Correlation Selection for Cluster Randomized Trials

• Activity Number: 447
• Type: Topic Contributed
• Date/Time: Wednesday, August 6, 2014 : 8:30 AM to 10:20 AM
• Sponsor: Biometrics Section
• Author(s): Philip Westgate*+
• Companies: University of Kentucky
• Keywords: Cluster randomized trial ; Correlation Selection ; Empirical Covariance ; Generalized Estimating Equations ; Power ; Test Size

Abstract

Cluster randomized trials (CRTs), also known as Group Randomized Trials (GRTs), randomize clusters, or groups, of people to intervention or control arms. Of primary interest is to test if the intervention has an impact on an outcome. In order to obtain the most efficient estimate of the intervention effect, therefore yielding the greatest statistical power, accurate modeling of the correlation structure is required. The correlation induced is typically exchangeable and measured using the intra-cluster correlation (ICC). We discuss two ways in which estimating any given working correlation structure can impact the variance of the estimated treatment effect. Furthermore, we propose a novel method that accounts for both of these impacts in order to select the correlation structure that is estimated to produce the least variable treatment effect estimate. We demonstrate the performance of the proposed method in a simulation study and application example.