I have seen an earlier version of this book recommended here on HN, and thought that some might be delighted to know that there is a revised version available to download.
Thank you, I was not aware of this book. I am currently working my way through the Counterfactuals and Causal Inference (2nd), and I would definitely recommend this book (arguably over JP's Causality) to those who are looking for a comprehensive, introductory text.
Nah, Robins started (asterisk) all this causal/propensity-scored/pretend-experimental mania.
It's what drove me back to hard interventional experiments and eventually to adaptive clinical trial design.
(Asterisk): ok more like Pearl started it and Robins made it more practical with doubly robust designs etc. I still find most of the mechanics rather shady.
Look up confounding by indication and some of the power/sensitivity studies for so-called doubly robust estimators. Counterfactuals are reasonable. A lot of "let's turn this observational study into a designed experiment with math" approaches turn out not to be. That's the gist of it; if you want rigor, you should read the papers. I don't have a reference to hand at the moment (on phone) but it shouldn't take more than a few minutes of searching google scholar to hit the appropriate vein. The bottom line is simply TANSTAAFL.
From the abstract: “Millions of times each year, judges must decide where defendants will await trial—at home or in jail. By law, this decision hinges on the judge’s prediction of what the defendant would do if released. … Yet comparing the algorithm to the judge proves complicated. … We only observe crime outcomes for released defendants, not for those judges detained. This makes it hard to evaluate counterfactual decision rules based on algorithmic predictions. … We deal with these problems using different econometric strategies, such as quasi-random assignment of cases to judges. … A policy simulation shows crime can be reduced by up to 24.8% with no change in jailing rates, or jail populations can be reduced by 42.0% with no increase in crime rates. Moreover, we see reductions in all categories of crime, including violent ones. Importantly, such gains can be had while also significantly reducing the percentage of African-Americans and Hispanics in jail. … While machine learning can be valuable, realizing this value requires integrating these tools into an economic framework: being clear about the link between predictions and decisions; specifying the scope of payoff functions; and constructing unbiased decision counterfactuals.”
they seem to be fixated on how shiny and new L1 penalties are, in 2014. Greg Ridgway started using gradient boosting machines (GBM) for propensity scoring in the early 2000s, and I didn't see them cite him, so I kind of hate them already. On the other hand, at least GBM works well.
I'm no economist, though. Perhaps this is novel at NBER. It's just odd to see someone acting like using an ensemble to enable data-driven model selection is something new.
nb. I didn't read the entire 76-page paper (partly because it's obscenely verbose). A quick skim and here are my from-the-hip remarks. If they suck, I'll refund every cent you paid me ;-)
Looks like Chapter 15 in the Causal Inference Book agrees with you:
“Outcome regression and various versions of propensity score analyses are the most commonly used parametric methods for causal inference. You may rightly wonder why it took us so long to include a chapter that discusses these methods. So far we have described IP weighting, the g-formula, and g-estimation–the g-methods. Presenting the most commonly used methods after the least commonly used ones seems an odd choice on our part. Why didn’t we start with the simpler and widely used methods based on outcome regression and propensity scores? Because these methods do not work in general. More precisely, the simpler outcome regression and propensity score methods–as described in a zillion publications that this chapter cannot possibly summarize–work fine in simpler settings, but these methods are not designed to handle the complexities associated with causal inference for time-varying treatments.”
Interesting. It was my impression that Donald Rubin was the main instigator of propensity-score methods. Do you have some paper references in mind to give me a better sense of the timeline?
Quite possible -- Robins was pointed out to me as one of the pioneers by Sander Greenland, around 2008 or so? I'm more of an experimental/clinical statistician at this point. The work I did with propensity scoring and doubly robust estimation convinced me that shrinkage was a less-bad approach for the things I needed to do. YMMV.
Haven't read the book, but the concept of causal inference from event data really deserves more attention. Automatic/assisted cause analysis in complex systems has huge potential.
Do you have any particular example or domain in mind?
Regarding the automatic analysis, Automatic Statistician (Tenenbaum & Ghahramani et al.), which is supposed to automate / assist the exploratory analysis, comes to mind. Not sure if it'd be fair but I tend to lump some of the probabilistic programming - such as BayesDB - platforms into the same bucket. While these are exploratory tools, they are ideally (pragmatically) meant to help identify associations and context whereby correlation is highly suggestive of causation.
