Partial identification bounds under drift or omitted modifiers
Statement
Manski-style partial identification bounds have not been published for PAIC or NMA under bridge drift or omitted effect modifiers, so the label ‘established in principle, underused in PAIC’ survives. The nearest published work is a sensitivity analysis rather than a bound set: Ren, Ren, Welton and Strong (doi:10.1017/rsm.2025.13) extend quantitative bias analysis to unanchored PAIC for unmeasured confounding, with algorithms, graphics and a colorectal cancer case study, and that is what analysts use instead. Bounds are informative only after explicit quantitative restrictions on drift or omitted-modifier effects; without them the identified set can be vacuous, and valid inference also needs a confidence region for the set rather than the set alone.
Why it is open
Deriving sharp bounds for PAIC is not a matter of importing econometric results: mixed individual and aggregate data, non-collapsible effect measures and network constraints each change the problem, and the sensitivity parameters need defensible calibration before a bound means anything. The institutional obstacle is real but secondary: bounds do not produce the single number that reimbursement pipelines, ranking outputs and economic models consume, so an honest interval is professionally more costly to report than a fragile point. Until a convention exists for acting on an interval, the point estimate keeps winning even when nobody defends its assumptions.
What has been tried
| Work | What it contributes |
|---|---|
| Ren, Ren, Welton & Strong, Quantitative bias analysis for unmeasured confounding in unanchored population-adjusted indirect comparisons, Res Synth Methods 2025;16:509-527 | Imports formal quantitative bias analysis for unmeasured confounding into unanchored PAIC and STC practice, with algorithms and a case study. It is sensitivity analysis rather than partial identification, and it establishes no convention for acting on an interval. |
Probable solution or research direction
Derive PAIC-specific bounds under stated omitted-modifier and bridge-drift restrictions, including the sampling-uncertainty procedure for the identified set, building on the quantitative bias analysis machinery already in use for unanchored comparisons rather than starting from scratch. Pair the result with decision-analytic conventions that let cost-effectiveness models and reimbursement processes act on an interval. Report the point estimate as the special case obtained by imposing a named restriction, so the strength of that restriction becomes visible.
Verification
Verdict. Partially addressed. Part of this problem is addressed by existing work; the entry names which part.
The source said. “Partial identification would report the set of treatment effects compatible with weak, defensible assumptions about bridge drift or omitted effect modifiers, instead of a point estimate that requires strong assumptions. The report labels it established in principle but underused in PAIC.”
What was wrong with it. The source said partial identification is underused without saying what is used instead; Grok 4.5 named Ren et al. (doi:10.1017/rsm.2025.13), a 2025 quantitative bias analysis for unanchored PAIC that partly fills the gap. GPT-5.6 Sol also rejected the claim that the remaining obstacle is purely institutional, since sharp bounds for mixed data, non-collapsible estimands and network constraints still require substantial derivation.
- Literature and prior-art check
- open / supported-with-caveat. I found no Manski-style partial identification bounds published for PAIC or NMA under bridge drift or omitted effect modifiers, so ‘established in principle but underused in PAIC’ survives. The caveat is that the nearest neighbour is now published and the entry does not acknowledge it: Ren, Ren, Welton & Strong, ‘Quantitative bias analysis for unmeasured confounding in unanchored population-adjusted indirect comparisons’ (Res Synth Methods 2025;16:509-527, doi:10.1017/rsm.2025.13), which simulates unreported covariates in unanchored STC and ‘enables the formal evaluating of the impact of unmeasured confounding in a quantitative manner without additional assumptions.’ Under this audit’s own calibration that is a sensitivity analysis, not identification, and it returns a robustness statement rather than a bound set, so it does not close ADJ-07; but an entry claiming partial identification is underused in PAIC should say what is being used instead. The claim that the obstacle is decision-analytic rather than technical is plausible but I found no source asserting it; treat that part as unverifiable.
- GPT-5.6 Sol, technical and source-code lens
- open / overstated. C1 is correct only after imposing explicit quantitative restrictions on bridge drift, omitted-modifier prevalence or effects, or related sensitivity parameters. Without such restrictions the identified set may be vacuous. C2 conflates an identified set with uncertainty: valid inference also requires confidence regions for the set and remains conditional on the sensitivity assumptions. C3 is generally correct, although a singleton set or an added decision criterion can produce a point. C4 is an unsourced utilization claim. The assertion that the remaining obstacle is institutional rather than theoretical is too strong; sharp PAIC bounds with mixed data, non-collapsible estimands and network constraints, plus defensible calibration of sensitivity parameters, can require substantial derivation.
- Grok 4.5, recency lens
- partially-solved / supported-with-caveat. Partial identification and sensitivity analysis for transport/generalization are established outside PAIC. Inside PAIC, Ren et al. (Res Synth Methods 2025; doi:10.1017/rsm.2025.13) extend quantitative bias analysis to unanchored PAIC/STC for unmeasured confounding, with algorithms, graphics, and a colorectal-cancer case study. That partially closes the underuse gap for omitted modifiers in unanchored settings, but is sensitivity/QBA rather than Manski-style partial identification under bridge drift, and does not create a decision-analytic convention for feeding intervals into CE models and reimbursement pipelines. C1–C3 remain faithful as statements of capability and institutional friction; C4 is slightly overstated after Ren but still directionally correct for routine PAIC practice and bridge-drift bounds. Identification problems stay untestable; sensitivity quantifies consequences only. Queries: partial identification PAIC MAIC omitted modifiers; Ren quantitative bias unanchored PAIC; bridge drift bounds ITC; NICE DSU sensitivity unanchored 2025–2026. Cites 10.1017/rsm.2025.13: Imports formal QBA sensitivity for unmeasured confounding into unanchored PAIC/STC practice. Does not establish reimbursement conventions for acting on partial-identification intervals under bridge drift..
- Decision path
-
R3:partially-solved(n=1) → R9:substantiated-single-overstatement:codex
How the evidence falls in time
1 paper in the reviewed corpus bears on this problem, published 2025. Every paper found here asserts the problem is open; none reports progress.
| Year | Says | Paper |
|---|---|---|
| 2025 | confirms it is open | Quantitative bias analysis for unmeasured confounding in unanchored population-adjusted indirect comparisons |
Source
Derived from C:L1646-1647, C:L1657 of the reviewed corpus.