No criterion says when ML-NMR should be preferred to MAIC or STC
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
ML-NMR combines IPD and aggregate data in one model and avoids the aggregation bias of aggregate-data meta-regression (doi:10.1111/rssa.12579), now extends to general likelihoods including time-to-event outcomes (doi:10.1093/jrsssa/qnaf169), and allows key assumptions to be assessed but not eliminated, as the 2023 psoriasis application showed (PMID 35997006), while practice remains dominated by MAIC and STC (doi:10.1016/j.jclinepi.2023.09.004). The auditors split on the recommendation that follows. The literature auditor verified every checkable claim and read the adoption gap as open; GPT-5.6 Sol held that partial IPD plus suspected effect-modifier imbalance is not sufficient to declare ML-NMR the default, because method choice also depends on the estimand, network structure, target-distribution information, overlap and model credibility, and that the absence of a designated default is not itself an unresolved methodological problem. The premise that ML-NMR is harder to specify, fit, diagnose and explain is plausible and widely repeated but unmeasured; neither auditor found a study of it.
Why it is open
ML-NMR is harder to specify, fit, diagnose and explain than MAIC or STC, so the barrier to adoption is practical rather than theoretical and default behavior has not shifted. Conditional constancy, shared effect-modifier assumptions, overlap and target-population relevance remain only partially testable inside ML-NMR, so switching buys coherence rather than assumption relief and the case for the extra effort is not self-evident to analysts or reviewers. On the weaker reading the missing thing is not a default but the material that would let an analyst make the comparison at all: diagnostics, training and reporting templates that state what has and has not been relieved.
What has been tried
| Work | What it contributes |
|---|---|
| dmphillippo/multinma | R package implementing Bayesian NMA, NMR and ML-NMR with IPD plus aggregate data, time-to-event support, integration checks and target-population prediction. |
| Phillippo et al. 2020, Multilevel network meta-regression for population-adjusted treatment comparisons, JRSS-A (PMID 32684669) | Defines ML-NMR: an individual-level regression model whose aggregate data are fitted by integrating over the covariate distribution, avoiding the aggregation bias of standard meta-regression. |
| Phillippo, Dias, Ades & Welton 2025, JRSS-A (preprint arXiv:2401.12640) | Removes the requirement that the aggregate-level likelihood have a known closed form, which had prevented application to time-to-event outcomes. This is the general-likelihood and survival support the entry claims. |
| Phillippo et al. 2023, psoriasis application, Medical Decision Making 43(1):53-67 (PMID 35997006) | Demonstrates for the first time how ML-NMR allows key assumptions to be assessed, finding little evidence that conditional constancy or shared effect-modifier assumptions were invalid. Assumptions were assessed, not eliminated. |
| Serret-Larmande et al. 2023, Journal of Clinical Epidemiology (PMID 37717707) | Methodological review of published PAIC analyses, overwhelmingly MAIC and STC, which is the evidence that practice has not moved to ML-NMR. |
| NICE quizartinib appraisal (PMID 39627603) | Cited by the literature auditor as one of the isolated applied appraisals using ML-NMR, consistent with adoption being sparse rather than absent. |
Probable solution or research direction
Build the operational layer rather than legislating a default: standardized integration and fit diagnostics, worked training material, and reporting templates that state which assumptions ML-NMR has and has not relieved. Set out explicit criteria for when ML-NMR is preferable and when it is infeasible, expressed in terms of estimand, network structure, available target-distribution information, overlap and model credibility, so the method choice is argued rather than asserted. Measuring the specification, fitting and communication burden against MAIC and STC would also convert the entry’s central unsourced premise into evidence.
Our study
Question. In a star network with individual data in one trial, do quantities an analyst can compute before fitting (the transport gap of the aggregate contrasts, overlap, how much of the target law is known) mark where ML-NMR has lower error than MAIC and STC, does that map survive a violated shared effect-modifier assumption, and what fitting burden does multinma impose on the same datasets?
