A Comparison of Methods of Estimation for Binary Outcomes in Paired Cluster Randomized Trials

Abstract

The randomized controlled trial is often considered the “gold standard” for evaluating the effectiveness of treatments. However, when individuals are naturally aggregated into groups (e.g., families, clinics, schools) it can be challenging to allocate individuals within the same group to different treatments. To fix this, Cluster Randomized Trials (CRTs) have widely been applied in various fields. Additionally, by matching clusters in CRTs, between-cluster variation can be reduced, which increases the precision of estimation of the treatment effect. The objective of this thesis is to compare different methods in estimating the treatment effect for paired CRTs via a series of simulations, including Restricted Maximum Likelihood (REML), Maximum Likelihood (ML) and two-stage analysis. The simulation results indicate that REML and ML provide unbiased estimation overall. However, REML demonstrates better performance than ML in estimating standard error and in controlling the type I error. Additionally, the two-stage analysis is more robust in maintaining the type I error rate, suggesting that it may be a good choice for controlling the type I error for large cluster size.