Title:Non-Robustness of the Cluster-Robust Inference: with a Proposal of a New Robust Method

Abstract:The conventional cluster-robust (CR) standard errors may not be robust. They are vulnerable to data that contain a small number of large clusters. When a researcher uses the 51 states in the U.S. as clusters, the largest cluster (California) consists of about 10% of the total sample. Such a case in fact violates the assumptions under which the widely used CR methods are guaranteed to work. We formally show that the conventional CR methods fail if the distribution of cluster sizes follows a power law with exponent less than two. Besides the example of 51 state clusters, some examples are drawn from a list of recent original research articles published in a top journal. In light of these negative results about the existing CR methods, we propose a weighted CR (WCR) method as a simple fix. Simulation studies support our arguments that the WCR method is robust while the conventional CR methods are not.