Methodological and Statistical Issues in the Design and Analysis of Stratified Cluster Randomized Trials

Abstract

Background and Objectives While the number of adopting stratified cluster randomized trials (CRTs) is increasing, we have limited knowledge about the methodological and statistical issues pertaining to this design. Our objectives were to (i) survey the literature to assess the methodological and statistical issues and quality of reporting of stratified CRTs; (ii) examine the sensitivity of methods for analyzing data from stratified CRTs; (iii) evaluate the performance of methods for analyzing continuous data from stratified CRTs. Methods We conducted a systematic survey and identified the stratified CRTs from the database MEDLINE. Data were abstracted on several methodological and statistical issues including sample size, randomization, and method of analysis. Two empirical studies were conducted to examine the robustness of methods for analyzing continuous and count data from stratified CRTs. Furthermore, a simulation study was performed to evaluate the performance of methods for analyzing continuous data from stratified CRTs under different scenarios including number of clusters, and cluster sizes.Results and Conclusions There was significant deficiency in reporting and analysis of data from stratified CRTs. The majority of the studies did not adjust the primary method for both clustering and stratification to assess the intervention effect. The results from the empirical studies indicated that the methods for analyzing continuous and count data yielded similar conclusions. However, these methods varied in terms of magnitude of the effect sizes and widths of the 95% confidence intervals (CIs). Moreover, these studies demonstrated that, widths of the 95% CIs were narrower, and pvalues were lower when adjusted for stratification compared to without adjusted for stratification. The results from the simulation study showed that, performance of all methods improved as the number of clusters and cluster sizes increases. However, the performance of these methods deteriorated as the value of intra-cluster correlation coefficient (ICC) increases. Generalized estimating equations (GEE) and meta-regression yielded type I error rate of approximately 10% for small number of clusters. Meta-regression was the least powerful and efficient method compared to GEE, mixed-effects, and cluster-level linear regression methods. The contributions of this thesis will guide the researchers to make informed decision about assessing the intervention effect and reporting of stratified CRTs.