Free scoping read. A written assessment of your manuscript in two business days, at no cost.

Start one
Contents

Academy · Research Data, Analytics & AI · NMA Statistical Models

Theoretical and Statistical Approaches in Network Meta-Analysis: Model Families and Inconsistency Testing

Network meta-analysis (NMA) can be fitted as a contrast-based model, which works entirely with relative treatment effects, or as an arm-based model, which works with the absolute outcome of each trial arm. Under the same real-world conditions the two approaches produce different estimates, in ways that remain actively debated in the methodological literature.

Pubrica AcademyWritten by a biostatistics methodologist Reviewed byan evidence synthesis statistician Published10 October 2026 Last checked10 October 2026 Reading time9 minutes Sources10 peer-reviewed

Underneath every network meta-analysis sits a specific statistical model, and the two broad theoretical families used to build one, contrast-based and arm-based, make different assumptions about what information a trial contributes and can produce different results even from identical data. A contrast-based model, the more common approach and the one Lu and Ades formalized in 2004, works entirely with within-trial relative effects (such as a log odds ratio between two arms of the same trial), preserving randomization, while an arm-based model works with each arm's absolute outcome directly, which some methodologists argue can introduce confounding that randomization was specifically designed to avoid. Most practicing analysts need not choose between the two explicitly, since the standard software tools default to a contrast-based implementation, but understanding that the choice exists, and why the field has settled where it has, helps an author recognize and justify the modeling decision when a reviewer or methodological co-author queries it.

Learning objectives

  • The difference between contrast-based and arm-based statistical models for network meta-analysis
  • How fixed-effect and random-effects specifications change what a network meta-analysis assumes about heterogeneity
  • What node-splitting and the design-by-treatment interaction model test statistically

Common misconceptions

  • “There is basically one accepted statistical model for network meta-analysis, and the only real choice is Bayesian versus frequentist estimation.” The choice between contrast-based and arm-based modeling is a separate, and still actively debated, theoretical decision that underlies both Bayesian and frequentist estimation: a contrast-based model preserves within-trial randomization by working only with relative effects, while an arm-based model works with each arm's absolute outcome and, critics argue, can reintroduce the kind of between-trial confounding randomization exists to prevent.

Contrast-based and arm-based models

A contrast-based model, the dominant approach in the field and the structure underlying Lu and Ades's foundational 2004 formulation, uses only the within-trial relative effect between arms of the same trial as its unit of data, which means randomization's protection against confounding is preserved exactly as it was in the original trial1. An arm-based model instead treats each trial arm's absolute outcome as a separate data point and estimates treatment effects by comparing across arms, including across arms from different trials, an approach that avoids some technical difficulties with multi-arm trials and can more directly accommodate trials with a missing arm, but that critics, including Lu and Ades themselves in later methodological work, argue can introduce between-trial confounding that a contrast-based model structurally cannot12. Arm-based proponents counter that their approach can more naturally handle a treatment arm that is missing an obvious comparator within a given trial, and that it integrates more smoothly with some prediction and extrapolation tasks, which is why the debate between the two camps has continued in the methodological literature rather than being resolved in favor of either side. The practical consequence is that the two model families are not simply two ways of computing the same answer; under certain network structures they can produce materially different estimates, which is why most major methodological guidance, including NICE's Decision Support Unit technical documents, recommends contrast-based models as the standard choice3. Dias and Ades's own 2016 methodological comparison of the two approaches examines specific scenarios in which an arm-based model's handling of baseline risk across trials can introduce exactly the kind of ecological confounding randomization is designed to prevent, which is the central technical argument behind the field's general preference, even though arm-based approaches retain legitimate uses in specific modeling situations involving missing arms or certain types of indirect prediction2.

Fixed-effect and random-effects specifications

Within either model family, authors must also choose whether to fit a fixed-effect specification, which assumes a single true effect size for each comparison across all contributing trials, or a random-effects specification, which assumes true effects vary across trials and estimates that variation as an additional heterogeneity parameter4. In network meta-analysis specifically, the random-effects heterogeneity parameter is typically assumed common across all comparisons in the network under a standard specification, an assumption that itself requires justification and that some extended models relax by allowing comparison-specific heterogeneity4. Higgins and colleagues note that this common-heterogeneity assumption is closely tied to the consistency assumption discussed elsewhere in network meta-analysis methodology, since both concern whether trials across the network can be treated as estimating a shared, comparable underlying quantity5. Higgins, Thompson, and Spiegelhalter's broader re-evaluation of random-effects meta-analysis cautions that the heterogeneity variance parameter itself is often poorly estimated when the number of studies is small, which applies with particular force inside a network meta-analysis, where some individual comparisons may be supported by only a handful of trials even when the network as a whole is large4.

