What Is Network Meta-Analysis and How Does It Work?

Network meta-analysis is a statistical method that lets researchers compare multiple treatments against one another at the same time, even when some of those treatments have never been tested head to head in a trial. It extends the logic of traditional meta-analysis, which pools results from studies that all test the same pair of treatments, by linking chains of evidence through shared comparators. The technique has become a cornerstone of clinical guideline development and drug-reimbursement decisions worldwide, in large part because head-to-head trials of every possible treatment pair are rarely feasible or funded.

The Problem It Solves

Imagine three drugs for the same condition: A, B, and C. Several trials compared A against placebo, several compared B against placebo, and a couple compared A against B. But no trial ever compared B against C directly. A traditional meta-analysis can tell you how A stacks up against placebo, and how B stacks up against placebo, but it cannot formally say anything about B versus C. Clinicians, patients, and insurers still need to choose among all three, though, so someone has to make that comparison. Network meta-analysis does it in a principled way rather than leaving it to informal judgment.

The method works by treating every trial in a body of evidence as a node in a connected web. As long as the network of trials is linked through at least one common comparator, the technique can generate estimates for every possible treatment pair in the web, including pairs that were never directly compared in any study.¹ That is the core appeal: you get a complete picture of how treatments relate to each other, not just the fragments that individual trials happened to test.

How Indirect Comparisons Actually Work

The simplest form of indirect comparison, sometimes called the Bucher method, is straightforward in concept. Suppose trial evidence tells you the effect of drug B relative to drug A, and separately, the effect of drug A relative to drug C. You can combine those two pieces to estimate the effect of drug B relative to drug C. The math amounts to taking the difference between the two known treatment effects; the uncertainty around that indirect estimate is the combined uncertainty of both contributing comparisons.² Because you are stacking two layers of uncertainty, indirect estimates are generally less precise than direct ones. But they are far better than no comparison at all.

When the network grows beyond a simple triangle of three treatments, the analysis gets more complex but the principle remains the same. Each link in the network contributes information, and the statistical model borrows strength across links. If both direct and indirect evidence exist for the same treatment comparison, the two can be combined into a single “mixed” estimate that is more precise than either alone, provided they agree with each other. That blending of direct and indirect evidence is what gives network meta-analysis its power.

The Transitivity Assumption

Every indirect comparison rests on a crucial assumption: the studies being linked through a common comparator are similar enough that combining them makes sense. This is called transitivity. In practical terms, it means there should be no systematic differences in the types of patients, disease severity, doses, or other factors that modify how well a treatment works across the different sets of trials feeding into the network.³

Think of it this way. If the trials comparing A versus placebo enrolled mostly younger patients with mild disease, and the trials comparing B versus placebo enrolled mostly older patients with severe disease, then linking those results through placebo is comparing apples to oranges. The placebo group in one set of trials is not interchangeable with the placebo group in the other, so the indirect comparison between A and B would be misleading. Transitivity requires that the shared comparator genuinely functions as a common reference point.

This is the assumption that makes or breaks a network meta-analysis. Researchers evaluate it by examining the clinical and methodological characteristics of the included trials before running the analysis. If the trials that inform different parts of the network differ substantially in patient demographics, outcome definitions, follow-up length, or disease stage, the indirect comparisons may not be trustworthy.¹

Checking Whether Direct and Indirect Evidence Agree

If both direct trial evidence and indirect evidence exist for the same treatment pair, researchers can check whether they tell the same story. When they disagree, it is called inconsistency, and it is a red flag that something may have gone wrong with the transitivity assumption or with the underlying data.

Several statistical tools exist for spotting inconsistency. One graphical approach, the net heat plot, works by temporarily removing one study design at a time from the network and measuring how much the remaining evidence contributes to disagreement. The result is a color-coded matrix that highlights which parts of the network are causing trouble.⁴ Other methods add “inconsistency factors” to the statistical model and then test whether those factors improve the fit, essentially asking whether the data behave better when we allow certain comparisons to deviate from what the rest of the network predicts.⁵

The process is not always clean. Researchers have pointed out that some widely used methods for splitting evidence into direct and indirect components do not actually separate the two as cleanly as they should, which can lead to misleading conclusions about whether inconsistency exists.⁶ Improved approaches continue to be developed, but the broader point for anyone reading a network meta-analysis is this: the authors should have explicitly tested for inconsistency and reported the results. If they did not, treat the findings with extra caution.

Ranking Treatments

One of the most appealing outputs of a network meta-analysis is a ranking of all the treatments from best to worst. This is also one of its most misunderstood features. The rankings are probabilistic, not definitive. A treatment ranked first does not necessarily beat all the others by a meaningful margin; it may have landed there because its confidence interval happened to tilt slightly in its favor.

