Difference-in-differences inference with few treated clusters

Inference using difference-in-differences with clustered data requires care. Previous research has shown that t tests based on a cluster-robust variance estimator (CRVE) severely over-reject when there are few treated clusters, that different variants of the wild cluster bootstrap can over-reject or under-reject severely, and that procedures based on randomization show promise. We demonstrate that randomization inference (RI) procedures based on estimated coefficients, such as the one proposed by Conley and Taber (2011), fail whenever the treated clusters are atypical. We propose an RI procedure based on t statistics which fails only when the treated clusters are atypical and few in number. We also propose a bootstrap-based alternative to randomization inference, which mitigates the discrete nature of RI P values when the number of clusters is small.