Comparison of Methods for Analyzing Cluster Randomized Trials with Binary Outcomes When Cluster Size Varies and Intraclass Correlation Coefficients Are Unequal

• Activity Number: 124
• Type: Contributed
• Date/Time: Monday, August 4, 2014 : 8:30 AM to 10:20 AM
• Sponsor: ENAR
• Author(s): Sheng Wu*+ and Catherine Crespi and Weng Kee Wong
• Companies: University of California, Los Angeles and University of California, Los Angeles and University of California, Los Angeles
• Keywords: cluster randomized trials ; intraclass correlation coefficient ; power ; varying cluster size ; unequal cluster size

Abstract

A variety of statistical methods have been proposed for analyzing a two-arm cluster randomization trial (CRT) with a binary outcome. Little is known about how statistical power to detect the proportion difference between the two arms is affected when the two arms have unequal cluster numbers and cluster sizes and the true intraclass correlation coefficients (ICCs) in the arms are different. We conducted a simulation study to address these questions by comparing common analysis methods of CRTs including the cluster-level t test, adjusted chi-square test, mixed effects logistic regression (MELR) and generalized estimating equations (GEE). Simulation results suggest that MELR and GEE generally have higher power than adjusted chi-square and cluster-level t tests, and increasingly so as variation in cluster size increases. However, MELR and GEE also have inflated Type I error rates. A general conclusion is that when the true ICCs in the two arms are different but we assume they are equal, there is generally only a slight loss in power compared to working with arm-specific ICCs. More details on our recommendations will be given in the presentation.