Hypotheses tested by node-splitting

Node-splitting, introduced by Dias and colleagues in 2010, tests consistency for one specific comparison (one "node" in the network) at a time by splitting the evidence for that comparison into its direct component, from trials that tested it head-to-head, and its indirect component, inferred through the rest of the network, and then comparing the two estimates statistically6. A statistically significant difference between the direct and indirect estimate for that specific node indicates local inconsistency at that comparison, without necessarily implying the entire network is inconsistent6. Because node-splitting is run separately for each closed loop in a network, it can require many individual tests in a densely connected network, which introduces a multiple-comparisons consideration that authors must account for when interpreting which, if any, nodes show meaningful inconsistency7. Node-splitting also has a practical data requirement that limits where it can be applied: a loop can only be split if it contains at least one closed cycle of direct comparisons, so a star-shaped network, where every treatment connects only to a single central comparator with no direct trials between the outer treatments, has no loops available to split at all, and its consistency has to be assessed by other means, or assumed, with that assumption flagged explicitly. This is a structural limit of local inconsistency checking rather than a software shortcoming, and it is a further reason that a network’s geometry (the number and position of its loops) should be mapped before analysis begins, since it determines which consistency checks can be performed on the available evidence.

The design-by-treatment interaction model

Statistical models for network meta-analysis distinguish between a "design," the specific set of treatments compared within a given trial, and a "treatment contrast," the specific pairwise comparison a design provides evidence about; a three-arm trial comparing drugs X, Y, and Z is one design that supplies evidence for three separate contrasts at once, each of which may also be tested by other, two-arm designs elsewhere in the same network. Where node-splitting tests one comparison at a time, the design-by-treatment interaction model, developed by Higgins and colleagues, tests the network globally by partitioning the overall variation into components attributable to heterogeneity (variation within a design) and inconsistency (variation between designs testing overlapping sets of treatments), producing a single global test for whether the network as a whole shows evidence of inconsistency5. Veroniki and colleagues' review of inconsistency-detection methods notes that global and local approaches, like the design-by-treatment model and node-splitting respectively, can disagree about whether a given network shows meaningful inconsistency, since they test different statistical hypotheses, which is why current methodological guidance generally recommends reporting results from more than one approach rather than relying on either alone7. Rücker and Schwarzer's frequentist ranking method, developed within the same contrast-based modeling tradition, also provides a decomposition of the network's overall heterogeneity and inconsistency that can be computed without Markov chain Monte Carlo sampling, which has made the design-by-treatment approach practically accessible to teams working in a frequentist framework rather than requiring a Bayesian implementation specifically to run a global consistency check8. Taken together, node-splitting and the design-by-treatment interaction model provide complementary views of the same underlying question, namely whether the network’s direct and indirect evidence agree; reporting both, alongside the model specification choices discussed earlier in this article, enables a reader to judge the statistical foundation on which the headline results of a network meta-analysis rest. A methods section that specifies each of these choices does more than satisfy a reporting checklist: it gives the reader the information needed to judge independently how much weight the final ranked results deserve. Omitting any of these specification details, however minor it may appear to the analysts, transfers that judgment to a reader who has less information with which to make it and who cannot recover the missing detail from the published tables and figures.

Key points

Key points

Network meta-analysis rests on a layered set of statistical choices — contrast-based versus arm-based modeling, fixed versus random effects, and local versus global inconsistency testing — each of which can change the result under real-world data and each of which should be justified and reported explicitly rather than left as a software default.

Practical application

Practical application

  1. A contrast-based model should be the default unless a specific reason indicates otherwise. Preserving within-trial randomization is the primary theoretical advantage of contrast-based models, and most current methodological guidance recommends it as the standard choice.
  2. The heterogeneity specification should be justified explicitly. Authors should state whether a common heterogeneity parameter is assumed across the whole network, and whether that assumption is reasonable given the similarity of the comparisons in the network.
  3. Both node-splitting and the design-by-treatment interaction test should be performed. Results should be reported from both a local (node-splitting) and a global (design-by-treatment) inconsistency check, since the two test different hypotheses and can disagree.

Frequently asked questions

Arm-based models can simplify handling multi-arm trials and some missing-data scenarios, which is why they are still used in parts of the methodological literature, but most current guidance favors contrast-based models as the default for preserving within-trial randomization.

Not automatically. A fixed-effect model is appropriate only if it is reasonable to assume a single true effect across all contributing trials; where trials differ meaningfully in population or design, a random-effects specification is generally preferred, but it adds a heterogeneity parameter that itself needs justification.

A non-significant result means no statistically detectable inconsistency was found for that node with the available data; it does not exclude a true difference that is too small, or data that are too sparse, for the test to detect; this is a power limitation inherent to the test.

Heterogeneity describes variation among trials estimating the same comparison (within a design); inconsistency describes disagreement between direct and indirect evidence for the same comparison (between designs). The design-by-treatment interaction model statistically separates the two.

