeuroStats @analisereal) Thanks Manjari for posting this paper. We should have mentioned Tukey in #Bookofwhy as a singularity (among statisticians) who stated publically that causation is NOT a species of correlation and that statistics can learn something from an outsider. Any link to the whole paper?

12.8.18 @9:51pm - Glad to see @_MiguelHernan carrying the torch of C-word to the lion-den of #NeurIPS2018. But if I were asked: Do all students of #causalinference agree with this hierarchy, I would say: NO! Lumping all of CI into one level creates confusion and worse See

12.8.18 @7:13pm - (Replying to @foil26 @melodem_group) External validity has been mathematized and algorithmitized. So, there is no longer any need to urge researchers to "pay attention to it"; it is an integral part of any scientific study. What remains is to educate social scientists in the logic of causal inference. #Bookofwhy

12.8.18 @3:44pm - I am retweeting this post in the hope that people would pause before re-chanting "Design trumps analysis", or at least explain what it means, and why we cannot create symbiosis between the two. #Bookofwhy #EpiTwitter

12.8.18 @2:37am - (1/2) (Replying to @stevesphd @eliasbareinboim and 2 others) In light of our ecumenical understanding, can we dispose of the slogan "Design trumps analysis" which was tweeted here as an uncontested, divine dictum? I first saw this phrase in Rubin's "For objective causal inference.." (Ann. App. Stat. 2008) and, ironically, following...
12.8.18 @2:42am - (2/2) (Replying to @yudapearl @stevesphd and 3 others) 2/2 .. this paper Rubin makes three glaring blunders for lack of analysis. Those who swear by the Dictum "Design trumps analysis" should take a good look at where contempt for analysis may take us. [The blunders are discussed in this underground paper ]

12.7.18 @10:33pm - (Replying to @jaimiegradus @EpiEllie and 4 others) Excited to see your DAG workshop spreading truth all the way to Boston U. Dont forget to carry a do-operaor with you, for protection -- a formidable weapon. #EpiTwitter.#Bookofwhy.

12.7.18 @4:25pm - I am grateful to readers for making the #Bookofwhy part of their thinking about science, about artificial intelligence and, almost forgotten, about cause and effect.

12.7.18 @10:20am - (Replying to @stevesphd @eliasbareinboim and 2 others) If you believe in a division of labor you would appreciate that after you think hard about a real-life problem with unknown DAG there is still some work to be done, i.e, to store the fruits of your thinking, to interrogate it, to combine it with data, etc. this is our challenge.

12.7.18 @8:20am - (Replying to @StuartReid1929) I happened to befriend Herbert Simon in the early 1990's. Believe me, you would not want me after talking to a real intellect like Simon. #Bookofwhy

12.7.18 @3:05am - (Replying to @stevesphd @eliasbareinboim and 2 others) Thanks for correcting me. I did not have an explanation why Imbens could not solve the 4 toy problems, so I blamed textbooks. The real reason is linguistic; asserting conditional exogeneity in the language of ignorability does not give us a clue whether it holds in a toy problem.

12.7.18 @1:45am - (Replying to @statnav @EpiEllie and 3 others) To this one can add my commentary on Dawid's paper, in which I labor hard to convince him that questions about Causes of Effects must invoke counterfactuals, and that, contrary to Rubin's teachings, this does not make them less scientific. #Bookofwhy.

12.7.18 @1:12am - (Replying to @eliasbareinboim @stevesphd and 2 others) This conversation with Imbens was really illuminating. It reveals how flexible IV analysts are to tiny variations on their textbook models. E.g.,One can tell right away that the idea of turning a bad IV into a valid one never appeared in IV textbook, hence the silence.#Bookofwhy

12.6.18 @6:39pm - (Replying to @Farial04) "The Curse of Free Will" appeared here , then in a chapter discarded from #Bookofwhy : . For introductory material, I must resist modesty, succumb to honesty, and strongly recommend , in fact, very strongly.

12.6.18 @11:15am - (Replying to @statnav @EpiEllie and 3 others) You will go down history as the first health scientist to openly admit the existence of an "obliged to follow convention", which is suboptimal, yet coerces authors to follow, so as to pacify reviewers, funders, recruiters and gurus. I call it "cultural roots" to be polite .....

12.6.18 @5:08am - (Replying to @statnav @EpiEllie and 3 others) I fully agree, with one exception. I find clarity, not cloud, in notational distinction. In a commentary on Dawid's legal papers I show how attention to notation protects us from wrongly articulating a target quantity. See also

12.6.18 @4:16am - (1/2) (Replying to @mendel_random @stevesphd @jwbelmon) You keep demanding "real-world" conclusions, and I argue for division of labor: You discover a new drug and I will try to analyse your method and the reasons for its success, and express it analytically, so that others could use it in a totally different domain. Is it a deal?
12.6.18 @4:26am - (2/2) (Replying to @yudapearl @mendel_random and 2 others) Along the same vein, note that THREE decades after the giants Cornfield and Hill made their contributions epidemiologists were still confusing "confounding" with collapsibility. Why? Because the giants did not have an analyst to generalize their thoughts in mathematical form.

12.6.18 @3:52am - (Replying to @stevesphd @mendel_random @jwbelmon) Good Ans. But IV too relies on the assumption of ignorability, plus exclusion. Rubin's dictum? My rebuttal still stands. Supertransparent? What other causal models have you analyzed? Most importantly, we need analysis to handle 12 interacting IV's, with or without dictums.

12.5.18 @11:42pm - (Replying to @mendel_random @stevesphd @jwbelmon) What part of current fixation with IV stems from cultural habit and populist mentality, as opposed to informed choice among alternatives? The latter requires that researchers study causal inference and then choose a method. Will they return to IV? I am eager to see. #Bookofwhy

12.5.18 @11:12pm - (Replying to @EpiEllie @robertwplatt and 2 others) I am not aware of a crispier separation between "causal effects we can estimate from data and those we can't" than the do-calculus. Are you? So you should be the first to ask: How does it work? What have I missed by ignoring it? Ask, and your "useful separation" will be granted.

12.5.18 @9:52pm - (Replying to @jon_y_huang @robertwplatt and 3 others) Thanks for the papers. I have (delinquently) skipped this literature because I could not find a simple interpretation of the quantity being estimated, rNDE; I am still unable to find it in the papers. Perhaps you have one? Indeed its one way of approximating R-3 by R-2 substit.

12.5.18 @7:05pm - (Replying to @EpiEllie @robertwplatt and 2 others) Sounds crisp, assuming readers understand what Rubin meant. Your note implies there are counterfactuals that are NOT potential outcomes, and they are "used", not "avoided". Next step is to welcome notational distinction between them, and you are ahead with do-calculus - Welcome!

12.5.18 @6:43pm - (Replying to @jon_y_huang @robertwplatt and 3 others) I have not heard of "stochastic mediation", any pointers? Rung-3 is not there to "avoid" but to welcome and use under full awareness of assumptions and available tools. "Avoidance" sounds counterproductive.

12.5.18 @5:40am - (1/3) (Replying to @eddericu @EpiEllie and 2 others) Thanks for illuminating the Rung-2/3 discussion with pages from "PRIMER" . Indeed, Rung-2 contains queries that can be estimated by various "doors" while some Rung-3 queries (e.g., NDE, PN,PNS) may be non-estimable by ANY graphical means. Moreover,..
12.5.18 @5:44am - (2/3) (Replying to @yudapearl @eddericu and 3 others) ...Rung-2 queries are GUARANTEED to be estimable from experimental studies, not so RUNG-3; some of its queries will forever remain beyond experimental reach (on populations) and require extra scientific knowledge on individual behavior (eg. structural equations). We can see..
12.5.18 @5:50am - (3/3) (Replying to @yudapearl @eddericu and 3 others) ..the significance of reducing a query Q to a do-expression. Once we do this, we are guaranteed (1) Q is experimentally estimable, (2) Q is observationally estimable if do-calculus says so, (3) If it says so, it tells you how. Teaching from PRIMER makes everything clear. Thx

12.5.18 @6:24am - (Replying to @stevesphd @mendel_random @jwbelmon) Transparency means being able to judge whether your formal assumptions are plausible, given your knowledge of biology. Sensitivity analysis cannot repair misjudgment due to lack of transparency, nor can it test the veracity of the assumptions or their compatibility with data.

12.4.18 @1:49pm - (Replying to @EpiEllie @eddericu and 2 others) Everyone distinguishes these two beasts in one's mind, as well as in practice. The question is do we distinguish them in the mathematical notation? e.g., in the research question that we write down? The current confusion would have been avoided by good notation. eg, do(x) vs. Y_x

12.4.18 @5:17am - (Replying to @jwbelmon @mendel_random) MR is the application of IV method to problems in which genes act as instruments. As such, MR is part of CI, and it rests on untestable causal assumptions that can be seen in the DAG by those who understand graphs and are taken at faith value by those who do not. #Bookofwhy

12.4.18 @4:02am - A recent emotional attack on #Bookofwhy and the idea that observational studies can sometimes be more trustworthy than RCT (e.g., when selection bias is unavoidable) behooves me to re-tweet my answer (below) to the question: "Is RCT still our gold-standard?"

12.4.18 @3:46am - (Replying to @JDHaltigan @mendel_random) I believe the jury is still out, for we need to factor in " the additional virtue of observing people's behavior in their own natural habitat instead of a laboratory" as stressed in #Bookofwhy and ignored by its attackers. See a serious debate here:

12.4.18 @1:21am - (1/2) (Replying to @EpiEllie) The distinction between Rung-2 and Rung-3 is not between "the observed and the counterfactuals," but between hypothetical interventions and hypothetical counterfactuals. @_MiguelHernan paper ... for example, classifies tasks into the following tricotomy:
12.4.18 @1:32am - (2/2) (Replying to @yudapearl @EpiEllie @_MiguelHernan) 1) description, 2) prediction and 3) counterfactual prediction lumping interventions together with counterfactuals, both expressed in subscripts (not distinctly using Y_x and do(x)). I am glad you rose to the defense of the distinction; the next step: show it in notation.

12.3.18 @2:08am - (1/3) Readers ask: Why is intervention (Rung-2) different from counterfactual (Rung-3)? Doesn't intervening negate some aspects of the observed world? Ans. Interventions change but do not contradict the observed world, because the world before and after the intervention entails ...
12.3.18 @2:09am - (2/3) ... time-distinct variables. In contrast, "Had I been dead" contradicts known facts. For a recent discussion, see Remark: Both Harvard's #causalinference group and Rubin's potential outcome framework do not distinguish Rung-2 from Rung-3. (cont.)
12.3.18 @2:09am - (3/3) This, I believe, is a culturally rooted resistance that will be rectified in the future. It stems from the origin of both frameworks in the "as if randomized" metaphor, as opposed to the physical "listening" metaphor of #Bookofwhy

12.2.18 @7:34pm - (Replying to @junzhez @faoliehoek and 3 others) Interesting. Do you know a paper where the two representations are shown side by side on a toy example?

12.2.18 @4:45pm - (Replying to @eliasbareinboim @alexdamour and 2 others) It reminds me of Larry Wasserman's comment: "It is my impression that the "graph people" have studied the Rubin approach carefully while the reverse is not true." (2014, Gelman's blog). It is my impression (JP) that they are producing a 3rd-generation model-blind PhDs. Non-Stop.

12.2.18 @4:11am - (1/5) For readers wishing to sharpen their understanding of RL and CI, recall how we answered the question: "How can RL optimize policies without a causal model?". We said: RL is on Level-2 of the Ladder because it receives data from controlled interventions. #Bookofwhy
12.2.18 @4:11am - (2/5) Indeed, the model created by such control dictates a simple estimand for the query of interest -- the action-outcome association in the bare data; no adjustment is necessary. (Same as RCT). We also remarked that this model-blindness is acquired at a cost: evaluated policies ...
12.2.18 @4:11am - (3/5) ... must deploy the same action-set as the one used in training. A causal model is needed to go beyond, i.e., to actions not used in training. An interesting observation by by Elias notes that a model would be required even if we want to narrow down the action-set. ...
12.2.18 @4:11am - (4/5) For example, suppose the action-set used in training was {A1,A2} and we wish to evaluate, without retraining, a new policy over a smaller set {A1}. This may be needed either because physical limitations prohibit the activation of A2, or because the optimal policy requires ...
12.2.18 @4:11am - (5/5) .. requires that A2 be left uncontrolled. The tools of CI permit us to evaluate policies with either expanded or narrowed action-sets. They invoke do-calculus to convert the target policy to a data-compatible expression, as shown here #Bookofwhy

12.1.18 @1:56am - (1/2) (Replying to @CsabaSzepesvari @eliasbareinboim and 2 others) Testability is a property of the assumptions, not of how they are articulated. There is even a symbolic logic that unveils it . However, since ignorability is posited by a human scientist, what counts is the cognitive ability to recognize it in ...
12.1.18 @1:56am - (2/2) (Replying to @yudapearl @CsabaSzepesvari and 3 others) ... in one's model of reality, eg, the DAG. Here, Elias' example provides a glimpse into what it takes for an unaided human to do. Don't skip it. It is important to understand what darkness our PO colleagues are working under when they write (publish) "assuming ignorability"

12.1.18 @11:21pm - (Replying to @EricTopol @carlzimmer and 8 others) @EricTopol, Wondering: what has elevated my humble book #Bookofwhy to the Parthenon of the immortals? This is a mortal "what", not a "why" question.
12.1.18 @11:21pm - (2/3) (Replying to @yudapearl @CsabaSzepesvari and 2 others) things, and how. Folks in CI have spent their energy on those other setups, and have gotten a formal handle of the when/how question. To give you a glimpse, suppose the action set is {A1,A2} and we wish to evaluate, without retraining, a new policy over a smaller set {A1}.
12.1.18 @11:21pm - (3/3) (Replying to @CsabaSzepesvari @eliasbareinboim @mendel_random) [The shrunk set may be the optimal] a causal model is needed to do the evaluation and we can tell you what kind. The same goes for augmenting the action set to {A1,A2,A3}, again without retraining. You can see the potential. #Bookofwhy

12.1.18 @10:18pm - (1/2) (Replying to @CsabaSzepesvari @eliasbareinboim @mendel_random) Agree. One of the reasons RL literature focused on estimation, not identification, is that it was not needed in engineered setups, where one had full control over the action-set during learning. RL folks are probably wondering now in what setups, if any, they need to change.
12.1.18 @10:23pm - (2/2) (Replying to @CsabaSzepesvari @eliasbareinboim @mendel_random) things, and how. Folks in CI have spent their energy on those other setups, and have gotten a formal handle of the when/how question. To give you a glimpse, suppose the action set is {A1,A2} and we wish to evaluate, without retraining, a new policy over a smaller set {A1}.

12.1.18 @4:16pm - (1/2) (Replying to @CsabaSzepesvari @eliasbareinboim and 2 others) But if the DAG is the source of our knowledge, why not articulate knowledge using easily verifiables features of the DAG, eg, missing edges, and get testability for free. Connditional ignorability of treatment is nasty beast for the unaided intellect.
12.1.18 @4:14pm - (2/2) (Replying to @CsabaSzepesvari @eliasbareinboim and 2 others) Indeed, the test you suggest is what one would have to go through if one insists on representing knowledge in ignorability language. It is like solving an equation 3x-4= 2x+1 by trying out all values of x until equality is established. #Bookofwhy

12.1.18 @6:35am - (Replying to @HenningStrandin @mendel_random) Different views. This is the key. And this is what I am waiting for when people say "there is room for different views". Mention some. @Mendel_random tried and all he could mention were variants of SCM. What else? religion? Give me a candidate and you will see openmindedness galo

12.1.18 @6:25am - (Replying to @mendel_random @HenningStrandin) came into being ecause epidemiologitswere able to conduct their studies clear headed, unperplexed by puzzles like confounding, mediation, and selection bias. Recall, it took decades after the death of alchemy for modern chemistry to discover new elements. #Bookofwhy

12.1.18 @6:20am - (Replying to @mendel_random @HenningStrandin) The control of confounding, mediation, external validity and selection bias are extremely powerful results. I know, I know, you are waiting breathlessly for new drugs and new medical discoveries. They will come, and when they do, it would be hard to tell how many of them 1/2

12.1.18 @4:03am - (Replying to @HenningStrandin @mendel_random) On the contrary. Understanding the revolution and internalizing the power it unleashes would behoove you to join it -- you would not have time to do anything else. Look at all the powerful results people obtained lately. Cults? You are kidding! What have they produced ?

12.1.18 @2:52am - (Replying to @HenningStrandin @mendel_random) Constructive commentaries are always valid. They unveil where one's mindset and priorities are; recognizing and leveraging the potentials of the causal revolution, or hagglingwhen it started or who deserves less credit for its happening. #Bookofwhy

12.1.18 @1:01am - (1/4) Replying to @eliasbareinboim @alexdamour and 3 others) Elias example demonstrates vividly that there are objective criteria for transparency and, more generally, there are objective criteria for preferring some models over others. Potential outcome (PO) enthusiasts have been avoiding such demonstrations like a plague
12.1.18 @1:05am - (2/4) (Replying to @eliasbareinboim @alexdamour and 3 others) for the past two decades, and pluralists have accused me of monotheism for presenting them. But readers of #Bookofwhy expect me to explain honestly why I prefer graphical models over alternatives. So, please examine Elias example seriously and, in addition to answering
12.1.18 @1:07am - (3/4) (Replying to @eliasbareinboim @alexdamour and 3 others) whether the model has testable implications, try to determine if the set of assumptions are consistent, whether they are compatible with any story you know and, if so, whether they are sufficient for describing the story. It is examples of this sort that prompted
12.1.18 @1:14am - (Replying to @eliasbareinboim @alexdamour and 3 others) the publication of , which highlights three incurable deficiencies of PO [below] and questions whether PO, void of graphs is truly "an approach" or merely dismembered component of an approach. #Bookofwhy

11.30.18 @9:28pm - (Replying to @alexdamour @eliasbareinboim and 2 others) I agree with Elias. One of the most common questions I get from students and newcomers is: What are the strengths & limitations of each notation system & how do we combine them? The answers are not entirely subjective, but the objective component is lost once people get annoyed.

11.30.18 @8:26pm - (Replying to @alexdamour @eliasbareinboim and 2 others) DAGs and Ignorability statements are derivatives of SCM (Structural Causal Models) so, saying that they can be used together amounts to saying that several derivatives of SCM can be used together. No contest! Indeed, for beautiful symbiosis see #Bookofwhy

11.30.18 @8:00pm - (Replying to @mendel_random) I answered 7h ago. Good point! " its womb" should be: "returning my fascination with causality to its womb in AI". It will be corrected in the next printing. I reiterate my comment that understanding the revolution means joining it, not knit picking from the side.

11.30.18 @5:01pm - (Replying to @CsabaSzepesvari @eliasbareinboim @mendel_random) We still need to clear two issues. 1) translation to causal graph is NOT ONE'S SPORT. It is going to where knowledge resides, which is a MUST when knowledge is provided by humans. 2) Some say "RL literature" is "everything", it should be replaced by the specific task analysized.

11.30.18 @3:16pm - (Replying to @alexdamour @eliasbareinboim and 2 others) Reviewers suggest confounders if (1) the paper is about a specific domain and (2) they know how to translate "ignorability" to DAG condition. But if the paper is methodological, starting with "assume ignorability" means one thing: "I want to deal with statistical estimation only"

11.30.18 @1:21pm - (Replying to @mendel_random) Good point! The phrase " its womb" should read: "returning my fascination with the causal revolution to its womb in AI". It will be corrected in the next printing. I reiterate my comment that understanding the revolution means joining it, not picking from the side.

11.30.18 @12:56pm - (Replying to @eliasbareinboim @alexdamour and 2 others) I think the notion of transparency will become transparent if we take an extreme case of an author stating "I assume identifiability" instead of ignorability". It is equivalent to wishing the problem to belong to the huge class of dags that permit identification. But...But..

