Determining Valid Covariate Adjustment Sets for Socially Ascribed Exposures in Causal Mediation

Determine which covariate sets should be controlled for to estimate causal effects in causal mediation analyses when the exposure is a socially ascribed characteristic such as race or gender, so that the exposure ignorability assumptions required for identification are satisfied.

Background

In causal mediation analyses based on natural direct and indirect effects, identification typically requires exposure ignorability, i.e., no omitted confounding of the exposure–outcome and exposure–mediator relationships. When the exposure is a socially ascribed attribute like race or gender, selecting an appropriate set of control variables is especially fraught, raising unresolved questions about which variables should be adjusted for to justify causal interpretation.

The paper contrasts this difficulty with its proposed causal decomposition framework, which avoids requiring exposure ignorability by focusing on observed disparities rather than causal effects of social groups. Nonetheless, the open question of how to specify valid adjustment sets for socially ascribed exposures remains important for mediation analyses that aim to identify natural effects.