Primary Submission Category: Mediation
Unpacking subgroup differences in treatment effects: A causal decomposition approach for mediated moderation analysis
Authors: Xiao Liu,
Presenting Author: Xiao Liu*
Assessing differences between demographic subgroups (e.g., female and male) in treatment effect—or, moderation analysis—is important in behavioral sciences. In moderation analysis, besides quantifying how much subgroups differ in treatment effect (“total moderation”), it is often useful to examine why the effect differences between subgroups arise—such as by examining intermediate variables contributing to the effect difference—or, mediated moderation analysis.
For causal inference involving intermediates, causal mediation methods are fast-growing but have limited development for mediated moderation analysis; a particular challenge is that the subgroups are often defined by demographic characteristics, which are non-manipulable.
This study extends the causal decomposition approach, and develops methods for mediated moderation analyses with causal interpretation. Our methods decompose the total moderation into the causal estimands capturing how much the subgroup difference in the effect is attributable to the subgroup difference in intermediate variable(s) (“mediated moderation”) and how much is not (“remaining moderation”). We develop multiply-robust estimators (including cross-fitted one-step estimators and targeted minimum loss estimators), which facilitate using machine learning techniques in causal inference. We illustrate the applications in an empirical mediated moderation analysis to unpack gender differences in the effects of an intervention for behavior problems.