Another area of interest: one of the comments below mentioned adaptive clinical trial design. I know that Bayesian trials are intensively implemented in some cancer research centers such as MD Anderson (off-topic: curious to what extent IBM Watson was used as part of this protocol). I would love to know whether adaptive trials are increasingly popular or widely used in pharma industry.
I was thinking about debugging & failure modes, in distributed systems. There's potential to automatically test hypotheses by starting systems from known states.
There are very interesting and fairly recent results on causal discovery under additive noise models [1]. Although such models aren't universally applicable the underlying concept is intuitive and a good fit for many problem domains. On the time series front, google open sourced CausalImpact [2] for Bayesian structural time-series modeling a few years ago. Looks like RankScience [3] is putting that research to good use.
I'm surprised that causal discovery doesn't get more play. Aside from the direct applications to scientific research, causality is a strong consistency hint for assessing models/features learned from data.
The theory you mention about additive noise, is clever but doesn't strike me as robust at all. It's been awhile since I looked at that literature, but when I did it seemed really unrealistic, that at some level it reduced to assuming that all deviations from normality are interpretable in terms of the desired causal inferences. There might be some scenarios where you could reduce the variables involved to the point where that idea is feasible, but otherwise it seemed really unbelievable to me.
Part of me wants to dive into this causality modeling, because it seems up my alley, but I'm very sceptical of it showing anything definitive. I do observational research, but short of a priori randomization, I'm sceptical of any claims to causality. Even then, with experiments, I'm deeply sceptical unless something has been replicated across various secondary conditions by multiple distinct groups.
Modern causality theory and modeling has definitely raised the bar in terms of what we say about data, and I love it, but sometimes I wonder if causality is a red herring. Even with hard experimental evidence, I'm tempted to not interpret it beyond "when someone does X, this tends to happen."
It does seem too good to be true, but they've compiled a real-world dataset of causal relationships, and the additive-noise and fancier ML algorithms do seem to infer the right direction well above chance, so there's at least something there.
I read through about 10 pages of the first book ("without models"). It struck me as very nearly identical to current statistics practice. It clearly differentiated itself as discussing counterfactuals (the data needed to actually determine causality) but I could not find the section of the book that described how counterfactual data can be inferred from missing data (without it being "turtles all the way down").
Does anyone in this area have a succinct way to explain how counterfactual data can be inferred by these techniques - and how traditional statistics practice is not able to perform this inference?
I wish abbreviations like inverse probability weighting and marginal structural models were expanded in the second book. It's annoying to have to look up "IP weighting" only to discover "oh, it's IPW, god damn you [authors] to hell".
MSMs are interesting. Now I remember why Robins' name stuck in my head. The book shows all the math explicitly, which is nice, and it delves into causal inference for time-to-event data, which is also nice.
Now I'm curious whether they look at piecewise constant survival models for time-varying coefficients. It's mentioned, but I didn't read enough to see if it's treated in detail. If it is, everyone who does A/B testing should read the book, because this is one of those "little details" from biostatistics that becomes super important at big retailers (like, say, Amazon, where the principal economist at the time pointed it out to me).
Answered my own question -- Part III of this book is what I am waiting for. I was going to send the authors a note, which seems silly, but my name is in an awful lot of standard textbooks just because I sent in corrections or notes, and for some reason I find that satisfying. Like I made some sort of a difference to some students somewhere.
30 comments
[ 3.1 ms ] story [ 94.2 ms ] threadIt's what drove me back to hard interventional experiments and eventually to adaptive clinical trial design.
(Asterisk): ok more like Pearl started it and Robins made it more practical with doubly robust designs etc. I still find most of the mechanics rather shady.
Interesting, can you expand on this? I have no experience with causal inference and would like to learn more. Thanks!
I'm wondering if their methodology is reasonable?
