Direct comparison
Network vs. Pairwise Meta-Analysis
Network meta-analysis compares interventions never tested head-to-head; pairwise can't. What each answers, the transitivity assumption, and when to use which.
Written and maintained by CASRAI Editorial Board
Last updated
Ask CASRAI · included with Regulatory Radar
Ask about Network vs. Pairwise Meta-Analysis
Ask CASRAI answers research-administration questions and cites the passages behind every claim — and says so when the corpus does not cover something, instead of guessing. It comes with a Regulatory Radar subscription at $29 a month, alongside the daily digest of regulatory changes and the dashboard of what changed.
150 questions a day, on this site, over the API, or inside your own tools through the CASRAI MCP server.
Everything CASRAI publishes — this page, the dictionary, the guides and the news — stays free to read, with no account and no card.
How do Network Meta-Analysis, Pairwise Meta-Analysis compare side by side?
The table below compares Network Meta-Analysis, Pairwise Meta-Analysis across 13 procurement-relevant dimensions, from what it combines through relationship to a simple indirect comparison.
Side-by-side comparison
| Dimension | Network Meta-Analysis | Pairwise Meta-Analysis |
|---|---|---|
| What it combines | Direct evidence (trials that actually compared a given pair of interventions) plus indirect evidence, chained through one or more common comparators, in a single statistical model (Cochrane Handbook Ch. 11). | Only direct evidence: studies that compared the exact same two interventions head-to-head. No chaining, no common-comparator inference (Cochrane Handbook Ch. 10). |
| Minimum number of interventions | Three or more, connected into a network via at least one shared comparator (a "hub" trial arm, often placebo or usual care). | Exactly two. A third intervention, even if evidence exists for it, is simply outside scope. |
| What it can answer that the other can't | The relative effect -- and a ranking -- of any two interventions in the network, including pairs that were never directly trialed against each other at all. | Nothing beyond what it's built for: it cannot say anything about a comparison no included trial actually made, direct or indirect. |
| Core assumption it depends on | Transitivity: the different sets of trials feeding the network must be similar, on average, in every important effect-modifying factor other than the intervention compared ("joint randomizability" -- as if all interventions could have been randomized in one shared trial). | No transitivity assumption needed -- there's no indirect chain to validate. Still needs the ordinary pairwise assumptions (comparable populations/settings across the *same*-pair trials being pooled, and a defensible choice between fixed-effect and random-effects models). |
| How the assumption can fail | If an effect modifier (e.g. disease severity) is distributed unevenly across comparisons -- say every A-vs-B trial enrolled moderate cases while every A-vs-C trial enrolled severe cases -- the indirect B-vs-C estimate is built on a false-equivalence and is not trustworthy, even though every individual trial was well-conducted. | The equivalent risk is ordinary clinical/methodological heterogeneity across the pooled A-vs-B trials themselves (population, dose, follow-up) -- a narrower, single-comparison version of the same underlying problem, addressed with I²/heterogeneity diagnostics rather than a network-level check. |
| How it's checked statistically | Consistency (the Handbook's term "coherence") is the observable, testable counterpart of transitivity: direct and indirect estimates of the same comparison are checked for agreement, typically via node-splitting or the design-by-treatment interaction test. Disagreement is evidence transitivity may not hold. | There is no direct-vs-indirect agreement check to run, since there's no indirect estimate. Heterogeneity across the pooled studies (I², τ², prediction intervals) is the closest equivalent diagnostic. |
| Output | Relative effect estimates for every pair in the network, plus a ranking/hierarchy of all interventions (commonly summarized with SUCRA or P-score). | A single pooled effect estimate (with confidence or credible interval) for the one comparison being made -- no ranking, because there's nothing to rank against. |
| Reporting standard | PRISMA-NMA (Hutton et al., 2015) -- adds network-specific items on top of base PRISMA: a network diagram, a description of network geometry, and an explicit inconsistency assessment. | Base PRISMA 2020, with no network-specific extension required. |
| Software | Requires network-capable tooling: the netmeta package in R (frequentist, graph-theoretical), or a Bayesian MCMC setup via WinBUGS/OpenBUGS/JAGS/Stan, often called through gemtc. | Any standard meta-analysis tool -- RevMan, metafor in R, or comparable software running an inverse-variance fixed-effect or random-effects pool. |
