Network meta-analysis, with reviewer-owned assumptions

When to use NMA

Use network meta-analysis when a review compares more than two interventions and decision-makers need to see how all the options relate, as in HTA, guideline, comparative-effectiveness and treatment-selection questions.

• The review question is explicitly comparative across several interventions.
• Each network is tied to one outcome, timepoint and compatible effect measure.
• NMA does not paper over incompatible populations, interventions, comparators or outcome definitions.

The workflow

1. Extract each trial as randomised arms in Extraction, with the result of every arm and a source quote for every value, then validate. Example, step 1
2. Choose each outcome's settings in Network setup: model, effect measure, whether higher or lower is better, follow-up window and reference treatment. Example, step 2
3. Confirm treatment groups: which network node each arm counts as. Then refresh the validated evidence. Example, step 3
4. Check the network. The check judges the outcome as a run would and says what to fix. Example, step 4
5. Run and read the results: the network graph, league table, ranking and consistency. Example, step 5
6. Rate certainty for each network comparison, then lock the outcome. Example, step 6

Assumptions to review

• Connected network. Treatments must connect through direct or indirect comparisons. A network in separate parts is not fitted: the check names each part and its treatments, and results can't be compared between parts.
• Transitivity. Studies comparing different pairs should be similar enough in effect modifiers for indirect comparisons to be meaningful.
• Consistency. Direct and indirect evidence should not conflict without explanation. Treat warnings as prompts for methodological review, not verdicts.
• Multi-arm studies. Keep every randomised arm of a trial in one study. Don't split a three-arm trial into two-arm studies: comparisons that share an arm are correlated.
• Reference treatment. Choose and document it before presenting relative effects or league tables.

Data the network needs

Only validated results enter a network, and a run is refused while any result it draws on is not validated. Draft results for one outcome never block another.

• Arms and denominators. Each arm records its treatment, dose, schedule, route, co-interventions and, for a control, what kind of comparator it is. The randomised N stays separate from each outcome's N; a count from another outcome is never borrowed silently.
• Arm numbers first. Events and totals, or means and SDs, are used where at least two arms report them; a comparison the paper reported is used only when they are missing. A missing SD is taken from the standard error or the confidence interval of the mean.
• Arms of one study are never combined unless you say they are the same intervention (“Combine these arms”, with a reason). A result the paper reports for several arms together is entered once as “Pooled with other arms”.
• Arms not reported. Record why, against the source; the outcome is analysed on the remaining arms and carries a selective-reporting note.
• One trial, several reports. Link them as Reports of this trial so the network counts the trial once. Unlinked reports that share a registry number block the run.
• Follow-up window. Each outcome uses one timepoint per study from its window. Without a window of your own it is read from the outcome's primary timepoint (“12 months” reads as 52 weeks, with 8 either side).
• Visual acuity. A change is converted to ETDRS letters only when the unit or outcome name says which scale it is on; otherwise it is left as extracted and flagged.
• Earlier extractions. Results saved before arm-level entry start with Adopt reviewed study arms in Extraction, which keeps the original values and quotes in history but doesn't validate anything.
• Independence. Arm summaries need independent participants, or at most one eye per participant; clustered or both-eye data need an adjusted analysis.

Treatment groups

By default an arm is grouped only with arms of the same intervention: the same treatment, dose, schedule, regimen, route and co-interventions. You can group by treatment, by treatment and dose, with biosimilars beside their reference product, ignoring co-interventions, or make your own grouping. The grouping in use is recorded with every result.

• Each arm reads Needs you, Automatic only, Confirmed, Approved or Rejected. The analysis runs only once every arm it uses is confirmed.
• Before a change is saved, its confirmation says what it would do to the network: groups, separate parts, studies left out and arms combined.
• An arm whose name is only partly recognised (“anti-VEGF + laser”) stays its own combination for you to review. A confirmation counts for the arm as confirmed: if its details are corrected later, it asks to be confirmed again.
• A group you exclude leaves every outcome, and each result says how many comparisons were left out and why.

Merging doses into one node is an assumption. See how the example merges three ranibizumab regimens.

Reading the results

• Effects on the natural scale, with 95% confidence intervals and, when estimable, prediction intervals. τ² and I² appear in the run summary.
• Zero events. For a ratio measure, 0.5 is added to each cell of a study with no events in one arm; a two-arm study with none in either is left out and listed.
• League table. Network estimates in the lower-left half, direct estimates in the upper-right; bold intervals exclude no difference. Above ten treatments, or in a narrow pane, it shows as a list.
• Ranking. P-score and SUCRA, ordered by the outcome's direction, with rank probabilities per treatment.
• Consistency. Node-splitting for every comparison with direct and indirect evidence, and the design-by-treatment test for the whole network. Without an independent evidence loop, inconsistency is reported as not assessable, which is not evidence of consistency.
• Network graph, report and downloads. Circle size follows participants, line width follows studies, and shaded areas are multi-arm studies. Each panel can be pinned to the report and downloaded as CSV, SVG or PNG.
• Drafts and locks. Each run is a draft, and is marked out of date when its evidence or settings change. Locking an outcome, or finalising its certainty review, fixes the groups, settings and data it used.

Read the example's league table and ranking.

Ranking is not certainty

A network that can't estimate between-study variance, such as one with a single trial or no evidence left over, needs an explicit common-effect choice: random effects are refused. With one independent trial, prediction intervals, rankings and rank probabilities are withheld.

Certainty review

Pairwise GRADE is not forced onto network estimates. Axelium prepares one row per network comparison for a CINeMA-style review; the final certainty is yours.

• Six domains, each rated no, some or major concerns; a concern needs a reason.
• Beside each rating, what the numbers say: never a rating in itself, and unavailable diagnostics are never read as reassuring.
• A suggested overall certainty drops one level for each “some concerns” and two for each “major concerns”. You can choose another level with a reason.
• Finalising asks you to confirm, records who finalised and when, locks the outcome when your role may lock analyses, and counts it towards the analysis freeze.

Reporting responsibly

Report the review question, eligibility criteria, treatment-node definitions, effect measure, reference treatment, network structure, model outputs, diagnostics, and certainty judgements together. If the evidence network is sparse, disconnected, or inconsistent, document those limitations rather than presenting a ranking as a stand-alone answer.