Within a Bayesian framework, rankings are derived from the posterior distributions of all treatment effects. For each treatment, you can calculate the probability that it is the best option, the second-best, the third-best, and so on. These rank probabilities are then summarized into a single number called SUCRA, which measures the estimated fraction of competing treatments that a given treatment beats. A SUCRA of 100 percent means the treatment almost certainly outperforms everything else in the network; a SUCRA near 50 percent means it is average; and a low SUCRA means it is likely among the worst.⁷ The frequentist equivalent of SUCRA, called P-score, produces nearly identical rankings without requiring the Bayesian machinery.⁸

The danger is that readers latch onto the rankings without looking at the underlying treatment effect estimates. Two treatments with very similar SUCRA scores may not differ from each other in any clinically meaningful way, and the ranking could flip with a single new trial. Rankings should always be read alongside the actual effect sizes and their uncertainty intervals.

Bayesian and Frequentist Approaches

Network meta-analyses can be run using either Bayesian or frequentist statistical frameworks. In practice, Bayesian models have been more common, partly because the early methodological papers were written in a Bayesian style and the software tools matured first. Bayesian approaches allow researchers to incorporate prior information and produce intuitive probability statements, but they require setting prior distributions and checking that the model’s sampling chains have converged properly, steps that demand statistical expertise and are sometimes glossed over in published analyses.⁹

Frequentist methods use graph-theory-based approaches and tend to be faster computationally. When researchers have compared the two head to head in simulation studies, the results are generally similar. Coverage tends to be somewhat higher with Bayesian methods, meaning the confidence intervals are more likely to capture the true value, but those intervals are also wider. Bias stays small under both approaches when heterogeneity is low and plenty of trials inform each comparison. The practical takeaway is that the choice of framework matters less than choosing an appropriate model for the data at hand.¹⁰

The Antidepressant Example

Perhaps the most widely cited network meta-analysis in any field is Cipriani and colleagues’ 2018 comparison of 21 antidepressant drugs for major depressive disorder in adults. The study synthesized data from over 500 trials and found that all 21 drugs were more effective than placebo, though the range was wide. Amitriptyline showed the strongest effect on symptom reduction, while reboxetine showed the weakest. The picture looked different for tolerability: agomelatine, citalopram, escitalopram, fluoxetine, sertraline, and vortioxetine had the lowest dropout rates, while amitriptyline, clomipramine, duloxetine, fluvoxamine, reboxetine, trazodone, and venlafaxine had the highest.¹¹

This study illustrates several strengths of the approach. No single trial could have compared 21 drugs simultaneously with enough statistical power to draw firm conclusions. The network format allowed every drug-to-drug comparison to be estimated, giving clinicians a comprehensive menu rather than isolated pairs of options. It also showed how rankings depend on the outcome you care about: a drug near the top for efficacy might be near the bottom for acceptability.

A companion network meta-analysis looked at antidepressants for children and adolescents and reached a strikingly different conclusion. Most drugs did not clearly outperform placebo or psychotherapy in that age group, and fluoxetine was the only pharmacological option with consistent support.¹² A later analysis that included both drugs and psychotherapies found that fluoxetine combined with cognitive behavioral therapy was more effective than either therapy alone in young people, though fluoxetine on its own also outperformed placebo.¹³ These analyses collectively shaped prescribing guidelines in multiple countries.

How Network Meta-Analysis Shapes Policy

Beyond clinical guidelines, network meta-analysis plays a direct role in decisions about which drugs a healthcare system will pay for. England’s National Institute for Health and Care Excellence, for instance, routinely evaluates manufacturer submissions that rely on network meta-analyses to compare a new treatment against existing options. When head-to-head trial evidence is unavailable, the manufacturer may use a network meta-analysis to make the case that their product is cost-effective relative to specific comparators already in use.¹⁴ Independent review groups then scrutinize the analysis, checking the network’s structure, the transitivity assumption, and inconsistency results. Similar processes exist in other countries with centralized health-technology assessment bodies.

Reporting standards have evolved to keep pace. The PRISMA-NMA extension provides a checklist specifically for systematic reviews that incorporate network meta-analyses, covering items like how the network geometry was described, how inconsistency was assessed, and how treatments were ranked.¹⁵ The checklist is not a guarantee of quality, but it gives readers a roadmap for evaluating whether the authors did their due diligence.

Rating the Confidence in Results

Just as with any evidence synthesis, the conclusions of a network meta-analysis are only as strong as the underlying trials and the methods used to combine them. Two major frameworks exist for rating how confident you should be in the results. One, developed by the GRADE Working Group, evaluates each comparison for risk of bias, imprecision, inconsistency, indirectness, and publication bias. The other, called CINeMA, automates much of this process using the network’s statistical output.

