Difference-in-Differences in Mediation Analysis: A Solution to Mediator-Outcome Confounding

• In Event: Methodological Challenges in Multilevel Modeling

Abstract

Causal identification of direct and indirect effects requires a stringent, no-unmeasured-mediator-outcome-confounding assumption even when the treatment is randomized. This study develops a difference-in-differences in mediation (DiDiM) approach that extends difference-in-differences methods to mediation. This extension allows causal identification of direct and indirect effects in the presence of unmeasured mediator-outcome confounding. Although this approach requires its own identification assumption, this study also proposes a method to adjust the DiDiM estimators for the bias due to the violation of the key assumption.

Author

• Yongnam Kim, Seoul National University