From the abstract: “Millions of times each year, judges must decide where defendants will await trial—at home or in jail. By law, this decision hinges on the judge’s prediction of what the defendant would do if released. … Yet comparing the algorithm to the judge proves complicated. … We only observe crime outcomes for released defendants, not for those judges detained. This makes it hard to evaluate counterfactual decision rules based on algorithmic predictions. … We deal with these problems using different econometric strategies, such as quasi-random assignment of cases to judges. … A policy simulation shows crime can be reduced by up to 24.8% with no change in jailing rates, or jail populations can be reduced by 42.0% with no increase in crime rates. Moreover, we see reductions in all categories of crime, including violent ones. Importantly, such gains can be had while also significantly reducing the percentage of African-Americans and Hispanics in jail. … While machine learning can be valuable, realizing this value requires integrating these tools into an economic framework: being clear about the link between predictions and decisions; specifying the scope of payoff functions; and constructing unbiased decision counterfactuals.”
I'm no economist, though. Perhaps this is novel at NBER. It's just odd to see someone acting like using an ensemble to enable data-driven model selection is something new.
nb. I didn't read the entire 76-page paper (partly because it's obscenely verbose). A quick skim and here are my from-the-hip remarks. If they suck, I'll refund every cent you paid me ;-)
“Outcome regression and various versions of propensity score analyses are the most commonly used parametric methods for causal inference. You may rightly wonder why it took us so long to include a chapter that discusses these methods. So far we have described IP weighting, the g-formula, and g-estimation–the g-methods. Presenting the most commonly used methods after the least commonly used ones seems an odd choice on our part. Why didn’t we start with the simpler and widely used methods based on outcome regression and propensity scores? Because these methods do not work in general. More precisely, the simpler outcome regression and propensity score methods–as described in a zillion publications that this chapter cannot possibly summarize–work fine in simpler settings, but these methods are not designed to handle the complexities associated with causal inference for time-varying treatments.”
Regarding the automatic analysis, Automatic Statistician (Tenenbaum & Ghahramani et al.), which is supposed to automate / assist the exploratory analysis, comes to mind. Not sure if it'd be fair but I tend to lump some of the probabilistic programming - such as BayesDB - platforms into the same bucket. While these are exploratory tools, they are ideally (pragmatically) meant to help identify associations and context whereby correlation is highly suggestive of causation.
Another area of interest: one of the comments below mentioned adaptive clinical trial design. I know that Bayesian trials are intensively implemented in some cancer research centers such as MD Anderson (off-topic: curious to what extent IBM Watson was used as part of this protocol). I would love to know whether adaptive trials are increasingly popular or widely used in pharma industry.
I'm surprised that causal discovery doesn't get more play. Aside from the direct applications to scientific research, causality is a strong consistency hint for assessing models/features learned from data.
[1] https://news.ycombinator.com/item?id=8776582
[2] https://opensource.googleblog.com/2014/09/causalimpact-new-o...
[3] https://news.ycombinator.com/item?id=13552862
Part of me wants to dive into this causality modeling, because it seems up my alley, but I'm very sceptical of it showing anything definitive. I do observational research, but short of a priori randomization, I'm sceptical of any claims to causality. Even then, with experiments, I'm deeply sceptical unless something has been replicated across various secondary conditions by multiple distinct groups.
Modern causality theory and modeling has definitely raised the bar in terms of what we say about data, and I love it, but sometimes I wonder if causality is a red herring. Even with hard experimental evidence, I'm tempted to not interpret it beyond "when someone does X, this tends to happen."
Does anyone in this area have a succinct way to explain how counterfactual data can be inferred by these techniques - and how traditional statistics practice is not able to perform this inference?
I wish abbreviations like inverse probability weighting and marginal structural models were expanded in the second book. It's annoying to have to look up "IP weighting" only to discover "oh, it's IPW, god damn you [authors] to hell".
MSMs are interesting. Now I remember why Robins' name stuck in my head. The book shows all the math explicitly, which is nice, and it delves into causal inference for time-to-event data, which is also nice.
Now I'm curious whether they look at piecewise constant survival models for time-varying coefficients. It's mentioned, but I didn't read enough to see if it's treated in detail. If it is, everyone who does A/B testing should read the book, because this is one of those "little details" from biostatistics that becomes super important at big retailers (like, say, Amazon, where the principal economist at the time pointed it out to me).