Network Meta-Analysis: Methods and Reporting

Pairwise meta-analysis compares two interventions. Network meta-analysis compares several at once by combining direct evidence from head-to-head trials with indirect evidence through common comparators, so that treatments never tested against each other can still be ranked. It is the method behind most comparative effectiveness questions, and also the one where reviewers look hardest at the assumptions.

The central assumption is transitivity: the trials in the network must be similar enough in the factors that modify the effect that an indirect comparison is meaningful. It is a clinical and methodological judgement, made by comparing the distribution of effect modifiers across comparisons. Its statistical counterpart is consistency, the agreement between direct and indirect estimates where both exist, examined with node-splitting or design-by-treatment interaction models.

Models can be fitted in a Bayesian framework, for example with gemtc or multinma in R, or in a frequentist one with netmeta. Bayesian models make prior choices explicit and produce probabilistic rankings; frequentist models avoid priors and are often quicker to communicate. Ranking metrics such as SUCRA and P-scores summarise the results but are easy to over-read, because a high rank built on imprecise estimates means little. Confidence in each comparison is rated with GRADE for NMA or CINeMA, and reporting follows the PRISMA extension for network meta-analyses.

Reviewers of network meta-analyses ask a recurring set of questions. Is transitivity argued or merely asserted? Was inconsistency tested, and what was done where it appeared? Are rankings presented with their uncertainty, and is a treatment ranked first on sparse evidence flagged as such? Were prior distributions justified and their influence checked in a sensitivity analysis? Answering these in the manuscript rather than in the response letter shortens peer review considerably.

The notes in this topic work through building and drawing a network, assessing transitivity, testing for inconsistency, choosing and checking a model, presenting league tables and rankings without overstating them, and rating certainty comparison by comparison.