Strata Academy
Network Meta-Analysis Explained for Students
Indirect comparisons, transitivity, consistency, league tables, and when NMA beats pairwise meta-analysis
Quick answer
Network meta-analysis (NMA) combines direct and indirect evidence to compare multiple treatments simultaneously, even when not all treatments were compared head-to-head in trials. It requires a connected network of studies, the transitivity assumption (similar trials compare similar populations), and consistency between direct and indirect estimates. Report with PRISMA-NMA extension and interpret rankings cautiously.
- NMA extends pairwise meta-analysis to multiple treatments in one model.
- Transitivity — not statistical consistency alone — is the key clinical assumption.
- League tables show pairwise comparisons; rankograms show probability of being best.
- Low certainty evidence can still produce confident-looking league tables — read GRADE.
1. What network meta-analysis is
Standard pairwise meta-analysis pools trials comparing A versus B. Network meta-analysis (also called mixed treatment comparison or multiple treatments meta-analysis) links trials that share comparators to estimate relative effects between all treatments in the network — including pairs never directly compared.
Example: if Trial 1 compares Drug A vs placebo and Trial 2 compares Drug B vs placebo, NMA can estimate A vs B indirectly through the shared placebo node — provided assumptions hold.
NMA does not create evidence from nothing. It statistically combines existing trial data under explicit assumptions. Weak or biased primary trials propagate through the network as they would in pairwise meta-analysis.
NMA is common in NICE technology appraisals, Cochrane reviews with multiple pharmacological agents, and student dissertations comparing several treatment options where head-to-head trials are sparse.
2. Reading the network diagram
Network diagrams display treatments as nodes and trials as connecting lines (edges). Line thickness often represents number of trials or participants on that comparison.
A network must be connected — every treatment must link to every other through some path of comparisons. Disconnected networks require separate analyses.
Inspect each node: how many trials? How many participants? Sparse networks with one small trial per treatment produce unstable estimates with wide credible intervals.
Check for star networks centred on placebo — common in drug development. Indirect comparisons between active drugs depend entirely on how each drug performed against placebo in potentially different populations.
- Nodes — treatments or interventions
- Edges — direct comparisons with trial counts
- Network connectivity — all treatments linked?
- Sparsity — few trials per comparison increases uncertainty
- Placebo hub — indirect comparisons rely on shared comparator quality
3. Transitivity and consistency
Transitivity is the clinical assumption that trials comparing different treatment pairs are sufficiently similar that indirect comparison is meaningful. Formally: the distribution of effect modifiers (prognostic factors that modify treatment effects) should be comparable across comparisons.
If Drug A vs placebo trials enrolled sicker patients than Drug B vs placebo trials, the indirect A vs B comparison confounds severity with treatment effect. Transitivity fails — statistics cannot fix this.
Consistency is the statistical manifestation: direct evidence for a comparison should align with indirect evidence through the network. Global inconsistency tests and node-splitting examine local disagreement. Inconsistency may indicate transitivity violations, chance, or study-level bias.
Cochrane Handbook guidance emphasises assessing transitivity clinically before trusting consistency statistics. A non-significant inconsistency test does not prove transitivity holds.
Note: Never conclude 'Treatment X is best' from rankograms alone. Rankings ignore magnitude, certainty, and absolute benefit. A treatment ranked first may differ trivially from second with wide credible intervals.
4. League tables, rankograms, and effect measures
League tables present pairwise comparisons for every treatment pair with effect estimates and credible intervals. For dichotomous outcomes, odds ratios or risk ratios; for continuous outcomes, mean differences. Check which effect measure and model (fixed vs random) authors used.
Rankograms and surface under the cumulative ranking (SUCRA) curves show probability each treatment ranks 1st, 2nd, etc. These are descriptive summaries — not hypothesis tests. NICE and Cochrane increasingly caution against over-interpreting rankings.
Compare NMA pooled estimates with direct meta-analyses where direct evidence exists. Large discrepancies warrant investigation before accepting network results.
Forest plots in NMA may show network estimates per comparison rather than study-level rows. Locate supplementary material for study-level data.
| Output | Shows | Interpret cautiously because |
|---|---|---|
| League table | All pairwise comparisons | Multiple comparisons — multiplicity not always adjusted |
| Rankogram | Probability of each rank | Ignores effect size and certainty |
| SUCRA score | Summary ranking metric | Arbitrary scale; not clinically interpretable alone |
| Network plot | Evidence structure | Connectivity ≠ quality |
5. Statistical methods (conceptual overview)
Frequentist NMA uses multivariate meta-analysis models estimating treatment effects relative to a reference treatment (often placebo or standard care). Bayesian NMA provides probabilistic rankings and credible intervals via Markov chain Monte Carlo.
Random-effects models account for between-study heterogeneity within each comparison. Consistency models assume direct and indirect evidence agree; inconsistency models relax this for exploration.
Assessment of heterogeneity in NMA is less straightforward than pairwise I². Examine comparison-specific heterogeneity and whether effect modifiers explain variation.
You do not need to run NMA software to appraise a published NMA — but you must understand enough to evaluate assumptions, network geometry, and whether conclusions match evidence certainty.
6. GRADE, reporting, and common student errors
GRADE for NMA follows emerging frameworks: assess certainty for each pairwise comparison in the league table, considering risk of bias in contributing trials, inconsistency, indirectness, imprecision, and publication bias across the network.
PRISMA-NMA extends PRISMA 2020 with items specific to network geometry, model specification, inconsistency assessment, and ranking presentation. Appraise published NMA with PRISMA-NMA and AMSTAR 2 as for any systematic review.
Common student errors: treating NMA as more authoritative than pairwise meta-analysis without justification; ignoring transitivity because the network is connected; citing SUCRA rankings in conclusions without league table effect sizes; pooling incompatible study designs without subgroup analysis.
When conducting NMA for a dissertation, pre-specify the network model, reference treatment, effect measure, and inconsistency approach in the protocol. Contact a statistician early — NMA is not a default upgrade from pairwise pooling.
- Pre-specify NMA in PROSPERO before seeing network geometry
- Justify transitivity with comparison of study characteristics
- Report direct and indirect contributions per comparison
- Present absolute effects alongside relative effects where possible
- Acknowledge sparse networks as low-certainty evidence
Frequently asked questions
When should I use NMA instead of pairwise meta-analysis?
Use NMA when the clinical question requires comparing multiple treatments and the trial network is connected with sufficient direct or indirect evidence. If only two treatments matter and ample head-to-head trials exist, pairwise meta-analysis is simpler and carries fewer assumptions.
What is the transitivity assumption in plain terms?
Trials in the network should be similar enough that an indirect comparison is fair — comparable patients, settings, co-interventions, and outcome definitions across treatment comparisons. If trials differ systematically on effect modifiers, indirect estimates are biased regardless of statistical consistency.
Can I trust SUCRA rankings for clinical decisions?
Rankings alone are insufficient. A treatment may rank highest with low probability and wide credible intervals on key outcomes. Always read league tables for effect sizes, certainty (GRADE), and absolute benefit. NICE guidance explicitly cautions against rank-only decision making.
How is NMA different from subgroup meta-analysis?
Subgroup meta-analysis stratifies trials within one comparison by a covariate (e.g. age). NMA simultaneously models multiple treatment comparisons in a network. They answer different questions and require different assumptions.
Interactive walkthroughs and quizzes load when JavaScript is enabled — the checklist and tables above are fully readable without it.