Keywords: network meta-analysis, multivariate random effects model, large sample approximation, Bartlett correction, bootstrap adjustment
Network meta-analysis enables comprehensive synthesis of evidence concerning multiple treatments and their simultaneous comparisons based on both direct and indirect evidence. Although ordinary likelihood-based methods or Bayesian analyses with non-informative priors have been adopted for the inference in multivariate random effects models in network meta-analysis, validities of these methods are founded on large sample theory. As widely known in conventional pairwise meta-analyses, coverage probabilities of confidence intervals of these methods can be substantially below the target level (e.g., Brockwell and Gordon, Statist Med 2007, 26: 4531-43), and it generally follows in multivariate models. One of effective approaches to resolve these inferences is adopting improved methods based on higher order asymptotic theory. In this study, I develop the Bartlett correction-based confidence interval for multivariate random effects models in network meta-analysis, and evaluate its practical effectiveness via simulation studies. In addition, applications to a real data example is provided.