Design. Star network of one individual-data and two aggregate trials, binary outcome, external target; transport gap, overlap, target information and shared modification crossed, plus a null cell; MAIC, STC and ML-NMR by paired squared error, multinma fitted on a subsample for burden. 17 cells, 1000 replicates each, 10 multinma fits per cell.
The protocol is registered before any result is seen: read it.
Verification
Verdict. Partially addressed. Auditors agreed the problem is open and no counterevidence surfaced. Reopened by the full-text reading: 6 findings from 6 papers survived both the reader and an independent second reviewer, so the verdict moves from confirmed-open to partially-addressed.
The source titled this. “ML-NMR is not the default despite being the most coherent framework” The title declares ML-NMR the most coherent framework and presumptive default, which the audit rejected since method choice also depends on estimand, network structure and overlap.
The source said. “The report recommends making ML-NMR the default upgrade path when IPD exist for a subset of trials in a connected network and clinically meaningful effect-modifier imbalance is suspected, because it is the most coherent framework currently available and now has survival and general-likelihood support. Practice has not moved, and the report is explicit that ML-NMR does not solve the underlying scientific problem.”
What was wrong with it. The source made a default-method recommendation with nothing cited. The literature auditor sourced each factual component (doi:10.1111/rssa.12579, doi:10.1093/jrsssa/qnaf169, PMID 35997006, doi:10.1016/j.jclinepi.2023.09.004), so the claim is now sourced, and voted the entry open; GPT-5.6 Sol voted it not a problem, because no universal default follows from partial IPD plus suspected imbalance. Both readings are kept above.
Revised by the full-text reading. Applied 2026-07-26 full-text reading of the consolidated library (687 papers).
- Verdict changed
- confirmed-open -> partially-addressed 6 surviving findings from: Anchors Away: Navigating Unanchored Indirect Comparisons wit; Assessing the performance of population adjustment methods f; Methods for Indirect Treatment Comparison: Results from a Sy; Network meta-interpolation: Effect modification adjustment i; Reframing Population-Adjusted Indirect Comparisons as a Tran; Response to discussant comments on “NMA, the first 20 years”. New citations: 10.1002/jrsm.1608, 10.1002/jrsm.1745, 10.1002/sim.8759, 10.3390/jmahp12020006
- Literature and prior-art check
- open / supported. Every checkable claim verified against retrieved sources. ‘Extends to general likelihoods including survival’: Phillippo, Dias, Ades & Welton, JRSS-A 2025 (doi 10.1093/jrsssa/qnaf169, arXiv:2401.12640) states that ML-NMR ‘currently requires the aggregate-level likelihood to have a known closed form. Notably, this prevents application to time-to-event outcomes’ and then removes that restriction. ‘Integrates IPD and AgD and avoids aggregation bias’: Phillippo et al. 2020 JRSS-A and the 2023 Med Decis Making psoriasis paper, which says ML-NMR ‘coherently synthesizes evidence from IPD and AgD studies in networks of any size while avoiding aggregation bias and noncollapsibility bias’. The ‘2023 psoriasis application’ claim is exactly right: Phillippo et al., Med Decis Making 2023;43(1):53-67 (PMID 35997006) demonstrates ‘for the first time how ML-NMR allows key assumptions to be assessed’ and found ‘little evidence that the conditional constancy or shared effect modifier assumptions were invalid’ - assessed, not eliminated. ‘Practice has not moved’ is consistent with Serret-Larmande et al. 2023 (288 PAICs, overwhelmingly MAIC/STC) and with only isolated NICE applications such as the quizartinib appraisal (PMID 39627603). The one unsourced claim is that ML-NMR is harder to specify, fit, diagnose and explain; plausible and widely echoed, but I found no study measuring it.
