Assessing key assumptions of network meta-analysis: a review of methods
Version of Record online: 1 AUG 2013
Copyright © 2013 John Wiley & Sons, Ltd.
Research Synthesis Methods
Volume 4, Issue 4, pages 291–323, December 2013
How to Cite
Donegan, S., Williamson, P., D'Alessandro, U. and Tudur Smith, C. (2013), Assessing key assumptions of network meta-analysis: a review of methods. Res. Synth. Method, 4: 291–323. doi: 10.1002/jrsm.1085
- Issue online: 17 DEC 2013
- Version of Record online: 1 AUG 2013
- Manuscript Accepted: 14 JUN 2013
- Manuscript Revised: 11 JUN 2013
- Manuscript Received: 28 JUN 2012
- network meta-analysis;
- multiple treatments meta-analysis;
- mixed treatment comparison;
- consistency, transitivity
Homogeneity and consistency assumptions underlie network meta-analysis (NMA). Methods exist to assess the assumptions but they are rarely and poorly applied. We review and illustrate methods to assess homogeneity and consistency.
Eligible articles focussed on indirect comparison or NMA methodology. Articles were sought by hand-searching and scanning references (March 2013). Assumption assessment methods described in the articles were reviewed, and applied to compare anti-malarial drugs.
116 articles were included. Methods to assess homogeneity were: comparing characteristics across trials; comparing trial-specific treatment effects; using hypothesis tests or statistical measures; applying fixed-effect and random-effects pair-wise meta-analysis; and investigating treatment effect-modifiers. Methods to assess consistency were: comparing characteristics; investigating treatment effect-modifiers; comparing outcome measurements in the referent group; node-splitting; inconsistency modelling; hypothesis tests; back transformation; multidimensional scaling; a two-stage approach; and a graph-theoretical method.
For the malaria example, heterogeneity existed for some comparisons that was unexplained by investigating treatment effect-modifiers. Inconsistency was detected using node-splitting and inconsistency modelling. It was unclear whether the covariates explained the inconsistency.
Presently, we advocate applying existing assessment methods collectively to gain the best understanding possible regarding whether assumptions are reasonable. In our example, consistency was questionable; therefore the NMA results may be unreliable. Copyright © 2013 John Wiley & Sons, Ltd.