11.30.18 @5:27am - (Replying to @CsabaSzepesvari @eliasbareinboim @mendel_random) I think by "assumptions are hidden" Elias (and me) tried to say that, even if you declare 10 times that you "assume conditional ignorability" the assumptions are still hidden because no mortal can decide if they hold or not, even in toy problems. I have tested famous mortals too

11.30.18 @2:39am - (1/3) (Replying to @deaneckles @KordingLab @mioana) Notwithstanding 3 inaccuracies, presenting these methods using DAGs is a great service to neuroscience. Seeing the assumptions so vivid, readers will never go back to standard presentations, crafted for graph-averse audience -- an endangered species. @chethan @sweichwald
11.30.18 @2:46am - (2/3) (Replying to @deaneckles @KordingLab @mioana) 2/2 Viewing your RDD model, students of causality would immediately say: "Hey! this is a front-door model, so why not do it in full generality, free of assumptions of threshold and linear regression?" Indeed, why not? Detecting similarity among problems is graph's specialty.
11.30.18 @2:58am - (3/3) (Replying to @deaneckles @KordingLab @mioana) While DAGs do not display functional properties, like monotonicity, they display assumptions that other representations have difficulty with, e.g., exclusion, or exogeneity, not to mention conditional exclusion and conditional exogeneity. Economists are beginning to dig it.

11.30.18 @12:52am - (Replying to @mendel_random @eliasbareinboim) #Bookofwhy designates 1920 as year-zero for "the causal revolution"; the year when S. Wright dared express a causal assumption in mathematical form. Researchers who appreciate the magnitude of this step are busy reaping the benefits of the revolution, not picking on priorities.

11.29.18 @3:54pm - (Replying to @mendel_random) science benefits from a division of labor. Some work on generating and substantiating "background knowledge", some work on representing and utilizing that knowledge once it is substantiated and some, luckily a minority, engage in trivializing what others are doing. #Bookofwhy

11.29.18 @9:57am - (Replying to @theophaneweber) Is this the reason why @Bookofwhy suddenly jumped to # 1 on Amazon best seller list in neural networks?

11.29.18 @9:40am - (Replying to @decodyng) 90% of scientific writing is trying very hard not to call in causal modeling.#Bookofwhy #causalinference

11.28.18 @10:24pm - (Replying to @Emaasit @akelleh and 7 others) Seminal papers may be extremely time consuming to read. If you are in compiling mood, of an immediate benefit would be a map of sub-problems and their associated tools, cast in one unifying vocabulary. #Bookofwhy

11.28.18 @12:39pm - (Replying to @alex_peys @CsabaSzepesvari and 8 others) The phrase "We know how actions are generated" needs to specify what actions are those; in training or at the target policy evaluated. If actions are generated differently in these two regimes we have a neat transportability problem on our hand which needs be solved, not avoided.

11.28.18 @2:00am - (Replying to @fhuszar @eliasbareinboim @CsabaSzepesvari) Counterfactuals are never terribly difficult to describe. If they were then causal inference would be difficult, but it is easy. I remember the Botou paper and classified it as an instance of process control problems (Causality pp. 74-76). Again, not hard to describe or solve .

11.27.18 @11:41pm - (Replying to @CsabaSzepesvari @junzhez and 8 others) The phrase: Imitation learning, human acts, robot should copy" triggered an old paper (Fig. 14 and 15), which I bring here only to show how problems like that should be formulated. Nothing is left to guessing; from target quantity, to information available

11.27.18 @10:21pm - (Replying to @RealUrso) Sure! It is all here

11.27.18 @9:56pm - (Replying to @fhuszar @eliasbareinboim @CsabaSzepesvari) It is not drawing or not drawing the model that counts but whether you correctly interpret the assumptions in the model and what you do with them. I am waiting for Csaba to clarify whether his graph represents the engineer or mother nature. Each leads to a different conclusion.

11.27.18 @9:42pm - (Replying to @CsabaSzepesvari @AIforHI and 6 others) I think we are heading here towards some mutual understanding. Can you exemplify the engineering problem in which "we know how actions are generated"?? Just describe the setup in which "we know" and, then, what we can conclude from this knowledge, and what the final result is.

11.27.18 @4:57pm - (Replying to @GaryMarcus @ylecun and 6 others) Appropos. I am sharing my interview with Martin Ford, and, as you can see, I am still hoping for ML folks to liberate themselves from data-centricity and other mind-suppressing substances #Bookofwhy

11.27.18 @1:44am - Having noticed that my twitter account swelled to 12K followers and 640 tweets, I created a searchable file of all my tweets . I find it useful for retrieving old discussions by topics, and digging out old gems. You may also find it useful. #Bookofwhy

11.27.18 @1:28am - Good News! #BookofWhy has survived the paper shortage and is now back on the shelves. Moreover, it contains all revisions and corrections detected by October 27. Make sure you get the 6th printing (#6 on last line of copyright page). May your reading be enlightening and rewarding

11.26.18 @1:17am - (Replying to @AIforHI @dustinvtran and 6 others) A question from a future partner. To most folks in causality research the words "assuming ignorability" mean stripping a problem from its causal content and solving a standard statistical problem instead (#Bookofwhy page 283). Must you really assume that?

11.25.18 @4:05pm - (Replying to @CsabaSzepesvari @junzhez and 8 others) This feeling is mutual. Plus, I know that RL is saturated with creative researchers who can recognize powerful ideas directly from a tool-map, before building an ant-eating robot. I have unveiled my tool-map is here: and I do not charge rental #Bookofwhy

11.25.18 @2:28pm - (Replying to @junzhez @CsabaSzepesvari and 8 others) HMMM, but if this is the case, where is the "learning"? It sounds more like "copying". Are the environments different?

11.25.18 @1:37pm - (Replying to @CsabaSzepesvari @eliasbareinboim and 7 others) We are on the same page indeed, and the next step is to draw a more complete map of problem types and tools available as I have started to do in the last re-tweets. I hope we are in agreement on my initial draft, mapping problems and tools, not labels. #Bookofwhy

11.25.18 @12:20am - (1/4) (Replying to @CsabaSzepesvari @eliasbareinboim and 7 others) I used to tell my students: "How do you find out if your ideas are worth a damn? When your colleagues start saying they used them all along, they just never called them like that". Today I tell students "We made it!" Csaba says: "RL is CI, we just never called them like that"
11.25.18 @12:35am - (2/4) (Replying to @yudapearl @CsabaSzepesvari and 8 others) Reiterating the relationship between RL and CI. Rl has a causal component, provided by the actions used during policy learning. In contrast, CI starts with no actions, just observations, and tries to infer effects of future policies. This task was not addressed by RL so far.
11.25.18 @12:47am - (3/4) (Replying to @yudapearl @CsabaSzepesvari and 8 others) In recent years, CI research embraced new tasks. (1) What if we have both actions AND passive observations, (2) What if we have actions and observations and we wish to predict effects of new policies, not those used in training, (3) many more combinations. What about RL ?
11.25.18 @1:03am - (4/4) (Replying to @yudapearl @CsabaSzepesvari and 8 others) Evidently RL folks were not interested in those other tasks (no hard feelings) and have not developed the logic needed for their solution. We are told they could have done so if they really wanted, which I do not doubt. A prerequisite however is to learn that new logic.

11.24.18 @1:43pm - (Correcting a crucial typo: "development in CI (causal inference) are way beyond "theoretical discussion", not MI. Indeed, these developments are shining in proofs and demonstrations and, most importantly, in the minds of the intellectually curious. #Bookofwhy.

11.24.18 @1:24pm - (Replying to @alexdamour @dustinvtran and 7 others) The causal revolution is happening already, with or without me. I am glad that ML folks are waking up to ask: "Are we missing the boat?" No, you are not. Because developments in MI are way beyond "theoretical discussion"; shining and ready for enlightened researchers to catch up.

11.24.18 @4:05am - (Replying to @alexdamour @dustinvtran and 7 others) I checked Google recruitment ad - no "causal inference" mentioned. UCLA graduates 1/3 PhD per year in CI. Who will shape future education in this country if they go work for Google? AI counts on the smart scientists at Google to foresee where the next revolution is heading

11.24.18 @3:42am - (Replying to @dustinvtran @eliasbareinboim and 6 others) The methodological revolution in health and social sciences should make any foresighted scientist jump into generating this preliminary evidence, not wait for it to show up in the NYT. Especially in a fad-driven era where 99% of PhD's hear nothing about CI. #Bookofwhy

11.24.18 @3:25am - (Replying to @stuz5000) Caution. CI is not an instant of probabilistic graphical models, but of causal graphical model. The difference is critical. The former cannot reason about interventions. See latter of causation #Bookofwhy

11.24.18 @3:08am - (Replying to @viettran86 @larosaandrea and 8 others) There are no modeling assumptions needed for RCT, except the randomness of the randomizing coin.

11.24.18 @3:01am - (Replying to @learnfromerror) Lord paradox is a causal problem. Each statistician proposes a method of estimating the causal effect of diet on weight gain. To discuss the notion of "valid" we must invoke causal vocabulary. We cannot do it in the vocabulary of variances and covariances no matter how intricate.

11.24.18 @1:00am - (Replying to @CsabaSzepesvari @eliasbareinboim and 6 others) Are you protesting the notion that RL has limits? or perhaps that RL (as practiced) is different from CI? or from AI?.. I am willing to declare CI a tiny subset of RL, if this would entice 5% of RL researchers to adopt 5% of the tools developed in CI, eg.,

11.23.18 @6:15am - (Replying to @ukmlv) Instrumental variables are treated in #Bookofwhy page 249, in the context of John Snow's investigation (1854) of the cause of Cholera.

11.23.18 @4:53am - (Replying to @PrannayKhosla @dustinvtran and 7 others) Good link (I hope) to solving toy problems in causal analysis: . #Bookofwhy. #causalinference. The key: start with what you know and act lazy, let mathematics do the rest.

11.23.18 @2:32am - (Replying to @c_prohm) No way !!! Both RL and CI seek actions and policies. CI is a bit more open-minded in that it leverages extra-data knowledge to infer the effects of new policies, while RL (thus far) has excluded such knowledge, at the cost of being "knob-blinded". #Bookofwhy

11.23.18 @2:21am - (Replying to @dustinvtran @eliasbareinboim and 6 others) "Strong empirical success" will appear when Google starts advertising for "PhD in ML with in-depth understanding of causal inference." Education comes first, empirical success, second. How many ML Phd's can solve the toy problems in ?? or even #Bookofwhy ?

11.23.18 @12:44am - The most common example is IV. We train by randomizing the instrument Z, and we wish to infer the effect of action do(X) that was not accessible in training. For that we need a causal model, assuring us that Z is not directly affecting Y (the outcome), an untestable assumption.

11.22.18 @11:33pm - The relation between RL (Reinforcement Learning) and causal inference has been a topic of some debate. It can be resolved, I believe, by understanding the limits of each. RL authors call this limit "off-policy". I like "off-knobs", imagining a machine with finite number of knobs.

11.22.18 @11:18pm - (Replying to @eliasbareinboim @jasonhartford and 6 others) Is RL an exercise in causal inference? Of course! Albeit a restricted one. By deploying interventions in training, RL allows us to infer consequences of those interventions, but ONLY those interventions. A causal model is needed to go BEYOND, i.e., to actions not used in training

11.22.18 @10:57pm - (Replying to @jasonhartford @eliasbareinboim and 6 others) By invoking interventions, RCTs provide us causal information that can be used, of course, in a restricted causal inference, i.e., inferring consequences of those interventions. A causal model is needed to go BEYOND, and infer consequences of actions that were not randomized.

11.22.18 @6:42pm - (Replying to @jtrecenti @analisereal @shell_ki) AN interesting question. To answer it I would need more grounding, because different statisticians mean different things by the "likelihood principle". Can you ground it in the form of a claim about two variables, X and Y.

11.21.18 @5:34pm - (Replying to @dustinvtran @zacharylipton and 5 others) I am not arguing, nor judging, just trying to appreciate what you do. Thanks for explaining, rather than sending me to "all the papers of XYZ". Given that you start with a model, do you compute an estimand before processing data? If not, what guidance is provided by the model?

11.21.18 @4:51pm - (Replying to @zacharylipton @eliasbareinboim and 4 others) I would be very curious to know how "all the recent papers" escape a theoretical impediment. We need conceptual guidance from someone who read them (who?) Do they "assume ignorability" like Imbens&Co.? Do they invoke a causal model? Do they discover equivalence model classes?

11.21.18 @6:53am - (Replying to @amine_ouazad @shell_ki) Keep going? And spoil the fun for readers who enjoy finding out for themselves how controversial questions get resolved by simple definitions? Sure, ignorability is causal and it's implied by backdoor (despite 2 pages of hand-waving in Imbens-Rubin book.) Now try PS for more fun.

11.21.18 @6:39am - (Replying to @gchierico @MeaningLifeTV and 2 others) I also got lost in this conversation. Compatibilists should focus their arguments on one question: How would I program a robot to have a sensation of free-will and harness this sensation to exhibit moral behavior. Once the conversation quits the robot, it gets lost in metaphysics

11.21.18 @5:06am - (Replying to @DavidAOliverJr @shell_ki) Confounding? Try to express it as a property of the joint distribution. Generations of epidemiologists and philosophers tried (#Bookofwhy), some economists are still trying, and potential outcomes seduce them a PC criterion called "ignorability" that no mortal can apply.

11.21.18 @4:53am - (Replying to @amine_ouazad @shell_ki) Right! But explaining this to an economist is a lifetime endeavor. And, if you try, you create an army of outraged enemies; from "How dare you ...?" to "We knew it all along...!" #Bookofwhy

11.21.18 @4:46am - (Replying to @shell_ki) But saying "they are all causal, because I use them" does not allow you to trace back the assumptions and distinguish concepts that can be inferred from data from those that rest on assumptions outside the data. E.g., generations believed "confounding" is testable from data.

11.21.18 @4:30am - (Replying to @IntuitMachine @shell_ki) We need it in order to exclude totally irrelevant sentences from the label "causal". For example, "Cinderella's hair is blue", which cannot be verified from the distribution of the observables, yet would hardly be deemed "causal" in a scientific context. #Bookofwhy

11.20.18 @11:54pm - (Replying to @yudapearl @shell_ki) The many "likes" received by this simple definition jolt me to entertain readers with a few examples: Which of the following is a causal concept? Spurious correlation, Granger causality, Confounding, Endogeneity, Control for (eg., age), Randomized, Instrument, Propensity score,

11.20.18 @6:04am - (Replying to @shell_ki) Contrary to expectations, the definition of "causal modeling" is fairly easy to articulate. To me, "causal model" is a set of assumptions about the data generating process, which cannot be expressed as properties of the joint distribution of observed variables. #Bookofwhy

11.20.18 @3:40am - (Replying to @DToshkov Not necessarily. Prospective "causal effects" can be defined and identified at the population level, without ever thinking about an individual. E.g., (assuming RCT), #(cured under treatment-1) vs. #(cured under treatment-2). This is level 2, invoking only do-expressions, no PO.

11.20.18 @2:20am - (Replying to @DToshkov) "Counterfactual" is short for "contrary to facts." If we take this definition as a guideline, there is nothing counterfactual in RCT. Inferring Joe's behavior from Jane's is not contrary to fact, because there is no logical contradiction between Joe's and Jane's treatments.

11.20.18 @12:21am - (Replying to @yudapearl @DToshkov) blindness and opacity in the past 2 decades (See the "three incurable bullets" of ). I am concluding with a question: How do you shake a framework from stagnation, if not by a bold an honest rhetorical strategy? #Bookofwhay #causalinference

11.20.18 @12:16am - (Replying to @yudapearl @DToshkov) 4. True, "the potential outcomes (PO) framework has been central to such concerns", and the #Bookofwhy (p.269-) salutes Rubin's achievement in chapter and verse. But "concerns"do not produce progress, and the centrality of PO did not prevent it from stagnating into blindness

11.20.18 @12:15am - (Replying to @yudapearl @DToshkov with those concerns? Virgil (29 AD) was also "concerned" when he said: "Lucky is he who understands the causes of things". Now what? I did not find anything about omitted variables developed in 1994 that was not known to Duncan in 1970. 4. Now to potential outcomes (PO) ...`

11.20.18 @12:12am - (Replying to @yudapearl @DToshkov) Duncan, Blalock, Goldberger) who adopted structural equations. Not very charitable to those who abandoned the causal interpretation of structural equations in the 1980's. Moreover, what good is it "to be concerned" with causal questions if you cant develop a methodology to deal

11.20.18 @12:07am - (Replying to @yudapearl @DToshkov) 2. Glad we agree that statistics deserves a rhetorical buff of neglect, though most statisticians would plead "not guilty" since "I have been concerned with causal questions all my life". 3.#Bookofwhy is full of admiration of early social scientists (Burks, Simon, Haavelmo,

11.20.18 @12:04am - (Replying to @yudapearl @DToshkov) It does not tell us anything about whether joe, who recovered under drug would have recovered without it. It is unfortunate that this critical distinction between the intervention and counterfactuals is blurred in most textbooks - a cultural neglect. #Bookofwhy #causalinference

11.19.18 @11:57pm - (Replying to @DToshkov) Appreciating your recommendation of #Bookofwhy and defending my rhetorical strategies. 1. Can we have interventional knowledge w/o counterfactual knowledge? Yes, an RCT provides one. It tells us how likely joe is to recover with and without a drug, but no more.

11.19.18 @4:10am - (Replying to @PHuenermund @PogrebnyakE and 8 others) The glory of this invention and the pain of the wrong turns are narrated here . The #Bookofwhy briefly decries, and mostly repairs the latters.

11.18.18 @12:04am - (Replying to @GoAbiAryan @nandanpc Causal utterances do not have parallels in classical logic. ~X->Y is a logical formula and stands for: If we find ~X we can conclude Y, which is not the same as "if we ban".

11.18.18 @11:00am - (Replying to @GoAbiAryan @nandanpc) Not really. You cant have causal words such as "affect" in an associational sentence. And you cant have counterfactual phrases such as "if there were no" in an interventional sentence. See examples on the Ladder of Causation in #Bookofwhy

11.17.18 @11:43pm - This AAAI-why-19 workshop promises to be unique. The interface between machine learning and causal inference will be given an informed, in-depth examination, guided by the Ladder of Causation and #Bookofwhy

11.16.18 @7:36pm - (Replying to @Lizstuartdc @Megtron9 and 3 others) This new paper on generalizability identifies three incurable limitations of the potential outcomes approach: . Questions about study-population vs. target-population would yield immediate answers when cast in graphical language, e.g., #Bookofwhy (p.354).

11.16.18 @8:25am - (Replying to @statwonk @0xeinar and 2 others) I would love to agree on everything, but I need to understand what the grievances are: "blindspots" "systematic errors", "land grab""ridicuouls claims" "absurd" "classic case" "pushing back". What is all this commotion about? Can we first do things right and then debate sources?

11.16.18 @5:38am - (Replying to @0xeinar @johnmyleswhite @statwonk @Oseinar), I have had the same question in mind. I tried to enter and calm down the emotional discussion about my land grab crimes, but I could not quite make out what the grievances are. #Bookofwhy

11.16.18 @5:28am - (Replying to @stephensenn) Your example resembles Wright's guinea-pigs diagrams. Can we draw one? I am having difficulty understanding what affects what and, more seriously, what the research question is. (how can "variance" be a "source of variation"). Dempster never believed in causal analysis -it shows.

11.15.18 @3:11am - (Replying to @stephensenn) What does it mean "overlooked"? Does it mean that some results in #Bookofwhy are wrong? or that more opportunities are available if we take advantage of hierarchical error structures? I would love to see how, but I am handicapped in my reliance on toy examples. Can you share one?

11.15.18 @2:16am - (1/2) I find this discussion rather illuminating, for it faithfully depicts the state of mind of 21st Century statistics.. A guy writes: "regression model is just a causal models," and then complains: "Why is Pearl bashing statistics?" Many rush to his defence because, they too, .

11.15.18 @2:16am - (2/2) when they do regression, they really think causation. Bashing does not help, because: "How can a whole glorious field be so wrong for so long? Pearl must be missing something." My hope is that #Bookofwhy reaches statistics students before it is banned as undesirable.

