No data-driven criterion can rank candidate bridge families
Source note. This entry derives in part from a source document whose inline citations were ChatGPT interface tokens rather than references, and so resolve to nothing. Claims below are attributed to that document and were checked independently where possible; anything that could not be independently sourced is marked in the verification trail.
Statement
Several bridge families can supply a cross-gap contrast: baseline synthesis and reference prediction, aggregate-level matching, dose-response MBNMA, treatment-class effects, component NMA, external controls and cross-design synthesis, mechanistic bridges, and expert elicitation with external priors, plus at least one the source’s eight-family list omits, multivariate synthesis across treatment lines with target-trial emulation. Among bridge models that target the same estimand and induce the same observed-data distribution within every disconnected component, no statistic computed only from those data can identify the correct cross-gap restriction, so selection requires external scientific information, validation data, or transparent sensitivity analysis. The families are overlapping options rather than mutually incompatible answers: expert priors can overlay any other bridge, and class, dose-response, baseline, and external-evidence structures can coexist. Published selection guidance exists, notably Goring et al. 2016’s framework for choosing among disconnected-network models and Ruecker, Schmitz and Schwarzer 2021’s head-to-head comparison of two families on the same disconnected myeloma network, but no accepted general procedure ranks them.
Why it is open
When candidate bridges induce the same observed-data distribution inside every component, they differ only across a gap that contains no evidence, so nothing computable from the dataset separates them. Some bridges do impose within-component restrictions that can be checked, and external or bridge-deletion data can discriminate among them, so the untestability is narrower than a blanket claim, but that leverage disappears for exactly the restriction that carries the gap. What remains missing is a general procedure that takes a network and a decision question and says which bridges it can support: the linking devices sit in scattered literatures, the published guidance is qualitative or case-specific, and the choice determines the estimand as well as the estimate.
What has been tried
| Work | What it contributes |
|---|---|
| Beliveau et al. 2017 | examines random baseline treatment effects in disconnected networks and finds their use appropriate for the two datasets considered, while cautioning that normality and exchangeability could be inappropriate in datasets it did not observe |
| Thom et al. 2022 | extends baseline-effect and aggregate-level matching methods to disconnected RCTs and single-arm studies and compares them by simulation |
| Petropoulou et al. 2023 | simulates additive CNMA and recommends it for disconnected networks only with strong clinical arguments for additivity |
| Perren et al. 2026 | provides a practical class-effects guide and four-step model-selection algorithm, including the case of networks disconnected at the treatment level but connected at the class level |
| netmeta (discomb) | provides discomb, additive network meta-analysis for combinations of treatments in disconnected networks, in the frequentist netmeta framework developed at github.com/guido-s/netmeta |
| Goring et al. 2016 | presents a theoretical framework to guide the choice between modeling approaches for disconnected networks, which is the selection guidance the source said did not exist |
| Ruecker, Schmitz and Schwarzer 2021 | compares a component NMA bridge against a matching bridge head-to-head on the same real disconnected network |
| Bridging disconnected networks of first and second lines of biologic therapies in rheumatoid arthritis with registry data (2022) | a bridge device outside the source’s eight-family taxonomy: Bayesian synthesis across treatment lines using registry data with target-trial emulation |
Probable solution or research direction
Build on the existing qualitative guidance rather than restarting: turn Goring et al. 2016’s framework and the published case comparisons into a documented selection procedure that names, for each candidate family, the parameter carrying the gap and the within-component restrictions it does imply and that can therefore be checked. Report the cross-gap contrast under several admissible bridges with partial-identification bounds across them rather than committing to one, noting that discomb in netmeta already supplies a frequentist component bridge for disconnected combination networks. Where the answer stays fragile under every admissible bridge, design a bridging study.
Verification
Verdict. Overstated. 2 independent auditors judged the claim as stated too strong. No criterion computable from a disconnected dataset can rank candidate bridge families, because the gap contains no evidence against which they could be scored; qualitative selection frameworks (Goring et al. 2016) and case-study comparisons of two families (Ruecker et al. 2021) exist, but there is no accepted general procedure. Disconnected networks are a recurring and growing problem in oncology and rare disease HTA, though no study establishes that they are the majority of unmet evidence-synthesis problems.
The source titled this. “No principled procedure for choosing among bridge families” The title says no principled procedure exists, but the audit found a published selection framework and a published head-to-head comparison of two bridge families.
The source said. “At least eight bridge families can supply a cross-gap contrast: random baseline effects and baseline synthesis, reference prediction and aggregate-level matching, dose-response MBNMA, treatment-class effects, component NMA, external controls and cross-design synthesis, mechanistic or pharmacometric bridges, and expert elicitation with external priors. Which family is defensible for a given problem is unresolved, and in practice disconnected evidence is linked by whichever workaround is easiest to implement.”
What was wrong with it. The source said there is no accepted criterion for choosing among bridge families and that most unmet HTA evidence-synthesis problems now sit in disconnected networks. The literature auditor found a published selection framework (doi:10.1002/jrsm.1204) and a published head-to-head comparison (doi:10.1002/bimj.201900339), and found no bibliometric study supporting the prevalence claim, which is dropped; GPT-5.6 Sol added that the untestability holds only among bridges that are observationally equivalent within the components, and that the families overlap rather than each answering a different question. The prior_work gloss on Beliveau et al. 2017 is also corrected here, for the reason recorded under DIS-03.
Revised by the full-text reading. Applied 2026-07-26 full-text reading of the consolidated library (687 papers).
