Randomized Trials with Clusters: Design and Analysis
In clustered data, each cluster consists of multiple subjects, called subunits, and observations are collected from the subjects of each cluster. Since the subjects in each cluster share some common frailties, their outcome data are usually positively correlated. In a randomized study, we can consider one of two randomization strategies. The first strategy, called cluster randomization, is to randomize clusters among treatment arms. In this case, all subjects within each cluster receive the same treatment, so that dependency of the outcome data exists within each treatment arm only and the outcome data among different arms are independent. The second strategy, called subunit randomization, is to randomize the subjects within each cluster between different treatment arms. In this case, subjects within each cluster receive different treatments as randomized, so that dependency of the outcome data exists both within each treatment arm and across different treatment arms. In comparing the mean outcome level among different arms, we have to appropriately account for the dependency in both the sample size calculation and the subsequent statistical analysis. In this chapter, we investigate some estimators of binary response rate and test statistics for comparing the response rates between two arms using these estimators. We consider both cluster and subunit randomization strategies and we assume that the number of subunits is possibly random among clusters.
