Instrument Validity, Not Network Depth, Bounds Heterogeneous Effects in Confounded EHR Data

Question: How do deep instrumental-variable and doubly robust policy learning methods estimate heterogeneous treatment effects from observational electronic health records when treatment assignment is unmeasured-confounded?

Answer

Taken together, the retrieved papers support a deflationary answer: deep IV and doubly robust policy learning each solve a nuisance-estimation problem, and neither, in its current form, estimates a heterogeneous treatment effect from unmeasured-confounded EHR data. Deep IV methods are now provably rate-optimal nonparametric structural estimators, but they identify average structural functions or complier-local effects, not the conditional effects a clinician needs; doubly robust learning delivers fast welfare-regret rates and protection against misspecification of either nuisance model, but its guarantees are conditional on sequential ignorability, and confounding robustness comes from separate machinery (proxies, sensitivity sets, partial identification). What actually buys identification in the confounding-hostile EHR setting is added structure — bridge/proxy functions, instrument validation, transportability corrections, and robust sets — and the empirical EHR papers that report credible heterogeneity are the ones that supply that structure rather than the ones that scale the network.

Deep IV is now rate-optimal, but it recovers a structural or complier-local estimand rather than a CATE

Deep feature IV regression is proven minimax optimal when the structural function lies in a Besov space, and is provably more data-efficient than fixed-feature kernel or sieve IV in the first stage [23]. The failure modes of the dominant spectral-feature approach are now characterised as a taxonomy driven by spectral alignment and eigenvalue decay, with a data-driven diagnostic for which regime a problem is in [22], and augmented, outcome-aware spectral features can survive spectral misalignment where outcome-agnostic features fail [24]. Deeper architectures mainly attack bias: DML-IV constructs a debiasing objective to remove first-stage regularisation bias and attains the same O(N^{-1/2}) suboptimality guarantee as learning from an unconfounded dataset [27], while pretrained transformers can emulate a bi-level optimisation converging exponentially to the 2SLS solution in-context [25]. None of these results changes the estimand: with non-compliance and effect heterogeneity, a binary instrument identifies the local average treatment effect among compliers rather than the population average treatment effect [32], nested instruments identify effects among always-compliers and switchers with formal tests of generalizability [2], and when the treatment is an aggregate of finer unobserved components the IV estimand is itself ambiguous absent contrived conditions on the intervention distribution [8].

Doubly robust orthogonality protects against nuisance misspecification, and is silent about hidden confounding

Efficient plug-in learning matches the oracle efficiency of DR- and R-learning while avoiding non-convex losses and the instability of pseudo-outcomes that violate bounds [18]. Distributionally robust policy learning built on cross-fitted doubly robust scores retains the doubly robust property — the O(n^{-1/2}) DRO regret rate holds if either the outcome or the propensity model is consistent, even at slow nonparametric rates — and attains fast rates under a margin condition on the treatment effect [12]. Classification-based dynamic treatment regimes maximise an AIPW policy value by backward induction and achieve n^{-1/2} welfare regret, explicitly under sequential ignorability [17], and sequential model doubly robust estimators give root-N inference under high-dimensional covariates as long as one nuisance model is correct at each exposure time [19]. DR guarantees for distribution-valued outcomes under Wasserstein barycenter rewards follow the same pattern, with regret governed by policy-class complexity [16]. The boundary is stated plainly in a Bayesian causal-forest paper whose effect estimators are unbiased with nominal coverage yet fail "only under unmeasured confounding, which we flag" [34]; the robustness being purchased is against misspecification and observed distribution shift, not against a latent confounder.

Recovering population or subgroup heterogeneity from an instrument requires proxy or bridge structure

The one retrieved framework that moves from local IV effects to population effects under latent heterogeneity does so by adding proxies for the latent factors driving compliance and effect heterogeneity, identifying the ATE through an outcome bridge or a compliance bridge whose combination is doubly robust, with efficient influence functions, orthogonalised moments under estimated instrument propensity, and three-way cross-fitting [32]. On EHR data specifically, a vector of recorded proxies (vitals and similar) can be compressed by factor analysis into covariates capturing the unmeasured confounder, and this is benchmarked against proximal causal inference including deliberate misuse of a confounder or instrument as a proxy, then applied to hospital admission for chest pain in older emergency-department patients [4]. Proximal reasoning also enables fusion of experimental and confounded observational data for long-term effects using proxies of the latent confounder with influence-function-based estimators [6], and doubly robust inference with matrix-completion cross-fitting delivers parametric-rate Gaussian limits in latent-factor designs with many units and outcomes [33]. Tree-based conditional complier effect estimation — DRRF-IV and a GRF-based IV adaptation against BCF-IV — has been applied to whether prompt ICU admission reduces 28-day mortality, where prior work found no overall effect and the question becomes whether identifiable subgroups benefit [35].

