Network Meta-Analysis Explained: Indirect Comparisons for HTA

A new medicine is usually tested against placebo or one standard treatment. A health technology assessment (HTA) body needs to know how it compares with each treatment patients would otherwise receive. When no head-to-head trial exists, network meta-analysis fills the gap.

The Cochrane Handbook notes that NMA usually gives more precise estimates than a single direct or indirect comparison, covers contrasts no trial has made and allows treatments to be ranked. Each benefit rests on assumptions, which this guide explains.

Why HTA needs indirect treatment comparisons

HTA compares a new technology with the treatments it would replace in practice, while pivotal trials are often designed against placebo or a single comparator. An indirect treatment comparison (ITC) bridges that gap.

The comparator is fixed by the decision problem, usually a PICO question (population, intervention, comparator, outcomes), covered in Module 1: what HTA is and how PICO frames the decision. Our guide to what health technology assessment is explains why HTA asks a different question from regulators.

A competitor can reach the market after a trial starts, standard care differs between countries, and an EU joint clinical assessment can contain several PICOs, each with its own comparator. One trial rarely answers them all.

Direct vs indirect evidence in a network meta-analysis

Direct evidence comes from trials that randomise patients between the two treatments of interest. Indirect evidence estimates the same contrast through a third treatment that both were tested against, such as placebo.

If drugs B and C have each been compared with placebo A, the two sets of trials can be combined to estimate B versus C. When some trials also compare B and C directly, NMA combines both sources into one network (mixed) estimate.

The rule is to combine relative effects from each trial, never the raw results of single arms: randomisation protects comparisons within a trial, not between arms of different trials.

The Bucher method: an adjusted indirect comparison step by step

The Bucher method, published in 1997, estimates B versus C as the difference between the B versus A and C versus A effects on a log scale, with a variance equal to the sum of the two variances. The common comparator A anchors the comparison, so randomisation is preserved within each trial.

1. Express each trial’s effect on the log scale (log hazard ratio, odds ratio or risk ratio).
2. Subtract the C versus A effect from the B versus A effect.
3. Add the variances: the standard error (SE) is √(SE_BA² + SE_CA²).
4. Take the estimate ± 1.96 × SE and exponentiate to get the 95% confidence interval.

A worked example with hypothetical trials

Take a hypothetical new drug B and an established competitor C, each tested against placebo A in a separate trial, with round, illustrative numbers.

The point estimate is −0.511 − (−0.288) = −0.223, a hazard ratio of 0.80 (the same as 0.60 ÷ 0.75). The standard error is √(0.10² + 0.12²) = √0.0244 = 0.156. On the log scale the interval is −0.223 ± 1.96 × 0.156, or −0.529 to 0.083, a hazard ratio of 0.59 to 1.09.

Both trials show a clear benefit over placebo, yet the indirect interval includes 1, because uncertainty from both trials adds up. And the estimate holds only if the two trials are similar in effect modifiers. Hazard ratios and confidence intervals are explained in Module 3: Biostatistics.

Network geometry: how to read an NMA network diagram

Network geometry describes which treatments are connected by trials and how much evidence sits on each link. It decides which comparisons are informed directly, which only indirectly, and whether inconsistency can be checked at all.

In a network diagram, each node is a treatment and each line a direct comparison; node size can reflect participants and line thickness the number of trials. A network of T treatments has T − 1 basic comparisons, and every other contrast follows from them.

The Cochrane Handbook separates the decision set, the options being chosen between, from a supplementary set, such as placebo, added to connect or strengthen the network.

Key NMA assumptions: transitivity and consistency

An NMA is only as valid as two linked assumptions. Transitivity means the trials behind different comparisons are similar, on average, in everything that modifies treatment effects. Consistency means direct and indirect estimates of the same comparison agree.

Transitivity (similarity)

An effect modifier changes the size of a relative treatment effect: disease severity, prior lines of therapy or dose, for example. If the B versus A trials enrolled sicker patients than the C versus A trials, and severity modifies the effect, the indirect B versus C estimate is biased.

Transitivity cannot be proven statistically. Reviewers tabulate likely effect modifiers across comparisons and apply clinical judgement; the Cochrane Handbook asks whether all treatments could, in principle, have been randomised in one multi-arm trial. Unreported effect modifiers cannot be checked. The EU guideline calls the overarching assumption exchangeability, assessed through similarity, homogeneity and consistency.

Consistency: how inconsistency is checked

Inconsistency, also called incoherence, can only be detected in closed loops. NICE DSU Technical Support Document (TSD) 4 and the Cochrane Handbook describe three approaches:

1. Loop-specific comparison: within a loop, report the difference between direct and indirect estimates (the inconsistency factor) with its confidence interval.
2. Node-splitting: for each comparison, contrast the direct estimate with the indirect estimate from the rest of the network.
3. Global checks: compare the fit of consistency and inconsistency models, or use network-wide inconsistency statistics.

Return to the example. Suppose a small hypothetical head-to-head trial reports B versus C at HR 0.95 (log-scale SE 0.15). The ratio of indirect to direct hazard ratios is 0.84 (95% CI 0.55 to 1.29; p ≈ 0.43). No significant inconsistency, yet the interval is too wide to rule out a clinically important conflict.

TSD 4 warns that inconsistency tests are inherently underpowered and that failing to detect inconsistency does not show consistency. NICE expects inconsistency to be reported, explained and, where possible, resolved; TSD 4 advises reviewing the whole network for effect modifiers.

