A0666
Title: On "sandwich" variance estimation: Bayesian versus frequentist
Authors: Cy Sin - National Tsing Hua University (Taiwan) [presenting]
Abstract: It is well known the Eicker-Huber-White variances are not only heteroskedasticity-robust and nonlinearity-robust but also nonnormality-robust. In a recent paper, some of the Eicker-Huber-White variances are reviewed. Among other things, they conclude: (a) Simulation studies suggest HC(4), a variant of robust variance estimator proposed by another study, does not over-reject or mildly under-rejects even in cases of non-normal distributions; (b) The original robust variance (denoted by HC(0) and its variants considered by the prevalent statistical software (such as R and STATA), are all asymptotically equivalent. The purpose is to take a Bayesian approach and consider the balanced loss function (BLF) proposed by a recent study. Unlike the conventional inference loss function (ILF), this function strikes a balance between estimation error and lack of fit. This function is, in turn, generalized upon what Zellner first proposed, which confines attention to normality-type likelihoods. The Bayesian estimator of the variance-covariance matrix is asymptotically equivalent to the frequentist estimator. Non-normal likelihoods are covered. Simulation studies that compare the Bayesian estimator with the conventional estimators are performed.
