Uncovering the Mechanisms of Policy Interventions: A Difference‐In‐Differences Framework for Mediation Analysis

Although mediation analysis is widely used in human resource management to explain how practices shape outcomes, most cross‐sectional approaches rely on the strong assumption of sequential ignorability. We develop a two‐period difference‐in‐differences framework that accommodates time‐invariant unobserved confounders of the mediator and outcome. Rather than requiring independence in levels, identification rests on restrictions on change, including parallel untreated trends and change‐based mediator–outcome exchangeability. The framework clarifies the assumptions needed to identify total and mediator effects as distinct from those required for controlled direct and natural indirect effects. With observed heterogeneity, the natural indirect effect equals the treated‐population average product of the mediator's effect on the outcome and the treatment's effect on the mediator; the full natural indirect effect adds an unidentified residual covariance term, assessed through sensitivity analysis. We extend the method to multi‐period staggered designs and provide estimation, bootstrap inference, diagnostics, and an open‐source R package.