No proximal ITC method works with an aggregate-only comparator
Statement
Proximal causal inference for indirect comparison is no longer speculative. Su, Rytgaard, Ravn and Eriksson, ‘Proximal indirect comparison’ (doi:10.1093/biomet/asaf044; arXiv:2405.10773), give identification and a doubly robust estimator for anchored ITC under unmeasured shifted effect modifiers, using an adjustment proxy in both trials and a reweighting proxy in the source, and Song et al. (arXiv:2603.04544) carry out a full proximal analysis of an external-control comparison in HIV prevention using baseline sexually transmitted infection status as a negative control outcome and geographic region as a negative control exposure. Both require individual data on each side, so the routine PAIC setting, where the comparator is a published baseline table, is untouched: no proxy’s negative-control role can be established from a summary table, and completeness and bridge-function conditions remain untestable.
Why it is open
Proximal identification requires proxies occupying specific structural roles plus bridge-function existence, completeness and positivity, and those conditions are not nonparametrically testable. The published methods assume individual patient data in both trials; in the usual PAIC setting the comparator is a published baseline table, from which no candidate proxy’s role can be checked. Unanchored comparisons and multi-arm networks also lie outside the two-trial construction that has been solved.
What has been tried
| Work | What it contributes |
|---|---|
| Su, Rytgaard, Ravn & Eriksson, Proximal indirect comparison, Biometrika 2025 | Proximal identification and a doubly robust estimator for anchored indirect comparison under unmeasured shifted effect modifiers, using an adjustment proxy in both trials and a reweighting proxy in the source. Assumes individual data in both trials. |
| Su et al., Proximal indirect comparison (preprint, 2024) | Preprint of the Biometrika paper; same ITC-specific proximal method. |
| Song, Wu, Landovitz, Shi, Gao, Donnell, Janes & Ye, Proximal Learning for Trials With External Controls: A Case Study in HIV Prevention (2026) | Applied proximal analysis of an external-control (unanchored) comparison using ordinary baseline variables as negative control outcome and exposure. Does not cover the aggregate-comparator PAIC case. |
Probable solution or research direction
Characterize which trial data structures could support proximal transport at all, and extend the anchored two-trial result to unanchored and network settings rather than re-deriving the anchored case. Probe the completeness and bridge-function conditions with sensitivity analysis instead of treating them as satisfied by construction, and test whether anything useful survives when one side is aggregate data only.
Verification
Verdict. Partially addressed. Part of this problem is addressed by existing work; the entry names which part.
The source titled this. “Proximal causal inference with proxies for unmeasured modifiers” The title presents proximal causal inference for unmeasured modifiers as the open problem, but the audit found a published method paper and an applied analysis for exactly this setting.
The source said. “Proximal causal inference is proposed as a route to identification in ITC settings with unmeasured effect modifiers or confounders, using proxy variables in place of the unmeasured quantity. The report labels it speculative for this field.”
What was wrong with it. The source called proximal causal inference speculative for this field, said suitable proxies are rarely available and cited nothing. Grok 4.5 found a method paper for exactly this setting (doi:10.1093/biomet/asaf044) and the literature auditor found an applied external-control analysis (arXiv:2603.04544); both weaken the rarity claim, though both need individual data on each side.
Dissent. literature dissented (overstated): A full applied proximal analysis of an unanchored indirect comparison now exists and the entry cites nothing. Song, Wu, Landovitz, … Shi, Gao, Donnell, Janes & Ye, ‘Proximal Learning for Trials With External Controls: A Case Study in HIV Prevention’ (arXiv:2603.04544, Mar 2026) applies proximal causal inference to estimate counterfactual placebo outcomes in an active-controlled trial using an external placebo-controlled trial, with baseline sexually transmitted infection status as negative control outcome and geographic region as negative control exposure. That is an external-control/unanchored ITC, and the proxies are ordinary trial baseline variables, not exotic instruments. This materially weakens ‘Suitable proximal proxy variables are rarely available in trial publications’ and ‘That makes the method inapplicable in most real ITCs.’ The narrow version survives: proxies were available because the authors had IPD from both trials; from a published baseline table you cannot verify a negative-control role. The completeness-condition claim is correct and untestable, as the entry says.
- Literature and prior-art check
- partially-solved / overstated. A full applied proximal analysis of an unanchored indirect comparison now exists and the entry cites nothing. Song, Wu, Landovitz, … Shi, Gao, Donnell, Janes & Ye, ‘Proximal Learning for Trials With External Controls: A Case Study in HIV Prevention’ (arXiv:2603.04544, Mar 2026) applies proximal causal inference to estimate counterfactual placebo outcomes in an active-controlled trial using an external placebo-controlled trial, with baseline sexually transmitted infection status as negative control outcome and geographic region as negative control exposure. That is an external-control/unanchored ITC, and the proxies are ordinary trial baseline variables, not exotic instruments. This materially weakens ‘Suitable proximal proxy variables are rarely available in trial publications’ and ‘That makes the method inapplicable in most real ITCs.’ The narrow version survives: proxies were available because the authors had IPD from both trials; from a published baseline table you cannot verify a negative-control role. The completeness-condition claim is correct and untestable, as the entry says. Cites arxiv:2603.04544: Covers the external-control (unanchored ITC) case with IPD on both sides: proximal identification using baseline STI status as negative control outcome and geographic region as negative control exposure. Does not cover the aggregate-comparator PAIC case..
- GPT-5.6 Sol, technical and source-code lens
- open / supported-with-caveat. C1 is correct only under an explicit proximal ITC or transport causal structure with treatment-inducing and outcome-inducing proxies or their transport analogues, bridge-function existence, completeness, positivity, and compatible source and target data. Arbitrary correlates of an unmeasured modifier are not enough. C2 is essentially correct, although completeness is a property of conditional expectation operators and distributions, not of a proxy variable in isolation, and is generally not nonparametrically testable. C3 is plausible for aggregate trial publications but is an uncited empirical prevalence claim that this technical and source-code audit cannot verify. C4 accurately reports the label. The narrow adaptation and data-requirement problem remains open; Medium is proportionate because the method is inapplicable when the required proxy structure is absent.
- Grok 4.5, recency lens
- partially-solved / supported-with-caveat. A direct method paper now exists for this exact setting: Su, Rytgaard, Ravn, Eriksson, Proximal indirect comparison (Biometrika, published 2025 / volume dating into 2026; arXiv:2405.10773), extending proximal causal inference to anchored ITC with unmeasured shifted effect modifiers via adjustment and reweighting proxies, with identification results and a doubly robust estimator under IPD in both trials. That means C1 is no longer purely speculative: proximal ITC is a formalized option in some IPD-rich anchored cases. The problem remains only partially closed because (i) completeness / bridge-function existence and structural proxy roles remain unverifiable assumptions, (ii) the paper itself requires suitable proxies (reweighting proxy in the source, adjustment proxy in both) and IPD, which most AgD-only trial publications do not supply, and (iii) unanchored and multi-arm network settings are not solved by this two-trial construction. C2–C3 and the practical inapplicability claim therefore still hold for routine HTA ITCs. Cites 10.1093/biomet/asaf044: Provides proximal identification and a doubly robust estimator for anchored indirect comparison under unmeasured shifted effect modifiers using reweighting and adjustment proxies (IPD setting).. Cites arXiv:2405.10773: Preprint of the Biometrika paper; same ITC-specific proximal method..
- Decision path
-
R3:partially-solved(n=2) → R9:substantiated-single-overstatement:literature
Source
Derived from C:L1646-1647, C:L1658 of the reviewed corpus.