Network model purpose
Network meta-analysis estimates relative effects among more than two interventions by combining direct and indirect evidence. A direct estimate comes from studies comparing two treatments. An indirect estimate compares treatments through one or more shared comparators.
In a simple A-B-C network, an indirect B versus C estimate can be derived from A-C and A-B contrasts, with orientation handled consistently. On an additive analysis scale, a consistency relationship is delta_BC = delta_AC - delta_AB when contrasts share reference A.
Contrast-based NMA
For contrast-based input, each comparison row supplies TE and seTE with treatment labels treat1 and treat2. The model uses netmeta with the selected summary measure, reference group, tau estimator, and common/random settings.
Ratio measures should be supplied on the log scale. Difference measures should be supplied on their natural additive scale. The app can display exponentiated estimates for OR, RR, and HR.
Binary arm-based NMA
For binary arm-level NMA, treatment arms supply events and total participants. Coventra builds pairwise comparisons from within-study arms and uses netmetabin for supported methods.
The binary NMA default method is LRP where available, with fallback behavior to MH when the required package or fit fails. MH, NCH, and IV sensitivity paths are available. The IV sensitivity path uses pairwise contrasts with a 0.5 continuity correction and includes all studies.
Double-zero comparisons are excluded from the primary binary NMA because they do not identify relative treatment effects.
Changeable NMA settings
- Reference treatment is user-changeable.
- Effect measure is user-changeable where the outcome type and NMA path support it.
- Tau estimator defaults to REML for inverse-variance contrast-based NMA and IV sensitivity paths.
- Users can request only selected modules, such as model, network graph, forest, rankings, league table, small-study effects, inconsistency, or confidence diagnostics.
- Ranking direction is user-changeable so lower values or higher values can be treated as better depending on outcome meaning.
NMA method-source profiles
Contrast-based NMA defaults to Recommended settings, which send REML explicitly while keeping the summary measure explicit. Native netmeta 3.5-0 defaults remain available as an advanced reproducibility option; package-default mode omits the tau-estimator argument and records the effective estimator returned by netmeta.
Sparse arm-level binary NMA does not offer package-default model selection. LRP, IV, MH, or NCH must be selected explicitly because these are substantively different analysis models.
Assumptions to verify
- Network connectivity: disconnected networks cannot estimate all comparisons.
- Treatment harmonization: clinically identical treatments should not be split accidentally, and clinically different interventions should not be merged without justification.
- Transitivity: effect modifiers should be sufficiently balanced across comparisons.
- Consistency/coherence: direct and indirect evidence should not show unexplained disagreement.
- Multi-arm studies: correlations among contrasts must be handled by the network model rather than double-counting arms in separate pairwise analyses.
Reporting checklist
- Report the network model, effect measure, reference treatment, tau estimator, and common/random setting.
- Report treatment grouping decisions.
- Report connectivity and direct evidence counts.
- Report transitivity assessment and effect modifiers considered.
- Report inconsistency diagnostics where loops exist, and state when they are not applicable.