Performance of Generalized Estimating Equations Using Bias-corrected Sandwich Variance Estimators with Count Outcomes

The sandwich estimator is used for estimating the asymptotic variance of regression coefficient in generalized estimating equations (GEE), it has been widely-known that proposed by [1]. This estimator is biased downwards and underestimated for variances in small sample settings. Various alternative bias-corrected sandwich variance estimators have been proposed by [2], [3], [4], [5] and [6]. Evaluation of the GEE using of the bias-corrected sandwich variance estimators with count outcomes data is limited. In this paper, the performance of the six sandwich variance estimators is compared in the GEE under various scenarios with correlated count outcomes. Two working correlation structures (exchangeable: EX and autoregressive of order 1: AR-1), the true correlation parameters of AR-1 and EX (alpha = 0.3, 0.5, 0.7), sample sizes (n = 20, 40, 60), and cluster sizes (m = 2, 4, 6) are considered. Under small sample size, the six sandwich variance estimators have more variability and are not robust in the covariance of regression coefficients if the working correlation structure is misspecified. Results reveal that the GST estimator outperforms other sandwich estimators.