Network Meta-Analysis: A Beginner’s Guide

A standard, pairwise meta-analysis pools studies that directly compare the same two interventions -- treatment A versus treatment B, across multiple trials. Network meta-analysis extends this to situations where multiple interventions have been studied across a web of different head-to-head comparisons, allowing you to estimate how interventions that were never directly compared against each other in any single trial likely compare, by combining direct and indirect evidence across the entire network.

Why this matters in practice

Clinical and research questions rarely involve just two competing options. If five different interventions exist for a condition, but trials have only ever compared some pairs directly -- A versus B, B versus C, C versus D -- a standard pairwise meta-analysis cannot tell you how A compares to D, since no trial ever tested that comparison directly. Network meta-analysis can estimate that indirect comparison by combining the chain of direct evidence that does exist, provided certain assumptions hold.

Direct and indirect evidence

Direct evidence comes from trials that actually compared two specific interventions head-to-head. Indirect evidence is inferred by combining direct evidence along a chain -- if A has been compared to B, and B has been compared to D, you can estimate an indirect comparison of A versus D through B as a common comparator. Network meta-analysis combines all available direct and indirect evidence across the network simultaneously, rather than relying on any single indirect chain in isolation, which generally produces more precise estimates than looking at any one indirect pathway alone.

The transitivity assumption

The validity of any indirect comparison rests on an assumption called transitivity: that the trials contributing to each direct comparison in the chain are similar enough, in terms of patient populations, outcome definitions, and study conduct, that combining them indirectly produces a meaningful estimate rather than comparing fundamentally different contexts. If the population enrolled in A-versus-B trials is meaningfully different from the population in B-versus-D trials -- younger patients, different disease severity, different concurrent treatments -- the indirect A-versus-D estimate may not be trustworthy, even if the individual pairwise meta-analyses feeding into it are each methodologically sound.

Assessing transitivity is a qualitative, substantive judgment about the clinical and methodological similarity of trials across the network, not something a single statistical test resolves on its own, and it should be discussed explicitly and specifically in your methods and results, not simply assumed.

Consistency: checking direct against indirect evidence

Where both direct and indirect evidence exist for the same comparison, a well-conducted network meta-analysis checks whether they agree, a property called consistency. A node-splitting analysis or a design-by-treatment interaction test are commonly used to formally assess this. Meaningful inconsistency between direct and indirect estimates for the same comparison is a signal that the transitivity assumption may not hold somewhere in the network, and it should prompt closer scrutiny rather than being averaged over or ignored.

Ranking interventions: SUCRA and its limits

Network meta-analyses often report a ranking of all included interventions by their probability of being the best, second-best, and so on for a given outcome, commonly summarized using SUCRA, the Surface Under the Cumulative Ranking curve. SUCRA rankings are useful for communicating relative performance at a glance, but they are frequently over-interpreted -- a small difference in SUCRA score between two interventions does not necessarily reflect a meaningful or statistically significant clinical difference, and rankings should always be reported alongside the actual effect estimates and their confidence intervals, not as a standalone headline finding.

PRISMA reporting for network meta-analysis

PRISMA-NMA is the network meta-analysis extension of the PRISMA checklist, adding reporting items specific to this methodology, including a network diagram showing which interventions have been directly compared, explicit reporting of the transitivity assessment, and consistency checking results. A network meta-analysis submitted without a network diagram or without an explicit transitivity discussion is missing reporting elements a methods reviewer will specifically expect, given how central both are to the method's validity.

When network meta-analysis is not the right choice

If your network of evidence is sparse -- very few trials, very few connections between interventions, or comparisons that only ever connect through a single, thin chain -- the indirect estimates produced may carry so much uncertainty that they are not clinically useful, even though the statistical machinery will still produce an output. A sparse or poorly connected network is a legitimate reason to present findings more cautiously, or to focus on the direct pairwise comparisons that do exist rather than pushing indirect estimates further than the underlying evidence can actually support.

Visualizing the network itself

A network diagram, typically presented early in the results section, shows each intervention as a node and each direct head-to-head comparison as a line connecting two nodes, with line thickness often representing how many trials contributed to that specific comparison. This diagram does real analytical work beyond illustration -- a thin, sparsely connected network with long chains between key interventions is a visual warning that some of your indirect estimates rest on limited direct evidence, while a densely connected network with many trials per comparison supports more confidence in the indirect estimates it produces. Star-shaped networks, where a single common comparator connects to every other intervention but interventions are never compared directly to each other, are a particularly common real-world pattern, and they mean nearly all of your comparisons between non-central interventions are indirect -- worth flagging explicitly when interpreting the ranking table, since the network geometry itself is telling you how much of your conclusion rests on inference rather than direct trial evidence. Reading the network diagram before reading the ranking table is a good habit, since it tells you how much weight the ranking actually deserves.

Network meta-analysis is a genuinely powerful extension of standard meta-analysis, but its added value depends entirely on whether transitivity and consistency actually hold across your specific evidence network -- assumptions worth investigating rigorously, study by study and comparison by comparison, before trusting the ranking table at face value.