Frequentist vs Bayesian network meta-analysis

Both frameworks can fit a valid NMA, and the EU methodological guideline notes that frequentist graph-theoretical NMA gives results similar to Bayesian NMA. They differ in how uncertainty is expressed and whether prior information is used.

The NICE DSU evidence synthesis series uses Bayesian models in its worked examples, and one document (TSD 6) covers embedding evidence synthesis in probabilistic cost-effectiveness analysis. That matters because NMA outputs usually feed Markov and other economic models and probabilistic sensitivity analysis.

Fixed vs random effects

A fixed effect model assumes one true effect per comparison, with differences between trials due to chance alone. A random effects model lets the true effect vary between trials around an average. NMAs usually assume the same heterogeneity for every comparison, so sparse comparisons borrow information from the rest of the network. The EU guideline calls the fixed effect assumption commonly implausible and says random effects should generally be used.

Treatment rankings and SUCRA: what they show and what they hide

SUCRA, the surface under the cumulative ranking curve, condenses a treatment’s ranking probabilities into one number between 0 and 1 (often shown as a percentage). A SUCRA of 1 means the treatment always ranks first; 0 means it always ranks last.

Ranking probabilities are often plotted as a rankogram. SUCRA can be read as the average proportion of competing treatments that a treatment beats, and the P-score is its frequentist equivalent. Rankings are easy to misread:

• They ignore the size of differences. A treatment can top the ranking with a clinically trivial advantage, or on low-certainty evidence.
• They mostly mirror point estimates. Rücker and Schwarzer found that ranking by SUCRA or P-score has no major advantage over ranking by point estimates.
• They can reward imprecision. The Cochrane Handbook warns against inference based only on the probability of being best.
• They are outcome-specific. The best treatment for efficacy may rank poorly on safety.

Certainty of evidence in NMA: GRADE and CINeMA

Certainty must be judged comparison by comparison, not once for the whole network. Two frameworks do this: the GRADE Working Group approach for NMA and CINeMA, which builds on GRADE.

The GRADE approach for NMA (Puhan and colleagues, 2014) rates the direct, indirect and network estimate for each comparison in four steps. An indirect estimate takes the lower rating of the two direct comparisons in its most influential loop. Without important incoherence, the network estimate takes the higher of the direct and indirect ratings. A single rating for a whole network, the authors conclude, is uninformative and likely to mislead.

CINeMA (Confidence in Network Meta-Analysis, 2020) judges six domains: within-study bias, reporting bias, indirectness, imprecision, heterogeneity and incoherence. Each is rated as no, some or major concerns, giving high, moderate, low or very low confidence. Its free web application weights each study’s risk of bias by its contribution to each estimate.

Vocabulary differs in HTA: the EU methodological guideline speaks of uncertainty and notes that this overlaps with, but is not the same as, GRADE’s certainty of evidence.

MAIC and STC: population-adjusted indirect comparisons

When trials differ in effect modifiers and a company has individual patient data (IPD) for its own trial but only published summaries for a competitor’s, population-adjusted methods try to correct the imbalance. NICE DSU TSD 18 (Phillippo and colleagues, December 2016) sets out how to use them.

• Matching-adjusted indirect comparison (MAIC) reweights IPD patients so their weighted baseline characteristics match the comparator trial’s published averages, much like inverse propensity score weighting.
• Simulated treatment comparison (STC) fits a regression model to the IPD and predicts outcomes in the comparator trial’s population.
• Multilevel network meta-regression (ML-NMR), a newer method covered in the EU guideline, combines IPD and aggregate data across a network.

TSD 18 adds two cautions. Results apply to the comparator trial’s population, which may not be the decision population, so the target population must be stated. And reweighting shrinks the effective sample size (ESS), which should be reported with the distribution of weights.

The tovorafenib JCA, the first JCA report to be published (June 2026), shows the stakes. The only comparison included was an unanchored MAIC against dabrafenib plus trametinib, with an ESS of about 6 to 14 patients. Assessors judged the risk of bias greater than for randomised evidence, with residual confounding of unknown magnitude and direction.

How network meta-analysis is used in HTA submissions and the EU JCA

In HTA, network meta-analysis supplies the relative effects of a new technology against every relevant comparator. Those estimates drive the clinical case and the QALY gains behind the incremental cost-effectiveness ratio, so agencies scrutinise the network, its assumptions and its uncertainty.

NICE

NICE’s manual (PMG36, section 3.4) prefers the NMA methods of the DSU evidence synthesis series. The network should ideally include every technology in the scope, randomisation must be preserved, and heterogeneity and inconsistency must be reported. Population adjustment can be considered in connected networks, with reference to TSD 18. See how NICE technology appraisals work for the wider process.

The EU joint clinical assessment

As of October 2026, the European Commission’s HTA pages publish a methodological and a practical guideline on direct and indirect comparisons, adopted by the Member State Coordination Group on HTA on 8 March 2024. They are not legally binding, but they set expectations for JCAs:

1. Exchangeability is the core assumption, with effect modifiers prespecified.
2. Random effects are the default; a fixed effect model needs rigorous justification.
3. Frequentist and Bayesian NMA are both described as usable.
4. Naive comparisons should not be performed, and anchored comparisons are preferred to unanchored ones.
5. Non-randomised comparisons can be tested against a shifted threshold rather than zero, with the threshold left to Member States.

Our guide to the HTA process step by step shows where evidence synthesis fits, and Module 12: Regulation and Reimbursement places the JCA in the wider reimbursement picture.