Network meta-analysis
What question does this help answer?
When several treatments exist for the same condition and few trials compare them head to head, how do they compare with each other, and how certain is each comparison?
When to use it
Use it within a systematic review when the trials of the candidate treatments form a connected network through common comparators, such as placebo or usual care, and the populations, outcomes and settings are similar enough that the comparisons can be combined.
When not to use it
Do not use it to rescue comparisons the evidence cannot support. If trials of one treatment were done in a different population, disease severity or era of background care than trials of another, the indirect comparison between them can be biased however sophisticated the model. A network with a treatment connected only through a single small trial gives that comparison very little support.
How it works in plain language
If treatment A has been compared with placebo and treatment B has also been compared with placebo, the two sets of trials give an indirect estimate of A against B. A network meta-analysis combines every such direct and indirect estimate across the network at once, so each comparison borrows strength from the others. That borrowing is only legitimate when the trials are similar enough, in everything that modifies the treatment effect, to be treated as parts of one body of evidence.
What a treatment network looks like
Key assumptions and design requirements
- Transitivity: the trials behind different comparisons are similar in the characteristics that modify treatment effects. This is judged by comparing study and participant characteristics, and it cannot be proven statistically.
- Consistency: where direct and indirect evidence for the same comparison both exist, they should agree. Disagreement can be examined only in closed loops of the network, and a lack of detected inconsistency is not proof of consistency.
- Heterogeneity within each comparison is assessed as it is in a pairwise meta-analysis, and assumptions about how it is shared across comparisons are stated.
- Risk of bias and certainty are judged for each comparison, including the indirect ones, rather than once for the network.
- Rankings, if reported, are presented with their uncertainty and never as the sole result.
Why Mallard does not run it
Mallard does not perform network meta-analyses or size systematic reviews. This page explains what such an analysis assumes so that a planned review can be specified, and a published one read, with those assumptions in view.
Clinical example
Several drugs for a chronic condition have each been compared with placebo, and two have been compared with each other. A review team plans a network meta-analysis, first tabulating the disease severity and background therapy in each trial to judge transitivity, prespecifying how inconsistency will be examined in the one closed loop, and planning to grade certainty for every pairwise comparison. The example illustrates planning choices only.
What to watch out for
A network meta-analysis is only as valid as its transitivity assumption: that the trials behind each comparison are similar enough in their effect modifiers that an indirect comparison stands in for a direct one. That assumption is judged from study characteristics, not proven by a statistical test, and a ranking of treatments presented without the uncertainty around it overstates how different they are.
Before using this in your workspace
Record the treatments and the network of comparisons expected, the effect modifiers used to judge transitivity, how inconsistency will be examined, the heterogeneity assumptions, how certainty will be graded for each comparison, and whether and how rankings will be reported.
- Indirect and mixed-treatment comparison, network, or multiple-treatments meta-analysis: many names, many benefits, many concerns for the next generation evidence synthesis tool (opens in a new tab) · Res Synth Methods · doi 10.1002/jrsm.1037
- Conceptual and technical challenges in network meta-analysis (opens in a new tab) · Ann Intern Med · doi 10.7326/0003-4819-159-2-201307160-00008
- The PRISMA extension statement for reporting of systematic reviews incorporating network meta-analyses of health care interventions: checklist and explanations (opens in a new tab) · Ann Intern Med · doi 10.7326/M14-2385