When can we trust cluster-robust inference?

Although cluster–robust standard errors are widely used, they can sometimes yield very unreliable inferences. Tests and confidence intervals based on the usual (CV1) standard errors are known to work poorly in certain circumstances, such as when there are few clusters, few treated clusters, or clusters that vary greatly in size or other features. Numerous methods have been proposed to obtain more reliable inferences. These include alternative standard errors, such as ones based on the cluster jackknife (CV3), nonstandard critical values, and bootstrap methods. I discuss what we have learned from the recent literature and attempt to provide some guidance for how to deal with cases where alternative methods yield conflicting results. The talk focuses on linear regression models, but logit models will also be discussed.