CRTgeeDR: an R Package for Doubly Robust Generalized Estimating Equations Estimations in Cluster Randomized Trials with Missing Data

Abstract

Semi-parametric approaches based on generalized estimating equation (GEE) are widely used to analyse correlated outcomes. Most available softwares had been developed for longitudinal settings. In this paper, we present a R package CRTgeeDR for estimating parameters in marginal regression in cluster randomized trials (CRTs). Theory for adjusting for missing at random outcomes by inverse-probability weighting methods (IPW) based on the use of a propensity score had been largely studied and implemented. We exhibit that in CRTs most of the available softwares use an implementation of weights that lead to a bias in estimation if a non-independence working correlation structure is chosen. In CRTgeeDR, we solve this problem by using a different implementation while keeping the consistency properties of the IPW. Moreover, in CRTs using an augmented GEE (AUG) allow to improve efficiency by adjusting for treatment-covariate interactions and imbalance in baseline covariates between treatment groups using an outcome model. In CRTgeeDR, we extend the abilities of existing packages such as geepack and geeM to allow such data augmentation. Finally, one may want to combine IPW and AUG in a Doubly Robust (DR) estimator, which lead to consistent estimation when either the propensity score or the outcome model corresponds to the true data generation process (Prague, Wang, Stephens, Tchetgen Tchetgen, and De gruttola 2015). The DR approach is implemented in CRTgeeDR. Simulations studies demonstrate the consistency of IPW implemented in CRTgeeDR and the gains associated with the use of the DR for analyzing a binary outcome using a logit regression. Finally, we reanalyzed data from a sanitation CRT in developing countries (Guiteras, Levinsohn, and Mobarak 2015a) with the DR approach compared to classical GEE and demonstrated a signiffcant intervention effect.