11.14.18 @1:12am - (Replying to @Megtron9 @eliasbareinboim and 3 others) Another discussion of "Generalizing experimental findings," highlighting the basic limitations of ignorability-based thinking, is: . It also distinguishes transportability from selection bias. #Bookofwhy @causalinference

11.14.18 @12:15am - In view of increasing attention to problems of generalizability, I dare claim that there is no generalization without encoding differences, i.e. selection diagrams (#Bookofwhy p. 354). For a recently posted summary and common confusions in the literature:

11.12.18 @12:23pm - Remember Lord's paradox (#Bookofwhy, chpt 6)?? Stephen Senn just posted a "statistical" solution here: aradox-guest-post/ to which I have added a commentary showing how it ought to be solved in our millennium. Enjoy the contrast. #causalinference

11.11.18 @1:12am - (Replying to @VoxBec) Glad to have you join the army of commonsense; there are many challenges ahead! #Bookofwhy #causalinfernce

11.10.18 @6:26pm - (Replying to @kerinalthoff @ProfMattFox) My mini-advice: Assure epidemiologists that they can now express their research questions in their mother tongue, the language of cause and effect; they no longer need to deform their questions to fit the molds of outdated mathematics. #Bookofwhy #causalinference

11.10.18 @2:57pm - (Replying to @jwbelmon) Regression analysis is the opiate of the masses: Think dirty, act clean and believe no one would notice. #Bookofwhy #causalanalysis

11.10.18 @1:12pm - (Replying to @juli_schuess) Thanks for sharing your teaching experience. #Bookofwhy

11.10.18 @10:50am - (Replying to @jwbelmon) Are you suggesting that #bookofwhy is not serious? That it does not offer useful tools for making medical decisions? That there is such a thing as "statistical causality" which offers alternative tools? Anxious to hear what you would rather see that is not in #bookofwhy?

11.10.18 @12:25am - (Replying to @Research_Tim) The word "approaches" suggests that we have a choice. But this is not the case. If we seek explanations we have no choice but to use explanatory models. Fortunately, today we know how to tailor models to research questions, thus avoiding many decades of confusion #Bookofwhy

11.8.18 @11:37pm - (Replying to @MelMitchell1 @adibzaman and 2 others) "Schema Networks" are quite powerful languages. What happens if we run them on some toy problems in #Bookofwhy? Do you think they would give the correct answer? Put differently, do you think they could be useful in answering some questions in #causalinference ?

11.8.18 @12:21am - (Replying to @MelMitchell1 @adibzaman and 2 others) I look forward to reading about causal models in your upcoming book. Can you share basic ideas on how you envision the building of those models from data, and how they are encoded, once built. Have recent developments in #causalinference been of some help to you? #Bookofwhy

11.6.18 @10:17pm - (Replying to @MelMitchell1 @adibzaman and 2 others) Granted that causation is a necessary (not sufficient) ingredient of intelligence, I was surprised to see how few ML researchers were aware of the Ladder of Causation (#Bookofwhy), the limitation it imposes, and of the tools now available to overcome them. #causalinference

11.6.18 @1:20am - (Replying to @DrChandraFord @UCLA) I was as excited and delighted as the many students whom I had the chance to address in this talk. Let me share my last slide here (edited) "The peak of this revolution is still ahead of us, and you now have the essential tools to be at its epi-center" #Bookofwhy #causalinference

11.6.18 @1:10am - (Replying to @adibzaman @eliasbareinboim) You touched on a peculiar side of model-blind ML. They complain of a stomach pain (eg, "Barrier of Meaning"), ignore the medicine and justify it by saying (still in pain): "No one heard of it! Read the NYT!" @melmarienitch @tdietterich #Bookofwhy #causalinference

11.4.18 @5:27pm - (Replying to @PHuenermund) True, but how would an open-minded economist find out about these theorems? Publication on NBER is reserved for club-members only, and editors of econometric journals lack leadership to import knowledge from outside the bubble. Its tough to be an economist these days. #Bookofwhy.

11.3.18 @11:37pm - (Replying to @fennell_p) Glad you opened #Bookofwhy on page 98, showing Bayes' billiard tables-- my favorite. I have read hundreds of papers on Bayes Rule and Bayesian statistics, etc etc yet, putting ego aside, I recommend this chapter over everything I read,, including the learned gurus of Bayesianism.

11.3.18 @11:14pm - (Replying to @JaimieGradus @EpiEllie) When done correctly, meta-analysis pools samples over areas of agreement and discard samples in areas of disagreement (See #Bookofwhy p. 355). Once awakened, data-analysts will rejoice the option of doing things correctly, e.g., ,

11.3.18 @5:47am - (Replying to @jtrecenti) I am delighted to re-read this quote and, if it was not mine, I would have retweeted it over and over. I am especially delighted to see it coming from a Statistician -- they usually deem this quote either trivial or wrong. Taking it seriously unveils how true it is. #Bookofwhy

11.1.18 @5:40am - (Replying to @smakelainen) Hillarious, thanks, a hymn to counterfactual thinking. Let me know where I can get the lyrics, if available. If I could sing like him... Hmmm....I would not have written #Bookofwhy, perhaps another book, perhaps a jingle ... an anthem! Yes! ..if only.....

11.1.18 @5:26am - (Replying to @NeatWitTweet @MadaGadol) Ein Ma Lefached, Chamudim, Bou V'Nikra. Hakol Mada, Hakol Chadash, Akaveh SheTehanu, Yehuda , Hamechune Judea or Yuda

11.1.18 @5:19am - (Replying to @MereteKonnerup) We seem to be living in two parallel universes, one is model-based (#Bookofwhy), the other model-blind (@DinaPomeranz). My preference for the former is based on comparing the scope of questions each can answer, and the transparency of the answers e.g., Replying to @NeatWitTweetol)

10.31.18 @1:33am - (Replying to @IMourifie @PHuenermund) Thanks for the paper. It's nice to see a 25-year-old problem come to a happy solution, round and polished. May all problems end up this way. #Bookofwhy

10.31.18 @1:09am - (Replying to @Chris_Auld @PHuenermund @HolgerSteinmetz) I would add "almost everywhere" to the statement that overidentification tests are tests of instrument validity. Imagine two invalid IV's which are identical twins (ie each having the same direct effect on Y). I believe they would pass the overidentification test, #Bookofwhy

10.30.18 @3:53am - (1/2) (Replying to @PHuenermund) IV's were invented in 1928, by Philip and Sewall Wright, It is amazing indeed that 90 years later many economists still refuse to accept the non-testability of IV's and, more generally, refuse to learn what is testable in a model. See #Bookofwhy
10.30.18 @3:59am - (2/2) (Replying to @yudapearl @PHuenermund) One extension is worth mentioning: IV's do have testable implications when X is discrete, as captured by the Instrumental Inequalities: #Bookofwhy These are closely related to Bell Inequalities in quantum physics.

10.29.18 @2:53am - For LA residents and LA travelers, I will be giving a book-talk at UCLA on Nov. 5. See . This time, to Law School, Economics and Computer Science, a partnership formed to deal with ML, social systems, and the causal forces that drive the latter. #Bookofwhy

10.26.18 @10:59pm - (Replying to @ehsan_hoque @rochci) Glad to see #Bookofwhy reviewed by another independent observer. Hoping to see a summary shared, when available -- it is important to protect simplicity from the wrath of traditional mystifiers #causalinference

10.26.18 @1:26am - (Replying to @shyamal_chandra) My website will lead you to many slides, videos, lectures, interviews and other goodies. But it is hard for me to guess what "material" you wish to understand and what backgrounds your friends and family have. #Bookofwhy

10.25.18 @1:02am - (1/3) (Replying to @shyamal_chandra) Your question sent me back to serious counterfactual thinking. Factually, the option of making my course public did not occur to me in 2010, when I stopped teaching the course; perhaps because public courses were not as popular at the time. #Bookofwhy
10.25.18 @1:10am - (2/3) (Replying to @yudapearl @shyamal_chandra) Fortunately, UCLA now offers two courses in this direction, one by Prof. Arah, and one by Prof. Hazlett. I hope they would make them public - causality needs to be learned. Intellectually, however, writing #Bookofwhy was more challenging that teaching "Causality", which has
10.25.18 @1:23am - (3/3) (Replying to @yudapearl @shyamal_chandra) already penetrated the academic elite (with 13K citations). It would have required mainly effort of dissemination, not of introspection. Writing #Bookofwhy on the other hand, forced me to view all this work from a different angle -- historical, philosophical and computational

10.25.18 @12:55am - (Replying to @XamilyGuy) We surely care about both controlled and uncontrolled under-representedness. The former indicts Harvard, the latter indicts society. But the word "control" may be misused. To some it means "condition on" to others "hold constant". Only the latter should enter mediation.#Bookofwhy

10.24.18 @1:32am - Enjoyed my talk #PyData, thanks, and now sharing my opening motto: "Data is our window to reality, and data-science is the eyeglasses that enable us to see through that window. It is not a mirror, in which data see themselves interpolated under makeup. #Bookofwhy #causalinference

10.21.18 @5:51am - Evidently #Bookofwhy is currently out of stock. Our publisher assures us this does not reflect a sudden rush of buyers, nor a breakdown of the printing press but a temporary U.S. shortage of paper. Delivery date now stands at Nov. 15. Apologies to all knowledge-hungry readers.

10.20.18 @8:57pm - (Replying to @gileshooker @tdietterich) I especially liked your question: "are you trying to understand how a particular function arrives at a prediction, or are you trying to say something about the underlying causes of that prediction?" It relates to the slide I shared earlier, motivated by Simpson paradox.#Bookofwhy

10.20.18 @6:10pm - (Replying to @yudapearl @tdietterich and 4 others) In the wake of our discussion on Explainability, I am sharing a slide that I used at USC, aimed to show the disparity between explaining why the system arrived at a given decision vs. explaining the logic of that decision. #Bookofwhy #causalinference

10.20.18 @5:58pm - (Replying to @tdietterich @DKedmey and 3 others) Tom, my criteria is not who is using my work, but whether the enterprise called "explainable ML" has produced criteria or principles that would allow an outsider like me to enter the field and use those principles, vs. starting from scratch. I honestly dont know the answer. Help?

10.20.18 @4:17pm - (Replying to @DKedmey @marypcbuk and 3 others) Terrific pointer! Thanks, David. What we need now is for DARPA to issue a new BAA for "Science-based Explainable ML" and exclude old PI's from reviewing new proposals. #Bookofwhy #causalinference

10.20.18 @3:54pm - (Replying to @MariaGlymour) I have no problem with this interpretation of "interpretation". My shrinkage reflected a long debate with economists and Rubinists who kept on asking: "Under what condition does a structural coefficient have causal interpretation?" or "When would a duck be a duck?"#Bookofwhy

10.20.18 @3:39pm - (Replying to @tdietterich) My helplessness reflects genuine yearning for a set of criteria by which we can distinguish "explainable" from "interpretable" from "dumb". I am still not sure if such a set is available to "explainable ML" insiders, or they are just very good at hiding it from outsiders.

10.20.18 @3:49am - (Replying to @marypcbuk @RandomlyWalking @tdietterich) I do not doubt the quantity of work that has gone into it, not only recently, but for the past 30 years. I am craving now for a set of principles that distinguish a useful explanation from a dumb one, like weights and parameters. Are we in possession of such a set?#Bookofwhy

10.20.18 @12:17am - (Replying to @RandomlyWalking @tdietterich) I am not insisting on a formal definition of "explainability" but there ought to be some common understanding why a copy of the code would not be accepted as an "explanation" of how the output was produced from the input. Has anyone articulated why? or what else is needed?

10.19.18 @11:28pm - (Replying to @tdietterich) I listened to her talk and could not understand a thing, probably because no one has explained to me what "explainable" or "interpretable" is. I know something about causal explanations, which made me ask a naive question: Is there explainability w/o a causal model? #Bookofwhy

10.19.18 @9:53pm - (Replying to @MariaGlymour) I shrink when I read a sentence such as: "estimates from ... can be interpreted causally". I thought only regressionists ask whether something has causal interpretation, not scientists. If your causal question is estimable then its estimate must have causal interpretation.

10.19.18 @12:13am - (Replying to @DrewLevy) If statistical mis-thinking rules the bio-marker literature, causal mis-thinking is even worse, especially when it comes to "surrogate end-points". See , section 4.2, and #Bookofwhy #causalinference #fharrell

10.17.18 @5:02am - (Replying to @ThomasVConti) Your idea of #Bookofwhy reading group to economics students is of enormous importance -- you must catch them before they get contaminated by their textbooks. I would take the textbook examples and ask questions from the three levels of the Ladder, where textbooks are dead silent.

10.17.18 @4:48am - (Replying to @zacharylipton) You dont get any tinge of anxiety when you get a tinge of thrill each time you can do something today that you could not do yesterday. #Bookofwhy #causalinference

10.17.18 @3:56am - (Replying to @CT_Bergstrom) It reminds me of the question: "Should I study arithmetic? No application in mind" @Bookofwhy @causalinference

10.17.18 @2:17am - (Replying to @TomCrellen) Thanks for catching. I hope we can fix it before the revised edition comes up, in November. #Bookofwhy #causalinference

10.17.18 @2:10am - (Replying to @oacarah) And thank you Onji for hosting this talk. I thoroughly enjoyed it, seeing so many students with this spark of curiosity in their eyes. I must retweet what I told them at the talk: "You are the elite force in the army of commonsense. You're fortunate!" #Bookofwhy #causalinference

10.15.18 @10:43pm - (Replying to @ResearchSmarter @JDKun @MediaMetrics) Thanks for the photos. I had great fun today at USC, perhaps because the audience was 90% students and 10% professors. I felt like I was talking to the next generation, free of cultural hangups, open-minded and curious. Like readers of #Bookofwhy.

10.15.18 @10:33pm - (Replying to @quantadan @hadleywickham @bearloga) The first, first thing I would like data-scientists to learn: Data science deals with the INTERPRETATION of data, not summarization of data nor re-visualization of data. "Interpretation" is what the data tells us about world outside the data. #Bookofwhy #causalinference

10.15.18 @1:18am - (Replying to @jasonroy) Great photo! Showing how easy it is for street cleaners to understand the do-operator, and how hard it is for statisticians. I heard it is still not taught in Stat 101 (not sure about #causalinference.) Glad it is in #Bookofwhy

10.14.18 @4:48pm - (Replying to @BreskinEpi @KatieMollan and 3 others) The structural theory of counterfactuals tells us precisely when individual level effects can be estimated, and from what types of data. For example, probabilities of sufficiency and necessity (#Bookofwhy ch. 8) are individual-level attributes, and can be bounded tightly 1/2
10.14.18 @5:14pm - (Replying to @yudapearl @BreskinEpi and 4 others) as shown graphically, in . These bounds might even collapse to point estimates. In short, the structural theory gives us yes/no answers (not opinions) about what is possible or impossible about ICE. The de-mystification of #causalinference is not a slogan.

10.13.18 @1:27pm - (Replying to @_eleanorina @EngineerDiet) The acolytes who draw up diagrams are doing so to explicate and communicate the limitations of scientific claims. More such acolytes are needed. #Bookofwhy

10.12.18 @8:10pm - (Replying to @zacharylipton) Agree, and the few papers I read from the FAT literature do not indicate that the authors know that the various shades of "bias" - statistical, causal, and counterfactuals -- can be given precise characterization, liberated from confusion, and even estimated from data. #Bookofwhy

10.11.18 @5:23pm - (Replying to @USCBiostat @malco_barrett) I am going to be there Monday, your part of town, I am told there is room for everyone.

10.11.18 @4:04pm - If you happen to be in LA Monday or Tuesday, you are invited to attend two lectures of mine. The first is at USC and the second at UCLA . I will be summarizing #Bookofwhy with slides and poetry. See you there. #causalinference

10.11.18 @3:43am - (Replying to @michael_nielsen) When it comes to individual cognitive development, I am willing to adopt a more nuanced version of Whorfian theory. But speaking of scientific progress, I am a hard-liner: no notation - no progress. Causal inference in 20th century is a living proof. #Bookofwhy #causalinference

10.11.18 @2:34am - (Replying to @blattnerma) For a conceptual starter, the #Bookofwhy will be ideal. For a more mathematical starter, I would recommend the PRIMER #causalinference

10.10.18 @9:37pm - (Replying to @andreabellavia @PaoloHead88 @StefanoRenzett1 Not sure Stephen stigler would agree with my account of the history of statistics, but he will agree that it's about time to write this history from a causal perspective, and that it looks quite interesting (albeit less admirable) from this angle. #Bookofwhy #causalinference

10.10.18 @8:30pm - (1/2) The do-operator is described in #Bookofwhy and in this public lecture For a deeper mathematical treatment, see or Causality (2009).It is, indeed, the basis of causal reasoning, shunned only by some PO's #causalinference 1/2
10.10.18 @8:40pm - (2/2) ... shunned only by potential-outcome enthusiasts, fearing light. Even Heckman confessed that you can't do economics without it, so he took "do" and called it "fix", to make eocnomists feel more comfortable with it. It didn't work; poor "fix" died of loneliness. #Bookofwhy

10.10.18 @1:44am - (Replying to @wlog @EpiDancer @EpiEllie) Even if we differentiate between "observed data" and "underlying data generating process" there is still a need to distinguish the operator "see" from "do". As to confusions in economics, see No causation without notation (just coined, but true)#Bookofwhy

10.9.18 @10:49pm - (Replying to @EpiDancer @EpiEllie) Glad to see consensus emerging. "Control" is a relic of the days when we could not distinguish between see(X=x) and do(X=x), and the only formal language we had was probability theory, with one operator: Bayes conditioning ie "see". #Bookofwhy #causalinference

10.9.18 @10:12pm - (Replying to @thoughtfulflyer) Hume's two quotes appear together, almost in same breath, on page 265 of #Bookofwhy, and it is really amazing that he did not realize that they rest on two totally different rungs of the Ladder of Causation. Strange, I can't recall a philosopher screaming "Hey!!"#causalinference

10.9.18 @12:06am - (1/3) The publisher of #Bookofwhy informs us that a revised and corrected 7th printing will be available early Novemeber. Hold your breath and, in the meantime, all errata are marked here , in red, for a smooth and flaw-less reading. #causalinference.
10.9.18 @12:13am - (2/3) Special beneficiaries of the corrected printing will be champions of RCT, some of whom were offended by my labeling the front-door method a "serious competitor to RCT". It's been changed to "useful alternative", which should invite everyone to examine its merits and to start 2/3
10.9.18 @12:23am - (3/3) to start "triangulating" RCT against observational studies. In SCM, "triangulation" is not a romantic aspiration but a working methodology. #Bookofwhy #causalinference. 3/3

10.8.18 @11:38pm - (Replying to @dynarski @EpiEllie) As important and necessary as descriptive analysis is to causal research, one must deploy causal models to bring the two together. Description and causes speak two different languages and causal models (eg, SCM #Bookofwhy) take inputs from both to cook up a #causalinference soup.

10.8.18 @4:06am - (Replying to @mendel_random @eliasbareinboim and 3 others) Articulate your causal model and don't fit any schema nor dogma nor popular estimators, just see if your causal question is estimable. If it is, you've got a "real-life plausible example," since you yourself proclaimed it to be YOUR model. #Bookofwhy #causalinference

10.6.18 @2:18pm - (Replying to @learnfromerror @LucasJfriesen @stephensenn) Just tell us what relations you wish to assign responsibility to, what data you have available and "mirror mirror on the wall" will tell you if it is doable or not, what else you need to complete the assignment, etc, etc. Articulate what you WANT to know, not what we dont know.

10.6.18 @2:10pm - (Replying to @learnfromerror @stephensenn) Ans. What Pearl says can only be understood by trying to "represent what might be going on" in any mathematical language, since "what is going on" is beyond statistics. "How do we use data?" Patterns of arrows do have testable implications, screaming loud when violated in data.

10.6.18 @1:00pm - (1/3) (Replying to @mendel_random @MariaGlymour) Idea-driven students, grad-schools and book clubs #causalinference, should view your quest for "actual real life example" as evidence of an opportunity to transform an underutilized theory into a practical and powerful methodology. They will. Here is why:
10.6.18 @1:12pm - (2/3) (Replying to @yudapearl @mendel_random @MariaGlymour) The reason I included (first front-door paper) is that it outlines the requirements as an "isolated" mediation system X--->M--->Y, shielded from external influences, with only input (X) and output (Y) exposed to confounded influences. #causalinference 2/3
10.6.18 @1:28pm - (3/3) (Replying to @yudapearl @mendel_random @MariaGlymour) If I were swayed by "give me one real-life example" voices in 1993, I would still be thinking regression today, perhaps sprinkled with ignorability justifications, but I would be scared of graphs, just writing books on why they spoil the youth. #Bookofwhy #causalinference 3/3

10.5.18 @9:15pm - (Replying to @UCSF_Epibiostat @MariaGlymour) My, My, I did not realize you started a book club at ucsf. I've got a nice example for your front-door magic, co-written by your sister Madelyn. Study Question 3.4.2 from our Primer book, It should entertain and challenge most students #causalinference

10.5.18 @3:37pm - (Replying to @swanderingf) A quick answer: We know what causality is, but we do not exactly know what interpretability is, ie, we do not have an agreed on mathematical definition of the criterion by which we judge interpretability. e.g. would the answer: "Because you programmed me that way" be acceptable?