In EHR the instrument and the sampling process are as fragile as the outcome model

There is direct empirical evidence that ICU EHRs are subject to unobserved confounding, established by eliciting confounders from clinicians through paired comparisons and by using clinical text notes as a proxy for physicians' knowledge in a semi-synthetic testbed with known ground truth [28]. Even a valid instrument fails when selection is an effect modifier or caused by one: IV estimates are biased under effect heterogeneity, with inverse probability weighting offered as a remedy [10]. Two-sample IV breaks when compliance is not transportable across the claims and survey populations that supply the first stage and reduced form, and reweighting the first stage recovers most of the bias, reported as an 88% reduction in the primary scenario [3]. Physician prescribing preference has been used as an instrument for multi-valued vasopressor comparisons in MIMIC-IV with Bayesian feature selection, yielding an ordered effectiveness hierarchy [26] — a design whose exclusion restriction the retrieved papers do not test. The sampling layer is itself informative: EHR analyses face lack of true negatives, differential monitoring, and informative missingness that can produce narrow intervals around biased estimates [1], and biomarker recording decomposes into a visit process and an observation process, each with distinct drivers across All of Us, Yale New Haven, and the Michigan Genomics Initiative [7]. Missingness of this kind demonstrably biases off-policy value estimates, with an inverse-probability-weighted estimator proposed to restore consistency [20].

Where confounding cannot be removed, the defensible output is a robust set or a partial identification bound

Sensitivity analysis via coarse relaxation of the confounding uncertainty set yields overly conservative policy values; a convex-programming estimator gives a sharp lower bound and reformulates as empirical risk minimisation, enabling robust policy learning [11]. Treating propensity weights under unobserved confounding as an uncertainty set defined by observable data and a domain-informed degree of confounding, combined with Wasserstein DRO, improves risk-score calibration by up to 29.2% over standard benchmarks [13]. Omitted-variable-bias theory now extends to nonparametric, machine-learned causal parameters including policy effects from covariate shifts, with simple judgments about the maximum explanatory power of omitted variables sufficient to bound bias and with valid inference on the bounds [37], and bounds for sequential decisions remain valid under arbitrary unmeasured confounding [14]. Confounding-robust deep RL formalises the analogous move by optimising the worst-case environment consistent with the observations [15].

Conflicts and gaps

A large EHR policy-learning literature evaluates learned treatment policies with weighted importance sampling and fitted Q evaluation under a behaviour policy estimated from observed state, reporting that the learned sepsis resuscitation policy exceeds clinician return [36]; that protocol presumes no latent confounding in the logging process, which the ICU confounding evidence directly contradicts [28]. The proxy literature disagrees internally on implementation and fragility, with factor-analytic proxies compared against proximal causal inference and both stress-tested under misspecification and variable misuse [4], while the proximal IV route rests on bridge existence and completeness conditions that no retrieved paper tests on real data [32]. Estimation-side problems compound the identification problems: in a large benchmark, 62% of CATE estimates had higher MSE than a trivial zero-effect predictor and orthogonality-based models won only 30% of the time [5], counterfactual metrics did not reliably recover the estimators preferred by observable metrics, and benchmark rankings did not transfer to real datasets [38], while leaf-averaged causal forests attenuate heterogeneity toward the average effect by a slope that recentering does not fix [31]. Not one retrieved paper demonstrates an end-to-end deep IV plus doubly robust policy pipeline on EHR data with both a validated instrument and confirmed effect heterogeneity; the closest production-scale EHR system relies on measured-confounding adjustment via collaborative targeted learning rather than any IV or proxy identification [9]. Finally, supplementary experimental data can falsify observational estimates but cannot validate them without a smoothness assumption on the correction function [21].

