When should I use a network meta-analysis?
Use one when your question involves three or more competing interventions and you need them compared against each other, including pairs that no trial has directly compared. It requires a connected network of trials and the assumption that the trials are similar enough to be combined indirectly. If you only have two interventions, or the network is disconnected, you do not need it and should not use it.
What it buys you
Guideline and reimbursement questions are rarely "does this work?" They are "which of these five should we recommend?" — and the trials almost never line up to answer that, because each new agent was tested against placebo or against whatever was standard at the time.
A network meta-analysis uses the shared comparators to connect those trials, so that A versus C can be estimated even if nobody ran that trial, provided both were compared with B. Where direct and indirect evidence both exist, it combines them, usually producing a more precise estimate than either alone.
The conditions that must hold
- A connected network
- Every intervention must link to the others through some chain of shared comparators. If your evidence splits into two islands, no amount of modelling will compare across the gap — and analysing only the largest connected component silently changes your question.
- Transitivity
- The trials contributing to each comparison must be similar in anything that modifies the effect — population severity, dose, era, follow-up. If the A-versus-B trials enrolled much sicker patients than the B-versus-C trials, the indirect A-versus-C estimate is confounded.
- Coherence
- Where direct and indirect evidence both exist for a comparison, they should agree within chance. Disagreement is a signal that transitivity has failed somewhere, and it must be investigated rather than pooled through.
Transitivity is judged from the study characteristics table, before you fit anything. Statistical incoherence tests have limited power, so a passing test is weak reassurance and a failing one is a real problem.
When a pairwise review is the right answer
If the decision is genuinely between two options, run the pairwise review — it is simpler, its assumptions are weaker, and it will be easier to defend. If the interventions are too dissimilar to sit in one network, splitting the question is more honest than lumping them.
Network meta-analysis buys you comparisons you could not otherwise make, and pays for them with an assumption you cannot test directly. Make that trade deliberately.
Where this answer stops
The indirect comparisons rest on transitivity — an assumption about the similarity of the trials that cannot be verified from the data, only argued from the study characteristics. Statistical checks can detect some failures; they cannot confirm the assumption holds.