Performance of Network Meta-Analysis in Large Networks: A Simulation Study

• Activity Number: 656
• Type: Contributed
• Date/Time: Thursday, August 7, 2014 : 10:30 AM to 12:20 PM
• Sponsor: Section on Statistics in Epidemiology
• Author(s): Binod Neupane*+ and Joseph Beyene
• Companies: and McMaster University
• Keywords: Network meta-analysis ; multiple treatments comparison ; meta-analysis ; type 1 error ; power ; validity

Abstract

Background: Network meta-analysis (NMA) has become increasingly popular in recent years. Transitivity and consistency are required for validity of NMA. However, when a large number of treatments are compared, differential patients characteristic, treatment doses and formulations, administration routes, duration of follow up, biases, and chance, etc. can lead to the violation of these basic assumptions. Methods: We carried out a simulation study under consistency assumption (same direct and indirect effects) with similar trials (no covariate modifying treatment effects) at different levels of heterogeneity for large networks. Data were simulated under the null hypothesis of no treatment effect and under the alternative to allow estimation of type I error and power, respectively. We assessed performances of Bayesian MNA with direct pairwise meta-analysis in terms of bias, mean square error (MSE), coverage probability, type I error rate, and power. Results: Both NMA and direct approaches are unbiased, but Bayesian NMA outperformed pairwise MA with smaller MSE and in general better type I error rates and power. Conclusions: Network meta-analysis in general performs better than direct