Network Meta-Analysis Overview

Overview

Network meta-analysis, NMA, compares several treatments simultaneously.

It combines:

• direct evidence from head-to-head trials;
• indirect evidence through common comparators.

For example, A versus C can be informed by A versus B and B versus C.

Why this matters

Indirect comparison is valid only when treatment comparisons are sufficiently exchangeable.

Differences in effect modifiers can violate transitivity.

Treatment rankings can appear precise even when the evidence is sparse or biased.

When to use NMA

Use NMA when:

• more than two relevant treatments exist;
• a connected trial network is available;
• effect measures are compatible;
• transitivity is plausible;
• trial and treatment definitions are coherent.

A disconnected network cannot provide all relative comparisons without additional assumptions.

Research question and variables

If treatment B is the common comparator:

$$\theta_{AC} = \theta_{AB} + \theta_{BC}$$

on an additive effect scale.

Consistency means direct and indirect evidence agree within sampling variation.

Data structure

Network meta-analysis inputs

simulated data

Variable	Meaning
Study	Trial identifier
Arm	Treatment
Outcome	Arm-level or contrast-level result
Effect modifiers	Baseline risk, dose, follow-up
Multi-arm indicator	Shared trial correlation

Statistical assumptions

Transitivity

Transitivity is a clinical and design assumption.

Ask whether a patient in an A-versus-B trial could reasonably have been randomised to C in another trial.

Important modifiers include:

• disease severity;
• previous treatment;
• dose;
• outcome timing;
• baseline risk;
• co-intervention.

Consistency

Consistency can be examined using:

• node splitting;
• design-by-treatment interaction;
• inconsistency models;
• comparison of direct and indirect estimates.

A non-significant inconsistency test does not prove consistency.

Worked example

Worked analysis in R

network_data <- read.csv(
  "data/part-16-network-meta-analysis-data.csv"
)

network_data

# Contrast-based illustration with netmeta:
# library(netmeta)
# net <- pairwise(
#   treat = treatment,
#   event = event,
#   n = n,
#   studlab = study,
#   data = arm_data,
#   sm = "RR"
# )
# fit <- netmeta(
#   TE,
#   seTE,
#   treat1,
#   treat2,
#   studlab,
#   data = net,
#   random = TRUE
# )
# netgraph(fit)
# netsplit(fit)

Selected analysis output

The simulated four-treatment network is connected.
The A–B comparison contains the largest number of direct studies.

Visualising the treatment network

Rankings

Rank probabilities, SUCRA and P-scores summarise relative ordering.

They can be misleading when:

• effects are clinically similar;
• uncertainty is large;
• outcomes differ in importance;
• safety is omitted;
• evidence quality varies.

Report relative effects and confidence or credible intervals before rankings.

Multi-arm trials

Multi-arm trials contribute correlated comparisons.

Treating each contrast as independent double-counts shared groups.

NMA software must model the covariance.

Interpreting the output

Presenting the analysis

Suggested reporting

Use PRISMA-NMA and assess certainty for each comparison [1], [2].

Report network geometry, transitivity, inconsistency, heterogeneity, rankings and risk of bias.

The netmeta implementation and graphical tools for evaluating network geometry provide practical extensions [3], [4].

Common mistakes

Knowledge check

What is transitivity?

Answer and reasoning

The assumption that indirect comparisons are valid because relevant effect modifiers are sufficiently comparable across trial contrasts.

What is consistency?

Answer and reasoning

Agreement between direct and indirect evidence for the same treatment contrast.

Why can rankings mislead?

Answer and reasoning

They may conceal small effect differences, uncertainty and evidence quality.

What is needed for all pairwise comparisons?

Answer and reasoning

A connected network and valid transitivity assumptions.

Why model multi-arm correlation?

Answer and reasoning

Contrasts sharing one trial arm are statistically dependent.

Key take-home messages

References

1. Hutton B, Salanti G, Caldwell DM, Chaimani A, Schmid CH, Cameron C, et al. PRISMA extension for network meta-analysis. Ann Intern Med. 2015;162(11):777–784. https://doi.org/10.7326/M14-2385
2. Salanti G. Indirect and mixed-treatment comparison, network, or multiple-treatments meta-analysis. Res Synth Methods. 2012;3(2):80–97. https://doi.org/10.1002/jrsm.1054
3. Rücker G, Krahn U, König J, Efthimiou O, Davies A, Papakonstantinou T, et al. netmeta: network meta-analysis using frequentist methods. R package.
4. Chaimani A, Higgins JPT, Mavridis D, Spyridonos P, Salanti G. Graphical tools for network meta-analysis. PLoS One. 2013;8(10):e76654. https://doi.org/10.1371/journal.pone.0076654