10.5.18 @2:34am - (Replying to @stephensenn According to Don Rubin (see his Fisher's Lecture) part of the controversy was over notation, and Fisher paid dearly for refusing to use Neyman's potential outcome notation. It could have spared him the blunder over the mediation fallacy. #Bookofwhy page 441. #causalinference

10.4.18 @8:37pm - (Replying to @rogierK @KristinaLerman @JohnHolbein1) A snippet remark on the re-kindled interest in Simpson's Paradox (SP). Page 211 of #Bookofwhy reminds us what makes SP "interesting", namely, what phenomenon SP is a symptom of, arguably worthy of being searched for -- one of the three trends is non-causal. #causalinference

10.3.18 @11:32pm - (Replying to @analisereal @ItsAStatLife @eliasbareinboim) Agree! The discussion were necessary for dispelling the myth that DAGs are good for single treatment problems but, when it comes to treatments over time, potential-outcomes should take over. The cost of this myth : a safe ticket to outdated results #Bookofwhy #causalinference

10.3.18 @3:55am - (Replying to @ItsAStatLife @eliasbareinboim @analisereal) SWIGS are an expansion of DAGs to those who need to touch Y(a). FINE. But when marketed and taught as a replacement to DAGs, stripped of do-calculus, they act as handcuffed mutations of DAGs, unable to tell even what they can or can't do with varying treatments. #causalinference

10.2.18 @8:23pm - (1/3) Well said. Multiple angles are great to highlight different aspects of the problem, when needed. This is the beauty of SCM: All tools are welcome, and each is given a tag of what it's good for. We use SWIG, for example, in the 2nd Rule of do-calculus #causalinference 1/n
10.2.18 @9:06pm - (2/3) to authorize exchangeability, and we use another twist of the DAG when another need arises. It sounds strange therefore that one who feels comfortable with SWig will feel uncomfortable with do-calculus, unless he/she is not given true exposure to the latter. #causalinference 2/n
10.2.18 @9:30pm - (3/3) Finally, let's ask what one can get from SWIG which is not gotten by Rule 1. It's important to tell in advance which problems can be submitted to SWIG analysis and which are too hard for SWIG, but solvable by other means. Can we tell? Yes we can #causalinference @BreskinEpi

10.2.18 @7:08pm - (Replying to @analisereal) Important point which I dared call: "The first principle of causal inference": Y(a)=Y(M_a) Strantely, students brought up with PO education are rarely aware of it, partly because Rubin resists it vehemently and partly because it takes away from the glory of SWIGs. #Bookofwhy

10.2.18 @3:29pm - (Replying to @analisereal @ItsAStatLife) It is not only a matter of complexity. The obsession to " touch Y(a) or else" may lead us to mis-formulate problems and miss solutions. In our case, a missing-data problem was mis-formulated as a Swig. Reason? The pleasure of touching Y(a) has costs. #Bookofwhy #causalinference

10.2.18 @5:56am - Sorry, the correct link to my Berkeley talk is here: … I checked, and it is still active. #Bookofwhy

10.2.18 @4:19am - (Replying to @martisamuser) I am glad you linked to my Berkeley talk ... I have enjoyed it a lot. It was given to theoretical computer-scientists, not my usual audience, but one that understands the roles of theory and models in processing data. Thanks. #Bookofwhy #Causalinference

10.1.18 @9:06pm - (1/2) For time-varying treatments, I highly recommend a recent Letter in Epidemiology which contrasts the simplicity and transparency of working with DAGS (where knowledge resides) with the cost of craving to touch a node named Y(a). #causalinference.
10.1.18 @9:23pm - (2/2) This contrast reflects a deeper issue: What comes first? Causal connections in our world, or dependencies of Y(a) which are derivable from those connections? Is Y(a)||A a more valid justification for adjustment than path blocking? Habit? Cult? or Science? #causalinference

10.1.18 @3:50pm - (Replying to @_MiguelHernan @analisereal) The description of sequential backdoor in terms of "action avoiding paths" (Causality p.352) requires inspection of a single graph, and avoids chasing independncies of Y(a), which are implied by, rather than explain exchangeability. Once tried - always used. #causalinference

9.30.18 @8:36pm - (1/n) Indeed, what happened to sequential backdoor (SBD)? It is a puzzle not only for readers of Miguel's book, but all folks doing "time varying treatment". I am sure if they were reminded how easy SBD is, they would never go back to the torture of chasing Y_g||Ak|Lk conditions. 1/n
9.30.18 @8:42pm - (2/n) The one fault Miguel's book finds in SBD is: "It does not DIRECTLY show the connection...[to] exchangeabitlity. I deem it a great benefit to skip the laborious chase after this "direct" connection when we can establish it effortlessly by backdoor. See 2/n
9.30.18 @9:03pm - (2/4) Moreover, I doubt whether the counterfactual independence Y_a||A|L contributes to the understanding of how exchangeability comes about. It was formulated in this convoluted way merely because Rubin (1978) knew no other language, but probability, for expressing dependencies. 2/4
9.30.18 @9:27pm - (4/4) Today we have more advanced languages to capture causal dependencies, so our students can be spared the agony that Rubin went through. "Conservation of Agony" isn't a physical law.The sequential backdoor can show #causalinference students how to turn pain into fun.#Bookofwhy

9.29.18 @9:25pm - (Replying to @_MiguelHernan) Looking at Chapter 19, I dont understand why it spends so much energy checking exchangeability conditions when we can skip them altogether and solve the problems directly from the DAG, using sequential backdoor. E.g., compare your Fig 19.5 to #Bookofwhy p. 241. #causalinference

9.29.18 @5:06pm - (Replying to @KirkD_CO @muscovitebob) It is not clear to me what Braingap considers to be "Having a clear understanding of the data". Simpson's paradox is a causal, not statistical dilemma,, the resolution of which requires an understanding of the causes behind the data, not the data. #Bookofwhy #causalinference

9.29.18 @9:21am - (Replying to @rationalexpec @stevenstrogatz @camelsincaves) I had a quiet dinner with Clive Granger in Uppsala, Sweden, 1991. He confessed to me that he feels embarrassed by the name "Granger causality", since it has nothing to do with causality, but he can't stop people from using it; they need some way to express what is needed.

9.29.18 @8:58am - (Replying to @camelsincaves @stevenstrogatz) Scientists, like children, succumb to two weaknesses: (1) infatuation with new toys, and (2) addiction to old toys. Today, progress in causal inference is held back by the latter. #Bookofwhy #causalinference

9.28.18 @11:44pm - (Replying to @PHuenermund) Most economists understand that curve fitting is not causal inference, so they excuse themselves with: "we assume ignorability" which makes them feel less guilty. In contrast, Most machine-learning researchers do not see a reason to apologize. #Bookofwhy #econbookclub

9.28.18 @5:44am - (Replying to @juli_schuess @cdsamii and 2 others) All I want to establish at this point is that the finding: "Sex affects hiring" has clear policy implications (not stylized), regardless of whether "sex" is determined by employer's suspicion, birth certificate, an entry in a form, or genetic analysis. #Bookofwhy #causalinference

9.28.18 @5:01am - (Replying to @cdsamii @eliasbareinboim and 2 others) Evaluating job-training and anti-discrimination programs are costly. The decision of what to fund first is also a policy decision. Corollary: studies concluding with: "sex has strong direct effect on salary" are not "ill-defined"; they DO have clear policy implications.#Bookofwhy

9.28.18 @2:41am - Adnan Darwiche article "Human-level intelligence or animal-like abilities?" just came out in CACM … and should be of interest to ML readers, as well as to other soldiers of artificial intelligence. #Bookofwhy #causalinference

9.28.18 @12:03am - Replying to @mattmasten @_MiguelHernan) Heckman agrees with Pearl when Pearl praises econometric models over model-blind statistics. Humorously, things change when Pearl presents economists with tools to maintain their edge. See , #Bookofwhy #causalinference #ecobookclub

9.27.18 @11:52pm - (Replying to @eliasbareinboim @cdsamii and 2 others) It seems that what really provokes manipulationists objections are studies void of policy implications. But studies concluding with: "sex has strong direct effect on salary" do have policy implications, e.g,. What to fund? Anti-discrimination programs? or job-training?#Bookofwhy

9.27.18 @7:42am - A friend sent me this 3 years old video It gives a fairly faithful overview of the landscape in which #Bookofwhy was hatched. Enjoy. #causalinference #ecobookclub

9.27.18 @6:20am - (Replying to @zack_weber) I do the same thing, but for a different purpose -- to escape from debates and to make peace with truth. #Bookofwhy

9.26.18 @7:52pm - (Replying to @yudapearl @cdsamii and 2 others) Reminds me again of DeMoivre (1657-1764) asking: Would I be permitted to TEMPORARILY write down exp(inx), just for the sake of calculating cos(nx), in return for my solemn oath never never to say i =sq(-1) #Bookofwhy #causalinference

9.26.18 @7:33pm - (Replying to @cdsamii @eliasbareinboim @deaneckles) We are waiting to hear from manipulationists whether they would permit us to TEMPORARILY write down do(blood-pressure), just for the sake of calculating do(drug), in return for our solemn oath never never to say: "causal effect of blood-pressure". #Bookofwhy #causalinference

9.26.18 @8:22am - (Replying to @_MiguelHernan) Recall provision (2). Causal effect of A is well defined IFF A is in your model of the world. Otherwise, A could be the size of unicorn - non scientific. By inviting A into your model you bestow upon it scientific meaning and, in our case, causal effects as well. #Bookofwhy

9.26.18 @7:10am - (Replying to @_MiguelHernan) Delighted to accept the invitation to estimate "the effect of obesity on death" if you just tell me where in your causal diagram I can find the variable 'obesity'. If it is not there, perhaps we can do it with "gender", which is surely in your model? #Bookofwhy #causalinference

9.25.18 @5:08pm - (Replying to @yudapearl @_MiguelHernan) Moreover,, this definition has an objective element to it, modulo our world knowledge; i.e., all people sharing your model of the world will also share your estimate of the causal effects as defined in the model. Great feature!!! Not hypothetical!! #Bookofwhy #causalinference

9.25.18 @4:58pm - (Replying to @_MiguelHernan) Recall condition 2): Your model represents what you believe about the world. This means that properties of your model reflect properties of the world. Corollary (and it ain't circular): Once defined, definitions are, by definition, well defined. #Bookofwhy #causalinference 1/2

9.25.18 @2:10pm - (1/2) (Replying to @_MiguelHernan Not exactly! What Pearl says is: "The causal effect of A on B" is well defined if (1) A is a variable in your causal model, and (2) your model represents what you believe about the world. (He sometimes add: why else would you add A to your model? - for show?) #Bookofwhy
9.25.18 @2:19pm - (2/2) (Replying to @yudapearl @_MiguelHernan) For two entertaining yet unambiguous accounts of what Pearl says see: and espeially Appendix II: Causes vs. Enablers. #Bookofwhy #causalinference

9.25.18 @1:18am - (Replying to @AllenLaneBooks @TheTLS) Very perceptive reviewer: "circumnavigating leading figures who have somehow come to hold up progress". May the Gods of counterfactuals protect us from becoming up-holders of progress. Any trick of accessing the whole article? #Bookofwhy

9.24.18 @3:48am - (Replying to @ProfMillennial1) Very interesting observation. To get this level of accessibility I will need to write another book. Will try. But, honestly, when I picked up his book I fell asleep on every page, perhaps because in my funny mind his many principles follow from just one -- structure. #Bookofwhy

9.23.18 @10:56pm - Agree. It was more than 6 years ago that we wrote this page, and every time I read it I feel like thanking the authors. Evidently, Campbell was so highly revered, that his disciples never got to suspect that threats can be disarmed #Bookofwhy #causalinverence

9.23.18 @8:15pm - (Replying to @melodem_group) Confessedly, having been trained in different trenches, on different fronts, I often fail to appreciate difficulties faced by some of my comrades. I would therefore value a typical example of an assumption you may fail to make if it were not for a helpful "check-list" #Bookofwhy

9.23.18 @4:30pm - (Replying to @melodem_group) Some also ask why #Bookofwhy does not highlight the influential frame of Campbell etal. My short answer: I really tried hard, but could not find the "frame". Frames solve problems -- Campbell lists them. My long answer: See #causalinference

9.23.18 @3:40pm - (Replying to @AlloriMD) "How" has two connotations: 1. How nature works, 2. How to achieve what we want. The former is perfect for mediation. The latter sits on rung-2. "Why" is often used to cover the former, as in "Why does lemon cure scurvy", which is almost synanymous to "How does lemon cure,,"

9.23.18 @3:01am - (Replying to @raquelrguima @she_knows_a_key) If we have big data, we can estimate PS for each stratum with high precision, with no modeling assumptions. At this point the c-equivalence Theorem kicks in and says: Yes, the bias will be exactly the same, for any set of covariates. #Bookofwhy, #causalinference Nice Theorem !!

9.23.18 @12:51am - (Replying to @she_knows_a_key) My Japanese is still rusty, but I hope you told readers that, asymptotically, the bias-reducing capacity of PS is identical to ordinary adjustment for the same set of covariates. I can't understand why people still refer to PS as "design", not "routine".

9.22.18 @10:41pm - A recent paper by Steiner etal sheds new light on how RCT, IV, RD and other model-free "designs" can be understood and unified from a graphical model perspective. #Bookofwhy #causalinference . A must for quasi-experimentalists and potetial-outcomists

9.21.18 @7:03am - Replying to @CateyBunce @WiringTheBrain) When done correctly, "meta analysis" was renamed "data fusion", because we could not find any meta-analyst who could do it correctly, i.e., neutralize causal disparities, not seen in the data I am curious if there are other ways of doing it correctly.

9.21.18 @3:04am - My UCLA colleague (and Chair), Adnan Darwiche, gave a talk on AI education at Tsinghua University (the MIT of China) a few days ago. It is only 12 minutes and ML researchers and educators should find it interesting: #Bookofwhy #datascience

9.21.18 @2:51am - (Replying to @ESteyerberg @tmorris_mrc and 2 others) Kudos to Els for speaking truth to power. I did not realize however that causal reasoning needs defending; in my little village, it is model-blind learning that people try to defend. #Bookofwhy #causalinference #datascience #causalai

9.19.18 @9:53pm - (Replying to @bzaharatos) Here it is: . And, as you see, the author complains about the results but does not address the basic issue: What is meta-analysis supposed to estimate, and why is it a valid answer to the research question? #Bookofwhy #causalinference

9.19.18 @8:47pm - (1/2) (Replying to @ROrellana4 @EpiEllie and 4 others) Thank you Robert for summarizing students' reaction to chap. 7 of #Bookofwhy and noting the interest in do-calculus. Epi classes de-emphasize it because it is not easily translatable to PO notation, to which most Epi teachers are culturally wedded. It is a forgivable neglect
9.19.18 @9:42pm - (2/2) (Replying to @yudapearl @ROrellana4 and 5 others) as long as you are satisfied with the low-hanging fruits of backdoor identification. But as soon as you move to questions of transportability, data-fusion, and even selection-bias, I dont envy your Epi peers when deprived of do-calculus. #Bookofwhy,#causalinference

9.19.18 @8:19pm - When people ask how meta-analysis differs from data-fusion I say: Meta-analysis averages apples and oranges to get properties of bananas. Here is why: , Science Magazine story, Sep 18, slips on the banana. #Bookofwhy #causalinference

9.18.18 @4:19pm - ML researchers looking into counterfactual reasoning should ask: "What caused the fire, the match or the oxygen?". It is derived here and just been updated with two new appendices, on abduction and enablers. Enjoy the oxygen. #causalinference #Bookofwhy

9.18.18 @1:38am - (Replying to @ranilillanjum @EpiEllie) OK, let's call it "ontology". Please identify your currently most favorite ontology of causation; we will compare it to the "listening" ontology, and see which of them answers the criteria that you think a good ontology should satisfy. Fair? #Bookofwhy #causalinference

9.17.18 @7:43pm - (1/2) (Replying to @lonnej) Thanks for an illuminating review of my article "Seven Tools.." . I need to clarify why RL falls short of covering the entire level of interventions. The effect of doubling our price depends on the intent behind recorded price increases in the data.
9.17.18 @7:52pm - (2/2) (Replying to @yudapearl @lonnej) Whether those increases were taken to compensate for production costs or to spite the competition makes a difference on what a price increase will do today, when production costs are different and we could not care less about the competition. We eloborate on it in #Bookofwhy

9.17.18 @3:32pm - (Replying to @EpiEllie @ranilillanjum) When does a philosopher stop asking: "And what is the meaning of XYZ?". Ans. When you ask him/her: "What is the meaning of meaning?" and you show him/er that XYZ satisfies all the requirements. "Listening" would satisfy your requirements, please check.#Bookofwhy #causalinference

9.16.18 @10:48pm - (1/3) (Replying to @EpiEllie) Thank you @EpiEllie for summarizing @ranilillanjum and @SCMumford book so succintly and creatively. Great thread! I must confess though that from all the allgories, parables, analogies and metaphors that our ancestors have devised to capture causation, I find the "listening"
9.16.18 @10:54pm - (2/3) (Replying to @yudapearl @EpiEllie @ranilillanjum) "listening" metaphor most useful for modern #causalreasoning. It goes: "X is a cause of Y if Y listens to X and decides its value in response to what it hears." The formal logic behind it is the same as Lewis logic for "counterfactual dependence", but
9.16.18 @10:58pm - (3/3) (Replying to @yudapearl @EpiEllie
@ranilillanjum) but its metaphorical power is what drives our decision whether to draw an arrow between X and Y in a causal diagram. Since the presence/absence of such arrows affects critical steps in the analysis, "listening" must carefully be listened to. #Bookofwhy #Econbookclub

9.16.18 @1:01am - This paper just came to my attention (and I can't forgive myself for not seeing it earlier) ... A friendly, comprehensive and unifying (no hangups) roadmap to epidemiological methodology. Highly recommended. #Bookofwhy #causalinference @EpiEllie @miquelporta

9.15.18 @7:15pm - (Replying to @ProfMattFox @EpiEllie) For freshmen class, I would use Ellenberg's "why my handsome dates are jerks?" or the two coins experiment (#Bookofwhy Page 200). But for more advanced class I would ask: "How come it is so hard for us to come up with an example?" The answer is on page 199, Reichenbach dictum.

9.15.18 @6:54pm - (Replying to @ClimentQD) Thank you for retweeting this simple truth, so often forgotten by its beneficiaries, so often ridiculed by the priests of RCT, and so liberating in its clarity when "experimentalism" hide the sun. Even structural economists (eg. Heckman) tend to forget it. #Ecobookclub #Bookofwhy

9.14.18 @11:05pm - (Replying to @Noahpinion) The fit-to-data practice that you are mocking is long gone from modern structural modeling, surviving only in econometric textbooks and few outdated circles. Modern structural models, unlike "natural experiments' tell us precisely how to test compatibility with data.#Bookofwhy

9.14.18 @9:05pm - (Replying to @yudapearl @Noahpinion) When you see a paper that relies on a natural experiment, your first question should be "Could the assumptions be depicted in a structural model?" If the answer is yes, check their plausibility. If the answer is no, recommend a major revision. #Bookofwhy #ecobookclub.

9.14.18 @8:38pm - (Replying to @Noahpinion "Natural experiments" are studies in structural models where some of the assumptions are justified by semblance to RCT and others are kept implicit, for lack of language, at the mercy of the "experimentalist". See #Bookofwhy #causalinference #ecobookclub.

9.12.18 @8:17pm - (Replying to @DKedmey) I would start with a simple story, like the firing squad, and ask the five questions that we asked here: , going from beliefs to actions to counterfactuals. For desert, I would also ask an explanation question, e.g, does oxygen explain the fire? #Bookofwhy

9.12.18 @7:34pm - (Replying to @wattenberg @GoAbiAryan and 3 others) Readers of #Bookofwhy would believe that counterfactuals are impossible without "opening the black box". So, either #Bookofwhy is wrong, or the explanations generated by this system are not the same as those labeled "counterfactuals" in the Book. E.g., Does Oxygen explain Fire?

