Background: Recent advances in meta-analysis have seen increased application of multivariate methods to evidence synthesis involving multiple outcome measures or multiple treatment effects. These methods are particularly appealing in evaluating the effectiveness of healthcare interventions because many studies and systematic reviews of individual studies typically focus on broader health effects and therefore usually report on multiple treatment and outcome measures. Analyses of such data should take into account the correlation structure between effect estimates from the different outcomes.

Objective: To (i) extend the standard network meta-analysis (NMA) model for simultaneous comparison of multiple intervention effects from the univariate to a multivariate outcome setting and in so doing (ii) enable intervention effects to be extrapolated across outcomes.

Methods: The standard NMA model is first described and then extended to multivariate outcome settings. These random effects multiple outcome NMA models allow appropriate modelling of the correlation structure between outcomes through inclusion of the within and between-study correlations. Then using a strategy first proposed by DuMouchel and Harris for combining evidence from human and animal studies, the multiple outcome models are adapted to enable extrapolation of intervention effects across the related outcome measures. Analyses are conducted using Markov Chain Monte Carlo (MCMC) techniques implemented within the WinBUGS software.

Results: The models are applied to binary outcome data investigating the effectiveness of seven home safety education interventions on four fire injury prevention measures in households with children.

Conclusion: In the absence of individual trial evidence on all outcomes, extrapolation of evidence across related outcomes can enable estimates of intervention effects for all outcome measures to be obtained, including intervention effects on outcomes where no trial evidence is available.