Recently the Nobel Prize in Economics was announced, and two of its three winners are pioneers in the discipline of causal inference.
Guido Imbens and Joshua Angrist are two econometricians who developed the “potential outcomes” framework for causality. Let’s say you’re studying the effect of studying on test scores. As a reader of this blog, you probably know that you can’t simply regress test scores on the amount of studying each student did, as there may be confounding factors. For example, students from a higher socioeconomic background may have had better prior education, and therefore would have a higher baseline of test scores, and they may also have more time to study.
The potential outcomes framework imagines that there are two versions of reality for every unit of analysis. (In this example, the unit of analysis is the student, but in other problems it may be the school, or the business, or the classroom, and so on.) One version of reality corresponds to receiving the treatment (studying more) and another version to the baseline (studying the normal amount). If you could somehow observe both versions of reality, there would be no problem with confounding factors: you’d simply subtract the baseline score from the score where the student studied more, and you’d have the effect that studying more had on each student. The fundamental problem of causal inference, though, is that you can only ever observe (at most) one of the potential outcomes for every unit of study.
The potential outcomes framework casts the whole discipline of causal inference into a missing data problem; the missing data here are all of the potential outcomes that we did not observe. It’s an elegant approach that fits nicely with our intuitive understanding of causality.
As someone whose career is essentially to take methods from academia and to apply them in industry, I’m of course excited to see these two share the Nobel Prize. (Although I was equally excited when Al Roth won, because he was kind enough to answer some hare-brained email I sent him long ago.)
But, to be honest, my first reaction was disappointment. I would have liked to see Judea Pearl share the award along with Imbens and Angrist. Although he is not an economist, he has pioneered a parallel path of causal inference, one with a wholly separate foundation, notation, and mathematical techniques: that of structural causal models (SCMs), or the do-calculus.
SCMs and the potential outcomes framework do not contradict each other (at least as far as I know). They are different lenses on the same problems. Different disciplines have landed on different frameworks for causality; Economics for the most part is dominated by the potential outcomes framework, and I understand that Epidemiology uses SCMs predominantly. Given that the award is for Economics, this may be why he was left off the award (which would have been reasonable); or maybe the committee wanted to recognize the contribution of the potential outcomes framework independent of the larger world of causal inference.
In my opinion, though, the SCM framework is more elegant, and my prediction is that it will have more staying power. Whereas the potential outcomes framework begins with a simple conceit, its application spins out into a confusing morass of various ad hoc estimators for specific conditions. The SCM framework allows you to first write down your model of reality (qualitatively), and then, from there, to build a model of causality.
In previous blog posts here I’ve used the SCM framework, which helped us choose which variables to control for in a regression, and to illustrate why qualitative research is critical for effective quantitative research. Finally, I argued that causal inference was one of the three “revolutions” in data science, and I used the SCM framework to illustrate why. I think I would have had a hard time writing the first two posts using the potential outcomes framework. So, while the Nobel Committee and I agree that causal inference is a paradigm-shifting area of research, we disagree on which parts of it are destined to last for the next century.