More generally, I don't like to take any observational studies based on cohort data like this to be useful for attributing any sort of causes. Whenever you've

More generally, I don't like to take any observational studies based on cohort data like this to be useful for attributing any sort of causes. Whenever you've got people self-selecting into treatment populations, you need to be extremely careful about causality. For example, let's say that you've got a positive association between delinquent behavior and divorce, and that it seems to be stronger when the custodial parent stays single than it is when the custodial parent remarries. One story is that an unstable family situation causes delinquent behavior. Another equally plausible one is that difficult-to-rear kids stress their parents out, which causes them to divorce, and makes it harder for them to remarry. It is possible to at least poke at these kinds of tangles, but you've got to be extremely diligent and careful about it. More careful than I think is possible in the "giant scattershot regression model" genre of observational study. When I asked, "Is that true?", I meant it as a rhetorical question. I doubt anyone actually knows the answer. There's an idea, which I got from Green & Gerber's textbook on field experimentation, that some questions are just fundamentally unanswerable.