When meta-analysis of clinical trials indicates substantial heterogeneity between studies, summarization by a pooled treatment effect may not be meaningful. In the first phase of this grant, we used Bayesian hierarchical regression models to explore sources of this heterogeneity. Speculating that the size of heterogeneous treatment effects might be related to the control rate, the underlying baseline risk, we developed algorithms for regressing treatment effects on control rates that properly adjusted for measurement error in the control rate and correlation between the treatment effect and control rate. Analyzing 115 meta-analyses with binary outcomes, we found significant correlation with the odds ratio in 15 percent of cases. To further explore the hypothesis that clinical trial heterogeneity derives from differences in treatment efficacy across subpopulations defined by the trials, we seek in this continuation to broaden our study of control rate regression by achieving the following aims: 1) Extend the application of control rate regression to continuous and survival time outcomes, to nonlinear and non-normal models and to meta-analysis of trials with few events; 2) Determine if control rate regression can improve estimation of pooled treatment effects; 3) Determine how control rate differences across trials correlate with differences in reported risk factors; 4) Search for new ways to report summaries of risk factors that might correlate better with differences in outcomes; 5) Study the generalizability of control rate regression in meta-analyses derived from different types of data sources including different medical specialties published in print and electronic formats; and 6) Develop tools to apply control rate regression in the prospective design of clinical trials. Heterogeneity among clinical trial results provides an opportunity to optimize treatment benefit by revealing sources of variation. Control rate regression can be an effective tool for exploring this heterogeneity.