When researchers applied both frameworks to the same network meta-analysis of opioids for chronic pain, the two systems disagreed on the confidence rating roughly a third of the time for pain relief and about one in six times for physical functioning. All disagreements were off by one level, and the CINeMA framework tended to give lower confidence ratings, largely because it flagged transitivity and heterogeneity concerns more aggressively.¹⁶ For a reader encountering a network meta-analysis, this means the stated confidence level can depend partly on which rating system was used, not just on the data itself.

Individual Patient Data and Component Networks

Standard network meta-analyses work with summary-level data: the average treatment effect, the overall dropout rate, and similar aggregated numbers reported in each trial. A more resource-intensive version uses individual participant data, meaning the raw, patient-by-patient records from each trial. This allows researchers to examine how treatment effects vary by age, sex, disease severity, or other patient-level characteristics, something that aggregate data cannot do reliably because of the risk of aggregation bias.

Simulation studies have found that individual-patient-data network meta-analysis consistently outperforms the aggregate-data version in both accuracy and precision. In one comparison, the typical error was about half as large when individual data were used, and the estimates were roughly five times more precise. The advantage was especially pronounced in small or sparse networks, where every bit of extra information matters. In large, densely connected networks the benefit shrank, particularly when only a small proportion of trials contributed individual data.¹⁷ The practical barrier is obvious: getting multiple trial teams to share their raw data is a major logistical and political undertaking. But when it works, the payoff in analytical quality is substantial.¹⁸

Another extension, component network meta-analysis, addresses treatments that are themselves combinations of smaller parts. In psychotherapy research, for example, a “treatment” might be cognitive behavioral therapy plus a medication plus regular follow-up calls. Component network meta-analysis breaks these complex interventions into their individual ingredients and estimates the contribution of each component, including potential interactions between components.¹⁹ This approach can even handle disconnected networks, where some treatments have no common comparator at all, as long as the composite treatments in different subnetworks share common components.²⁰

Visualizing the Evidence

A network meta-analysis typically comes with a network plot: a diagram where each node represents a treatment and each line connecting two nodes represents direct trial evidence. Thicker lines usually mean more trials or more patients, and larger nodes mean more total participants received that treatment. These plots are generated using graph-theory algorithms that try to place nodes at distances reflecting the amount of evidence connecting them.²¹ Even at a glance, a network plot tells you which comparisons are well supported and where the network relies heavily on indirect evidence.

Newer visualization tools go further. The Vitruvian plot, for instance, displays multiple outcomes simultaneously so that clinicians and patients can see how a treatment performs across different dimensions, such as efficacy and side effects, alongside the confidence in each estimate.²² Visualization matters because raw numbers from a network meta-analysis can be overwhelming, with dozens of treatment pairs each carrying effect sizes, confidence intervals, and ranking scores. A well-designed plot compresses that information into something a decision-maker can actually use in a clinic visit.

Software for Running a Network Meta-Analysis

Most network meta-analyses today are run in R, the open-source statistical programming language. The two main packages mirror the two statistical frameworks. For Bayesian analysis, the gemtc package provides a comprehensive toolkit that links to external sampling software and handles binary, continuous, count, and survival outcomes. It can model heterogeneity and inconsistency, produce forest plots, and generate rankograms. For frequentist analysis, the netmeta package uses graph-theory methods, accepts contrast-level data for any outcome type, and includes the net heat plot for inconsistency detection.²³ A third package, pcnetmeta, offers a simpler Bayesian interface for binary outcomes but lacks inconsistency-testing features.²⁴

For researchers without programming experience, web-based tools and point-and-click interfaces have emerged, though the fully featured analyses still require coding. The relative accessibility of the software has helped network meta-analysis proliferate across fields, but it has also lowered the barrier for poorly executed analyses. Knowing which package was used does not tell you whether the model was specified correctly, the priors were sensible, or the convergence diagnostics were checked.

Use Beyond Medicine

Although clinical medicine is the field where network meta-analysis gained its footing, the method is not inherently medical. Any domain with a body of comparative studies that share common reference points can use it. Researchers have recently introduced the framework into agroecology, for example, to compare the effectiveness of different types of farmland margins, such as hedgerows, wildflower strips, and beetle banks, in conserving natural enemies of crop pests across grain crops like wheat, maize, soybean, and rice.²⁵ The logic is identical to the medical case: individual studies compared some margin types against conventional fields but not against each other, so the network fills in the gaps.

The technique also appears in education research, veterinary science, and environmental policy, anywhere that many interventions exist but head-to-head comparisons are incomplete. As the method spreads, the same assumptions apply: transitivity must hold, inconsistency must be checked, and rankings must be read alongside the precision of the underlying estimates. The audience changes, but the intellectual machinery does not.