What would change the answer

A randomized encouragement or nudged-assignment instrument embedded in a health system — the design that makes the first stage known rather than assumed — combined with the nested-IV overidentification tests that permit machine-learned nuisances and formally test generalizability [2] would convert complier-local estimates into testable population claims. Benchmarking deep IV, proximal bridge estimators, and DR learners on a semi-synthetic ICU EHR with expert-elicited ground-truth confounders and clinical-text proxies [28] would establish which method recovers heterogeneity when the truth is known. Systematic instrument audits using negative control outcomes and exposures, including the case where a single negative control outcome fails and a negative control exposure restores identifiability [29], plus deliberate "confounder-as-instrument" misuse stress tests [4], would quantify exclusion-restriction violations for preference-based instruments. Linked EHR-and-claims analyses already show that data-source choice flips an effect estimate from protective to null [30] and could serve as a bias oracle for such audits. Reporting omitted-variable-bias bounds and the minimum confounding strength needed to overturn a reported subgroup effect alongside every deep IV CATE [37], together with Gaussian-process interval construction anchored on experimental data [21], would make the uncertainty from unmeasured confounding a visible output rather than an unstated assumption.

References

[1] Butzin-Dozier, Z., Ji, Y., Wang, L.-C., Anzalone, A. J., Hurwitz, E., Patel, R. C., van der Laan, M., Colford, J. M., Hubbard, A. E., on behalf of the N3C Consortium,. "Causal Inference via Electronic Health Records in the National Clinical Cohort Collaborative: Challenges and Solutions in Long COVID Research". medRxiv. 2025-06-11T00:00:00.000Z. URL: https://doi.org/10.1101/2025.06.06.25329168

[2] Rui Wang, Ying-Qi Zhao, Oliver Dukes, Bo Zhang. "Nested Instrumental Variables Analysis: Switcher Average Treatment Effect, Identification, Efficient Estimation and Generalizability". arxiv. 2026-10-12T00:00:00.000Z. URL: https://arxiv.org/abs/2405.07102

[3] Qian, Y., Song, Y.. "Two-Sample Instrumental Variables under Population Mismatch: A Transportability Framework with Bias Diagnostics". medRxiv. 2026-06-26T00:00:00.000Z. URL: https://doi.org/10.64898/2026.06.15.26355602

[4] Haley Colgate Kottler, Amy Cochran. "A proxy-based approach for unmeasured confounding in electronic health records research". arxiv. 2026-03-23T00:00:00.000Z. URL: https://arxiv.org/abs/2506.12177

[5] Haining Yu, Yizhou Sun. "Do Contemporary Causal Inference Models Capture Real-World Heterogeneity? Findings from a Large-Scale Benchmark". arxiv. 2025-02-21T00:00:00.000Z. URL: https://arxiv.org/abs/2410.07021

[6] AmirEmad Ghassami, Chang Liu, Alan Yang, David Richardson, Ilya Shpitser, Eric Tchetgen Tchetgen. "Combining Experimental and Observational Data for Identification and Estimation of Long-Term Causal Effects". arxiv. 2025-09-30T00:00:00.000Z. URL: https://arxiv.org/abs/2201.10743

[7] Yang, C.-H., Salvatore, M., Lu, H., Zhu, Z., Tennant, P., Shi, X., Ohno-Machado, L., Khera, R., Gross, C., Li, F., Mukherjee, B.. "Who seeks care, and what gets measured? Understanding the distinct mechanisms behind visit and observation processes in multi-center electronic health records". medRxiv. 2026-08-17T00:00:00.000Z. URL: https://doi.org/10.64898/2026.08.12.26360236

[8] Danielle Tsao, Krikamol Muandet, Frederick Eberhardt, Emilija Perković. "Lost in Aggregation: The Causal Interpretation of the IV Estimand". arxiv. 2026-08-04T00:00:00.000Z. URL: https://arxiv.org/abs/2601.1212

[9] Coyle, J., Shah, N., Hubbard, A., Mukerji, A., Chappelka, M., Sanghavi, N., Gombar, S.. "High-Throughput Observational Evidence Generation Using Linked Electronic Health Record and Claims Data". medRxiv. 2026-09-18T00:00:00.000Z. URL: https://doi.org/10.64898/2026.04.07.26350300

[10] Gkatzionis, A., Hartwig, F. P., Tilling, K.. "Selection without Collider Stratification can Bias Instrumental Variable Analyses in the Presence of Causal Effect Heterogeneity". medRxiv. 2026-09-30T00:00:00.000Z. URL: https://doi.org/10.64898/2026.09.27.26364122

