Comparison of GEE-based methods in cluster-randomized trial with missing data when outcome depend on other patients covariates

• Activity Number: 537
• Type: Invited
• Date/Time: Wednesday, August 6, 2014 : 2:00 PM to 3:50 PM
• Sponsor: ENAR
• Author(s): Melanie Prague and Rui Wang and Eric Tchetgen and Victor DeGruttola*+
• Companies: Harvard School of Public Health and Brigham & Women's Hospital and Harvard School of Public Health and Harvard School of Public Health
• Keywords: cluster-randomized trials ; permutation tests ; IPW ; variable selection

Abstract

We investigate the use of semiparametric regression for estimation of intervention effects on clustered outcomes motivated by cluster-randomized trials (CRTs) of effectiveness of HIV prevention. Interest lies in inference for the marginal parameter describing these effects in presence of data that are subject to being missing at random (MAR). We focus on complex correlation structure designs with patients nested in sub-clusters, and with outcomes and missingness patterns that may depend on other subjects' covariates. Through simulation studies, we investigate the impact of the choice of the correlation structure in unweighted and inverse probability-weighted (IPW) GEE and demonstrate biases that arise unless a working independence structure is used for the GEE. To make use of knowledge of the correlations, we propose the use of an augmented term that depends on other patients' covariates whose selection may be guided by the correlation structure; this term is intended to improve the efficiency of the estimation of the treatment effect.