- closure-upheld
- 10 findings from the reading assert this problem is open; the closure was re-examined and stands Put to an independent reviewer with the closing auditors’ reasoning, the work they relied on, and the new findings in view together. The new literature directly restates the closure’s residual: fit and output comparisons do not identify the correct cross-gap restriction, and selection remains deliberative, assumption-based, or sensitivity-based. The 2021 classification review and the 2023 migraine case study reinforce the absence of an accepted general selector, but do not supply one or show that the closure’s impossibility condition is wrong. Additional bridge families based on multi-outcome borrowing widen the taxonomy but do not alter the statement’s core claim, which is explicitly conditional on candidates inducing the same observed-data distribution within disconnected components. Neither the qualitative frameworks nor descriptive head-to-head comparisons contradict that conditional identification argument. (reviewer confidence: high)
- Flagged for reopening
-
10 findings from the reading assert this problem is open while the registry carries
overstatedFlagged, not applied: reversing an earlier audit’s closure needs the closure’s own evidence re-examined alongside this, which the reading did not do. See documentation/audit/reading/PROPOSED.md. - Literature and prior-art check
- open / overstated. I verified the netmeta claim: discomb’s CRAN documentation reads ‘Additive network meta-analysis for combinations of treatments (disconnected networks)’ and says it is particularly intended for disconnected networks. The core claim, that no data-driven criterion can rank bridges when the gap contains no evidence, is sound. But two things are overstated. First, ‘there is currently no accepted criterion for choosing among bridge families’ ignores Goring et al. 2016, which explicitly ‘present[s] a theoretical framework to guide the choice between modeling approaches’, and ignores Ruecker, Schmitz and Schwarzer 2021, who compare two bridge families head-to-head on the same disconnected myeloma network. Second, and more seriously, ‘Most unmet HTA evidence-synthesis problems now sit in disconnected networks, especially in oncology and rare disease’ is a bald prevalence claim with no source; I found no bibliometric study establishing it, and publishing it under a named byline is a liability. The eight-family taxonomy is a claim about the source report, not about the literature, and cannot be audited. Cites 10.1002/jrsm.1204: Partially narrows the ‘no criterion for choosing’ claim: supplies a published framework for choosing among disconnected-network models. Cites 10.1002/bimj.201900339: A published head-to-head comparison of two bridge families on the same disconnected network, absent from prior_work. Cites 10.1016/j.jclinepi.2022.06.011: A ninth bridge device in use (multivariate/bivariate synthesis across treatment lines with target-trial emulation) not in the entry’s eight-family taxonomy. Cites cran:netmeta/discomb: Confirms the discomb claim as written.
- GPT-5.6 Sol, technical and source-code lens
- open / overstated. The eight-family list in C2 is a reasonable descriptive taxonomy, but the families are neither mutually exclusive nor guaranteed to target different estimands. Expert priors can overlay any other bridge, while class, dose-response, baseline, and external-evidence structures can coexist. C3-C4 are therefore too strong: different bridges impose different restrictions and vulnerability patterns, but do not necessarily import bias or answer different questions. C6 holds only when candidate bridges are observationally equivalent on the observed components. Some models also impose within-component restrictions that can be checked, and external validation or bridge-deletion data can discriminate among them. C11 is source-confirmed locally: netmeta exports and implements discomb for disconnected combination networks. C12 and C14 are sound. C13 is speculative because ensemble weights cannot be learned from an evidence-free gap, and averaging is incoherent if estimands differ. C1, C5, C7, C9, and C10 are sourcing, practice, acceptance, prevalence, or universal-absence claims outside this technical audit. Bridge choice remains decision-critical, but a universal data-driven selector is not an attainable solution.
- Decision path
-
R5:default-open → R8:overstated-majority(n=2)
How the evidence falls in time
10 papers in the reviewed corpus bear on this problem, published 2014 to 2025. Every paper found here asserts the problem is open; none reports progress.
This problem is recurrent: three or more papers spanning at least eight years assert it independently. Sustained restatement by authors who mostly do not cite each other is the strongest evidence this reading can offer that a gap is real rather than one group’s framing.
| Year | Says | Paper |
|---|---|---|
| 2014 | confirms it is open | Network meta-analysis of multiple outcome measures accounting for borrowing of information across outcomes |
| 2018 | confirms it is open | The use of single armed observational data to closing the gap in otherwise disconnected evidence networks: a network meta-analysis in multiple myeloma |
| 2019 | confirms it is open | Joint Estimation of Remission and Response for Methotrexate-Based DMARD Options in Rheumatoid Arthritis: A Bivariate Network Meta-Analysis |
| 2020 | confirms it is open | Multivariate network meta-analysis incorporating class effects |
| 2021 | confirms it is open | Component network meta-analysis compared to a matching method in a disconnected network: A case study |
| 2021 | confirms it is open | Classifying information-sharing methods |
| 2021 | confirms it is open | Joining the Dots: Linking Disconnected Networks of Evidence Using Dose-Response Model-Based Network Meta-Analysis |
| 2022 | confirms it is open | Bayesian network meta-analysis methods for combining individual participant data and aggregate data from single arm trials and randomised controlled trials |
| 2023 | confirms it is open | Comparison of indirect treatment methods in migraine prevention to address differences in mode of administration |
| 2025 | confirms it is open | Methods for information-sharing in network meta-analysis: Implications for inference and policy |
Source
Derived from A:L49, A:L108, A:L218, A:L223, C:L52, C:L214-215, C:L236, C:L1088-1152 of the reviewed corpus.