Electronic Thesis and Dissertation Repository


Doctor of Philosophy


Epidemiology and Biostatistics


Dr. Allan Donner and Dr. Neil Klar


Cluster randomization trials have become increasingly popular when theoretical, ethical or practical considerations preclude the use of traditional trials that randomize individual subjects. Although some methods for analyzing clustered ordinal data have been brought to wide attention, these are less developed as compared to methods for analyzing clustered continuous or binary outcome data. The aim of this thesis is to refine existing strategies which may be applicable to clustered ordinal data as well as extensions which have been previously considered only for clustered binary responses. The approaches include adjusted Cochran-Armitage tests using an ICC estimator, and correction and modification strategies to improve the small-sample performance of the Wald test and score test in GEE for clustered ordinal data. The type I error and power for these test statistics are investigated using a simulation study.

Simulation results show that kappa-type estimators had less bias than ICC estimators when cluster sizes were fixed and small for ρ = 0.005 or ρ = 0.01. Conversely, ANOVA ICCs had relatively smaller bias in the case of variable cluster sizes. In addition, small-sample performance of GEE robust Wald tests are improved by using adjustments and corrections. The adjusted test WBC1 is recommended in terms of type I error and power. The discussion is illustrated using data from a school-based cluster randomization trial.