Tests Of The Effect Of Treatment In Stratified Cluster Randomization Trials

It is becoming increasingly common for epidemiologists to consider randomizing intact social units (e.g. families, schools, communities) rather than individuals in experimental trials. Reasons are diverse, but include administrative convenience, a desire to reduce the effect of treatment contamination and the need to avoid ethical issues that might otherwise arise. Dependencies among cluster members typical of such designs must be considered when determining sample size and analysing the resulting data.;The primary focus of this thesis is on comparisons of tests of the effect of treatment in trials where clusters are randomly assigned to treatment groups after stratifying on cluster-level baseline risk factors (e.g. cluster size). Particular attention is paid to the analysis of binary outcome data.;Tests of the effect of treatment for such trials range in complexity from adaptations of standard statistical methods performed using the cluster as the unit of analysis to extensions of logistic regression adjusted for clustering. The validity of such extensions was shown to be assured if the average correlation among cluster members is fixed. This assumption can be relaxed by using robust variance estimators. Test statistics using these different approaches were shown to be asymptotically equivalent when there is no variability in cluster size.;Simulation studies were used to examine the small sample properties of test statistics assuming an average cluster size of 100 subjects and either two or four strata. These simulation studies indicated that exact permutation tests should be used to make inferences about the effect of treatment if there are 20 or fewer clusters per treatment group. Approximate test statistics using cluster-level analyses or extensions of the Mantel-Haenszel test statistic are appropriate if there are more than 20 clusters per treatment group. Valid rejection rates for methods using robust variance estimates can not be assured even if there are 40 clusters per treatment group. There is little need to employ such techniques, however, since the simulation studies also showed that typical violations of the common correlation assumption have no effect on the validity or power of test statistics