Reading a Network Meta-Analysis and the Transitivity Assumption

A network meta-analysis compares several treatments at once by combining head-to-head trials with indirect comparisons made through a shared comparator. The whole structure rests on transitivity, the assumption that the trials making up each comparison are similar enough in the things that modify the effect that borrowing evidence across them is fair. When you read one, check that the network is connected, that the trials look comparable, and that the authors tested whether direct and indirect evidence agree before you trust a ranking.

What a network meta-analysis is doing

A conventional meta-analysis pools trials that all compared the same two things. A network meta-analysis is more ambitious. It takes a whole web of trials, where different studies compared different pairs of treatments, and estimates how every option compares with every other, including pairs that no single trial ever put side by side.

Picture the treatments as points and the direct comparisons as lines joining them. If drug A was tested against B, and B against C, the two lines meet at B. That shared point is what lets the analysis reach an answer for A versus C even though no trial randomized patients to A against C directly. The reward is a single coherent picture, and often a ranking, across many competing options.

Direct, indirect, and mixed evidence

Three kinds of evidence live inside a network. Direct evidence for A versus C comes from trials that actually compared them. Indirect evidence for A versus C is built from A-versus-B trials and B-versus-C trials, anchored on the common comparator B. Where both exist, the analysis blends them into a mixed estimate.

The indirect step is powerful precisely because it manufactures comparisons that were never run. That is also where the danger lives. An indirect estimate is only as trustworthy as the assumption that the A-versus-B trials and the B-versus-C trials were comparable enough to be chained together.

Transitivity: the assumption that makes the indirect step legitimate

Transitivity is the idea that you could have swapped patients between the different sets of trials without changing the treatment effect. Put plainly, the trials contributing to each comparison should be similar in the characteristics that modify how well a treatment works, things like disease severity, age, dose, or length of follow-up.

Suppose the A-versus-B trials enrolled much sicker patients than the B-versus-C trials. Now the common comparator B is not really playing the same role in both sets, and the indirect A-versus-C estimate inherits that mismatch as bias. Transitivity cannot be proven from the numbers. It is argued from the clinical and methodological features of the trials, which is why a careful review lays out a table of trial characteristics across comparisons and asks whether they line up.

Consistency: the statistical check on transitivity

When a comparison can be estimated two ways, directly and indirectly, transitivity predicts that the two answers should broadly agree. That agreement is called consistency, and its absence is inconsistency. Methods such as node-splitting, also called side-splitting, separate the direct and indirect evidence for a comparison and test whether they clash. A design-by-treatment interaction test checks the network as a whole.

The important nuance is that consistency can only be examined where the network contains a closed loop, meaning a comparison supported by both direct and indirect paths. A star-shaped network, where every treatment was only ever compared with one central option, offers no loops and therefore no way to check consistency at all. In that situation transitivity carries the entire weight and cannot be verified statistically, so the case for it has to be made on clinical grounds.

Reading the rankings without being fooled

Network meta-analyses often end with a ranking: a probability that each treatment is best, or a summary score such as the SUCRA. These are seductive, because a single ordered list feels like a verdict. It usually is not.

Rankings ignore how far apart the treatments actually are and how uncertain each estimate is. A treatment can sit at the top of the list on the strength of one small, imprecise trial, edging out a better-established option by a margin that means nothing clinically. Read the rank alongside the effect sizes and their confidence intervals, and treat a top ranking built on sparse evidence as a hypothesis rather than a conclusion.

A reader's checklist

Start with the network diagram. Is it connected, and where are the loops that would let inconsistency be checked? Thin lines carrying only one or two small trials are weak joints in the structure.

Next, look for the transitivity argument. Did the authors compare the distribution of effect modifiers across the different comparisons, or simply assume similarity? Then find the inconsistency assessment and the certainty ratings. Reporting guidance for these reviews and the GRADE adaptation for networks both ask authors to rate each comparison rather than wave at the network as a whole, so a review that hands you one confident ranking with no discussion of transitivity, consistency, or certainty has skipped the steps that decide whether the answer can be believed.