Finding clarity in complexity: an introduction to the concept of network meta-analysis

Network meta-analysis (NMA) is a method for comparing three or more treatments in one analysis by combining direct head-to-head trial evidence with indirect comparisons drawn through a shared comparator. That matters because clinicians and patients often face “which treatment should I choose?” questions, yet the evidence base rarely contains a single trial that compares every option. NMA builds a connected evidence network, allowing relative effectiveness estimates and treatment rankings even when some pairings were never tested directly.

The presentation starts by grounding NMA in the familiar hierarchy of evidence. A systematic review collects studies that meet prespecified eligibility criteria using explicit methods to minimize bias. Pairwise meta-analysis then statistically combines results from studies that compare two interventions (for example, drug A versus drug B). NMA extends this logic: if trials compare A versus B and B versus C, the method can infer an indirect A versus C comparison through B, while also incorporating any available direct A versus C trials. The result is a “mixed” treatment effect estimate that reflects both evidence pathways.

The core requirement for credible indirect comparisons is the transitivity assumption: the different sets of randomized trials must be similar on average in all important factors other than the interventions being compared. The talk illustrates how transitivity can fail when effect modifiers differ across comparisons—for instance, when placebo is administered in different forms (injection versus pill) or when participant age ranges vary substantially between trial sets. A related concept is coherence (also called consistency), which checks whether direct and indirect estimates agree statistically. Coherence can be assessed globally across the whole network or locally within specific loops, helping identify “hot spots” where assumptions may break down.

Because NMA adds complexity, the protocol and analysis plan must be more explicit than in a standard pairwise review. The protocol should justify why NMA is needed, state whether ranking is an objective, define inclusion/exclusion criteria with transitivity in mind, and specify how interventions will be handled (including doses, administration routes, and whether to lump or split network nodes). Search strategy also needs care: relying on existing systematic reviews is usually inadequate for capturing the full network. Data analysis sections should name the statistical framework (Bayesian or frequentist; fixed or random effects) and address multi-arm trial correlation and heterogeneity assumptions. If Bayesian methods are used, prior distributions and Markov chain Monte Carlo convergence details should be reported.

Quality and reporting considerations receive sustained attention. The talk notes an extension to PRISMA for NMA reporting and highlights that many published NMA protocols underreport key elements such as transitivity and model choice. For presenting results, it recommends concise formats—league tables, direct/indirect/network estimates, and clear handling of uncertainty—often moving nonessential figures to online repositories. Treatment ranking is possible using metrics such as SUCRA (surface under the cumulative ranking curve), but ranking should not replace relative effect estimates and must match the clinical question and appropriate ranking metric.

A worked example centers on uterotonics for preventing postpartum hemorrhage, including network diagrams, league-table-style results, coherence checks, and summary-of-findings tables structured by outcome rather than by comparison. The talk also addresses what to do when NMA becomes infeasible: keep the protocol, transparently report methods, and assess risk of bias so the review can be updated later if the evidence base grows. Finally, it points to key resources, including the Cochrane Handbook chapter on NMA, Cochrane training modules (protocol considerations, GRADE, CINeMA, and more), and software options such as WinBUGS, JAGS, NMAstudio, and web tools like meta insight.