[11] Kei Ishikawa, Niao He, Takafumi Kanamori. "A Convex Framework for Confounding Robust Inference". arxiv. 2026-01-13T00:00:00.000Z. URL: https://arxiv.org/abs/2309.1245

[12] Debmalya Mandal. "Distributionally-Robust Policy Learning with Observational Data". arxiv. 2026-10-20T00:00:00.000Z. URL: https://arxiv.org/abs/2610.25591

[13] Ryan Edmonds, Yingxiao Ye, Sina Aghaei, Andrés Gómez, Çağıl Koçyiğit, Phebe Vayanos. "Learning Risk Scores Robust to Unobserved Confounders". arxiv. 2026-09-24T00:00:00.000Z. URL: https://arxiv.org/abs/2609.27144

[14] Muhammad Faaiz Taufiq. "Uncertainty Quantification and Causal Considerations for Off-Policy Decision Making". arxiv. 2025-02-11T00:00:00.000Z. URL: https://arxiv.org/abs/2502.06011

[15] Mingxuan Li, Junzhe Zhang, Elias Bareinboim. "Confounding Robust Deep Reinforcement Learning: A Causal Approach". arxiv. 2025-10-27T00:00:00.000Z. URL: https://arxiv.org/abs/2510.2111

[16] Yiyan Huang, Cheuk Hang Leung, Qi Wu, Zhiheng Zhang. "Wasserstein Policy Learning for Distributional Outcomes". arxiv. 2026-06-18T00:00:00.000Z. URL: https://arxiv.org/abs/2606.19117

[17] Shosei Sakaguchi. "Policy Learning for Optimal Dynamic Treatment Regimes with Observational Data". arxiv. 2025-05-21T00:00:00.000Z. URL: https://arxiv.org/abs/2404.00221

[18] Lars van der Laan, Marco Carone, Alex Luedtke. "Combining T-learning and DR-learning: a framework for oracle-efficient estimation of causal contrasts". arxiv. 2026-03-20T00:00:00.000Z. URL: https://arxiv.org/abs/2402.01972

[19] Yuqian Zhang, Weijie Ji, Jelena Bradic. "Dynamic treatment effects: high-dimensional inference under model misspecification". arxiv. 2025-01-31T00:00:00.000Z. URL: https://arxiv.org/abs/2111.06818

[20] Han Wang, Yang Xu, Wenbin Lu, Rui Song. "Off-Policy Evaluation Under Nonignorable Missing Data". arxiv. 2025-07-10T00:00:00.000Z. URL: https://arxiv.org/abs/2507.06961

[21] Jake Fawkes, Michael O'Riordan, Athanasios Vlontzos, Oriol Corcoll, Ciarán Mark Gilligan-Lee. "The Hardness of Validating Observational Studies with Experimental Data". arxiv. 2025-03-20T00:00:00.000Z. URL: https://arxiv.org/abs/2503.14795

[22] Dimitri Meunier, Antoine Moulin, Jakub Wornbard, Vladimir R. Kostic, Arthur Gretton. "Demystifying Spectral Feature Learning for Instrumental Variable Regression". arxiv. 2025-11-27T00:00:00.000Z. URL: https://arxiv.org/abs/2506.10899

[23] Juno Kim, Dimitri Meunier, Arthur Gretton, Taiji Suzuki, Zhu Li. "Optimality and Adaptivity of Deep Neural Features for Instrumental Variable Regression". arxiv. 2025-01-10T00:00:00.000Z. URL: https://arxiv.org/abs/2501.04898

[24] Dimitri Meunier, Jakub Wornbard, Vladimir R. Kostic, Antoine Moulin, Alek Fröhlich, Karim Lounici, Massimiliano Pontil, Arthur Gretton. "Outcome-Aware Spectral Feature Learning for Instrumental Variable Regression". arxiv. 2026-06-01T00:00:00.000Z. URL: https://arxiv.org/abs/2512.00919

[25] Haodong Liang, Krishnakumar Balasubramanian, Lifeng Lai. "Transformers Handle Endogeneity in In-Context Linear Regression". arxiv. 2025-05-13T00:00:00.000Z. URL: https://arxiv.org/abs/2410.01265

[26] Qian, Y., Ma, B.. "Bayesian Feature Selection for Multi-valued Treatment Comparisons: An Electronic Health Records Study of Vasopressor Effectiveness". medRxiv. 2024-12-20T00:00:00.000Z. URL: https://doi.org/10.1101/2024.12.19.24319363

