Comparing Bayesian and Classical Methods in the Analysis of a Cluster Randomized Trial (the Community Hypertension Assessment Trial)

Abstract

Cluster randomized controlled trials are increasingly used to assess the effectiveness of life-style interventions in improvement of health services or prevention of disease. However, statistical methods in the analysis of cluster randomized controlled trials are not well established especially for analyzing binary outcomes. This project is motivated by the Community Hypertension Assessment Trial (CHAT) to assess the effectiveness of a 12-month community-based blood pressure management program in improving the management and monitoring of high blood pressure (BP) among older people. The study is a paired cluster randomized controlled trial, where the family physicians' practices are the clusters randomly allocated to CHAT intervention or usual practice, and a random sample of 55 patients 65 years and older were selected from the 14 practices in each study arm for health record review. The primary outcome was controlled BP over 12 months defined as systolic BP c:; 140 and diastolic BP c:; 90 for patients without diabetes or target organ damage or systolic BP c:; 130 and diastolic BP c:; 80 for patients with diabetes or target organ damage. Secondary outcomes include frequency of BP monitoring and average BP over a 12 month period. The clinical objective of this project is to evaluate the effectiveness of the CHAT intervention. The statistical objective is to compare Bayesian and classical methods of analyzing cluster-randomized trials using CHAT study as an example. We compared the results of different cluster-level analysis methods: i) un-weighted regression, ii) weighted regression, iii) random-effects meta-analytic approach, and different individual-level analyses: i) standard logistic regression, ii) robust standard errors approach, iii) generalized estimating equations, iv) random-effect logistic regression, v) Bayesian random-effect regression. We find that there is no sufficient evidence in support of the effectiveness of the CHAT intervention on all outcomes. For BP control, odds ratio (95% confidence interval) is 1.14 (0.72, 1.80) from generalized estimating equations. This result remains robust under different methods. We also find that the results from different statistical methods are different. The results from cluster-level analysis methods are quite different, while the results from the individual-level analysis methods are similar. We conclude that using various methods to analyze the trial provide good sensitivity analyses to help in interpreting the results of cluster randomized trials. Extensive simulation studies comparing the statistical powers of the different methods in different situations are required.