Skip to content
Back to search results

Network Meta-Analysis

Introduced by
Lumley (2002)
Year
2002
Type
Method
Family
Advanced Meta-Analysis
Font size

What does it do?

Network meta-analysis (NMA) is a systematic method for comparing multiple interventions simultaneously within a single analytical framework, incorporating both direct evidence (head-to-head trials) and indirect evidence (comparisons via common comparators).

First formalized by Lumley in 2002, NMA allows researchers to rank treatments and quantify comparative effectiveness even when some treatment pairs have never been directly studied.

The intuition behind the method

Traditional pairwise meta-analysis compares treatment A versus B using only trials that directly tested this comparison. However, if many interventions exist but few direct comparisons, researchers cannot rank all options. NMA solves this by creating a 'network' of evidence—if trials exist for A vs B and B vs C, NMA can synthesize an indirect comparison of A vs C through B, provided the assumption of consistency (transitivity) holds. This ranks all treatments on a single scale and identifies which is likely most effective, tolerable, or cost-effective.

History

Network meta-analysis emerged in the early 2000s as Bayesian methods and computing power made simultaneous modeling of multiple treatment comparisons feasible. Lumley's 2002 publication formalized the Bayesian approach and introduced the consistency framework. Caldwell et al. (2005) and later Dias & Ades refined methods for assessing inconsistency. The Cochrane Collaboration adopted NMA methods and now publishes multiple Cochrane NMA reviews annually. Software development (BUGS, WinBUGS, JAGS, Stan) and R packages (netmeta, gemtc, BUGSnet) have made NMA accessible beyond specialist statisticians. Living NMA (continuously updated as new trials publish) is an emerging frontier.

When to use it

Use NMA when (1) multiple interventions for the same condition exist, (2) researchers and clinicians need to rank or choose between them, (3) direct head-to-head trials are sparse or absent for some comparisons, (4) a connected network exists (treatments linked directly or indirectly), and (5) you can assess and justify the assumption of consistency. Common applications include drug comparisons (antidepressants, antihypertensives), surgical techniques, rehabilitation interventions, and preventive strategies. NMA is particularly valuable in health technology assessment (HTA) to inform reimbursement decisions across multiple options.

Limitations and considerations

  • Requires strong assumption of consistency (transitivity): that treatment effects are comparable across different trial populations and contexts. Violation invalidates rankings.
  • Results depend on the evidence base; sparse networks or unconnected subgraphs reduce precision and may lead to unreliable indirect estimates.
  • Computational complexity increases with network size, and Bayesian NMA requires careful specification of prior distributions, which can influence results.
  • Assumes all available trials are identified; missing studies (especially negative ones) can bias results via small-study effects.

Source conditions do not replace independent scientific review or validation by an analysis engine.

Strengths

  • Enables ranking of all treatments on a single efficacy scale, even when direct trial evidence is absent for some pairs.
  • Increases statistical power by borrowing strength from indirect comparisons, improving precision of estimates.
  • Provides a comprehensive evidence synthesis for guideline and policy development, reducing reliance on arbitrary or incomplete pairwise meta-analyses.
  • Identifies network inconsistencies that may signal heterogeneity, study design bias, or unmeasured effect modifiers.
  • Integrates multi-arm trials naturally, avoiding the need to break trials arbitrarily into separate pairwise comparisons.

Applications

  • Comparing multiple antidepressants, antipsychotics, or other drug classes to rank efficacy and tolerability for guideline development.
  • Evaluating surgical techniques (e.g., different approaches to knee ligament repair) where head-to-head trials are rare.
  • Health technology assessment (HTA): ranking new oncology drugs, immunotherapies, or interventions for reimbursement decisions.
  • Behavioral interventions: comparing psychotherapy modalities, weight-loss programs, or digital health tools where multi-armed trials are limited.
  • Rehabilitation and physical therapy: ranking exercise protocols or devices for stroke recovery, chronic pain, or sports injury.

How it works

How it works

1

Begin by conducting a standard systematic review of randomized controlled trials (or other high-quality designs) comparing any pair of the interventions of interest. Create a network diagram showing which treatments were directly compared and how many trials exist for each comparison. Check whether your network is connected (all treatments linked, directly or indirectly) and identify treatment pairs with sparse or absent evidence. Prepare a dataset with trial-level data: study identifiers, arms, sample sizes, events (for binary outcomes) or means/SDs (for continuous outcomes), and outcome measurements. For network geometry analysis, assess whether the network is simple (each trial compares exactly two arms) or complex (multi-arm trials present).