[27] Daqian Shao, Ashkan Soleymani, Francesco Quinzan, Marta Kwiatkowska. "Learning Decision Policies with Instrumental Variables through Double Machine Learning". arxiv. 2025-06-25T00:00:00.000Z. URL: https://arxiv.org/abs/2405.08498

[28] Drago Plecko, Patrik Okanovic, Torsten Hoefler, Elias Bareinboim. "Confounder Detection via Treatment Intent: A New Observational Study Design". arxiv. 2026-05-27T00:00:00.000Z. URL: https://arxiv.org/abs/2605.26413

[29] Yong Wu, Yanwei Fu, Shouyan Wang, Yizhou Wang, Xinwei Sun. "Discovering Causal Relationships using Proxy Variables under Unmeasured Confounding". arxiv. 2025-10-21T00:00:00.000Z. URL: https://arxiv.org/abs/2510.17167

[30] Makkar, S. R., Bogumil, D., Estepp, T., Keshet, A., Hotaling, N., Hansen, K., Gersing, K. R., Sidky, H., Tripler, T. N.. "Enhancing Scientific Validity by Integrating EHR and Claims Data in the National Clinical Cohort Collaborative". medRxiv. 2026-10-15T00:00:00.000Z. URL: https://doi.org/10.64898/2026.10.12.26365348

[31] Harry Aytug. "Attenuated Heterogeneity in Fixed-Effects Causal Forests, and a Cross-Fitted Correction". arxiv. 2026-07-28T00:00:00.000Z. URL: https://arxiv.org/abs/2607.22896

[32] Zixuan Yao, Guosheng Yin. "Bridging Local and Population Causal Effects: A Proximal Instrumental Variable Approach". arxiv. 2026-09-22T00:00:00.000Z. URL: https://arxiv.org/abs/2609.22383

[33] Alberto Abadie, Anish Agarwal, Raaz Dwivedi, Abhin Shah. "Doubly Robust Inference in Causal Latent Factor Models". arxiv. 2024-10-30T00:00:00.000Z. URL: https://arxiv.org/abs/2402.11652

[34] Danilo A. Sarti. "Factorial Multivariate Bayesian Causal Forests: heterogeneous main and interaction effects of multiple treatments on correlated outcomes". arxiv. 2026-09-28T00:00:00.000Z. URL: https://arxiv.org/abs/2609.31332

[35] Karolina Gliszczyńska-Schroeder. "Estimating Heterogeneous Causal Effects with Tree-Based Methods under Imperfect Compliance". arxiv. 2026-09-29T00:00:00.000Z. URL: https://arxiv.org/abs/2609.31769

[36] Marc Pérez-Roig, David Fernández-Narro, Carlos Sáez. "Offline Reinforcement Learning for Hemodynamic Management of Sepsis in the ICU: a MIMIC-IV Study with Dual Off-Policy Evaluation". arxiv. 2026-08-27T00:00:00.000Z. URL: https://arxiv.org/abs/2608.16482

[37] Victor Chernozhukov, Carlos Cinelli, Whitney Newey, Amit Sharma, Vasilis Syrgkanis. "Long Story Short: Omitted Variable Bias in Causal Machine Learning". arxiv. 2026-09-22T00:00:00.000Z. URL: https://doi.org/10.1162/REST.a.1705

[38] George Panagopoulos. "Real vs. Semi-Simulated: Rethinking Evaluation for Treatment Effect Estimation". arxiv. 2026-05-26T00:00:00.000Z. URL: https://arxiv.org/abs/2605.1043

How this survey was made

This survey answers one question: “How do deep instrumental-variable and doubly robust policy learning methods estimate heterogeneous treatment effects from observational electronic health records when treatment assignment is unmeasured-confounded?”

We keep our own collection of recent preprints and refresh it every day from bioRxiv, medRxiv, arXiv, PsyArXiv, SocArXiv and ChemRxiv. For each survey, we search that collection for recent work and keep the strongest matches: papers that use the same terms as the question, and papers on the same topic in different words.

The text was written from the titles and abstracts of those papers; the full texts were not read. Statements that draw on a particular paper are marked with a number in brackets. The reference list just above is the 38 papers whose work this text draws on, chosen from the 131 papers those searches returned; the rest of the shortlist is not printed. It is not a complete review of everything published on the question.

Treat this as a starting point and a reading list, not a peer-reviewed review. The searches behind this survey ran on October 22, 2026 and covered work from the previous two years.