Joint Meta-Regression Methods Accounting for Postrandomization Variables Principal Investigator: Haitao Chu, M.D., Ph.D. Summary The rapid growth of interest in comparative effectiveness research and evidence-based medicine has led to dramatically increased attention to systematic reviews and meta-analyses, which synthesize and contrast multi- ple randomized clinical trials. T o examine the impact of covariates on study-specific treatment effects, meta- regression methods are available for conventional meta-analysis comparing two treatments and for network meta-analysis simultaneously comparing multiple treatments . While there is broad consensus on methods for examining study-level covariates  which are similar across a study's treatment arms because of randomization  it is much more challenging to adjust for postrandomization variables, which are expected to differ between treatment arms within a study. Examples include differential noncompliance, measured as the proportion of premature treatment discontinuation or drop out, loss to follow-up, or change to an alternative therapy. To the best of our knowledge, existing meta-regression methods only focus on the impact of study-level covariates, which are assumed to be fixed, while postrandomization variables are generally considered random. Thus, ex- isting meta-regression methods cannot account for postrandomization variables. Because postrandomization variables such as differential noncompliance can induce bias in estimating the effect of treatment plans, in responding to PA-16-161 this proposal's overall goal is to develop cutting-edge joint models to account for postrandomization variables in meta-analysis, and to integrate them into publicly available, easy-to-use software to enhance the reproducibility, validity, and generalizability of meta-analyses. Specifically, we will apply Bayesian hierarchical models in these three specific aims: 1) develop joint meta-regression meth- ods to adjust for postrandomization variables in conventional meta-analysis; 2) develop multivariate joint meta- regression methods to adjust for postrandomization variables in network meta-analysis; and 3) objectively eval- uate the proposed methods and develop an open-source R package. We will evaluate the strengths and weaknesses of these methods compared to existing meta-analysis meth- ods, through real data applications and extensive simulations. The proposed statistical methods will be broadly applicable to many meta-analyses. Completing these aims will substantially advance comparative effectiveness research and evidence-based medicine through innovative meta-analysis methods. It will improve public health by facilitating treatment selection for various cancers and for cardiovascular, infectious, and other diseases.

PROJECT NARRATIVE This project's overall goal is to develop innovative joint meta-regression methods adjusting for postrandomization variables and free, open-source, easy-to-use software implementing them. As postrandomi- zation variables (such as differential levels of compliance with treatment) generally differ in a study's different treatment arms, existing meta-regression methods cannot account for postrandomization variables. The proposed statistical methods and software will be broadly applicable to many meta-analyses and will improve comparative effectiveness research and public health by facilitating treatment selection for various cancers and for cardiovascular, infectious, and other diseases.