A natural experiment is an experiment (an AB-test if you will) that occurs by chance rather than conscious design. For example, two neighboring countries contem

A natural experiment is an experiment (an AB-test if you will) that occurs by chance rather than conscious design. For example, two neighboring countries contemplate banning smoking in restaurants, but in one the bill fails with 49% of the vote, in the other the ban goes through with 51% of the vote. It's not perfect, but you could argue that these countries can now be used to estimate the effect of a smoking ban on mortality and health in a way that is almost as good (but not quite as good) as a randomized clinical trial, whereas you can't just compare two arbitrary countries with differential rates of smoking, because they might be different on so many other counts as well and there is no pre-intervention data to serve as a baseline. More broadly, ML does not really answer questions like "was this death caused by smoking, or rather by a hundred other things associated with smoking like lower income and bad health insurance?", though it is excellent at predicting who is likely to die prematurely. So it's great for prediction, but not so great if you want to learn more about the underlying structure of a phenomenon. Statisticians are sometimes surprised to see so much interest in machine learning given that its view of the world is not open to inspection (though there's https://github.com/interpretml/interpret https://github.com/interpretml/interpret I guess) so we as humans learn nothing, but it turns out that in many cases we really don't care all that much about the underlying mechanism, as long as we can make accurate predictions.