I wrote two posts in the past couple of months about when we can be confident that social science results apply in specific cases (or not). Specifically, the posts were about when we can trust average causal (treatment) effects, given that the average effect might mask any number of hidden interaction variables or, put another way, differences in treatment effect across group members.
I’m pleased with how those two posts came out, but they both assumed that average treatment effects were all we had. I focused on practical questions you might ask to decide whether to trust the social science, not on how the social science itself could answer the question.
I saw a good presentation recently on machine learning and heterogeneous treatment effects, so I thought I should post something mentioning that social scientists do often try to go beyond just main effects to consider how the effect might differ for different parts of the population. A bit more on heterogeneous treatment effects is here, and a good paper on how machine learning can help estimate them is here.