- GPT-5.6 Sol, technical and source-code lens
- not-a-problem / overstated. C3 through C5 are broadly correct with important conditions. ML-NMR combines IPD and AgD coherently and can avoid conventional aggregation and non-collapsibility errors by integrating an individual-level model, but only when the outcome model, modifier structure, covariate distributions and numerical integration are adequate. It does not remove overlap, transportability or shared-modifier assumptions. C2 and C6 are comparative judgments rather than statistical results, and C1 is not a substantive claim. The presence of some IPD plus plausible modifier imbalance is insufficient to declare ML-NMR the default: method choice also depends on the estimand, network structure, target-distribution information, overlap and model credibility. C7 is a paper-specific claim outside this technical and cpaic-source audit. Training and implementation guidance may be useful, but failure to designate a universally preferred default is not an unresolved methodological problem, and the asserted very-high priority is inflated.
- Decision path
-
R5:default-open → R9:substantiated-single-overstatement:codex
How the evidence falls in time
15 papers in the reviewed corpus bear on this problem, published 2016 to 2026. Progress and a fresh assertion of openness land in the same year.
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.
It is also being closed in pieces: partial results come from more than one paper in more than one year, each covering a different part.
| Year | Says | Paper |
|---|---|---|
| 2016 | confirms it is open | A scoping review of indirect comparison methods and applications using individual patient data |
| 2020 | partly addresses it | Assessing the performance of population adjustment methods for anchored indirect comparisons: A simulation study |
| 2022 | confirms it is open | A Response to: Letter to the Editor Regarding “Network Meta-analysis of Food and Drug Administration-approved Treatment Options for Adults with Aquaporin-4 Immunoglobulin G-positive Neuromyelitis Optica Spectrum Disorder” |
| 2023 | partly addresses it | Network meta-interpolation: Effect modification adjustment in network meta-analysis using subgroup analyses (Harari, Soltanifar, Cappelleri, Verhoek, Ouwens, Daly, Heeg; Research Synthesis Methods 2023;14(2):211-233) |
| 2023 | confirms it is open | Mapping the characteristics, concepts and methodologies of matching-adjusted indirect comparison studies assessing pharmacological therapies in oncology: a scoping review protocol |
| 2023 | confirms it is open | The Acceptance of Indirect Treatment Comparison Methods in Oncology by Health Technology Assessment Agencies in England, France, Germany, Italy, and Spain |
| 2024 | partly addresses it | Methods for Indirect Treatment Comparison: Results from a Systematic Literature Review |
| 2024 | partly addresses it | Response to discussant comments on “NMA, the first 20 years” |
| 2024 | confirms it is open | Advancing unanchored simulated treatment comparisons: A novel implementation and simulation study |
| 2024 | confirms it is open | Twenty years of network meta-analysis: Continuing controversies and recent developments |
| 2025 | confirms it is open | Unanchored simulated treatment comparison on survival outcomes using parametric and Royston-Parmar models with application to lenvatinib plus pembrolizumab in renal cell carcinoma |
| 2026 | partly addresses it | Reframing Population-Adjusted Indirect Comparisons as a Transportability Problem: An Estimand-Based Perspective and Implications for Health Technology Assessment (Chandler & Ishak; arXiv:2602.17041v3). The batch file points at media/media/Table_1_v2.pdf; the full LaTeX source Transportability_PAIC.tex sits in the same arXiv package and was read. |
| 2026 | partly addresses it | Anchors Away: Navigating Unanchored Indirect Comparisons with Multilevel Unanchored Meta-Regression (ML-UMR) (Chandler & Ishak; arXiv:2606.20341v1). The batch file points at media/media/Table1.pdf; the full LaTeX source mlumr.tex sits in the same arXiv package and was read. |
| 2026 | confirms it is open | Population-adjusted network meta-analyses provide new insights into the efficacy of treatment alternatives for metastatic castration-sensitive prostate cancer |
| 2026 | confirms it is open | Evidence Synthesis via Indirect Treatment Comparisons in the European Framework of Joint Clinical Assessment |
Source
Derived from A:L9, A:L47, A:L215 of the reviewed corpus.