An ‘all-in-one’ meta-analysis model: joint synthesis of multiple outcomes to compare multiple interventions

Article type
Authors
Efthimiou O1, Dimitris M1, Andrea C2, Georgia S1
1Department of Hygiene and Epidemiology, University of Ioannina School of Medicine, Ioannina, Greece
2Department of Public Health and Community Medicine, Section of Psychiatry, University of Verona, Verona, Italy
Abstract
Background: The joint meta-analysis of two or more correlated outcomes, known as multivariate meta-analysis, is expected to improve precision compared to a series of independent meta-analyses and can, under circumstances, address selective outcome reporting bias. To date multivariate meta-analysis has been applied only to compare pairs of interventions. Further, its usefulness has been limited by the fact that within study-correlations are rarely reported in studies.

Objectives: To develop a model to synthesize multiple dichotomous outcomes over a network of studies that compare competing interventions and to introduce a simple procedure to elicit expert opinion for the within-study correlation.

Methods: We propose a flexible model for the meta-analysis of multiple outcomes and multiple treatments. By making plausible assumptions about the heterogeneity, we simplify the model and decrease the number of between-studies correlation parameters that need to be estimated. When within-study correlations are not available, we show how these can be estimated from a set of appropriate probability parameters for which expert opinion can be easily obtained. We fit the model within a Bayesian framework that offers maximum flexibility and allows incorporation of expert opinion.

Results: We apply our model to a network of 13 treatments for acute mania and we consider 2 correlated outcomes: efficacy and acceptability. Our multiple-outcomes network meta-analysis model produces considerably narrower confidence intervals compared to the simple network meta-analysis. The gain in precision is significant for the efficacy outcome because many studies report only on acceptability and our model allows for borrowing strength across outcomes.

Conclusions: The suggested model constitutes a viable candidate for performing network meta-analysis for multiple outcomes within a Bayesian framework and improves precision in the estimates and ranking. The suggested prior elicitation method is straightforward and easily understood by clinicians.