Weights to address non-parallel trends in panel difference-in-differences models

Abstract

Causal inference using the difference-in-differences (DD) method relies on the untestable assumption of parallel counterfactual trends across units that are assigned to different treatments. To facilitate the application of the method in settings where the parallel trends assumption is seemingly violated, I suggest weighting observations such that the conditional correlations between treatments and pre-treatment outcome trends are minimized, that is, weighted trends are parallel. I evaluate the performance of a weighted parallel trends (WPT) DD estimator in a Monte Carlo study and provide an application to a case study context in which a benchmark estimate exists. The WPT DD approach can be applied in settings with multiple continuous treatment variables as well as to estimating time-varying treatment effects.