Researchers often ask how an intervening variable contributes to observed group-based outcome disparities. Examples include the roles of education in generating gender wage gaps, racial health disparities, and intergenerational income mobility. We introduce a new nonparametric decomposition approach to put this enterprise on a firm causal footing. We make three contributions. First, we introduce a previously overlooked disparity-generating mechanism; second, we show how to isolate three distinct mechanisms under unstructured effect heterogeneity; third, we develop multiply robust and semi parametrically efficient double/debiased machine learning estimators with desirable properties. Empirically we show how college graduation plays multiple and nearly countervailing causal roles in intergenerational income mobility, including via our newly discovered generative mechanism. Joint work with Ang Yu.