Choice of Working Correlation Structure for a GEE-Based Analysis of Incomplete Longitudinal Binary Data

• Activity Number: 423
• Type: Contributed
• Date/Time: Wednesday, August 9, 2006 : 10:30 AM to 12:20 PM
• Sponsor: Biopharmaceutical Section
• Author(s): Priya Kulkarni*+ and Devan V. Mehrotra and Xiaoming Li
• Companies: Merck Research Laboratories and Merck Research Laboratories and Novartis Pharmaceuticals Corporation
• Keywords: longitudinal binary data ; GEE ; missing data ; optimal correlation structure ; compound symmetry ; dropouts

Abstract

Consider a randomized clinical trial in which subjects receive treatment A or B. For each subject, a binary response of interest (responder/non-responder) is measured at fixed time points. The objectives are to estimate the true proportions of responders at the scheduled end of the trial, p_A and p_B, and to test the hypothesis: p_A = p_B. The generalized estimating equations (GEE) approach can be used to address these objectives. The working correlation structure needed for the GEE-based analysis is either specified in advance (CS, AR(1), UN, etc.), or is chosen after looking at the data based on a pre-specified criterion. We will use simulations to compare the Type I error rate and power properties of these two competing strategies in the setting of incomplete longitudinal binary data, for both missing completely at random (MCAR) and missing at random (MAR) missing data mechanisms.