Recent literature has investigated the effects of more realistic data scenarios on the estimation of treatment effects via difference-in-differences (DiD). In particular, researchers now know that when treatment is staggered over time and the true treatment effect is dynamic and/or heterogeneous, then regressions of the form $$y_{it} = \alpha_i + \lambda_t + \boldsymbol{\gamma} D_{it} + \beta X_{it} + \epsilon_{it},$$ result in a (potentially largely) biased estimate of the average treatment effect on the treated (ATT). This bias can be so large that it results in an estimate of the opposite sign of the true effect. The bias is a result of DiD estimates incorporating "forbidden comparisons" when treatment is staggered, in which already treated observations are used as a "control" for estimating treatment effects.

We summarize the literature that led to this conclusion, and we discuss how these new results affect antitrust practitioners. We discuss the implementation of new estimators from the literature to address bias, along with providing a large list of sensitivities for applied economists to incorporate in their research going forward. We perform a Monte Carlo analysis to investigate some of the properties of staggered DiD models, where we find some interesting results about unbalanced panels providing an additional source of bias for estimation.

The papers that we summarize include Goodman-Bacon (2021), Sun and Abraham (2021), Wooldridge (2021), de Chaisemartin and d'Haultfoeuille (2023), Roth et al. (2023), Borusyak et al. (2024), and Gardner (2022). We also mention many other new papers throughout the footnotes of our article.