Statistical models and methods for network meta-analysis

Abstract

Meta-analysis, the methodology for analyzing the results from multiple independent studies, has grown tremendously in popularity over the last four decades. Although most meta-analyses involve a single effect size (summary result, such as a treatment difference) from each study, there are often multiple treatments of interest across the network of studies in the analysis. Multi-treatment (or network) meta-analysis can be used for simultaneously analyzing the results from all the treatments. However, the methodology is considerably more complicated than for the analysis of a single effect size, and there have not been adequate explanations of the approach for agricultural investigations. We review the methods and models for conducting a network meta-analysis based on frequentist statistical principles, and demonstrate the procedures using a published multi-treatment plant pathology data set. A major advantage of network meta-analysis is that correlations of estimated treatment effects are automatically taken into account when an appropriate model is used. Moreover, treatment comparisons may be possible in a network meta-analysis that are not possible in a single study because all treatments of interest may not be included in any given study. We review several models that consider the study effect as either fixed or random, and show how to interpret model-fitting output. We further show how to model the effect of moderator variables (study-level characteristics) on treatment effects, and present one approach to test for the consistency of treatment effects across the network. Online supplemental files give explanations on fitting the network meta-analytical models using SAS.