Reducing Mean Squared Error in the Analysis of Binary Paired Data

Abstract Matched pairs appear frequently in studies of the association between a dichotomous exposure and a dichotomous outcome. The standard conditional logistic regression estimator of the common odds ratio is consistent, but uses only information from the doubly discordant pairs. To gain efficiency, Liang and Zeger (1988 , Biometrics, 44:1145–1156) and Kalish (1990 , Biometrics, 46:493–499) proposed a weighted average estimator, which compromises between the standard estimator and the marginal odds ratio among the outcome discordant pairs. This estimator is useful in matched case–control studies, which by design generate outcome discordant pairs. In this article, we generalize the weighted average estimator, so that it incorporates information from doubly concordant pairs as well, whenever such pairs are available. We provide an R-function which implements the proposed weighted average estimators.