Table 1.

Summary of multivariate and network meta-analysis approaches

Question	Multivariate meta-analysis of multiple outcomes	Network meta-analysis of multiple treatment comparisons
What is the context?	Primary research studies report different outcomes, and thus a separate meta-analysis for each outcome will utilise different studies	Randomised trials evaluate different sets of treatments, and thus a separate (pairwise) meta-analysis for each treatment comparison (contrast) will utilise different studies
What is the rationale for the method?	• To allow all outcomes and studies to be jointly synthesised in a single meta-analysis model      • To account for the correlation among outcomes to gain more information	• To enable all treatments and studies to be jointly synthesised in a single meta-analysis model      • To allow indirect evidence (eg, about treatment A v B from trials of treatment A v C and B v C) to be incorporated
What are the benefits of the method?	• Accounting for correlation enables the meta-analysis result of each outcome to utilise the data for all outcomes      • This usually leads to more precise conclusions (narrower confidence intervals)      • It may reduce the impact of selective outcome reporting	• It provides a coherent meta-analysis framework for summarising and comparing (ranking) the effects of all treatments simultaneously      • The incorporation of indirect evidence often leads to substantially more precise summary results (narrower confidence intervals) for each treatment comparison
When should the method be considered?	• When multiple correlated outcomes are of interest, with large correlation among them (eg, > 0.5 or < −0.5) and a high percentage of trials with missing outcomes; or      • When a formal comparison of the effects on different outcomes is needed	• When a formal comparison of the effects of multiple treatments is required      • When recommendations are needed about the best (or few best) treatments
What are the potential pitfalls of the method?	• Obtaining and estimating within study and between study correlations is often difficult      • The information gained by utilising correlation is often small and may not change clinical conclusions      • The method assumes outcomes are missing at random, which may not hold when there is selective outcome reporting      • Simplifying assumptions may be needed to deal with a large number of unknown variance parameters	• Indirect evidence arises through a consistency assumption—ie, the relative effects of ≥3 treatments observed directly in some trials are (on average) the same in other trials where they are unobserved. This assumption should be checked but there is usually low power to detect inconsistency      • Ranking treatments can be misleading owing to imprecise summary results—eg, a treatment ranked first may also have a high probability of being ranked last      • Simplifying assumptions may be needed to deal with a large number of unknown variance parameters