Skip to content
All guides

Can I trust network meta-analysis rankings?

Network meta-analysis · Updated July 2026

Short answer

Not on their own. A ranking tells you the probability that a treatment is best given the model, and it says nothing about how much better it is or how certain that is. An intervention tested in one small imprecise trial can rank first, because ranking rewards a high point estimate and does not penalise a wide interval. Read the effect estimates and their intervals; treat the ranking as a summary of them, never as a result in itself.

What a ranking metric actually computes

Ranking statistics such as SUCRA, or the equivalent P-score, summarise the whole distribution of an intervention's possible ranks into a single number between 0 and 1. It answers "how likely is this to be among the better options, across the model's uncertainty?"

What it does not carry is magnitude. Two interventions can be separated by a rank and by an effect difference that no patient would notice. The ranking will still put one above the other, with no indication that the distinction is meaningless.

Why imprecise treatments rank well

This is the failure mode worth internalising. An intervention studied once, in eighty patients, with a wide confidence interval, has a genuine chance of being the best — the data cannot rule it out. Ranking metrics integrate over that uncertainty, and a wide distribution with a favourable centre produces a strong rank.

The practical tell: check how many trials and how many patients sit behind the top-ranked intervention. If the answer is "one small trial", the rank is a statement about ignorance, not about efficacy.

How to report rankings responsibly

Lead with effects, not ranks
Report the estimates and intervals against a common comparator first. Let the ranking support the narrative rather than supply it.
Pair every rank with its certainty
A rank order without certainty ratings per comparison invites exactly the misreading you are trying to avoid.
Show the evidence base per node
Number of trials and participants contributing to each intervention. It is the fastest way for a reader to discount a rank built on nothing.
Resist the league-table headline
"X was ranked most effective" is the sentence that gets quoted, and it is usually not what your analysis supports.

Where this answer stops

Rankings inherit every assumption of the network — transitivity, coherence, and the certainty of each contributing comparison. A confidently reported rank order built on low-certainty evidence is a presentation problem, not a statistical one, and no ranking metric flags it for you.