Reworking Wild Bootstrap Based Inference For Clustered Errors

Many empirical projects involve estimation with clustered data. While estimation is straightforward, reliable inference can be challenging. Past research has suggested a number of bootstrap procedures when there are few clusters. I demonstrate, using Monte Carlo experiments, that these bootstrap procedures perform poorly with fewer than eleven clusters. With few clusters, the wild cluster bootstrap results in p-values that are not point identified. I suggest two alternative wild bootstrap procedures. Monte Carlo simulations provide evidence that a 6-point bootstrap weight distribution improves the reliability of inference. A brief empirical example concerning education tax credits highlights the implications of these findings.