Meta-Analysis: Paired and Network Approaches

Search strategy · Data extraction · Fixed and random effects · Network meta-analysis · Consistency · GRADE

Published 2026 · Last updated 2026-09-27 · Hailu Zhou

Meta-analysis is the craft of turning a pile of studies into one estimate without laundering the heterogeneity underneath. Paired meta-analysis compares two treatments at a time; network meta-analysis compares many treatments at once by borrowing the indirect evidence that runs through common comparators. Both live or die on the search and extraction stage, which happens before any model runs.

A search strategy that can be reproduced

Report the search so that someone else can rebuild it. That means naming the databases, the exact search strings, the date of the last search, language restrictions, and whether grey literature was included. Combine controlled vocabulary with free-text terms across the population, intervention, comparator, and outcome facets. Register the protocol before starting — the registration is the only real defence against selective reporting of results you like.

Data extraction and the paired-comparison core

For a paired (pairwise) meta-analysis, each study contributes an effect estimate with a variance — a mean difference, a standardised mean difference, a risk ratio, or an odds ratio, depending on outcome type. Extract the 2×2 table or the group means and standard deviations, not the p-value; studying only the p-value throws away the effect size you are trying to pool. Handle multi-arm studies by selecting relevant arms or accounting for the correlation between arms that share a control group.

Once extracted, two models compete. The fixed-effect model assumes one true effect shared by all studies and gives the tightest confidence interval — which is exactly why it feels good and is often wrong. The random-effects model assumes study effects are drawn from a distribution and adds between-study variance to the weight, producing a wider, more honest interval when heterogeneity is real. Choose based on the question: a fixed effect answers "what is the effect here," a random effect answers "what is the average effect across settings."

Heterogeneity is the finding, not the annoyance

The I² statistic reports the proportion of total variation attributable to between-study differences rather than chance. A high I² does not invalidate pooling, but it demands an explanation: subgroup analysis, meta-regression on study-level covariates, or a decision that pooling was inappropriate. Leave-one-out and influence diagnostics show whether one study drives the result. Reporting a pooled estimate with I² above 75% and no investigation is a red flag, not a result.

Network meta-analysis: many treatments, indirect evidence

When several treatments have been compared but not all against each other, network meta-analysis estimates the missing comparisons through a connected graph of studies. A comparison like A versus C can be inferred from A versus B and B versus C, and the estimates can be combined with any direct A versus C trials. The network must be connected; a disconnected network cannot borrow information across components, and transitivity — the assumption that the indirect comparison is valid because the studies are sufficiently similar in effect modifiers — is the load-bearing assumption that no statistical test confirms.

Framework Style Strength
Frequentist NMAWeighted least squares / likelihoodFast, familiar p-values
Bayesian NMAMarkov chain Monte Carlo, prior distributionsRanking probabilities, flexible models

The frequentist route estimates relative effects by generalised least squares and is quick to run; the Bayesian route places priors on the relative effects and yields posterior probabilities that each treatment is best, which is why ranking tables are popular — and why they mislead, because a treatment's rank probability is unstable when the network is sparse or heterogeneity is high.

Consistency testing

The central check in any network is whether the direct and indirect evidence agree. A global test compares the fit of a model that assumes consistency with one that allows each comparison its own effect; a node-splitting or side-splitting approach isolates individual comparisons where direct and indirect effects diverge. When inconsistency is detected, the fix is not to average it away — it is to look for clinical or methodological explanations: differing definitions of the outcome, differing populations, or a trial that should not have been included. A well-fitting inconsistent network is a narrative, not evidence.

Certainty with GRADE

GRADE rates the certainty of evidence on four levels — high, moderate, low, very low — across domains including risk of bias, inconsistency, indirectness, imprecision, and publication bias. For network meta-analysis the framework extends with additional considerations such as intransitivity and incoherence. The point is that a precise pooled estimate from biased, indirect trials is still low-certainty evidence, and the grade belongs in the abstract, where a reader will actually see it. Reporting numbers without a certainty judgement is how meta-analyses end up more confident than their inputs.

References

  1. DerSimonian R, Laird N. Meta-analysis in clinical trials. Control Clin Trials. 1986;7(3):177–188.
  2. Higgins JPT, Thompson SG, Deeks JJ, Altman DG. Measuring inconsistency in meta-analyses. BMJ. 2003;327(7414):557–560.
  3. Page MJ, McKenzie JE, Bossuyt PM, et al. The PRISMA 2020 statement: an updated guideline for reporting systematic reviews. BMJ. 2021;372:n71.
  4. Salanti G, Ades AE, Ioannidis JPA. Graphical methods and numerical summaries for presenting results from multiple-treatment meta-analysis: an overview and tutorial. J Clin Epidemiol. 2011;64(2):163–171.
  5. Salanti G. Indirect and mixed-treatment comparison, network, or multiple-treatments meta-analysis: many names, many benefits, many concerns for the next generation evidence synthesis tool. Res Synth Methods. 2012;3(2):80–97.
  6. Chaimani A, Higgins JPT, Mavridis D, Spyridonos P, Salanti G. Graphical tools for network meta-analysis in STATA. PLoS One. 2013;8(10):e76654.
  7. Dias S, Welton NJ, Caldwell DM, Ades AE. Checking consistency in mixed treatment comparison meta-analysis. Stat Med. 2010;29(7–8):932–944.
  8. Guyatt GH, Oxman AD, Vist GE, et al. GRADE: an emerging consensus on rating quality of evidence and strength of recommendations. BMJ. 2008;336(7650):924–926.
  9. Higgins JPT, Jackson D, Barrett JK, et al. Consistency and inconsistency in network meta-analysis: concepts and models for multi-arm studies. Res Synth Methods. 2012;3(2):98–110.