A Comparison of Bias-Corrected Covariance Estimators for Generalized Estimating Equations

Abstract

Although asymptotically the sandwich covariance estimator is consistent and robust with respect to the selection of the working correlation matrix, when the sample size is small, its bias may not be negligible. This article compares the small sample corrections for the sandwich covariance estimator as well as the inferential procedures proposed by Mancl and DeRouen (2001), Kauermann and Carroll (2001), Fay and Graubard (2001), and Fan et al. (2012). Simulation studies show that when using a maximum likelihood method to estimate the covariance parameters and using the between-within method for the denominator degrees of freedom when making inference, the Kauermann and Carroll method is preferred in the investigated balanced logistic regression and the Mancl and DeRouen and Fan et al. methods are preferred in the investigated proportional odds model. A collagen-induced arthritis study is employed to demonstrate the application of the methods.