Yes. The PRISMA extension for network meta-analysis requires authors to specify their statistical model, including the contrast- or arm-based choice, fixed- or random-effects specification, and consistency-checking method used.

References

  1. Lu G, Ades AE. Combination of direct and indirect evidence in mixed treatment comparisons. Statistics in Medicine. 2004;23(20):3105-3124. read 10 October 2026. View on Wiley Online Library
  2. Dias S, Ades AE. Absolute or relative effects? Arm-based synthesis of trial data. Research Synthesis Methods. 2016;7(1):23-28. read 10 October 2026. View on Research Synthesis Methods
  3. Dias S, Welton NJ, Sutton AJ, Ades AE. NICE DSU Technical Support Document 2: A Generalised Linear Modelling Framework for Pairwise and Network Meta-Analysis of Randomised Controlled Trials. 2011 (last updated 2014). read 10 October 2026. View on NICE DSU (University of Sheffield)
  4. Higgins JPT, Thompson SG, Spiegelhalter DJ. A re-evaluation of random-effects meta-analysis. Journal of the Royal Statistical Society: Series A. 2009;172(1):137-159. read 10 October 2026. View on Journal of the Royal Statistical Society
  5. Higgins JPT, Jackson D, Barrett JK, Lu G, Ades AE, White IR. Consistency and inconsistency in network meta-analysis: concepts and models for multi-arm studies. Research Synthesis Methods. 2012;3(2):98-110. read 10 October 2026. View on Research Synthesis Methods
  6. Dias S, Welton NJ, Caldwell DM, Ades AE. Checking consistency in mixed treatment comparison meta-analysis. Statistics in Medicine. 2010;29(7-8):932-944. read 10 October 2026. View on Statistics in Medicine
  7. Veroniki AA, Vasiliadis HS, Higgins JPT, Salanti G. Evaluation of inconsistency in networks of interventions. International Journal of Epidemiology. 2013;42(1):332-345. read 10 October 2026. View on International Journal of Epidemiology
  8. Hutton B, Salanti G, Caldwell DM, et al. The PRISMA extension statement for reporting of systematic reviews incorporating network meta-analyses of health care interventions. Annals of Internal Medicine. 2015;162(11):777-784. read 10 October 2026. View on Annals of Internal Medicine
  9. Rücker G, Schwarzer G. Ranking treatments in frequentist network meta-analysis works without resampling methods. BMC Medical Research Methodology. 2015;15:58. read 10 October 2026. View on BMC Medical Research Methodology
  10. Jansen JP, Naci H. Is network meta-analysis as valid as standard pairwise meta-analysis? It all depends on the distribution of effect modifiers. BMC Medicine. 2013;11:159. read 10 October 2026. View on BMC Medicine

Every source above is a peer-reviewed article, a publisher’s record of one, or a standard issued by the body that owns it. Each carries the date it was last read. Where a standard moves, this article is revised rather than left up with a note.

How to cite this article

Pick your referencing style and copy it. Your university or target journal may vary the punctuation — check its guide if it has one.

Pubrica Academy. (2026, October 10). Theoretical and Statistical Approaches in Network Meta-Analysis: Model Families and Inconsistency Testing. Pubrica. https://pubrica.com/academy/concepts-definitions/what-are-the-theoretical-and-statistical-approaches-in-nma/

This article is written to be learned from. Cite it if you are referring to what it says; if you are using the evidence behind it, cite the sources listed above instead.

NMA Statistical Model Specification Checklist

Decide and document each of these model choices before fitting the network meta-analysis.

[ ] Chosen and justified contrast-based versus arm-based modeling
[ ] Chosen and justified fixed-effect versus random-effects specification
[ ] Checked whether a common-heterogeneity assumption across the network is reasonable
[ ] Run node-splitting on closed loops and noted any local inconsistency
[ ] Run the design-by-treatment interaction model for a global inconsistency check
[ ] Reported all model specification choices explicitly in the methods section

Pubrica AcademyWritten and maintained by Pubrica’s subject matter experts. Articles are revised when the underlying standards change, and each article states the date on which it was last reviewed.

How Pubrica can help

Pubrica’s biostatistics team specifies, fits and documents the full statistical model underlying a network meta-analysis, including model family, effects specification, and both local and global consistency checks, before results are reported in the manuscript.

Find something else

Search the Pubrica Academy

Access is free and requires no account. Enter a topic, or open the full library to filter by category, sub-topic or resource type.

Full specification and justification of the statistical model in network meta-analysis

A scoping assessment is free and takes two business days. It provides a written assessment of what the work requires and which specialist would undertake it.

Get a free quote

Regulations

Check your project against this rule

  1. 1Your work
  2. 2Your details
  3. 3Send
What do you need?

Material you send is held under signed confidentiality, with access limited to the assigned team and encrypted transfer. It is not used to train anything and is not passed to third-party tools. Ask for the agreement before you send anything you would rather cover first.