9.12.18 @3:34pm - (Replying to @LauraBBalzer @_MiguelHernan @societyforepi) My favorite source on #DataScience is (no dogs). It makes claims similar to @_MiguelHerman, but it separates interventions from counterfactuals. It deals with the "dynamic knowledge" issue via transportability accross time-shifting environments. #Bookofwhy

9.12.18 @3:08pm - (Replying to @FaustoBustos) "Reasoning by cases" underlies both decompositions, and it would be interesting to investigate the class of concepts that permit such decomposition. Recall also that humanoids are causal, not probabilistic machines, so the causal version would be more intuitive. #Bookofwhy

9.12.18 @2:47pm - (Replying to @JenMandelbaum @quantadan @MyaLRoberson) Arrows do not address, they describe. The question before us is how, after unpacking and re-unpacking all the multitude of factors, we describe in our model that some employers will let hiring decisions be influenced by the candidate race? An arrow can do that. Any alternative?

9.12.18 @3:46am - (Replying to @GoAbiAryan @wattenberg and 3 others) Anyone can translate this tool into the language of principles and ideas? What is the input? where does it come from? Data only? or some external knowledge? if the latter, in what form? etc. Thanks for helping. #Bookofwhy

9.12.18 @1:57am - Readers who asked to see a derivation of "What caused the fire, the match or the oxygen?" can find it here; and connect to PO-orthodoxy, Rothman's PIE and other topics of discussion. Enjoy the oxygen. @EpiEllie @_MiguelHernan #causalinference #Bookofwhy

9.12.18 @12:50am - (Replying to @DBetebenner @EpiEllie @_MiguelHernan) Complex numbers are math constructs that help us answer questions about real numbers (eg, deMoivre Theorem). They are "ill-defined" to my 12-year old grandson who needs to finish his math homework. The priests of "manipulation-based" causality are similarly impatient: act or else

9.12.18 @12:19am - (1/4) Here is a vivid example where we can estimate effects of a manipulable variables (eg, Smoking) only if we apply the do-operator to a nonmanipulable variable (eg, Tar). Sounds strange? Please glance at the derivation of the front-door formula (#Bookofwhy, page 236) #EpiEllie
9.12.18 @12:23am - (2/4) Watch how do(t) sneaks in, creates havoc, pushes Smoking under the do-operator, and leaves the scene unscratched, as if it has never been manipulated. Orthodox PO vice-squads would probably charge us with crime against "well-definedness" (more),,,
9.12.18 @12:27am - (3/4) Others will forgive this temporary violation of prudent science as a harmless mathematical gimmick that leads to a well-defined effects. I imagine that infinitessimals were forgiven in the 17th century on similar grounds.
9.12.18 @12:30am - (4/4) Gimmick or no gimmick, never tell your students that do(Tar) is "ill-defined". Complex numbers are also "ill-defined," but where would science be without them? @EpiEllie @_MiguelHernan #causalinference #datascience #Bookofwhy

9.11.18 @11:41pm - (Replying to @quantadan @JenMandelbaum) Surely, all our models are simplifications of our complex world. But there comes a point when we need to stop elaborating and submit a model to analysis. Would we be permitted at this point to draw an arrow from "race" to "hiring',keeping in mind everything said-written on it?

9.11.18 @12:33pm - (Replying to @JenMandelbaum) An arrow going from "race" to "hiring" says exactly what you expressed in words: preferential hiring based on race, and should not be forbidden from entering our model of society. #Bookofwhy #causalinference @DavidSFink

9.11.18 @3:48am - (Replying to @BerkOzler12 @dmckenzie001) I think it is an encouraging sign. One more push and Manski will be ready to join modern causal analysis and discover the "credible basis" that he and Goldberger have sought for so long. #Bookofwhy

9.10.18 @5:52am - (1/3) (Replying to @ProfMattFox @DavidSFink @EpiEllie) More than politics. Notation is what drives our thoughts and eventually our actions. If you forbid you students from saying: "race causes inequality" you also forbid them from drawing an arrow from "race" to "hiring" and they would seek remedies elsewhere e.g, rain-dancing.
9.10.18 @5:56am - (2/3) (Replying to @yudapearl @ProfMattFox and 2 others) To students of physics, semantics commands awesome respect. The idea that an electron responds to the "field" around it, not to the particles that created that field has had a revolutionary impact on science (eg Maxwell eqs). Why deprive your students from such innovations
9.10.18 @6:00am - (3/3) (Replying to @yudapearl @ProfMattFox and 2 others) Finally, there is damage in loss of opportunities (mentioned in a previous tweet). Forbidding some variables from receiving the do-operator, even temporarily, might prevent you from identifying effects of manipulable variables (by surrogates). #Bookofwhay #causalinference

9.10.18 @6:31am - (1/2) (Replying to @DavidSFink @ProfMattFox @EpiEllie) The reluctance of orthodox PO framework to attribute causal power to non-manipulable variables (eg sex, race) is a harmful cultural hangup, surviving like the monarchy, simply because Rubin spawned PO from RCT, and others did not shake it off yet. #Bookofwhy #causalinference
9.10.18 @6:35am - (2/2) (Replying to @yudapearl @DavidSFink and 2 others) I can't see how/why liberating epidemioligists from this hangup would prevent them from improving health. I tried to elaborate on it here #causalinference #Bookofwhy

9.8.18 @7:51pm - For our philosophically minded readers (and who isn't?), I am sharing an interview with 3:am magazine on what/how/if the Causal Revolution contributes to the philosophy of science. #Bookofwhy looks a bit different when viewed from a philosophical lens.

9.8.18 @5:12pm - (1/2) (Replying to @statsepi @EpiEllie) Twitter may be a waste of time in mature fields such as physics or statistics, but not in new fields such as #causalinference which are in their formative stage, and must shake off conflicting traditions that labor to mystify things to survive. I find twitter useful for ...
9.8.18 @5:20pm - (2/2) (Replying to @yudapearl @statsepi @EpiEllie) ... useful for calling attention to areas of demystification which traditional textbooks tend to suppress. Not out of malice, of course, but by lacking the vocabulary needed for demystification. PS is a good example: #Bookofwhy

9.8.18 @2:34pm - (Replying to @yudapearl @f2harrell and 5 others) I believe the discussion on PS could benefit from a Section titled "Understanding propensity scores" from Chapter 11 of Causality (2009). Here is a link which should clarify why PS is smoother, not a bias-reducer. #Bookofwhy #causalinference

9.8.18 @2:34pm - (Replying to @f2harrell @EpiEllie and 4 others) Moreover, let's not forget the PS is only as good as the covariates that enter into it. Asymptotically, the bias-reducing capacity of PS is identical to ordinary adjustment for the same set of covariates. That is, PS is a smoother, not a bias-reducer. #Bookofwhy #causalinference

9.8.18 @6:23am - Readers asked for a conceptual summary of my book Causality (2009), free of the mathematics that decorates most pages. I'VE FOUND ONE: -- a lecture given to AI audience in 1999, and its still fun to read. #Bookofwhy #Causalinference #ecobookclub

9.7.18 @7:45pm - (1/2) (Replying to @mendel_random) What would one call the mega-tons of research going into "observational studies" from Cochran (1965) till today, if not "continuous erosion of the supremacy of RCT". Cochran did not have the tools that the Causal Revolution is now providing us, the erosion IS continuing.
9.7.18 @8:00pm - (2/2) (Replying to @yudapearl @mendel_random) Suppose you have a choice between an RCT on a risky drug, for which you can barely recruit 10 homeless volunteers, and an observational study, full of back and front door conditions, on 12,000 samples from the target population. Wouldn't you pause? Isnt your pause "erosion"?

9.6.18 @10:30pm - OJ Arah has noticed that the non-manipulability issue is sipping into influential journals like the Lancet.... As an unbiased observer, I still prefer the general discussion given here #Bookofwhy #causalinference #ecobookclub

9.6.18 @10:06pm - (Replying to @preskill) It is not just you. Quite a few readers felt uncomfortable with my songs. I assume they misinterpreted my personal joy of doing science from other songs they must have heard in the past. For me, the whole thing was a miracle, and fun too. Enjoy #Bookofwhy

9.6.18 @3:07am - A new adage just came to mind: "Not all assumptions are equally indefensible." It came handy in a comment I wrote on Lars P. Syll's "Causal Interaction and External Validity: But I see its usefulness more broadly. #Bookofwhy #ecobookclub, #causalinf

9.5.18 @2:01am - (Replying to @aztezcan @ProfMattFox) I haven't given up, and twitter makes it easier: A DAG is a "KNOWLEDGE AMPLIFIER"; in comes qualitative knowledge and out go three new creatures: 1) Vivid photo of the input, 2) Logical implications of the input.3) Quantitative conclusions of data+input. Its a Miracle!#Bookofwhy

9.4.18 @9:27pm - Explaining KIKKKO - "Knowledge In, Triple Knowledge Out." 1st K - Your raw input knowledge. 2nd K - The logical implications of 1st K, which are now explicit (e.g., independencies in data). 3rd K - combined with data, 1st K turns from qualitative to quantitative. QED #Bookofwhy2:01am - (Replying to @aztezcan @ProfMattFox) I haven't

9.4.18 @3:21pm - GIGO means GIGO! That's what the wise men said when the calculator first came out. They forgot the other side of the coin: "KIKKKO means KIKKKO". "Knowledge In - Triple Knowledge Out." Always say KIKKKO, when you hear a wise man say GIGO. #Bookofwhy #causalinference #ecobookclub

9.4.18 @7:16am - (Replying to @MagellanDeStato @Altea_Lorenzo and 2 others) Gee, and I thought it is David Cox.

9.4.18 @6:53am - (Replying to @SchultenSimon @PHuenermund @eliasbareinboim) The 3 bullets hold in SCM; their negations hold in PO. And it is not a matter of mere visualization It is a matter of doing things correctly when you have no idea about "ignorability" (no mortal has). Complementary concepts? Yes! PO derives from SCM. But you cant skip the source.

9.4.18 @6:33am - (Replying to @SchultenSimon @PHuenermund @eliasbareinboim) There is nothing really wrong with the graph-less Potential Outcome framework, save for the three bullets at the end of this interview : *Interpretability, *Identifiability and *Testability. #Bookofway #causalinference @EpiEllie

9.4.18 @6:13am - (Replying to @totteh @PHuenermund @eliasbareinboim) To detemine effect heterogeneity in randomized experiments it is necessary to understand #causalinference, see for example . If you start by assuming ignorability, the problem becomes statistical: "Find a set of regressors such that..." . #Bookofwhy

9.4.18 @5:47am - (Replying to @Altea_Lorenzo @XihongLin) Thanks for the slide !!! The RSS must have changed since the last time I visited their Proceedings. Recall, in 1833 they vowed to publish only data, data and data, no opinion, no interpretation, and certainly no causation. A medal due to @XijongLin and @d_spiegel #Bookofwhy

9.4.18 @4:20am - (Replying to @epi_viborg @evpatora) Joking aside, I think you will find these two introductions to counterfactuals helpful in your PhD research: , and . If you are into #causalinference, you should also find them easy and intellectually challenging. Goodluck. #Bookofwhy

9.3.18 @10:10pm - (Replying to @thoughtfulflyer @davidhumeinst @PhilosophyMttrs) True, I do follow the thinking of David Hume, with one added twist: True, counterfactuals are behind causation, and behind most scientific thoughts but, to my delight, they can also be algorithmitized. Thus, next generation robots will be mini-scientists. #Bookofwhy

9.2.18 @11:36pm - (1/3) (Replying to @ngutten) RL can infer causal effects of those interventions only which RL can control and randomize. But, like old PO, it cannot infer effects of events (X=x) or states which are not directly manipulable. Whereas enlightened PO folks now accept some "hypothetical interventions", RL can't.
9.3.18 @12:41am - (2/3) (Replying to @yudapearl @ngutten) This comment on the role of reinforcement learning (RL) in the ladder of causation also brings to light what conservative PO is missing by rejecting the do-operator and insisting on physical, rather than conceptual interventions. @_MiguelHernan @Bookofwhy @causalinference
9.3.18 @12:47am - (3/3) (Replying to @yudapearl @ngutten and 2 others) The most obvious loss would be the use of surrogates experiments, i.e., finding manipulable variables that enable us to estimate the effects of non-manipulable variables. See for example: @_MiguelHernan #Bookofwhy @causalinference @causalityblog

9.2.18 @5:54pm - (Replying to @evpatora) Your summary of what you took from #Bookofwhy is very gratifying to me, as co-author. But the item that truly warms my heart is #5) "Counterfactuals are derivable." Because it is so crucial and so often ignored, even by causal analysts. Please post it high on your office door.

9.2.18 @3:43pm - (Replying to @autoregress) Note (Fig 1 #Bookofwhy) that the estimand is ALWAYS a statistical procedure, yet this does not make every problem "statistical". The fact that we can solve a regression problem using 2SLS, does not make it an IV problem. Its all in the description, not in the solution #ecobookclub

9.2.18 @5:24am - (Replying to @jonasobleser) Thanks for reminding me how words retain their truth even months after they were written. It is also relevant to the discussion of "data science" which I defined as the a 'two body" enterprise -- data and reality. Data alone is hardly a science. #Bookofwhy

9.1.18 @11:54pm - (Replying to @ehud) You may be right. Our quote "noticed this sudden shift in Galton's aims and aspirations: "What was silently missing was Darwin, the chutes, and all the `survival of the fittest.' . . ." may have been taken from Stigler 2012 The title "The Solution to Darwin" is surely misleading me how words retain their truth even months after they were written.

9.1.18 @5:40pm - (Replying to @dtweiseth) Not clear what you find objectionable in my treatment of Simpson's paradox or here . Is it in exposing its causal roots? or insisting on answering: "What do we do if we find it? (as in ) #Bookofwhy Puzzled ????

9.1.18 @5:32pm - (Replying to @jeremyfreese) I wish I could click five times on your "like" button. I normally tell publishers that I am "password illiterate". But they are getting bolder and bolder in converting our time and goodwill to profit and laziness. A "reviewers uprising" is in order.

9.1.18 @4:10pm - (Replying to @IntJObesity @_MiguelHernan) The Greeks invented formal logic because they got tired of debating endlessly, a habit that democracy stirred up . Honoring my Greek roots I propose we appeal to logic, rather than debated traditions in thinking obesity. @miquelporta @AlfredoMorabia @causalityblog #Bookofwhy

9.1.18 @2:06pm - (Replying to @mendel_random) The limitations of "statistical techniques" lie not in the "baggage of their history" but, rather, in their impoverished vocabulary. Sewall Wright extended that vocabulary by adding a new symbol: "an arrow", without which we can't define MR. #Bookofwhy #econbookclub

9.1.18 @1:57pm - (Replying to @mendel_random) Do you have the reference to the initial exposition of MR? I am curious to see how they justified the technique without causal logic, and without Sewall Wright.

9.1.18 @1:52pm - When a reviewer (of my paper in 2000) insisted that "instrumental variables" (IV) is a well-defined "statistical technique" I challenged him/her to define IV without invoking causal vocabulary. He couldn't! My paper was accepted. #Bookofwhy

9.1.18 @1:33pm - (Replying to @KordingLab @MariaGlymour and 2 others) What in the pre-causal days were called "causal insights" are today "causal assumptions", expressed in some formal model. If you know of a "causal insight" that has not yet been captured by SCM, let us know, and the next PhD project in my lab will be to capture it. #Bookofwhy

9.1.18 @1:18pm - (Replying to @ehud) I was under the impression that Stigler agrees with our interpretation that regression to the mean was not a solution but an abandonment of Darwin's question. Perhaps not as explicitly as we tried to make it, but he hinted that it was a cop out rather than a solution. #Bookofwhy

9.1.18 @12:57am - (Replying to @IntJObesity @_MiguelHernan) Glad to see @IntJObesity interest in our recent #Obesity debate. I wish to note that the link to my revised article is and that I believe the disagreement to be temporary, emanating primarily from Rubin's cultural roots in RCT. #Bookofwhy #causalinference

9.1.18 @12:15am - (Replying to @PHuenermund) Your discussion with Schaper brings to mind another useful motto: How can you tell if someone never used DAGs? Ans. He/she repeats slogans of potential-outcome loyalists; most typical: "The assumptions you find behind DAGs are stronger compared to those WE USE". #Bookofwhy

8.31.18 @8:40pm - (Replying to @MariaGlymour @KordingLab and 2 others) My tweet got lost: Cartwright's dictum "no causes in, no causes out" tells us that any approach that appears to be "better fit in "not really understandable systems" must make causal assumptions about what we presumably do not understand and is, if valid, a special case of SCM

8.31.18 @8:27pm - (1/3) (Replying to @KordingLab @alex__morley @dan_marinazzo) Cartwright's dictum "no causes in, no causes out" tells us that any approach that appears to be "better fit in "not really understandable systems" must make causal assumptions about what we presumably do not understand and is, if valid, a special cases of SCM#Bookofwhy 8.31.18 @8:29pm - (2/3) (Replying to @KordingLab @alex__morley @dan_marinazzo) The argument of "not really undersntandable systems" was used by "IV-experimentalists" to justify their model-blind approach. They stopped using this orgument (to the best of my knowledge), when someone showed them how easy it is in SCM to confess ignorance,. #Bookofwhy
8.31.18 @8:32pm - (3/3) One of my favorite motto in that discussion was: "It is only by taking models seriously that we learn when they are not needed." It never fails.

8.31.18 @11:00am - Judea Pearl Retweeted @KordingLab Aug 31 Replying to @alex__morley @dan_marinazzo The other problem is that @yudapearl #bookofwhy type causality approaches mostly help when there is hope for you to get at the network that represents reality. Other approaches such as RDD may be a better fit in "not really understandable" systems.

8.31.18 @7:07pm - (Replying to @eddericu @berglund_anita) Anyone who wishes to obtain the solution manual for the beautiful examples in should write to and indicate that it will be used for self-study and will not undermine instructors who assign these questions to grade students. #Bookofwhy

8.31.18 @1:54am - (Replying to @goodfellow_ian) The theoretical impediments listed in #Bookofwhy apply to any model-free approach to learning, not necessarily DL. If we define DL as a function-fitter then it cannot rise above rung 1 -- association. To be a causal reasoner DL must be guided by a model that tells it what to fit.

8.31.18 @1:37am - The Science magazine podcast containing our #Bookofwhy interview came out today! You can find it at . The book review segment starts at 18:48, and lasts a little over 5 minutes.

8.31.18 @12:53am - I was happy to find out today that my twitter has swelled to 10K followers. @berglund_anita deserves special recognition and a signed copy of #Bookofwhy for crossing the 10,000 mark. I will add an honorary plaque and a solution manual for all homeworks in

8.30.18 @12:41am - Many readers of #Bookofwhy were kind enough to send us typos and errors for consideration. We have now compiled them into one errata file For smooth reading, please mark your copy, and alert others to do the same. #causalinference #Epibookclub #ecobookclub

8.30.18 @2:21am - (Replying to @goodfellow_ian) The problem with today's DL is not its depth but its model-blindness. To appreciate, take your favorite DL program and run it for two weeks on data (X,Z,Y) generated by the smoking-tar story of Fig.7.1 #Bookofwhy. Would it ever conclude that smoking does not cure cancer?

8.30.18 @2:09am - (1/3) Papers often cited as combining DL and causality fall into three main categories. (1) Ignoring the impedements mentioned in . (2) Circumventing the impdiments by invoking approximate causal models, ... #Bookofwhy
8.30.18 @2:09am - (2/3) elicited either exogenously or from data (Pillar 7), and (3) Stripping problems from their causal content by assuming "ignorability," then using ML methods to estimate the left-over regressions. Only (2) deserve the title "combining", and we should watch them carefully.
8.30.18 @2:09am - (3/3) Note however that is we do not learn how to infer actions and counterfactuals from a given model, we are not likely to do so from data-extracted models. And this is the best commercial I can compose for #Bookofwhy

8.29.18 @5:13am - (1/2) If two different changes in the environment give rise to same change in the data, each calling for a different action (or a different way of handling the change) then it is hard to see how DL can look at the data only and act correctly. #Bookofwhy #causalinference
8.29.18 @5:17am - (2/2) Examples of how different environmental changes result in the same distribution are here: (Fig 3 a,b,c), which is also the gentlest introduction to transportability. Again, the meat is in the examples, not in ideologies. #Bookofwhy #causalinference

8.28.18 @4:07am - (1/3) I am elated and encouraged to see a record number of 450 'likes" on my recommendation to regression-minded data-scientists . It tells me that regression analysis is still the working horse of data science, that regression analysts are not too happy ...
8.28.18 @4:10am - (2/3) ...with the literature with which they were broght up, and that they are eager to enter the age of causation with renewed vigor. One word of advice: do not under-estimate the power of toy examples. If you really want to dig the new spirit of #causalinference and #Bookofwhy,
8.28.18 @4:17am - (3/3) solve the problem of poor Joe (Sections 4.4 & 4.5), who wonders what his salary would have been like had he had one more year of education. It is a simple exercise in pure counterfactual analysis, free of philosophical and cultural hangups -- dont skip. #Bookofwhy

8.28.18 @2:31am - (Replying to @LeaHilde) Glad to see that @Bookofwhy and @causalinference made it all the way to Amsterdam. Yes, the exercises should be super helpful. I can't think of a more exhilarating experience than solving in 2-3 lines problems that generations of philosophers deemed to be "meta-physical".

