Convenience functional forms and untested flexible alternatives
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.
Largely yes, the refuting sentence holds. Under a hinge in the treatment effect at a target 1.5 SD from the source, parametric STC was biased by -0.144 and a GAM with smooth treatment modification by -0.085 (coverage 0.93): about 40% less bias, short of the registered halving, with the same RMSE (0.314 against 0.318) because its variance was larger. Under a covariate-by-covariate modification every method was biased by about -0.18, since additive smooths cannot represent a product. A threshold prognostic term largely cancelled in all methods. Read the study below.
Statement
Functional form and interaction structure in population-adjusted outcome models are chosen for convenience, and conventional STC can be biased when that structure is wrong, particularly where transport requires extrapolation; unanchored MAIC can be biased when important adjustment variables are omitted from the matching. The simulation evidence the source invoked does not say what the source said. Hatswell, Freemantle and Baio (doi:10.1016/j.jval.2020.02.008) evaluated unanchored MAIC only and never STC, and found MAIC unbiased across a range of misspecification scenarios, biased only when important characteristics were left out of the matching and degraded when matched characteristics had low explanatory power or overlap was poor. The defensible open part is that flexible outcome models such as splines, BART, Gaussian processes, causal forests, super learner and TMLE have not been benchmarked against parametric baselines in PAIC under adversarial data-generating mechanisms.
Why it is open
Extrapolation to a target population probes the outcome model where the source data are sparse, which is where a wrong functional form does the most damage and where residual diagnostics have the least power, so poor overlap and misspecification compound. Not every omission bites: a common prognostic nonlinearity or covariate-covariate interaction need not bias a randomized identity-link treatment contrast, although it can bias nonlinear-link or standardized predictions. Flexible alternatives need an explicit route to valid inference and enough events to estimate an interaction surface, and single HTA-scale trials rarely have either, so flexibility can trade transparent misspecification for opaque misspecification. With no shared adversarial benchmark, each paper defines its own comparison and the interpretability cost of flexible models has no offsetting evidence.
What has been tried
| Work | What it contributes |
|---|---|
| Hatswell, Freemantle & Baio 2020, The Effects of Model Misspecification in Unanchored Matching-Adjusted Indirect Comparison: Results of a Simulation Study, Value in Health (PMID 32540233) | The study the source cited without a locator. Simulates unanchored MAIC against propensity score weighting and finds MAIC unbiased under a variety of misspecification scenarios, biased only when important characteristics are omitted from the matching. STC is not evaluated. |
| Phillippo et al. 2020, Statistics in Medicine (PMID 33015906) | Large anchored simulation study in which ML-NMR and STC eliminated bias when the requisite assumptions were met, while MAIC performed poorly in nearly all scenarios. Cuts against the source’s ‘especially in STC’ framing. |
Probable solution or research direction
Run head-to-head comparisons of parametric STC, flexible STC, BART and TMLE-style hybrids and doubly robust estimators on adversarial data-generating mechanisms containing splines, thresholds, nonlinear effect modification and covariate-covariate interaction, reporting misspecification robustness, calibration and the interpretability trade-off rather than friendly-mechanism performance. Pair any flexible fit with an explicit route to valid inference, but do not state cross-fitting as a universal requirement: parametric or Donsker nuisance classes, valid selective inference, debiasing and resampling that repeats the full tuning procedure are alternatives with their own conditions. Treat a flexible fit as a nuisance estimator with causal structure imposed rather than learned, never as a source of causal effect-modifier claims.
Our study
Flexible outcome models reduced but did not remove transport bias where the target left the data, and could not see a modification they were not built to see
The question put to it. Do flexible outcome models remove the transport bias a parametric STC incurs under threshold, hinge and covariate-interaction surfaces, and with valid intervals, as overlap worsens?
What it found. Largely yes, the refuting sentence holds. Under a hinge in the treatment effect at a target 1.5 SD from the source, parametric STC was biased by -0.144 and a GAM with smooth treatment modification by -0.085 (coverage 0.93): about 40% less bias, short of the registered halving, with the same RMSE (0.314 against 0.318) because its variance was larger. Under a covariate-by-covariate modification every method was biased by about -0.18, since additive smooths cannot represent a product. A threshold prognostic term largely cancelled in all methods.
Results.
- Bias grew with target distance for every method on the hinge and product surfaces.
- The structured GAM (smooth prognosis, linear modification) was no better than STC on the hinge surface.
- Coverage stayed near 0.90 to 0.93 in biased cells because intervals were wide (empirical SD about 0.3).
Design.
- Estimand
- Target-population marginal log odds ratio, A versus C
- Methods compared
- parametric STC, GAM with smooth treatment modification, GAM with linear treatment modification
- Data-generating mechanism
- Binary-outcome source trial with one adversarial surface feature at a time and targets at three distances; parametric STC, a GAM with smooth modification and a structured GAM with linear modification, standardized by G-computation. 12 cells, 500 replicates each.
- Replicates
- 500 per cell, 12 cells
What this does not answer. No BART, random forest, doubly robust or ML-NMR arm; logit link, one sample size; model-based intervals only. Peer review has not been done.
Peer review. Not yet done. The result stands on the study alone until it is.
Read it: PDF · Markdown · OpenDocument · Code and data: studies/MOD-02-adversarial-surfaces-flexible-outcome-models · Protocol: registered before the run
Verification
Verdict. Confirmed open. Auditors agreed the problem is open and no counterevidence surfaced. The source attributed the claim to the wrong work, which is corrected in the entry; the problem itself stands.
