Partial Homogeneity in Staggered Difference-in-Differences

Abstract

In staggered difference-in-differences (DiD) designs, units enter treatment at different calendar times, so the treatment effect is not a single number but a set of Cohort-Average Treatment effects on the Treated (CATTs), one per cohort-time cell. Estimating every CATT as its own parameter, as the standard fully flexible estimator does, is unbiased but inefficient when some of these effects are in fact equal, whereas pooling them all into a single two-way fixed effects (TWFE) coefficient is efficient but, whenever the heterogeneity is genuine, biased for the individual effects. We frame the choice between these extremes as a partition-selection problem on the cohort-time cells and address it with a Dirichlet Process (DP) mixture prior on the CATTs. The model favors parsimonious groupings without fixing their number, and a collapsed Gibbs sampler delivers point estimates, credible intervals that marginalize the unknown partition, and co-clustering probabilities for every pair of CATTs. With the error variance held fixed and a pairwise penalty placed on the partition, a maximum a posteriori (MAP) partition reduces to an $\ell_0$-penalized regression, connecting the Bayesian formulation to the homogeneity-pursuit literature. In a calibrated simulation, the model cuts the sampling variance of the cohort-time effects by 26--52\% relative to the fully flexible estimator, without the pooled estimator's bias, provided the distinct effects are separated enough to be recovered, and the posterior delivers near-nominal confidence-interval coverage by averaging over the unknown partition. In two applications the method recovers a precision-improving partial-homogeneity structure in one, where the cohort-time effects are genuinely heterogeneous, and reports that full pooling is adequate in the other, where they are not.