A simple regression model for network meta-analysis

Abstract

Introduction: The aim of this paper is to propose a transparent, alternative approach for network meta-analysis based on a regression model that allows inclusion of studies with three or more treatment arms.
Methodology: Based on the contingency tables describing the frequency distribution of the outcome in the different intervention arms, a data set is constructed. A logistic regression is used to determine the parameters describing the difference in effect between a specific intervention and the reference intervention and to check the assumptions needed to model the effect parameters. The method is demonstrated by re-analysing 24 studies investigating the effect of smoking cessation interventions. The results of the analysis were similar to two other published approaches to network analysis using the same data set. The presence of heterogeneity, including inconsistency, was examined.
Conclusion: The proposed method provides an easy and transparent way to estimate treatment effect parameters in metaanalyses involving studies with more than two arms. It has several additional attractive features such as not overweighting small studies as the random effect models do, dealing with zero count cells, checking of assumptions about the distribution of model parameters and investigation of heterogeneity across trials and between direct and indirect evidence.