The source said. “Population-adjusted outcome models typically assume linear prognostic terms and a small set of simple treatment-by-covariate interactions, which are convenience assumptions rather than scientific claims, and the transported effect inherits whatever form is imposed. Omitted treatment interactions, nonlinear prognostic effects, nonlinear effect modification, covariate-covariate interactions, non-proportional hazards, and an incorrect target joint distribution are distinct routes by which the convenience model biases the answer, and simulation evidence indicates unanchored MAIC and STC are highly vulnerable when the model is wrong.”
What was wrong with it. The source said simulation evidence indicates unanchored MAIC and STC are highly vulnerable to misspecification, attributing it to ‘Hatswell and colleagues’ with no locator. Both auditors judged it misattributed: the study is doi:10.1016/j.jval.2020.02.008, it evaluates MAIC and not STC, and its conclusion is that MAIC performed well; the largest anchored simulation study (doi:10.1002/sim.8759) likewise found that STC and ML-NMR eliminated bias when assumptions held and reserved its concerns for MAIC. The claim was unattributed in the source and is now sourced.
Revised by the full-text reading. Applied 2026-07-26 full-text reading of the consolidated library (687 papers).
- closure-upheld
- 2 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 2020 simulation finding is directly consistent with the closure’s conceded residual: functional-form misspecification can bias STC and ML-NMR when extrapolation occurs without overlap. It does not rehabilitate the earlier overstrong or misattributed claim about Hatswell, which studied unanchored MAIC rather than STC and did not find broad misspecification bias. The 2026 example shows an analysis declining to investigate interactions, but one practice example does not establish a field-wide convenience claim. Neither finding contradicts the closure’s acknowledgment that functional-form risk remains an analysis-specific open issue. (reviewer confidence: high)
- verdict-corrected
-
not-supported -> confirmed-open The adjudication rule let a misattributed citation set the published verdict to
not-supported, though every auditor on this entry voted the problemopen. A misattribution is a fault in the provenance, not in the claim. The rule is corrected in build/adjudicate.mjs and the verdict re-derived from this entry’s own audit record. Auditors agreed the problem is open and no counterevidence surfaced. The source attributed the claim to the wrong work, which is corrected in the entry; the problem itself stands. - Flagged for reopening
-
2 findings from the reading assert this problem is open while the registry carries
confirmed-openFlagged, 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 / misattributed. Retrieved the only Hatswell paper matching the description: Hatswell, Freemantle & Baio, ‘The Effects of Model Misspecification in Unanchored Matching-Adjusted Indirect Comparison: Results of a Simulation Study’, Value Health 2020 (PMID 32540233). Two problems. First, it studies unanchored MAIC only, against propensity score weighting; STC is not evaluated anywhere in it, so attributing an STC vulnerability finding to it is wrong. Second, its headline conclusion is the opposite of ‘highly vulnerable’: MAIC was unbiased across a variety of misspecification scenarios and only became biased when important characteristics were omitted from the matching. Separately, the largest anchored simulation study in the field (Phillippo et al. 2020, Stat Med) found STC and ML-NMR eliminated bias when assumptions held and reserved its ‘serious concerns’ for MAIC, which cuts directly against ‘especially in STC’. The flexible-model claim (BART/causal forests/TMLE not systematically benchmarked against parametric baselines in PAIC under adversarial DGMs) I could not falsify and vote open.
- GPT-5.6 Sol, technical and source-code lens
- open / misattributed. C1 is a provenance placeholder, not a technical claim. C2 is misattributed and conflates models: standard unanchored MAIC is a weighting estimator, not an outcome model, and the cited Hatswell study is about unanchored MAIC rather than STC. C3-C4, C6, C8, C10, C17-C19 describe real conditional risks, especially when extrapolation is material. C7 and C9 are only potential bias routes: omitted common prognostic nonlinearity or covariate interactions need not bias a randomized identity-link treatment contrast, although they can bias nonlinear-link or standardized predictions. C11 concerns marginally standardized outcome regression, not the conventional mean-centered STC estimand itself. C16 is false as an absolute: cross-fitting relaxes overfitting and empirical-process conditions, but valid inference can instead rest on parametric or Donsker nuisance classes, external training, selective inference, valid debiasing, or resampling that repeats the full tuning procedure. C12-C15 and the claim that existing comparisons are uniformly friendly are literature and priority assertions outside this audit lens. Functional-form risk remains open, but very-high severity is analysis-specific rather than universal.
- Decision path
-
R5:default-open → R6:counter-quote-uncontested
How the evidence falls in time
2 papers in the reviewed corpus bear on this problem, published 2020 to 2026. Every paper found here asserts the problem is open; none reports progress.
| Year | Says | Paper |
|---|---|---|
| 2020 | confirms it is open | Assessing the performance of population adjustment methods for anchored indirect comparisons: A simulation study |
| 2026 | confirms it is open | Comparing the efficacy of cipaglucosidase alfa plus miglustat with alglucosidase alfa for late-onset Pompe disease: an expanded network meta-analysis utilizing patient-level and aggregate data |
Source
Derived from A:L46, A:L64, A:L158, A:L178, A:L219, C:L214-215, C:L231, C:L947-961, C:L1009-1020 of the reviewed corpus.