Confidence Interval Estimation for Continuous Outcomes in Cluster Randomization Trials

Cluster randomization trials are experiments where intact social units (e.g. hospitals, schools, communities, and families) are randomized to the arms of the trial rather than individuals. The popularity of this design among health researchers is partially due to reduced contamination of treatment effects and convenience. However, the advantages of cluster randomization trials come with a price. Due to the dependence of individuals within a cluster, cluster randomization trials suffer reduced statistical efficiency and often require a complex analysis of study outcomes. The primary purpose of this thesis is to propose new confidence intervals for effect measures commonly of interest for continuous outcomes arising from cluster randomization trials. Specifically, we construct new confidence intervals for the difference between two normal means, the difference between two lognormal means, and the exceedance probability. The proposed confidence intervals, which use the method of variance estimates recovery (MOVER), do not make certain assumptions that existing procedures make on the data. For instance, symmetry is not forced when the sampling distribution of the parameter estimate is skewed and the assumption of homoscedasticity is not made. Furthermore, the MOVER results in simple confidence interval procedures rather than complex simulation-based methods which currently exist. Simulation studies are used to investigate the small sample properties of the MOVER as compared with existing procedures. Unbalanced cluster sizes are simulated, with an average range of 50 to 200 individuals per cluster and 6 to 24 clusters per arm. The effects of various degrees of dependence between individuals within the same cluster are also investigated. When comparing the empirical coverage, tail errors, and median widths of confidence interval procedures, the MOVER has coverage close to the nominal, relatively balanced tail errors, and narrow widths as compared to existing procedure for the majority of the parameter combinations investigated. Existing data from cluster randomization trials are then used to illustrate each of the methods