Network Meta-Analysis Explained: When and How to Compare Multiple Treatments

Standard pairwise meta-analysis compares two treatments at a time. But clinical decisions rarely involve just two options. When multiple interventions exist for the same condition, network meta-analysis (NMA) lets you compare all of them simultaneously — even when head-to-head trials do not exist for every pair.

This guide explains what NMA is, when to use it, and what the key concepts mean in practice.

What Is Network Meta-Analysis?

Network meta-analysis is a statistical extension of traditional meta-analysis that combines direct and indirect evidence to compare three or more interventions within a single coherent framework.

Direct evidence comes from trials that directly compare two treatments (A vs B). Indirect evidence is inferred through a common comparator — if trials compare A vs C and B vs C, NMA estimates the relative effect of A vs B through C.

The result is a network of evidence where every intervention is connected, and the analysis produces relative effect estimates for all possible pairwise comparisons, along with a ranking of treatments from most to least effective.

When to Use NMA Instead of Pairwise Meta-Analysis

NMA is appropriate when:

Multiple treatments exist for the same condition and you want to compare all of them
No single trial compares all options — indirect evidence is needed to fill gaps
Decision-makers need a ranking of available treatments
Clinical guidelines require comparative effectiveness across all relevant interventions

NMA is not appropriate when:

Only two interventions are being compared (use pairwise meta-analysis)
The interventions are too heterogeneous to form a meaningful network
The transitivity assumption cannot be met (see below)

Key Concepts

The Network Geometry

An NMA network is typically visualized as a graph where nodes represent interventions and edges represent direct comparisons from trials. The size of each node reflects the number of patients studied, and the thickness of each edge reflects the number of trials.

Well-connected networks (many direct comparisons, multiple loops) produce more reliable estimates than sparse networks with long chains of indirect evidence.

The Transitivity Assumption

Transitivity is the foundational assumption of NMA. It requires that the studies comparing different pairs of treatments are sufficiently similar in their clinical and methodological characteristics that the relative effects can be meaningfully combined.

In practical terms: if the patients, settings, outcomes, and follow-up periods differ substantially across comparisons, indirect evidence may not be valid. Before conducting NMA, you should verify that potential effect modifiers are reasonably balanced across comparisons.

Consistency

Consistency means that direct and indirect evidence agree. When a loop exists in the network (A-B, B-C, and A-C all have direct evidence), you can check whether the direct estimate of A vs C matches the indirect estimate derived from A-B and B-C.

Inconsistency can signal that the transitivity assumption is violated. Statistical tests (global and local) help identify and investigate inconsistency.

Treatment Rankings

NMA produces treatment rankings using measures such as:

P-scores (frequentist): a score from 0 to 1 indicating the probability that a treatment is better than competing treatments
SUCRA (Surface Under the Cumulative Ranking curve): the Bayesian equivalent of P-scores
Rankograms: bar charts showing the probability of each treatment occupying each rank

Rankings should always be interpreted alongside effect estimates and confidence intervals. A treatment ranked first with wide confidence intervals may not be meaningfully different from treatments ranked second or third.

League Tables

A league table is a matrix showing the relative effect estimate for every pair of interventions. It provides a complete summary of the NMA results and is often required by journals and guideline panels.

Statistical Methods

NMA can be conducted using either frequentist or Bayesian methods:

Frequentist NMA

The netmeta R package implements frequentist NMA using a graph-theoretical approach. It supports:

Fixed-effect and random-effects models
Binary, continuous, and generic outcome data
Network graphs, forest plots, and league tables
Consistency tests (global Q-statistic and net heat plots)
Treatment rankings (P-scores)
Comparison-adjusted funnel plots for small-study effects

Bayesian NMA

Bayesian approaches (using WinBUGS, OpenBUGS, or R packages like gemtc) allow incorporation of prior information and produce probability distributions rather than point estimates. They are more flexible for complex models but require more computational expertise.

Reporting NMA Results

The PRISMA-NMA extension provides guidance for reporting network meta-analyses. Key requirements include:

A network geometry plot
A description of how direct and indirect evidence were combined
Results of consistency/inconsistency assessments
A league table or equivalent summary of all pairwise comparisons
Treatment rankings with uncertainty measures
A flow diagram following PRISMA-NMA conventions

NMA-Specific Quality Assessment: CINeMA

For NMA, the standard GRADE framework is extended through CINeMA (Confidence in Network Meta-Analysis), which evaluates certainty across six domains:

Within-study bias
Reporting bias
Indirectness
Imprecision
Heterogeneity
Incoherence (replaces GRADE's "inconsistency" for network contexts)

CINeMA provides a structured way to rate confidence in each pairwise comparison from the NMA.

Conducting NMA with mapped

mapped supports network meta-analysis as a dedicated study type. When you create a project with the NMA study type:

The PRISMA-NMA flow diagram is generated automatically
The meta-analysis step uses netmeta for frequentist NMA
Network geometry plots, league tables, rankograms, and comparison-adjusted funnel plots are produced
CINeMA replaces standard GRADE for evidence quality assessment
Forest plots show results for all pairwise comparisons in the network

The entire pipeline — from literature search through NMA and CINeMA to manuscript — is integrated, so you do not need to switch between tools or export data to external statistical software.