Inference About Clustering and Parametric Assumptions in Variance Covariance Estimation

• Activity Number: 346
• Type: Contributed
• Date/Time: Tuesday, August 3, 2010 : 10:30 AM to 12:20 PM
• Sponsor: Business and Economic Statistics Section
• Author(s): Mikko Packalen*+ and Tony Wirjanto
• Companies: University of Waterloo and University of Waterloo
• Keywords: variance covariance matrix estimation ; clustering ; cluster-robust ; heteroskedasticity-robust

Abstract

An important step in applied work is to determine the appropriate level of robustness in estimating the variance covariance matrix. From less robust to more robust, the available choices include: Eicker/White heteroskedasticity-robust estimation, Newey and West heteroskedasticity-and-autocorrelation-robust estimation, and cluster-robust estimation. The motivation for testing the appropriate level of robustness in variance covariance estimation is that a less robust estimation method can have better size and power properties. However, both analytical and bootstrapped versions of the existing tests are known to have poor finite-sample properties, although to our knowledge the reason for the failure has not been understood. We first show why these tests perform poorly. We then propose an alternative strategy for testing the appropriate level of robustness, and show that it performs well.