8.28.18 @12:49am - (Replying to @HolgerSteinmetz @PHuenermund @Noahpinion) This takes us back to my heroine Barabara Burks (1926) (#Bookofwhy Chapter 9), who startled her colleagues with an innocent question: "What makes us think that a partial regression brings us closer to effect size than pair-wise regression.?" She could't get a job.#causalinference

8.28.18 @12:07am - (Replying to @HolgerSteinmetz @PHuenermund @Noahpinion) This paper has nothing to do with causation. They talk about "effect size" (as do many wishful) but if you replace "effect size" with "regression-coefficient size" every sentence will still be valid. My litmus test is notation: Diagram?, do(x) or Y(x) ? else it's pre-causal

8.27.18 @11:49pm - (Replying to @tdietterich @GaryMarcus) I have reasonable doubts that these problems will be solved withing the DL paradigm. Why? Because generalization is a causal notion, needing a model of what relations might be perturbed by changing environments. DL researchers are still treating it as a statistical problem.

8.27.18 @5:35pm - (Replying to @Jabaluck @cdsamii @analisereal) The "overfitting" metaphor, whereby "the more knobs there are the easier it is to get the desired result" is a relic of statistical thinking. (as is "letting the data speak for itself") Things are quite different in causal modeling, begging for a drastic paradigm shift.#Bookofwhy

8.27.18 @4:47pm - (Replying to @Jabaluck @cdsamii @analisereal) "Full models" do not exist. We are discussing modeling very few variables that you perceive to be relevant to your problem of interest. For example, can you convert a bad IV into a good IV by conditioning on another variable?. A repeated question that requires a tiny model yet...

8.27.18 @3:51pm - (1/2) (Replying to @Jabaluck @cdsamii @analisereal) You just hit on another virtue that economists can borrow from #causalinfernce and #Bookofwhy (it goes back to Haavelmo ): "Thinking" is done only ONCE - when the model is constructed - he rest is vividly displayed. The model spares you the torments..
8.27.18 @4:00pm - (2/2) (Replying to @yudapearl @Jabaluck and 2 others) The model spares you the torments of "thinking how X2 might be endogenous," and other thinking exercises that "IV-experimentalists" are tormented by, all for refusing to put down a model of how variables are affecting each other. #Bookofwhy #causalinference

8.27.18 @2:54pm - (Replying to @dtweiseth) I never assume "data" is just snapshot. See for example #Bookofwhy Fig. 7. 6, which depicts sequential treatments. Temporal information, if available, can significantly enhance our model's veracity, but it cannot replace causal information. eg the rooster precedes the sunrise.

8.27.18 @1:18am - (1/2) (Replying to @Jabaluck @ben_golub) @Jabaluck, I think you are right. The question: "What can economists learn from DAGs given that they already excel in 1,2,3 ..." should not be answered with litmus tests for 1,2,3,... This tends to offend professional prides and create eternal enemies. Instead ...
8.27.18 @1:22am - (2/2) (Replying to @Jabaluck @ben_golub) Instead, the answer should be: "This is where DAGs excel and, if it has any merit for economics, your bright students will smell it from 5 miles away and adopt it as obvious." The #Bookofwhy aims to do just that. Independenly, I invite them to enjoy: .

8.26.18 @8:11pm - (Replying to @eliasbareinboim @deaneckles and 6 others) The implication of Elias's note is: If someone gets an answer different from the one obtained by do-calculus, it must be that someone either erred or failed to solve it without the do-operator -- a risky practice of some potential-outcome circles. #causalinference #Bookofwhy

8.26.18 @7:10pm - (Replying to @ravithekavi) Paragraphs that withstood the test of time, and in which I still believe despite a stormy year in the trenches deserve a special status in my diary. Thanks for bringing it to my attention, Ravi. #Bookofwhy

8.26.18 @4:47pm - (1/2) (Replying to @Jabaluck @analisereal) It is a deeply cherished belief among economists that "econometrics has powerful results" for 1) 3) 5) ... The litmus test is to ask your brightest student to solve any of the toy problems presented here: . Have things improved since 2014? Lets see
8.26.18 @4:57pm - (2/2) (Replying to @yudapearl @Jabaluck @analisereal) I have chanced to read a very recent survey paper on econometric identification. Things have only gotten worse by the "IV-experimentalism" campaign, which now absolves economists from thinking about causal models, including about whether a variable is a proper IV. #Bookofwhy

8.25.18 @10:51pm - (1/2) (Replying to @eliasbareinboim @jpirruccello @_MiguelHernan) Apologies! You are absolutely right in objecting to my playful poetry about Harvard not welcoming DAGs. The poetry came from a discarded chapter of #Bookofwhy describing the funny schism between the noth side of the Charles River (Statistics) where diagrams are tabooed, ....
8.25.18 @10:56pm - (2/2) (Replying to @eliasbareinboim @jpirruccello @_MiguelHernan) ... and the south side (HSPH) where diagrams are encouraged. My reply to Kevin Gray referred to the North side, and I will correct it on the blog. ...#bookofwhy #causalinference #Epibookclub @MiguelHernan

8.24.18 @10:49pm - For data analysts who wish to understand causes and counterfactuals from linear regression viewpoint, I recommend this tutorial: . It illuminates key concepts in #Bookofwhy by the light of linear models: effects, mediation, robustness and generalization.

8.24.18 @3:39am - Sorry for the wrong link. The gentle introduction to counterfactuals can be found here which is Chapter 4 of the Primer Enjoy.

8.24.18 @2:53am - This gentle introduction to counterfactuals should clarify for everyone the perplexing questions that were tweeted here: do(x) vs. Y(x), structural vs. PO framework + more. Enjoy, and don't let clarity surrender to tradition #Bookofwhy #causalinference

8.24.18 @2:21am - (Replying to @PWGTennant @eliasbareinboim and 3 others) Likewise, save for two problems: (1) Where do we place "diagnosis" (eg Bayes' billiard Table) which is not "prediction" and it needs no intervention? (2) Suppose we can run RCT on every variable, can we estimate countefactuals? eg ETT, "but-for", CoE mediation, etc. #Bookofwhy

8.23.18 @10:24pm - (Replying to @rationalexpec) Causality is more advanced, requiring graduate level math, elementary probability theory and some logic. It goes through proofs, and deeper philosophical discussions. It still stands correct though, like a rock, despite the dancing and prancing. #Bookofwhy

8.23.18 @8:23pm - (Replying to @atatlas123) Sure! We can't let publishers slow down progress. If you write to my assistant we'll make sure you get the solution manual (for personal use). Truly nice questions and truly neat solutions! Enjoy. #Bookofwhy #Epibookclub #causalinference #causality

8.23.18 @7:49pm - (Replying to @KordingLab @melb4886 @fadinberg) An economist "careful about thinking about causality" should re-think if he/she is in the right profession, b/c causal policy makers rely on his/her thinking. As to Causal Revolutions, mine begins here: and here . Yours? #Bookofwhy

8.23.18 @6:49pm - The next reading to dive into would be which is the technical compendium of #Bookofwhy, and includes beautiful examples, non trivial exercises and software support. For a quick glance click on Pearce and Lawlor 2016 review # epibookclub #causalinference

8.23.18 @5:57pm - Normally the choice of taxonomy is arbitrary. But the more I think about it, the more I suspect the difference between Miguel's trichotomy and the Ladder of Causation may account for the hurdles we encountered while reading chapters 7-8 of #Bookofwhy. Any #Epibookclub thoughts.

8.23.18 @4:23am - (Replying to @HolgerSteinmetz @PHuenermund @Noahpinion) Tell us more about psychology. Any major article unveiling how far behind the field is? Any embarrassing statement by the leaders? I know that Psychometrika just rejected an article on graphical methods, but this is not unusual; editors are paid to impede progress. What's new?

8.23.18 @2:14am - (Replying to @wgeary @MSFTResearch @akelleh) This is great news, which should make causal inference a household item. It reminds me of Max Planck's quote: "A new scientific truth does not triumph by convincing opponents to see the light, but because a new generation grows up that is familiar with it." (1949) #Bookofwhy

8.23.18 @1:49am - (Replying to @swanderingf) This is another good question: doing discovery in potential outcomes (PO). The answer is OF COURSE, the two are equivalent, but why would anyone want to do arithmetic with addition only? The SCM framework is PO + DAGs, the former emerging from the latter. So SCM is PO. #Bookofwhy

8.23.18 @1:02am - (Replying to @yudapearl @_MiguelHernan @rdpeng) Note that Miguel's trichotomy is different from our Ladder of Causation, which behooves readers to ask: 1) Where is "diagnosis" (or "abduction") situated? and 2) Where is the barrier between intervention and counterfactuals situated?#Bookofwhy #causalinference

8.23.18 @12:45am - (Replying to @rjmori @PHuenermund @Noahpinion) This looks familiar!!! Its the definition of (nonparametric) identifiability from Causality (2000)!- Thanks for sharing! Sadly, even the most recent survey paper on econometric identification that I reviewed does not provide a definition, not to mention a solution. #Bookofwhy

8.22.18 @10:22pm - (Replying to @ShlomoArgamon) My deep-learning problem is much more modest. Forget solving, I am concerned merely with representing the sentence: "the barometer fall does not cause rain". I do not know how to encode it in any deep-learning program. Once encoded we can try solving, but not before. #Bookofwhy

8.22.18 @10:07pm - (Replying to @swanderingf) Great question! The answer is YES, under selection. Examine the graph: X---->Z<---U--->Y. If we select patients having Z=1 (say pain from treatment X) then the path between X and Y will be active, producing correlation w/o causation ie confounding. #Bookofwhy #causalinference

8.22.18 @9:19pm - (Replying to @PHuenermund @eliasbareinboim) I'd never give up hope; see #Bookofwhy. Several economists indeed Tweeted that things are waking up, and the paper by Abadie-Cataneo affirms it. Still, the editors of Econometrica do not see urgency in teaching readers what identification and endogeneity are from a causal lens.

8.22.18 @7:28pm - (Replying to @_MiguelHernan @rdpeng) I specially love: "Enough with platitudes like "gaining insights" & "extracting meaning". These were favorites of pre-causal statisticians. Many still fail to realize that "meaning" and "insights" have been transformed into formal questions and algorithmic answers #Bookofwhy

8.22.18 @3:59pm - (Replying to @PHuenermund @Noahpinion) Economists had to change the definition of "identification" because 95% of them do not know what it means (especially the PhD's). It is the only field that prides itself on embarrassment. The gurus admit it, but will do nothing about it #Bookofwhy

8.22.18 @2:04pm - (Replying to @djvanness @rdpeng) Hybrid process is indeed what you find today in #causalinference and #Bookofwhy but the Bayes option is a dangerous Siren song. As I explain here , why I am only Half-Bayesian. Spraying priors and waiting for posteriors to peak will never flip --> into <--

8.22.18 @6:58am - (Replying to @djvanness) The point I make in is that, even if we let a deep-learning algorithm digest the data for three years it still has no language to postulate a scientific hypothesis, such as "the barometer fall does not cause rain" #Bookofwhy @rdpeng no C in, no C out.

8.22.18 @6:26am - (Replying to @PHuenermund @eliasbareinboim) It is easier to convince the Pope to preach Voodoo than to teach an economist to speak cause-effect. But its not entirely their fault -- it has been so long since they last communicated with the outside world, and their journals would not publish well-spoken papers. #econometrics

8.22.18 @5:55am - Very well put! But I would go even further: "To be a science, data-science should start where science does." I tried to add it here but one reviewer got offended. #Bookofwhy #ecobookclub #Epibookclub

8.21.18 @6:15pm - (Replying to @lisabodnar @ja_labrecque_ and 5 others) The correct link to the revised "On non-manipulable causes" is Sorry.

8.21.18 @6:05pm - (Replying to @lisabodnar @ja_labrecque_ and 5 others) For passengers on board of the "Obesity" cruiser, I have just revised the paper "On non-manipulable causes" . Thanks for all your comments. Key message: Time to separate epidemiological research from its potential outcome roots #Bookofwhy #causalinference

8.21.18 @3:13pm - (Replying to @Undercoverhist @causalinf and 9 others) Frisch "thoughts experiments" may be imaginative and insightful (like "Maxwell's demon"). But Haavelmo's priority shines through the credo: "A formula is a baked idea. Words are ideas in the oven" #Bookofwhy p 335 Note, Granger and Sims are regressionists - no C in, no C out.

8.21.18 @2:40pm - (Replying to @ProfMattFox) No need to struggle. The "Probability of sufficiency" (#Bookofwhy) stands for what it says, i.e., the prob. that an event X=1 would be sufficient to produce outcome Y=1. It is informed by the data. What PIE stands for is still waiting for experts' interpretation in #Epibookclub

8.21.18 @3:31am - (1/3) Anyone interested in the historical origins of the "well-defined intervention" fixation can watch Don Rubin's talk on youtube (nov 2016), describing where potential outcomes came from, and how they can influence users choice of treatment. #Bookofwhy
8.21.18 @3:40am - (2/3) Rubin also stresses the importance of separating Science from what we do to learn about it. I mention it hoping to convince Rubin's disciples to separate causes (the Science) from interventions (what we do to learn about them) #Bookofwhy #causalinference #Epibookclub
8.21.18 @3:51am - (3/3) The link to Rubin's talk is here : Note also how the antecendant x in the potential outcome Y(x) is always a "treatment", never a state (e.g., temperature = 30) on event (e.g., earthquake). and this is 2017. #Bookofwhy #causalinference #Epibookclub

8.19.18 @11:23pm - (Replying to @thosjleeper @eliasbareinboim and 2 others) Our ability to define (race = black) has little to do with manipulation. Suppose we make the term very very precise, say by requiring a spectral content of skin pigments. Now what? Would the sentence "The cup would not have shot him if only he was white" be more "well-define"?

8.19.18 @6:48pm - (Replying to @robertwplatt @ProfMattFox) Hard to measure, or no way to measure are different from "not well-defined" . The structural framework makes this distinction crisp, the potential-outcome framework conflates the two. Why not choose clarity? If something has no effect then that effect must first be well-defined.

8.19.18 @6:19pm - (Replying to @jamessseattle) I am not familiar with Eliasmith's work. If you have the source, please try to summarize it within Twitter's constraints. Just what the input is and what the output we can expect. Thanks, JP

8.19.18 @5:22pm - (1/2) (Replying to @EpiEllie) Ellie, Congratulations on finishing the @Bookofwhy. It has been truly edicational for me to work with the #Epibookclub and learn all the obstacles that readers may have and which we assumed away. Please pass my hugs to all the survivors. Hey, how was the paradox party? 8.19.18 @5:36pm - (2/2) (Replying to @yudapearl @EpiEllie) Speaking of learning more #causalinference and going a step beyond the non-technical style of #Bookofwhy, I must suppress modesty and recommend Uniquely liberated from "potential outcomes" handcuffs, yet smiling in coherence and examples.#Datascience

8.19.18 @4:21pm - (Replying to @ProfMattFox) Sec 5: "The basis for rejecting the new drug is precisely your understanding that "obesity has no effect on outcome," the very quantity that some epidemiologists now wish to purge from science, all in the name of caring only about "what to do"#Bookofwhy

8.19.18 @3:09pm - (Replying to @thosjleeper @eliasbareinboim and 2 others) You hit it on the nail: "phenomena for which effects cannot be defined because the cause is poorly defined." Isn't it possible that such phenomena are fictional? or cultural habit?, a relic of Rubin's RCT roots which epidemiologists now should, but find hard to shake off?

8.19.18 @6:24am - (1/6) (Replying to @mendel_random) #Bookofwhy has two missions: 1. To lay before readers what can and cannot be done given a set of "defensible anchors" (DAs) that they may possess. 2. To make those DAs advertise their own defense by making them meaningful to and scrutizable by as wide a pool of peers
8.19.18 @6:28am - (2/6) (Replying to @yudapearl @mendel_random) a pool of peers and adversaries as possible. The fact that you are able today to find flaws in some of those DAs, articulate by your peers, proves that mission (2) has been accomplished. Had these DAs been made under the rug of ignorability or, worst yet,
8.19.18 @6:33am - (3/6) (Replying to @yudapearl @mendel_random) under the methodology that ruled epi prior to 1990, you would be unable to judge their plausibility -- they would be buried in hand waving. As to mission (1) it is purely mathematical, hence unassailable. Still, showing what is impossible saves many hours of futile research
8.19.18 @6:37am - (4/6) (Replying to @yudapearl @mendel_random) and showing what is possible tickles the imagination to find defensible opportunities for the unveiled possibilities. There remains the problem of what to do when a researcher finds himself with no DAs. You addressed it in Ref 1, and proposed a "broaden scope" approach.
8.19.18 @6:44am - (5/6) (Replying to @yudapearl @mendel_random) I then showed step by step in or that what your "broadened scope" proposal aspires to do is already done in DAGs. Including "triangulation" "reasoning to the best explanation" and other aspirations. #Bookofwhy
8.19.18 @6:54am - (6/6) (Replying to @yudapearl @mendel_random) The argument that powerful tools are dangerous because they invite misuse was made against the telescope in 17th century Italy. I dont think it applies to 21st-Century Epi. The priests are scientists, epidemiologists are smarter, and their tools are transparent. #Bookofwhy

8.18.18 @6:38pm - (Replying to @EpiEllie @eliasbareinboim) Happy that you pointed to the source paper. Note though that the paper starts quite friendly and provides ample examples. Importantly, the paper aims to be "anti-standardization", showing that standard "standardization" methods (ie, re-calibtation) must be revamped. #Bookofwhy

8.18.18 @6:17pm - I am sharing a new paper on non-manipulable causes which summarizes my position in the semi-heated discussion we have had here regarding Obesity and its consequences. May all discussions come to a good cause. #Bookofwhy #causalinference #Epibookclub

8.18.18 @5:39pm - (Replying to @eliasbareinboim @EpiEllie @leskocar) The PO perspective has advanced considerably in the past decade. It started with "DAGs are not helpful", and advanced to "DAGs are helpful but only to verify ignorability". I now see it heading towards: "Ignorability is not really needed" #Bookofwy #causalinference

8.18.18 @4:00pm - (Replying to @bobgrossman) There is also free access to Chapter 2, for those who care about the genesis of causal analysis, how statistics squandered it, and how my hero, Sewal Wright, single-handedly stood in: "And yet it moves!" Where did he get this Chuzpa? #Bookofwhy

8.18.18 @4:03am - (Replying to @fadinberg) It is not only a relic of the past. I am asking myself what big-data and deep learning would recommend to patients after analyzing megatons of medical records. #bookofwhy #epibookclub

8.18.18 @1:03am - (Replying to @BreskinEpi @eliasbareinboim and 5 others) Eager to hear what the restrictions are: Number of pies? content of pies? 2-SAT? Horn restriction? Moreover, given a restricted PIE, does it let you "visualize" things like synergism or interactions, or send you back to sweat over the various pies in the system? #Bookofwhy

8.17.18 @8:40pm - (Replying to @BreskinEpi @eliasbareinboim and 5 others) One reason computer scientists are baffled by PIEs is that, recognizing that "A and B only cause Y when both are present" (given a bunch of N PIEs) amounts to solving an N-variable SAT problem, which is hard. Can PIE do what math says we can't? #Bookofwhy #Epibookclub

8.16.18 @10:04pm - (Replying to @EconTalker @SpenceKjell) Russell, Rumor has it that you are trying to reach me. I am alive and reachable on Somewhat tired, but alive. Judea

8.16.18 @8:06pm - (Replying to @JaimieGradus @EpiEllie and 2 others) As far as #bookofwhy is concerned, the purpose of Fig. 7.5 is to show: (1) that even toy problems can present a formidable challenge to the unaided mind (and PO). (2) Mt. Intervention is conquered, (3) science can be fun. Readers who internalized these 3 points need not solve 7.5

8.16.18 @6:55pm - (Replying to @malco_barrett @EpiEllie and 3 others) Back to PIEs. The fact that so many smart people are in love with PIEs gives me 99% assurance that there is something good there. The fact that none was able to explain how to look at a bunch of pies and get the insight they swear by tells me that someone should do it. #Bookofwhy

8.16.18 @3:15am - (Replying to @statsepi @tmorris_mrc and 5 others) What most people do not realize is that MAR can now be tested, see So, I imagine 2 years from now your abstract will read:: "We used MI and passed our data through the famous MAR test." No editor will resist ! No reviewer will squeak! #Bookofwhy

8.16.18 @1:09am - (Replying to @julesgreig) Fascinating article, thanks for sharing. No, my math cannot prove causation, it is after all only math, namely going from assumptions to conclusions. But even this dry exercise is fun, because you find things you did not expect. For example, that Reichenbach forgot colliders.

