Doubly robust generalized estimating equations with consistent variance estimator when one auxiliary model is misspecified

Abstract

The doubly-robust generalized estimating equation (DR-GEE) is a popular analytic tool for repeated measurements with missing data. It requires two assumptions on auxiliary models for outcome and the observation indicator, and produces a consistent point estimator when either of the model assumptions is correct. Another advantage is an easy variance formula due to asymptotic independence between the estimator of interest and the estimators from the auxiliary models. However, this feature can be capitalized only when both models are correctly specified. In this paper, we propose an alternative DR-GEE that produces a consistent variance estimator even when one of the auxiliary models is misspecified. The main idea is to construct estimating equations for the auxiliary models so that the estimators of main interest and the estimators from auxiliary models are asymptotically independent. We illustrate the method on data from a clinical trial for hypertension.