Aiding Decision-Making and Trial Design using Multivariate Network Meta-Analysis

University Of Texas Hlth Sci Ctr Houston, Houston TX

Investigators

Linked publications & trials

Abstract

Systematic reviews of treatments for mental health disorders should be exploited in order to obtain accurate information about efficacy of current interventions, and to use existing data to plan future clinical trials. Most systematic reviews result a graphical networks of multivariate, multi-arm data, often with up to 50% missing outcomes. Missing clinical trial outcomes are frequently a result of outcome reporting bias (ORB), in which outcomes are unreported based on observed level of significance. Such bias causes pooled meta-analytic effect sizes to be biased. To obtain unbiased and precise network meta-analytic effect sizes, networks should be jointly analyzed using a multivariate network meta-analytic (MNMA) framework, which has not yet been proposed. Under a Bayesian paradigm powered by Markov chain Monte Carlo tools, the methods described in this proposal will exploit outcome correlation and mitigate effects of ORB via the development of the MNMA model, resulting in less biased and more precise pairwise estimates of treatment effects (even for treatments that have been weakly or never-compared). Based on these results, predictive distributions will be used to inform operating characteristics of new clinical trials. Goals: Multivariate NMA will be developed and apply it to 3 case studies: systematic reviews of randomized controlled trials of second-generation anti-depressants for the treatment of adult, adolescent, and older adult major depressive disorder, respectively, for which outcomes have been already shown to be subject to reporting bias. Comparisons with univariate NMA methods will be made. A methodology for future trial design will be developed utilizing Bayesian predictive inference informed by the multivariate network. This approach would refine power and sample size calculations resulting in optimally-powered and more efficient trials for weakly- or never-tested treatments. Software will be completely generalizable to networks arising from all clinical disciplines and will be disseminated freely.