8.15.18 @11:32pm - (Replying to @MaartenvSmeden) Even after data collection it is difficult to infer causes of unexpected missing data, or even expected missing data. But we found that the easier problem too is in total confusion: Suppose we postulate those causes. Can we estimate causal effects?

8.15.18 @10:19pm - (Replying to @causalinf @jakewertz and 9 others) Philosophy aside, to me, the causal revolution in econ begins with Haavelmo, and I justify it step by step in . The ET Editors asked me to comment on the "credibility hype." My conclusions (Section 4): they did not pass the causal litmus test. #Bookofwhy

8.15.18 @9:29pm - (Replying to @thosjleeper @TheBrettGall) In #Bookofwhy I praise Morgan and Winship for saving social science from the fate of economics. Indeed they cover both sides, missing one tiny step -- showing that the two sides are ONE. Namely, results obtained in one are obtainable in the other. A DAG generates all PO's.

8.15.18 @8:57pm - (Replying to @AndersHuitfeldt @MaartenvSmeden) The most crucial thing to learn from data is which variables are contaminated by missingness and which are not. We usually know this in advance. What this paper tells us is, assuming that we know which is which, can we estimate what we need? #Bookofwhy

8.15.18 @8:42pm - (Replying to @thosjleeper @TheBrettGall) You dont need to drop potential outcomes (PO). All you need to do is to see how PO's are derived from DAGs, then join the PO camp pretending you are one of them, unsure of where PO's come from -- you will perform miracles. #Bookofwhy #causalinference #Epibookclub

8.15.18 @6:37am - (Replying to @MaartenvSmeden) My faithful oracle #Bookofwhy says that "missing data mechanism" is a causal notion, hence it cannot be "inferred" by data collection. It sometimes has testable implications, and this paper tells us when/how . It even tests MAR, and some MNAR. New and fun

8.14.18 @6:45pm - Intrigued by the "Paradox of inevitable regret", ,some readers asked how Forney and Bareinboim managed (2017) to combine information from observational studies to improve on RCT performance. Here it is: #Bookofwhy

8.14.18 @3:29pm - (Replying to @AndersHuitfeldt @_MiguelHernan @analisereal) Whatever helps detect model misspecification is useful. But philosophical differences make a huge difference in practice. They lure whole communities to dismiss the need for model specification, or even for models per se. Look at all the model-blind armies around us. #Bookofwhy

8.14.18 @5:54am - (1/2) (Replying to @AndersHuitfeldt @_MiguelHernan @analisereal) First, "obesity" is not a "node", it is a vector of 17 factors that captures the word "obesity". Second, a model is not a 'rug", it is our only window to reality. Third, the 17 factors may be "misspecified", fine, but their effects remain intervention-neutral. #Bookofwhy.
8.14.18 @5:05am - (2/2) (Replying to @AndersHuitfeldt @_MiguelHernan @analisereal) Surely there is ambiguity in defining "obesity" and in handling side-effects of interventions. But the whole point of this discussion is to treat these sources of ambiguity as instances of "model misspecification" not of "undefined-ness" or "non-scientific-ness" #BookofwhyReplying to @AndersHuitfeldt @_MiguelHernan

8.14.18 @2:56am - (Replying to @_MiguelHernan @analisereal) This text makes me cringe: "different interventions often result in different effect estimates... " Are we are all doomed? No! In SCM, different interventions NEVER result in different effect estimates. X in P(y|do(x)) is not the Soda, it is Obesity. Breath safely #Bookofwhy

8.13.18 @8:14am - (Replying to @EpiEllie @l__ds and 3 others) Correction: These two sets are logically equivalent:
set1 (A=1,C=0) (B=1,C=1)
set 2 (A=1,B=0, C=0) (A=1, B=1) (A=0, B=1, C=1)
A and B interact in set 2, but not in set 1 Can you share "the rules of causal pies"? #Bookofwhy #causalinference #Epibookclub

8.13.18 @7:11am - (Replying to @EpiEllie @l__ds and 3 others) I used to think the way you do. But consider these three pies:
(A=1,C=0) (B=1,C=1) (A=0, C=1)
A and B do not interact. Now consider:
(A=1,B=0, C=0) (A=1, B=1) (A=0, B=0, C=1)
Here A and B do interact, right? Alas, the two are logically equivalent. #Bookofwhy #causalinference

8.13.18 @6:02am - (Replying to @EpiEllie @l__ds and 3 others) If a PIE helps "visualize how interaction could arise" then I expect someone to be able to look at the PIE and tell whether "interaction could arise". This amount to "detecting" interaction. So, how do we tell interactions (or some other relation) by looking at a set of PIEs ?

8.13.18 @2:27am - (Replying to @AndersHuitfeldt @eliasbareinboim) Put yourself in the shoes of a 1925 farmer: "The RCT is interesting because and only because it corresponds to what will happen if I use-Fertilizer1. If those two ever came apart (and Fisher did not proved they wont), the RCT is no longer interesting." Thus RCT comes 2nd to do(x)

8.13.18 @1:07am - (Replying to @evpatora) Thanks for sharing the arithmetics with other followers of #Bookofwhy. The three steps needed in computing potential outcomes are fundamental to understanding counterfactuals and, honestlyy, I dont believe anyone can understand counterfactuals without doing these steps ONCE. Thkx

8.13.18 @12:55am - (Replying to @AndersHuitfeldt) The decision makers that I know could not care less about "what would have happened if we ran RCT". They care about "what will happen if we use fertilizer F1 (vs F2) on the entire field" And this is what the do-operator simulates. #Bookofwhy #causalinference

8.12.18 @8:44pm - (Replying to @l__ds @EpiEllie and 3 others) I also thought that the relationships between SCC and PO are well formalized. Today I begin to suspect that they were left hanging. I dont know what "synergism" is, but formal definition of "Sufficient Causation" is on page 289, line 4. #bookofwhy Can Ellie read it from PIEs ???

8.12.18 @8:15pm - (Replying to @l__ds @EpiEllie and 3 others) In the SCM model, the function Child=f(Parents) is part of a complete specification of the model, usually kept implicit. The SCC model explicates this function. I am still waiting to hear from Ellie how she detects notions eg. "interaction" from a collection of PIEs.#bookofwhy

8.12.18 @7:18pm - Errata: A diligent reader of #Bookofwhy, anxious to verify the fallability of MATCHING in potential-outcome analysis, has found a typo on page 282. Line 22 should read: -$9,500, instead of $5,000. No change in conclusions. Please mark your copy. #Epibookclub

8.12.18 @6:19pm - Congratulations go to co-follower @eliasbareinboim, + his student and collaborator, who were just awarded the UAI best student paper award! The paper provides an interesting way of identifying causal effects from equivalence classes of models. #Bookofwhy

8.12.18 @4:48pm - (1/n) Heavens, No! If there is ONE thing I hope readers will take from #Bookofwhy it is that causal quantities (eg, causal effects) should not be left to the mercy of English or Potuguese, but be defined as a property of a formal model. Not back to Alchemy! #causalinference 1/n
8.12.18 @5:10pm - (2/n) I hope that #Bookofwhy convinces readers that defining "causal effect" in English won't let us verify if any proposed estimator yields the quantity we wish estimated. If doubts remain, then a more convincing book must be written dedicated entirely to this crucial p oint. 2/n

8.12.18 @5:48am - Before quitting #Bookofwhy chap 9 with counterfactual heart-breaking headache, recall, this is precisely the headache that causal models prevent. Modelers think only ONCE, when constructing the diagram. The rest is algorithmic, including mediation. Headache is for the birds.

8.12.18 @4:38am - RCT-ists beware: The important problem of "surrogate endpoint" has been lingering in decades of quasi-formal confusion. A new paper by Tikka and Karavanen ... now submits the problem to full causal analysis. Worthy of attention! #Bookofwhy

8.12.18 @3:40am - (1/2) (Replying to @swanderingf) Before we ask "how do we measure the goodness of the assumptions" we should ask "why should not the assumptions themselves advertise their goodness"? After all they came from the roots of OUR knowledge, so we should be able to judge their plausibility. Next we can assist our 1/2
8.12.18 @3:52am - (2/2) (Replying to @swanderingf) Next we can assist our plausibility-judging software by doing sensitivity analysis, which is a way of focusing this software on the most crucial assumptions first. Finally, we submit a PROVISIONAL policy recommendation coupled with the assumptions. Why numbers? #Bookofwhy 2/2

8.11.18 @3:32am - (Replying to @cygarde) Off hand, "predictive models that change causal variables" sounds like a violation of Cartwright's dictum. I would love to see and toy example. Perhaps the example will also help me understand your proposal of "conjoint modeling". #Bookofwhy

8.11.18 @1:31am - (Replying to @FriedrichHayek) Pragmatically, I hope you agree that econometric education and methodology need foundational reforms, that such reforms will not come from its current leadership, and that the democratization of causal inference e.g., #bookofwhy is a step in the right direction. @PHuenermund

8.11.18 @12:55am - (Replying to @LauraBBalzer @Ron_Wasserstein @AmstatNews) In reading this interview on @AnstatNews, please note my super emphasis on the "three bullets" at the end. They are meant not only for hard core PO analysts, but also for those tormented by split loyalty between PO habits and new structural thinking. #Causalinference #bookofwhy

8.10.18 @2:06am - (Replying to @l__ds @EpiEllie and 3 others) Indeed, that PIEs were re-discovered in philosophy, epi and law says a lot for their naturalness. That's why I am so eager to explore what makes them "natural" and what function they serve in causal modeling, as well as in teaching epi. Help me explore, What is it? #Bookofwhy

8.9.18 @8:03pm - (Replying to @EpiEllie @malco_barrett and 3 others) No miscommunication, just systematic explication. If one's goal is quickly verify absence of interaction, three questions arise: (1) What for? (2) Is PIE the best representation for doing it? (3) What's our input? i.e., where is the knowledge needed to construct a PIE? @Bookofwhy

8.9.18 @6:48pm - (Replying to @EpiEllie @malco_barrett and 3 others) Gladly. In DAGs, interactions are presumed to exist a priori, with no bells or chimes. Why? Because the theory of Boolean functions (1848) tells us that the percentage of non-interactions in the space of functions Child=f(Parents) is super infinitesimal. #Bookofwhy cont. 1/n
8.9.18 @6:57pm - (Replying to @EpiEllie @malco_barrett and 3 others) In DAGs, if you want to quantify a specific interaction, all you need to do is to estimate P(Y|do(x),do(z)) and find out to what degree the effect of X on Y is modified by Z. This is a do-expression, so it is estimable by do-calculus. Done. No bells, no fancy articles @Bookofwhy
8.9.18 @7:20pm - (Replying to @EpiEllie @malco_barrett and 3 others) Continuing with interactions. Note that I wrote a do-expression P(Y|do(x),do(z)) and NOT a conditional expression as habits and PIEs so seductively tempt us to do. Note also that PIEs tacitly insist on components being Parents(Y), not so X and Z above. #Bookofwhy 3/n
8.9.18 @7:35pm - (Replying to @EpiEllie @malco_barrett and 3 others) This is not to say that PIEs have no role in SCM. But as analysts we need to explicate this role carefully if we want robots to use PIEs, or even just to teach Epi students, and "help them think". What can they do with a PIE at hand, compared with PIE in the sky? #Bookofwhy n/n

8.9.18 @5:57am - (Replying to @yudapearl @malco_barrett and 4 others) Given that Causal PIE Models do not tell us whether Oxygen or Match was a sufficient cause for Fire, we conclude that additional information is needed to make this distinction. Where does it reside? The #Bookofwhy says: it resides in the pre-fire distribution, and given by PS.

8.9.18 @5:50am - (Replying to @yudapearl @malco_barrett and 4 others Thinking of PIEs. Suppose Joe is naive causal analyst who never heard about PIEs, but manages nevertheless to express Child=f(Parents) in some other convenient way (say conjunctive normal form). What would Joe not be able to do that a PIE lover can. #Bookofwhy #Epibookclub

8.9.18 @5:45am - (Replying to @yudapearl @malco_barrett and 4 others) Thinking about Causal PIEs. Suppose Joe is a naive causal analyst who never heard about causal PIEs, and is using a DAG in which ACE is identified (say by backdoor.) What would Joe gain by learning about PIE. #Bookofwhy #Epibookclub

8.9.18 @2:59am - (Replying to @OstlundOllie @AndersHuitfeldt) To rule out interpretations based on selection bias, let us assume that the 200 students were randomly selected; not "volunteers". Let us also assume that students did not know whether they are in the randomized group or in the free choice group. Let's keep it simple. #Bookofwhy

8.9.18 @12:08am - Rumours say that a party is planned in a bookclub to celebrate the successful crossing of Chapter 6 - Paradoxes Galore! I would love to spice up the party with a humble a paradox of my own, , discarded from the #Bookofwhy by space limitations. Enjoy!

8.7.18 @1:12pm - This is a great article, by Chris Hitchcock, both profound and comprehensive. Thanks for bringing it to our attention here, at the Tweeting trenches. It even mentions #Bookofwhy, despite of what the book says about the philosophy of "probability raising" in chapter one.

8.8.18 @4:50am - (Replying to @causalinf) Thanks for the paper. I am totally speechless, and stand in awe to celebrate the publication of the FIRST econometric paper since Herbert Simon to contain a graphical model. As we say in the trenches: there is hope to econometric after all. #Bookofwhy Correction: forgot Hal White

8.8.18 @1:42am - (Replying to @malco_barrett @NeuroStats and 3 others) Agree. The role that Rothman's PIE plays in causal modelling is simple: You zoom in on any family in the DAG and, if the function Child = f(Parents) is Boolean, you can give it visual description, called Disjunctive Normal Form (DNF). But are the slices "sufficient causes" ? 1/3 8.8.18 @1:46am - (Replying to @malco_barrett @NeuroStats and
3 others) Is it justified to call the slices of the PIE "sufficient causes" ? Consider the famous oxygen-match-fire problem: Was the Presence of Oxygen a sufficient cause for the fire?. More trouble: The DNF is not unique. If A is a sufficient cause, so is any event B=1. Why? see 2/3
8.8.18 @2:03am - (Replying to @malco_barrett @NeuroStats and 3 others) If A is a sufficient cause, so is any event B=1. Why? Because A is logically equivalent to (A & B=1) OR (A & B=0). [See Causality Section 10.1.4. for discussion]. The notion of "sufficiency" described in #Bookofwhy is invariant to syntax and, surely, prefers Match to Oxygen. 3/3

8.7.18 @11:43pm - (Replying to @DKedmey) My all time favorite economist is Joseph. In econometrics I vote for T. Haavelmo and H. Simon. Modern heretics? None! None of the priests can solve any of the toy problems posed to them in . But I see sparks of heresy among the young, even here on Twitter.

8.7.18 @10:53pm - (Replying to @MariaGlymour) Beautifully said. Now I understand why I find more common language with epidemiologists than economists. #Bookofwhy

8.7.18 @12:22pm - (Replying to @malco_barrett @NeuroStats and 3 others) Great. But since every potential outcome model is a derivative of a SCM (structural causal model) we can ask: What connects SSC and SCM? To see this, lets ponder (1) What this pie means to you? (2) What the pie would look like when Y=1 if and only if A=B. #Epibookclub. #Bookofwhy

8.7.18 @6:07pm - (Replying to @erikbryn) Sorry, I overlooked your Aug 5 tweet. No reason for Kleinberg etal to remind us that prediction is important for policy making; else why would politicians would be hiring expensive pollsters. We are now learning the surprising converse; even sworn pollsters should study causation

8.7.18 @5:52pm - (Replying to @malco_barrett @NeuroStats and 3 others) At this juncture, my college professor would say: Now that we understand what a model is and how it helps us answer hard questions on interventions and counterfactuals, lets go back to our favorite PIE and ask where it fits in the #Bookofwhy scheme of things. Do I see any hands?

8.7.18 @12:22pm - (Replying to @malco_barrett @NeuroStats and 3 others) Great. But since every potential outcome model is a derivative of a SCM (structural causal model) we can ask: What connects SSC and SCM? To see this, lets ponder (1) What this pie means to you? (2) What the pie would look like when Y=1 if and only if A=B. #Epibookclub. #Bookofwhy

8.7.18 @2:00am - (Replying to @malco_barrett @ken_rothman) I am sure there is a connection between the pie and the notion of "necessary causation" mentioned in #Bookofwhy but to explicate it one needs to know: 1. what this pie means to you? and 2. what the pie would look like when Y=1 if and only if A=B. #Epibookclub. @sharlagelfand @epi

8.6.18 @6:56pm - (Replying to @DataPuzzler) As the next technical introductionI beyond #Bookofwhy I strongly recommend the Primer It contains references to software (Daggity) and has entertaining homework problems that even the gurus will find hard to solve.

8.6.18 @12:31pm - (Replying to @eddelbuettel) Thanks for noting this oversight; it will be corrected. My skepticism of econometrics is not about its scientific foundations but about its current cultish and insular leadership (see ) #Bookofwhy. @ndoogan

8.6.18 @12:14pm - (Replying to @malco_barrett @ken_rothman) My difficulty w Macksey's INNUS condition ( Causality chap 8) stems from a simple principle: You can't express counterfactual concepts (rung-3) in propositional logic (rung-1), the "necessary" (and "sufficient") idea requires counterfactual specification. #Bookofwhy @EpiEllie

8.6.18 @2:55am - (Replying to @stephensenn) I hope you agree there is a difference between "not throwing information away" and "use baseline as covariate". If we blindly stratify on (or "control for") every measured baseline we will get into the M-bias fallacy (e.g., page 161 @Bookofwhy). Any qualification? @f2harrell
8.6.18 @11:59am - (Replying to @yudapearl @stephensenn @f2harrell) The M-Bias fallacy does not afflict RCT's, so my request for qualification is answered. Thanks. #Bookofwhy #Epibookclub

8.6.18 @2:12am - (Replying to @EpiEllie) Let us not conflate specific conditions needed for estimation with relationships that hold universally in every problem. The first law of causal inference is universal, and consistency follows from this law, regardless of debatable concerns over "well-defineded-ness" #Bookofwhy

8.4.18 @1:57pm - This is what we used to believe, until problems of transportability, robustness and explainability taught us that even purely predictive tasks should be model-guided. See #Bookofwhy Gush! It rains whenever I forget my umbrella (selection bias in big data).

8.4.18 @10:45pm - (Replying to @erikbryn) I think you are saying: Yes to commonsense, Yes to the Ladder of Causation. Yes to modern #causalinference. Join us at #Bookofwhy.

8.4.18 @1:15pm - Well put, with one twist: replace "a good empirical strategy" with "a provable strategy", because most big data guys think they "have a good empirical strategy for identifying causal effects", citing leading statisticians who play down the importance of identification. #Bookofwhy

8.2.18 @1:48pm - To all captains and sailors on this ship, The errata sheet for #Bookofwhy has been updated. See . Thank you all for your comments and suggestions. Sail safely and enjoy the rest of your (smooth) voyage. #epibookclub #causality #causalinference

8.2.18 @2:01pm - Sorry, the correct link to the errata sheet for #Bookofwhy is . Sail safely . #epibookclub #causality #causalinference

8.1.18 @12;59PM - My interview on #Bookofwhy has just come out on Amstat News . Please note my "three bullets" summary of what potential-outcome enthusiasts can learn from causal modeling. #epibookclub #causalinference #Datascience

7.29.18 @8:38pm - For those who are hooked on Y(x) go ahead and replace P(Y|do(x)) with P(Y(x)), no rule will fail. Just recall: P(Y|do(x), do(z), W=w) goes into P(Y(x,z)|W(x,z)=w). Also note: the #bookofwhy does not go into the do-calculus, except conceptually. #epibookclub

7.29.18 @8:16pm - Got it, thanks. It is ready to be corrected

7.29.18 @6:54pm - Econ should have striven to be the "crown" because it had such a huge head-start over other disciplines. We need to ask: "What happened!" if we are serious about removing the obstacles that resist working together in true openness.

