Alex Vasilescu points us to this new paper, “Towards Causal Representation Learning,” by Bernhard Schölkopf, Francesco Locatello, Stefan Bauer, Nan Rosemary Ke, Nal Kalchbrenner Anirudh Goyal, and Yoshua Bengio.
I’m glad to see more people working on these problems, as they are important no matter how the theory is labeled. I’ve written on occasion about how to use statistical models to do causal generalization (what is called “horizontal, strong, or out-of-distribution generalization” in that paper). My general approach is to use hierarchical modeling; see for example the discussions here and here.
Or here for an applied example in pharmacometrics.
In citing my own work, I’m not trying to devalue the work of others. There are lots of different ways to express the same idea—in this case, partial pooling when generalizing inference from one setting to another, within a causal inference framework—and it’s good that people are attacking this problem using a variety of tools and notations.
I didn’t find R.A. Fisher in the references and D. Rubin is mentioned once. It looks like computer science has a different evolutionary branch on this topic.
Probably more overlaps than it appears.
I brought some causality stuff up in a 2018 talk by Bengio and sent this follow up email afterwards “The comment I made about “randomizing assignment (to do something) being the only practical way yet known to avoid confounding” may have been unclear or even off target. You did seem to be talking about randomization to get a representative population rather than equal (in distribution) groups to compare.” and attached Andrew’s list of three challenges in statistics – generalizing from sample to population, generalizing from control to treatment group and generalizing from observed measurements to the underlying constructs of interest.
My guess is that others sent similar material over the years.