Title: Network meta-analysis for adverse events: A discrete multivariate Bayesian approach with Gaussian copulas
Authors: Rebecca Graziani - Bocconi University (Italy) [presenting]
Sergio Venturini - Bocconi University (Italy)
Abstract: A Bayesian multivariate network meta-analysis (NMA) model of multiple discrete correlated outcomes is proposed. An NMA makes it possible to combine all the direct evidence with all the indirect evidence coming from the studies included in the analysis. While the literature on univariate NMA is now extensive, few methods have been published for synthesizing evidence from studies reporting on multiple discrete outcomes for networks of competing treatments. We propose a new Bayesian copula-based method for multivariate NMA of multiple discrete correlated outcomes. The observed outcome in each study is assumed to be a realization of a multivariate discrete random variable whose elements are marginally distributed according to a binomial distribution. The dependence among the univariate outcomes is induced through a Gaussian copula. The probability to observe any of the individual outcome in each study is modeled as a logistic regression with study-specific baseline effects and arm-specific treatment effects. Estimation proceeds by Markov chain Monte Carlo methods using a mixed Gibbs and adaptive random walk Metropolis-Hastings update for the parameters. The correlation matrix of the Gaussian copula is instead updated through a two-stage parameter expanded Metropolis-Hastings algorithm. We compare the performance of our method with those of other published methods within a simulation study. We apply our proposal to a real data set of adverse events.