Free explainer2 min read
Network meta-analysis
Understand the model, recognise it in a stem, separate the look-alikes, then apply it.
Start with the mental model
How can a guideline panel say drug A beats drug B on efficacy when no trial ever put A and B head-to-head against each other? A network meta-analysisA study that statistically combines the results of several separate studies into one overall estimate.Read more → (NMA) is a statistical method that combines every available trial in a treatment area — even ones comparing different pairs of drugs — into a single connected network, so treatments that were never directly compared can still be ranked against each other. For example, imagine trial 1 compared drug A against placeboAn inactive treatment given to a comparison group so any real effect of the active treatment can be judged against it.Read more →, and trial 2 compared drug B against the same placebo, but no trial ever compared A against B directly. In practice, this means an NMA can still estimate how A and B compare, by using placebo as the shared bridge between the two separate trials, so long as the trials are similar enough to make that bridge fair.
01Core model
An NMA represents each trial's comparisons as an edge connecting two treatment 'nodes', so a whole treatment area becomes a network diagram: some treatments are linked by direct trials, others only by a shared comparator. Where two treatments have been compared head-to-head, the NMA uses that direct evidence. Where they have not, it derives an indirect estimate by comparing each treatment's effect against the shared comparator: if A beats placeboAn inactive treatment given to a comparison group so any real effect of the active treatment can be judged against it.Read more → by X and B beats placebo by Y, the model derives an indirect A-versus-B estimate from the difference between X and Y. Where both direct and indirect evidence exist for the same comparison, the NMA statistically combines them into one 'mixed' estimate, which is usually more precise than either source alone. This whole approach rests on the transitivity assumption: that trials feeding into the network are similar enough in their patients, dosing and setting that borrowing evidence through a common comparator is a fair like-for-like exchange, rather than comparing genuinely different populations. Where enough of the network forms closed loops (a treatment reachable by more than one path), the model can also test statistical consistency — whether the direct and indirect estimates for the same comparison actually agree; disagreement flags a possible transitivity violation.