2

Fit a NMA model—most commonly a Bayesian hierarchical model—that simultaneously estimates direct and indirect treatment effects and borrows strength across evidence pathways. The core insight is that all arm-to-arm treatment comparisons are estimated from a shared set of treatment effects. For binary outcomes, estimate log odds ratios; for continuous, estimate mean differences. Perform consistency checking by comparing the network estimate (combining direct and indirect paths) against the direct-only estimate for each treatment pair. Inconsistency may suggest heterogeneity, effect modification, or data quality issues. Calculate SUCRA (Surface Under the Cumulative Ranking Curve)—a probability-based rank for each treatment from 0 to 100, indicating position in the efficacy ranking. Conduct sensitivity analyses by excluding small/high-risk studies, restricting to specific populations, or assuming different heterogeneity structures.

3

Report league tables (matrices showing pairwise comparisons) and ranking plots (mean SUCRA with credible intervals for each treatment). Interpret SUCRA cautiously: a treatment with high SUCRA is probably the best, but credible intervals around ranks quantify uncertainty—overlapping intervals mean treatments cannot be clearly distinguished. If consistency is violated (e.g., direct and indirect estimates diverge meaningfully), report both and investigate reasons (subgroup effects, methodological differences, small-study bias). Communicate that NMA estimates are conditional on transitivity: they assume baseline characteristics and effect modifiers are balanced across the network. Highlight which treatment pairs have direct evidence vs rely heavily on indirect inference. For clinical application, balance rank with absolute effect sizes and side-effect profiles.

Common pitfalls

  • Interpreting SUCRA as absolute probability of superiority rather than relative rank within the study network. SUCRA 75% does not mean 75% chance of being best; it reflects ranking position, and uncertainty may be high.
  • Assuming consistency without testing. Always check consistency explicitly via consistency plots or models that allow inconsistency parameters.
  • Applying NMA to networks with very sparse comparisons (e.g., only one trial per edge). Indirect estimates become unreliable; report with appropriate caution.
  • Ignoring subgroup or meta-regression evidence that suggests treatment effects differ across populations. If consistency is violated, stratified NMA may be needed.

Frequently asked questions

What is the difference between direct and indirect comparisons?

A direct comparison comes from trials that randomized patients to both treatment A and treatment B within the same trial. An indirect comparison is inferred from a network: if A is compared to C in Trial 1 and B is compared to C in Trial 2, we can indirectly compare A vs B through their effects relative to C (the common comparator). NMA synthesizes both types.

What is the consistency assumption, and what happens if it is violated?

Consistency (transitivity) assumes that the relative effect of A vs B is the same regardless of which pathway you take through the network. If A is better than C in trials with young patients but C is better than A in trials with elderly patients, the network is inconsistent. Violations suggest that patient populations or study designs differ meaningfully. Check consistency using consistency plots or by fitting a model with inconsistency parameters. If violated, investigate causes and consider stratified or subgroup NMA.

How do I interpret SUCRA, and does a high SUCRA mean a treatment is definitely best?

SUCRA is the cumulative probability of being best, second-best, etc., across the posterior distribution. SUCRA 90% means, across posterior samples, the treatment ranks best about 90% of the time. However, this does not mean it is definitively best in practice. Always inspect credible intervals around point estimates and ranks. If two treatments have overlapping credible intervals or similar SUCRA values, they are not clearly different. SUCRA is useful for ranking but must be paired with absolute effect size interpretation.

When is NMA not appropriate?

Avoid NMA if (1) the network is disconnected (some treatments are not linked, directly or indirectly), (2) evidence suggests strong violations of consistency (e.g., different patient populations or substantially different study designs), (3) heterogeneity is very high and cannot be explained (may indicate a violated consistency assumption), or (4) the network is very sparse (only one or two trials per comparison) and indirect estimates are therefore highly uncertain.

Related methods

Sources

  1. Lumley, T. (2002). Network meta-analysis for indirect treatment comparisons. Statistics in Medicine, 21(16), 2313–2324.

    Open DOI reference
    Copied.
  2. Bucher, H. C., Guyatt, G. H., Griffith, L. E., & Walter, S. D. (1997). The results of direct and indirect treatment comparisons in meta-analysis of randomized controlled trials. Journal of Clinical Epidemiology, 50(6), 683–691.

    Open DOI reference
    Copied.
  3. Dias, S., Welton, N. J., Caldwell, D. M., & Ades, A. E. (2010). Checking consistency in mixed treatment comparison meta-analysis. Statistics in Medicine, 29(7–8), 932–944.

    Open DOI reference
    Copied.