Network Meta-Analysis
The Network Meta-Analysis (NMA) section compares three or more treatments simultaneously, pooling direct evidence (head-to-head studies) with indirect evidence (treatments connected through a common comparator) to estimate every pairwise comparison in the network -- even between treatments never compared head-to-head.
The workflow
| Step | What you do |
|---|---|
| 1. Enter the studies | For each study, supply the two treatments compared and the effect data |
| 2. Configure the analysis | Select effect measure and analysis options |
| 3. Run the analysis | Engine extracts direct comparisons, derives indirect comparisons, combines into network estimates |
The section provides a guided Study Entry Form, Column Mapper for imports, Network Results with rankings and league table, Results Interpretation view, and Methodology Section.
The two kinds of evidence
| Estimate | What it represents |
|---|---|
| Direct | Pooled effect from studies that directly compared A vs B |
| Indirect | Effect inferred via a connecting node (e.g. A vs C + C vs B) using transitivity |
| Network | Combined estimate using both direct and indirect evidence, weighted by precision |
The network estimate is typically reported; direct and indirect components are inspected for consistency.
Statistical methods
Direct comparisons
For each pair with head-to-head studies, standard pairwise pooling produces effect size, SE, 95% CI, z-statistic, and p-value.
Indirect comparisons
Under the transitivity assumption:
For path A -> C -> B:
Indirect_effect(A,B) = Direct_effect(A,C) + Direct_effect(C,B)
Indirect_SE(A,B) = sqrt(SE(A,C)^2 + SE(C,B)^2)
CI = Indirect_effect +/- 1.96 * Indirect_SE
z = Indirect_effect / Indirect_SE
When several connecting paths exist, indirect evidence from each path can be combined.
The transitivity assumption is the core requirement: participants and study characteristics in A-vs-C studies must be comparable to those in C-vs-B studies. If violated, the indirect estimate is unreliable.
Combined network estimates
When both direct and indirect evidence exist, they are combined by precision-weighted pooling -- typically dominated by whichever source has smaller SE.
Heterogeneity
A network-level heterogeneity result summarises between-study variability across the analysis.
Method options
The interface offers analysis-method options: traditional, frequentist, and bayesian.
Treatment rankings
| Quantity | Meaning |
|---|---|
| SUCRA | 0-100 score; 100 = best treatment in network, 0 = worst |
| Mean rank | Average rank position across ranking distribution |
| Ranking probabilities | Probability of being in each rank position |
Higher SUCRA indicates a treatment more likely to be among the best. SUCRA is sensitive to uncertainty -- a slightly better point estimate with wide CI may score lower than a marginally worse estimate with tight CI. Read ranking probabilities alongside SUCRA for a complete picture.
The league table
Displays the network estimate for every pair of treatments in a single matrix. Each cell shows the comparative effect of row treatment vs column treatment with 95% CI. The canonical NMA summary table.
Network geometry
The network plot visualises evidence structure:
- Nodes -- treatments in the network
- Edges -- head-to-head comparisons (thickness/size reflects number of studies/participants)
Use to inspect whether the network is connected (prerequisite for NMA) and whether it is dense or sparse.
Specialised tools
| Tool | Use |
|---|---|
| Dose-response calculations | Modelling effect across multiple dose levels |
| Economic calculations | Cost-effectiveness summaries (e.g. ICER-relevant) |
| Single-arm conversion | Incorporating single-arm data where appropriate |
Outputs
| Output | Contents |
|---|---|
| Direct comparisons | Per-pair pooled effect, SE, 95% CI, z, p |
| Indirect comparisons | Per-pair indirect effect, SE, 95% CI, z, p |
| Network estimates | Combined estimate for every pair |
| Treatment rankings | SUCRA, mean rank, ranking probabilities |
| League table | All pairwise comparisons in matrix form |
| Network plot | Visual network geometry |
| Heterogeneity | Network-level result |
| Methodology Section | Auto-generated methods description |
| Results Interpretation | Plain-language summary |
| Export | Results for reporting |
When NMA is the right tool
| Situation | NMA appropriate? |
|---|---|
| Two treatments, head-to-head only | No -- use pairwise meta-analysis |
| Three+ treatments, connected network | Yes |
| Three+ treatments, disconnected network | No -- cannot estimate across the gap |
| New treatment compared only to placebo, want ranking vs active comparators | Yes, if active comparators connect to placebo |
The most important precondition: the network must be connected and transitivity must be plausible.