Abstract

Difference-in-Differences (DiD) is a cornerstone of causal inference in observational studies, substituting the untestable "no-unmeasured-confounding" assumption with the parallel trends assumption. While DiD is widely used for total effect estimation, its application to causal mediation analysis—decomposing effects into natural direct and indirect pathways—remains methodologically constrained. Existing DiD mediation methods often rely on restrictive, unverifiable conditions such as monotonicity or counterfactual parallel trends across all strata, which may not hold in complex empirical settings. We propose a novel theoretical framework to identify the natural direct effect on the treated (NDET) and the natural indirect effect on the treated (NIET) within a DiD design. Our identification strategy introduces the "Mediator Common Trend" (MCT) assumption. Unlike traditional mediation assumptions, MCT is partially testable; researchers can empirically assess its validity by examining pre-treatment trends within mediator-stratified subsamples. We provide formal identification formulas for these effects and further extend the framework to accommodate multiple causally ordered mediators, allowing for a more granular decomposition of the average treatment effect on the treated (ATT) through sequential causal pathways. By targeting the ATT and utilizing design-based diagnostics, this framework aligns mediation analysis with the rigorous standards of modern DiD literature, relaxing the reliance on global exchangeability and providing a transparent basis for mechanism investigation. This study bridges a critical gap in the causal inference literature by providing a robust and testable framework for mediation in DiD designs, offering researchers a systematic tool to explore causal mechanisms without relying on overly restrictive assumptions.