Multivariate Network Meta-Analysis to Mitigate the Impact of Outcome Reporting Bias

Outcome reporting bias (ORB) occurs in a large percentage of medical studies, and it is a particular challenge in systematic reviews because it has the potential to undermine the validity of both pairwise and univariate network meta-analysis (UNMA). To effectively handle missing outcomes, multivariate meta-analysis has been increasingly used in the last decade as it can provide less biased estimates by borrowing information from all available outcomes. However, the utility and robustness of multivariate meta-analytic tools in a network setting, multivariate network meta-analysis (MNMA), has yet to be examined. To determine the ability of MNMA to reduce the impact of ORB on pooled effect sizes, we applied the Bayesian MNMA model to “true” multivariate outcomes (\u3e2 outcomes) and conducted an extensive simulation study, using 3 missingness scenarios. Simulation results demonstrated that MNMA substantially reduced the bias of effect sizes in missing at random scenario, and in missing not at random scenario to a lesser extent. Further, these results showed that MNMA improved the precision of estimates, producing narrower credible intervals. Thus, we demonstrated the applicability of the approach via application of MNMA to a multi-treatment systematic review of randomized controlled trials (RCTs) of antidepressants for the treatment of depression in older adults. Using the MNMA method, we also derived a framework for planning future RCTs and examined whether missing multivariate outcomes, likely due to ORB, impact power and sample size estimation in a prospective manner. Via thorough simulations, we showed that power estimates obtained using MNMA were different (higher or lower) from those obtained using UNMA, likely because the network is under pressure of ORB. Despite the advantages of MNMA methods, they are rarely used by medical researchers in systematic reviews due to the lack of available software packages and lack of understanding of the model. Development of the package for MNMA method would thus make MNMA methods more accessible to these researchers and facilitate their use. In the last part, we introduced the mnma package and showed how to run MNMA and UNMA using this package