| When it's the right choice | Multiple interventions are relevant to the same decision and at least one pair was never tested head-to-head -- the classic case being a health technology assessment body that needs to compare a sponsor's drug (tested only against placebo) with a competitor already on the market (also tested only against placebo, in a different trial). | The evidence base is genuinely closed: every relevant trial compares the same two interventions, with no third comparator to route an indirect estimate through. Simpler, fewer assumptions, easier to communicate -- and NMA would add complexity without adding any information. |
| Common pitfall | Running an NMA on a network that is technically connected but not really transitive -- e.g. splicing together an old trial generation with a modern one, or paediatric evidence with adult evidence, and treating the merged estimate as if it came from one coherent evidence base. | Forcing a pairwise pooled estimate across studies that are clinically or methodologically too different to combine meaningfully -- the single-comparison version of ignoring heterogeneity. |
| Network geometry | Drawn and reported as a network diagram: nodes are interventions, edges are direct comparisons, edge thickness typically reflects the number of trials. A sparsely connected, "star-shaped" network (everything only ever compared against one central comparator, with no closed loops) leans harder on transitivity than a densely connected one, because there's less direct evidence to cross-check indirect estimates against. | No network diagram applies -- there is exactly one edge, between the two interventions being compared, so there's nothing to map. |
| Relationship to a simple indirect comparison | Generalizes the older Bucher (adjusted indirect comparison) method, which handles only a single indirect comparison through one common comparator, to an arbitrary network of interventions and comparisons estimated jointly in one model. | Not applicable -- pairwise meta-analysis never constructs an indirect estimate in the first place, so there's no indirect-comparison method to generalize from. |
Common questions
Common questions about Network Meta-Analysis vs Pairwise Meta-Analysis
Does network meta-analysis replace pairwise meta-analysis?
+
No -- it generalizes it. A network meta-analysis with only two connected interventions and one shared comparator collapses to the same estimate a pairwise meta-analysis of that pair would give. NMA is the broader tool for when three or more interventions are relevant to the same question; pairwise meta-analysis remains the right, simpler tool when only two interventions are ever being compared.
What exactly is the transitivity assumption, in plain terms?
+
That the different groups of trials feeding a network meta-analysis are similar enough, on average, in everything except which intervention they tested, that it's reasonable to treat them as if they could all have been arms of one single, larger trial. If the A-vs-B trials and the A-vs-C trials enrolled systematically different kinds of patients, the indirect B-vs-C comparison you'd derive from them isn't trustworthy, even if each individual trial was well-run.
How do I tell whether an evidence base is 'genuinely closed' and pairwise is enough?
+
Map every trial in your evidence base by which two interventions it compared. If every trial in scope compares the same two interventions -- no third intervention appears anywhere in the eligible literature -- there is no indirect evidence to combine and nothing for a network model to add. The moment a third intervention with any connecting trial exists, you have a network, even a small one.
What happens if transitivity is violated but I run a network meta-analysis anyway?
+
The model will still produce numbers -- a relative effect estimate and a ranking -- but they describe a comparison built on a false equivalence between trial populations, not a real one. This is why consistency checking (node-splitting or the design-by-treatment interaction test) is treated as a mandatory step, not an optional extra: it's the closest available statistical evidence for whether transitivity actually held.
Is network meta-analysis the same thing as an 'indirect comparison' or 'mixed treatment comparison'?
+
Closely related, and the terms are often used loosely. A purely indirect comparison (Bucher method) uses only the indirect evidence for a single pair. Mixed treatment comparison (MTC) is an older name for what's now usually called network meta-analysis: combining direct and indirect evidence together across a full network, not just one indirect pair.
Do regulators and HTA bodies accept network meta-analysis, or only head-to-head trials?
+
Health technology assessment bodies routinely rely on network meta-analysis precisely because head-to-head trials against every relevant competitor rarely exist -- NICE's Decision Support Unit publishes dedicated Technical Support Documents on how to conduct and report NMA for submissions. It is an accepted, expected method in that context, not a workaround, provided transitivity and consistency are properly assessed and reported.
Going deeper
