Valid and non-conservative treatment-effect intervals from EHR under unmeasured confounding

• Sharp bounds for averages. Dorn, Guo and Kallus (arXiv:2112.11449; JASA 2025) derive sharp bounds on the average treatment effect under the marginal sensitivity model and estimators that are "doubly valid" (bounds valid even if most nuisance functions are wrong) and "doubly sharp" (efficient when they are right).
• Sharp bounds for conditional effects. Frauen, Melnychuk and Feuerriegel (NeurIPS 2023, arXiv:2305.16988) generalise the marginal sensitivity model and derive sharp bounds for CATEs, mediation and distributional effects with a scalable estimator; guarantees are asymptotic and rely on correctly specified or sufficiently accurate nuisance estimators.
• Individual effects. Jin, Ren and Candes (PNAS 2023; arXiv:2111.12161) give a conformal prediction method with marginal coverage under the marginal sensitivity model, and a Gamma-value per unit (the minimum confounding that explains the effect away). Their guarantees are marginal over covariates, not conditional on $x$, and a concurrent method by Yin, Shi, Wang and Blei (arXiv:2112.03493) is reported as equivalent.
• Anchoring to trials. RCT-DUPLICATE (JAMA 2023) emulated 32 trials in claims data: significance agreement was 75% originally and 72% after a published correction, with correlation about 0.94 for closely emulated trials; this supplies empirical but dataset-specific information about how large real-world confounding is.
• Not shown. No procedure combines conditional validity, near-oracle sharpness and a principled choice of $\Gamma$ in EHR-scale data with weak overlap in sources reviewed (October 2026).

Clinicians and regulators need to know how far a data-driven recommendation for a specific patient group can be trusted. Valid, informative intervals would support real-world evidence submissions, guide which subgroups merit a trial, and prevent harmful policy learning from confounded records.

A refereed method with theorems for items 1-4 and an empirical validation on at least five trial-anchored EHR emulations (where the RCT effect is the hidden truth) showing nominal coverage for the individual-effect and CATE intervals and widths within 1.2 times the oracle bound where computable, together with open software.