7.29.18 @6:39pm - Where exactly is the "land grab"? That I classify maximum likelihood in "traditional statistics"? or that I harness modern machine learning techniques to compute maximum likelihood in much larger models?

7.29.18 @6:22pm - The counterfactual Y_x (e.g, "Wet had it rained") is a unit property; it refers to identical circumstances, same season, same day, same grass etc. Do-expressions are population properties. Typical grass on a typical day will get wet if we do(rain). #bookofwhy #epibookclub

7.29.18 @3:02pm - The do(x) operator simulates what will happen if you actually apply the atomic action do(X=x). If you cant implement such action (eg. do(rain)) it still tells you what will happen if you could, and sometimes it does not matter, (eg, the grass will still get wet). #bookofwhy

7.30.18 @12:50am (Replying to @EpiEllie @eliasbareinboim @PHuenermund) Everything we do with do-calculus and counterfactuals, in the entire #bookofwhy is WELL-DEFINED, because "well-defineness" is a mathematical notion that we adhere to religiously by defining things unambiguously on the model. Estimation and implementation are orthogonal questions

7.29.18 @1:14pm - Absolutely, as long as it encode causal assumptions. And, BTW, I do not know of an economic model that is not a causal model, though many economists shake when you ask them.

7.29.18 @1:08pm - P(Y|do(x)) tells you that, to get ATE, you do not need to know counterfactuals (rung-3, Chapter 8), nor worry about "ignorability" magic before you adjust. Just make sure you adjust for variables satisfying the backdoor condition. #Bookofwhy #epibookclub #causalinference

7.29.18 @12:14am - #Bookofwhy #epibookclub. Chapter 6 should actually be most fun and rewarding Even the first section (the back-door) already walks you through territories that economists will find unchartered and leading statisticians will deem heretical, if not "strongly ignorable." Happy voyage

7.28.18 @8:38pm For those who are hooked on Y(x) go ahead and replace P(Y|do(x)) with P(Y(x)), no rule will fail. Just recall: P(Y|do(x), do(z), W=w) goes into P(Y(x,z)|W(x,z)=w). Also note: the #bookofwhy does not go into the do-calculus, except conceptually. #epibookclub

7.29.18 @11:36pm 1/2 (Replying to @_MiguelHernan @EpiEllie) This debate surfaces at least twice a year since Paul Holland's announced his model-blind dictum (1986) "No causation without manipulation". I have defended a model-based position in Causality (2009) Sections 11.4.3, 11.4.5, 11.4.6, and many more places. To summarize: 2/2

7.28.18 @1:45pm - Your review is excellent and comprehensive. Now that we dig the limitations of classic IV's, attention should perhaps be shifted to "mediating IV's" (frontdoor) which provide nonparametric point identification of ATE and ATT (#bookofwhy, p. 235-7), albeit under different setting.

7.27.18 @6:24pm - Lauren has discovered a typo in #bookofwhy. It should be P(no tar|do(smoking)). A few typos have already been corrected here: . A few more will be corrected soon. I suggest to mark your copy accordingly. Please let me know if any more. #Epibookclub

7.28.18 @1:28pm (Replying to @zongel @EpiEllie) This is another correction on our plate. Andrew Forbes alerted us to Elizabeth's major contribution and we will reflect that in the next revised printing. #Bookofwhy

7.30.18 @12:04pm (Replying to @yudapearl @_MiguelHernan @EpiEllie) Deja vu !!! I recognized this debate from the ancient past. Here it is again: "Does obesity shorten life?" . I hope it makes the reading of #Bookofwhy more enjoyable and less dusty: If you suspect side-effects, you must put them in the model - No cheating!

7.26.18 @11:12pm (replying to @IronEconomist) "... most important and underrated statistician". Hmm. The former is overrated and the latter makes me cry. But given that #Bookofwhy is #1 on Amazon best seller lists, in both stat. and biostat., I think statistics shows promising signs of awakening to modernity.

7.26.18 @9:24pm (Replying to @Jsevillamol) I just posted the answer on and invited you to view the full solution, which was posted here #bookofwhy yesterday. Glad mathematically inspired students see the challenge and beauty of graphical modeling.

7.26.18 @6:16am (Replying to @AndersHuitfeldt) It is not notational convenience but a semantical leap. To define Pr(Y_x) I need to define Y_x, not so in defining Pr(Y|do(x)); it is what we measure in a randomized experiment (rung-2). The fact that the two collapse under Pr does not make them equally meaningful (eg to computer

7.26.18 @4:13am (1/2) (Replying to @_MiguelHernan) I am over-flattered by your kind words about #Bookofwhy and thanks for posting our interview. We disagree on the role of intervention in the definition of "effect" and I explained my reasons for defining "effects" prior to intervention. (eg the moon affecting tides). 7.26.18 @4:42am (2/2) (Replying to @_MiguelHernan) We also differ on whether #Bookofwhy (p. 9) description of the do-operator is overrated in light of previous attempts to distinguish doing vs. seeing. I consider it pivotal, being the first to offer a knowledge-based counterfactual-free semantics for the distinction. (rung-2)

7.25.18 @5:37am (Replying to @aynumazi @Harvard) In the 1960-1980's, Simon, Lewis, Rubin and Robins devised various counterfactual theories (Chapters 8-9) which could distinguish actions from observations but did not become operational until the do-operator (1993) told us when the effects of actions can be estimated from data.

7.25.18 @1:18pm (Replying to @cdsamii @analisereal and 3 others) I like potential outcomes too, and use them as much as graphs. You know why? Because, given a graphical model, I know what Y_x means, where it is in the model, how it is derived from the model, etc. It is a good language to define the quantities we wish estimated. #bookofwhy

7.25.18 @7:38pm (Replying to @aynumazi @Harvard) Let me explain why your marked sentence is valid in its context. Prior to 1993 there was no formal way of taking what people understand about storms and barometers and proving them right on their intuitive predictions. Counterfactuals would not have gotten us very far.

7.25.18 @1:49pm (Replying to @_MiguelHernan @aynumazi @Harvard) Under certain conditions, the do-operator bears syntactic resemblance to g-formula ie, both are products of conditional probabilities. The resemblance ends here. The do-operator is defined on my knowledge -- a causal diagram, while the g-formula on an object I do not recognize.

7.25.18 @6:31pm (Replying to @AndersHuitfeldt @_MiguelHernan and 2 others) Are you familiar with Stalnaker-Lewis definition of "effect" in terms of "closest worlds"? It is also "not as easy to interpret" and formally isomorphic to DAGs and FFR.... Yet it is metaphysical, for it is not defined in terms of the way knowledge is represented by researchers.

7.25.18 @2:50pm (Replying to @eliasbareinboim @AndersHuitfeldt and 4 others) I would not insist on "solving", let's focus on "encoding," to see how the quantity "the effect of X on Y" is represented in "FFRCISTG" which is proposed as alternative representation of causal knowledge. We just want to be sure that the answer is not pre-encoded in the FF...

7.25.18 @6:19pm (Replying to @AndersHuitfeldt @analisereal and 5 others) It is not a matter of "user friendly". It is a matter of defining "effect" in terms of what you know (diagram) or defining it in terms of metaphysical quantities, eg counterfactuals, and later show formal isomorphism. In SCM, counterfactuals are defined BY THE Model not converse.

7.25.18 @2pm - Notation is only part of counterfactuals, semantics is no less important. The diagram is more than a visual aid; it DEFINES the do-operator. Without it I would not know what to cut and what to leave uncut. Without it I can't explain why tweaking the barometer will not stop rain.

7.27.18 @10:31pm (Replying to @aynumazi) Glad you liked the answers to your question. The main point is that the do-opeartor cannot be separated from the diagram on which it operates, and from the features of the diagram that legitimize its predictions. (e.g., which factors to wipe out). #Bookofwhy #causalinference

7.24.18 @10:27pm - You might be one of the enlightened few, but I have spoken to literally hundreds of microeconometrics graduates, across the country. Infinitesimally few can solve the toy problems presented in my ET paper and many many resent me for counting -- I stopped.

7.25.18 @2:09am (Replying to @fabinger) - Gee, Michal, I hope I do not sound "complaining". No way! I am describing the sad state of econometric education to entice students and researchers to study causal inference on their own, eg.through #bookofwhy rather than stale textbooks and conservative professors. It is working

7.26.18 @3:27am (Replying to @juli_schuess @fabinger) The Chen and Pearl survey has been updated recently by Chris Auld, . Moreover, some authors of econ. texts are calling us to review their upcoming revised editions. There is hope, but it is very hard to uproot a culture baked in confusion. #Bookofwhy

7.24.18 @10:54pm- (Replying to @rationalexpec) The correct link to the ET paper is . The title is "TRYGVE HAAVELMO AND THE EMERGENCE OF CAUSAL CALCULUS" and the toy problems are in Section 3.2. I have mentioned this section before to #bookofwhy --it is worth a second reading.

7.24.18 @3:46pm - Economics has had all the potentials for becoming the queen of social science. It blew it by fostering an orthodox and insular culture. Do you know that to post a comment on NBER Webserver you need to be an "approved NBER family member"? #bookofwhy is not very eco-charitable.

7.24.18 @3:46pm - And do you know the percentage of economists who can tell (in their own model) which parameter can be identified by OLS? Or which variable would qualify as a (conditional) instrumental variable? Or whether a model has testable implications? Or,...Or.. It could have been the queen

7.24.18 @10:10pm (Replying to @stewarthu) You made my day, Stewart, thanks. And, as our sages said: "One convert weighs more than ten born-righteous" (free translation from the Hebrew). Welcome to the global march on commonsense. #bookofwhy

7.24.18 @6:25pm (Replying to @intensivemargin) I am not so sure. Readers of Morgan and Winship (2014) and a good chunk of political science (eg #bookofwhy p.228-230) can answer these elementary questions. The enlightened camp of econometrics is still dormant/suppressed. And it shows in the literature, in paper after paper.

7.24.18 @11:47pm - (Replying to @PHuenermund @vhranger) Thanks, Paul. But, to escape charges of bias, I would not state a problem in graphical terms. The problem is Eco. 101. You take ANY economic model and ANY parameter BETA in the model, and you ask an innocent question: Can BETA be estimated using OLS with ANY control whatsoever?

7.25.18 @ 6:07am - (Replying to @vhranger @PHuenermund) I am not speaking graphical models. I am speaking econ 101 where you teach simultaneous equation models and you explain why OLS does not always work. A curious student asks: when can we estimate a parameter, say beta, using OLS. This is not a provocative question. Why fight it?

7.24.18 @2:59pm - My opinion on @aaronecarroll article is echoed here: . Trialists and writers need to know that mathematical tools for managing and interpreting "pragmatic trials" are now available, hence they can safely enter the folds of science.

7.24.18 @1:47am (to Paul Hunermund @PHuenermund) Well put, Paul. I often say it in less poetic terms: "The assumptions embedded in the model act like a catalyst to steer the data to support the conclusion that we seek to obtain". This clumsy sentence may be the formal definition of "Extract insight". Well put. #Thebookofwhy

7.23.18 @9:14pm - I can think of (at least) two ways of achieving model-free behavior: (1) rely on data only, and (2) match any new situation to one of your prestored models, then use the latter. Which one do you advocate? Which one is currently practiced in ML? #Bookofwhy #DataScience

7.23.18 @8:19pm - Caution: "garbage in garbage out" is sometimes used against causal inference, saying: "your conclusions are not more informative than your assumptions". Wrong! Our conclusions are quantitative, informed by data, our assumptions are qualitative, reflecting prior knowledge.

7.23.18 @12:42am - Hector Geffner just gave an interesting talk at IJCAI titled "Model-free, Model-based, and General Intelligence" . He likens these two modes of reasoning to Kahneman's System-1 and System-2. #bookofwhy #Epibookclub #DataScience #causalinference

7.23.18 @12:57am (Replying to @corbyjerez @AlexVasilescu) I hope this work:

7.15.18 @11:32am - This is tyranny galore!! Just because some path analysts disavow the causal meaning of their models you were forced to do the same? By AJE? We truly need a Hall of Shame and we truly need to educate those "satisfied referees" that this is the 21st century! #bookofwhy should help.

7.15.18 @11:19am - I agree with your conclusion, but not with your first sentence. Do say "X causes Y", with the understanding that it is an abbreviation for the rest. Science has known many abbreviations before (eg, density=weight/volume) why not this one? It is not fair to ask us to repeat.

7.15.18 4:24am - "Causal inference" is benign, it goes from causal assumptions to causal conclusions, never claiming to "infer causality" from naked data. Making this explicit, and justifying some of the assumptions, should cure editors allergy. Mentioning the "causal revolution" might also help.

7.15.18 - I am surprised that, in the 21st century, submissions to medical journals still require special justification for causal inference. I propose starting a new blog, "Hall of Shame", to post reviews and editorial letters from journals that do require such justifications. #bookofwhy.

7.14.18 @7:08pm - For the curious, the second part of Cinelli's tutorial has been posted on: It demonstrates the logical equivalence of Rubin's potential-outcome framework and the structural causal model used in #bookofwhy. A conclusion in one, is a conclusion in the other

7.15.18 @12:50pm (Replying to @robinberjon) Sorry, this is all I have. My Login was free and easy.

7.14.18 @4:01pm - It is not "forbidden" physically. If we do not observe a person more than once we escape the criticism: "Once treated, it is not the same person " or, "Once operated on, you cannot undo the surgery." #bookofway #epibookclub It makes the results applicable to undoable treatments.

7.14.18 @3:35pm - This paper would not have been published were it not for Phil Dawid's editorial leadership to overrule all the authoritative reviewers and move the field forward. Today,such editors are an endangered specie. And that is why, Kuhn explains, paradigm shifts take forever. #bookofwhy

7.14.18 @11:25am - So what says the Jury? (1) The heterogeneity tests proposed have been considered before and found to be useless, or (2) they havn't been considered and the jury is still out, or (3) they haven't been considered but are judged useless apriori? Where are we? #bookofwhy #epibookclub

7.14.18 @10:28am - No radical steps allowed. Every person is observed only ONCE. The observational study is conducted separately, in the general population, not involving subjects chosen for the experimental study.#Bookofwhy #epibookclub @stephensenn @AlexJohnLondon @f2harrell

7.14.18 @9:46am - This uncertainty in response holds for an isolated RCT, true. But if you also run a separate observational study, then the two studies give us an inequality that, if violated, rules out the second option, namely, the population cannot be homogeneous. #BookOfWhy #EpiBookClub

7.14.18 @9:08am - Using the inequalities in we can only tell that there are substantial differences in the population, that is, some patients benefit much more than others, but we cannot tell who those are. #bookofwhy #epibookclub

7.13.18 @3:13pm - I had to pacify a reviewer's demand to remove the reference to Babylon, possibly to protect the honor of ML. I truly recommend Toulmin's "Foresight and Understanding" (1961). Published papers must compromise to the lowest denominator. Hoping #BookOfWhy will see an uprising.

7.14.18 @4:17pm (Replying to @aesopesque) It is not unusual. After all, reviewing is a boring job unless you see it as part of a mission to defend your ideology, or your profession. That is why the professional literature is so slow to catch up with progress (see econometrics). That is why I wrote #Bookofwhy #epibookclub

7.13.18 - The heterogeneity tests derived in only require one observation per individual. Did I derive the impossible? I doubt it! But I used counterfactual logic and causal models, which may have not been available to more senior analysts. #BookOfWhy #EpiBookClub

7.13.18 - A reader asked if there is a test for heterogeneity, e.g., whether a drug is uniformly beneficial or kills some and saves more. Remarkably, three such tests have been uncovered in . Can someone in #BookOfWhy or #EpiBookClub check if the FDA is aware?'

7.13.18 - Speaking about Reichenbach, the #BookOfWhy (p. 47-51) is highly critical of "probabilistic causality" - a branch of philosophy that emerged from Reichenbach's work, and which was misguided from day one. See Section 33.5 in . Philosophers can make mistakes.

7.13.18 - Readers ask if the #BookOfWhy describes the work of Hans Reichenbach, the first to attempt a connection between causation and probabilities. Briefly. Noting that his failure to account for colliders represents our mind's tendency to assume:"no correlation without causation."

7.11.18 - Perhaps "Let's fake it" is too harsh. Anyone who tells us: "Let's assume ignorability" is really saying: "Let's move to where I feel more comfortable - statistical estimation - and leave causal inference to others. #BookOfWhy

7.11.18 - Why I can't trust "ignorability" assumptions? Because I am yet to find an ordinary mortal who can discern ignorability claims in his/her head, even in toy problems. The mental task is formidable so, anyone who tells us: "Let's assume ignorability" is saying: "Let's fake it". #Boo

7.11.18 - Fashion pushes researchers to say they are combining causality with ML. So, watch for imposters. How? They hide assumptions. Some start: "Let's assume ignorability"-- suspect them. Others will not even say that -- shun them. I trust those who climb the 3-step ladder. #BookOfWhy

7.10.18 - In the spirit of Twitting, I recently completed a paper that summarizes the #BookOfWhy in seven words, each standing for a principle or a tool: Once you acquire these seven tools you would qualify as top Causal Inference Expert - a champion of commonsense.

7.8.18 (1/4) - 10 days ago, when I decided to go on Twitter, I did not know what I am getting into. Today, with over 5,000 followers, I feel a sense of obligation to give you a progress report, and to reflect on this experience. This will probably take 4 Tweets - stay tuned #bookofwhy
(2/4) - First, I find the intellectual exercise of compacting pages of technical analysis into 3-4 sentences challenging, educational and empowering. Second, the genuine quest for understanding I see among you gives me the hope that my primary goal of writing the Book is realizable.
(3/4) - My goal was and is: the democratization of causal inference . By that I mean empowering students and researchers to understand causal inference on their own, without waiting for professors or journal editors to catch up. Toward that end, I will continue to bring up ...
(4/4) - for discussion topics that have been ignored or suppressed by mainstream literature and, simultaneously, try to be as flexible, honest and open-minded as I can in responding to your questions and comments. Unortunately, I can only Tweet once or twice a day. #bookofwhy

7.7.18 - The logical equivalence between the structural vs. potential outcome frameworks is one of those issues that researchers are taking for granted but rarely discuss. For those interested, Carlos Cinneli is giving a tutorial on this topic here: #bookofwhy

7.7.18 - My dream is "No Reader Left Behind". What's the problem? Conceptual or mathematical? #Bookofwhy

7.6.18 - Cartwright"s dictum is so simple and so universal, like the conservation of mass and energy. Yet it is so often ignored, especially by Bayesians. Bayesianism today is a license to spray priors on whatever you can't estimate, then wait for the posteriors to peak, in case they do.Judea Pearl added,

7.6.18 - Kudos to Jonathan Kirsch for a most insightful review of #BookOfwhy, which makes even humble me understand what the Book is all about. Who said you have to be a rocket scientist to understand cause and effect?

7.3.18 - Someone asked: Why a revolution and not a dialogue? T. Kuhn lamented the same, and concluded that, given scientists ego and investment, no dialogue can avert a revolution. I tried it, for the past 25 years - and wrote the book becasues I failed. #BookOfWhy #Epibookclub

7.1.18 - Someone asked (can't find who): Doesn't correlation always imply some form of causation? Answer: Almost! This was Reichenbach's dictum (1956), But he forgot about colliders (See The Book of Why, page 199, and the Monty Hall paradox). #BookOfWhy

7.1.18 - Where do we fit? For a physicist, the equations of physics are sufficient. Not so for a robot who would be easily tempted to tweak the barometer to prevent rain. Structural equations operationalize some of the implicit information a physicist must use before applying the eqs.

7.1.18 - "Only focused on robots?" God forbids! I am focused on science, and thinking "robots" helps us uncover all the implicit information that a scientist must bring to bear on a problem, before applying the explicit information encoded in the equations. It's a good thinking device.

6.28.18 - An interviewer asked me today what I believe will be the most important practical implication of the Causal Revolution. My answer: Serious erosion of the hegemony of RCT's. See @deatonangus

6.30.18 - I am not sure what is meant by: "response schedule/counterfactual etc controversy in statistics". From what I recall, "response schedule" was David Freedman's round-about way of saying "function", and "counterfactuals" were made controvertials by Dawid (2000), but no longer are.

6.27.18 - Hi everybody, the intense discussion over The Book of Why drove me to add my two cents. I will not be able to comment on every tweet, but I will try to squeak where it makes a difference.

5.15.17 - I exist therefore I tweet!