{"id":"77e8454e-1762-4a7b-bcbc-4cb807dfe275","arxiv_id":"2506.12215","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"Two debiased estimators, one based on linear programming solutions and one on entropic smoothing, provide asymptotic confidence intervals for covariate-dependent partial identification bounds and support policy learning.","lead":"This paper develops statistical methods for estimating bounds on quantities that data cannot pin down exactly, using linear optimization problems that depend on covariates. It offers two debiased estimators with confidence intervals and applies them to Medicaid's effect on emergency department visits.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 2's pointwise η condition is incompatible with Assumption 3 unless α>4; for α≤1 Corollary D.1 has no admissible η sequence, so root-n normality for the entropic estimator is unsupported.","rationale":"I read the paper in good faith. The framework is a genuine contribution: debiased BFS estimation avoids vertex enumeration, the entropic estimator provides a smooth computable alternative, the simulation study shows good finite-sample RMSE and coverage in the configurations tested, and the application to the Oregon Medicaid experiment is carefully framed. The reader's verdict of CONDITIONAL is appropriate. My stress-test sharpens the reader's weakest assumption rather than replacing it: the margin condition and the data-dependent η condition interact in a way that is not merely a verification nuisance. Theorem 2's pointwise condition η≥(R1+RH)/Δ_L(X_i) is not implied by Assumption 3 and, for small α, cannot be satisfied simultaneously with the growth requirements n^{-1/4}η→0 and n^{1/2}e^{-η}→0. Corollary D.1, which is the paper's own fallback, shows that α>1 is needed for any η sequence to work at all, and the pointwise version effectively requires α>4 under the least favorable tail. The paper's text acknowledges the possibility of a too-fast η requirement but does not reconcile it with Assumption 3, so the abstract-level claim that asymptotic normality holds when nuisance estimates converge fast enough and the margin condition holds is stronger than what is proved. This does not make the BFS results suspect, nor do I question the simulations; it means the entropic estimator's Wald intervals require either a stronger, verifiable margin condition or an explicit sensitivity analysis over α. Because the reader already recommended a conditional verdict, my analysis does not change the verdict, but it identifies the precise condition that should be addressed in a revision.","tokens_in":40019,"tokens_out":9151,"duration_ms":122193,"concrete_test":"Analytical check: in Corollary D.1 set η=n^β and ask whether any β satisfies β(1+α)>1/2 and β<1/4. For α=0.5 the feasible set is empty; for α=2, β=0.3 violates the second inequality and β=0.2 violates the first. If no β exists for α≤1, Theorem 2's normality claim cannot be rescued without strengthening Assumption 3. A complementary simulation would engineer Δ_L(X) with tail P(Δ≤t)=t^α and report Wald coverage for α∈{0.5,2} across β∈{0.1,0.2,0.24,0.3}; failure to achieve nominal coverage in the α=0.5 case confirms the gap.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Assumption 3 only controls P(0<Δ_L(X)≤t)≲t^α. In an iid sample, the smallest positive gap has order n^{-1/α} under a Pareto tail consistent with this condition. Theorem 2 conditions on every observed Xi and requires η≥(R1(Xi)+RH(Xi))/Δ_L(Xi), so η must typically grow at least like n^{1/α}. But Theorem 2 also requires n^{-1/4}η→0 and n^{1/2}e^{-η}→0. These requirements are simultaneously satisfiable only if n^{1/α}=o(n^{1/4}), i.e., α>4. Assumption 3 permits any α>0. The fallback result, Corollary D.1, removes the pointwise condition but requires β(1+α)>1/2 and β<1/4 for η=n^β; no β exists when α≤1, and for 1<α≤4 the minimal η consistent with the pointwise condition will typically exceed n^{1/4}. Thus for a large part of the parameter space permitted by the stated assumptions, no choice of η yields the claimed root-n Gaussian approximation. The authors themselves flag this in Section 3.2, noting that 'depending on the tail behavior of the sub-optimality gaps ΔL(Xi), it may be possible that under some data generating processes the minimal size of η grows too quickly.' The Oregon application chooses η=100 without estimating Δ_L or checking the pointwise condition, so the displayed Wald intervals are outside the theorem's verified regime. This does not invalidate the BFS estimator or the simulations, but it is the load-bearing gap in the entropic normality claim.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper develops a unified framework for estimating and drawing inference about partially identified parameters whose sharp bounds are expectations of covariate-conditional linear programs. The estimand class includes functions of the joint distribution of potential outcomes, inequality-aware collective utility functions, and instrumental-variable settings. The paper proposes two debiased estimators: a plug-in basic feasible solution (BFS) estimator and an entropic-regularized estimator, and states asymptotic normality results for both (Theorems 1 and 2, Corollaries 1--2 and D.1) as well as excess-regret bounds for policy learning (Theorems 3 and 4). The methods are illustrated with simulations and an Oregon Medicaid application. The BFS estimator is computationally attractive because it avoids vertex enumeration, and the paper contains detailed appendix proofs and extensive simulation comparisons.","tokens_in":40350,"tokens_out":4231,"duration_ms":56838,"significance":"If the main theorems hold, the paper provides a substantial unification: it handles several important partial-identification problems with a single estimation and inference recipe, and it shows that debiasing can be done using only the output of standard LP solvers or entropic dual solutions, without closed-form bound derivations. The detailed appendix proofs, the explicit treatment of nuisance estimation via Riesz representers, and the simulation study comparing against log-sum-exp approximations are clear strengths. The main weakness is that the entropic estimator's root-n normality is only established under a pointwise, unverifiable condition on the regularization parameter η whose compatibility with the stated margin condition is much more restricted than the paper acknowledges; as a result, the claim that both estimators are asymptotically normal is not fully supported over the parameter space permitted by Assumption 3.","major_comments":[{"comment":"Theorem 2 conditions on the realized sample: it requires η ≥ (R1(Xi)+RH(Xi))/Δ_L(Xi) for every observed Xi with Δ_L(Xi) > 0. Under Assumption 3, the smallest positive sub-optimality gap in an iid sample has order n^{-1/α} under a Pareto-tail model that saturates the margin condition, so this condition forces η to grow at least like n^{1/α} with probability tending to one. The same theorem also requires n^{-1/4}η → 0. These requirements are simultaneously satisfiable only if n^{1/α} = o(n^{1/4}), i.e., α > 4, whereas Assumption 3 only states α > 0. The discussion in Section 3.2 acknowledges that 'the minimal size of η grows too quickly' may occur in some data generating processes, but the theorem as stated claims normality under Assumptions 1--3 without this restriction. The statement needs either an explicit lower-tail condition on Δ_L, a restricted range for α, or a different asymptotic regime for the entropic estimator.","section":"Theorem 2"},{"comment":"Corollary D.1 removes the pointwise η condition but requires n^{1/2}η^{-(1+α)} → 0 and n^{-1/4}η → 0. Writing η = n^β gives β > 1/(2(1+α)) and β < 1/4. For α ≤ 1 these inequalities cannot both hold, so no admissible η sequence exists. For 1 < α ≤ 4 the admissible interval is nonempty only for β near 1/4, while the minimal η forced by the pointwise condition (n^{1/α}) is incompatible with β < 1/4 unless α > 4. Thus for a substantial part of the parameter space permitted by Assumption 3, neither Theorem 2 nor Corollary D.1 supplies a choice of η under which the entropic estimator is root-n normal. The same issue affects the second claim in Theorem 4, which requires the pointwise η condition uniformly over all π ∈ Π.","section":"Appendix D.1, Corollary D.1"},{"comment":"The empirical section reports that for 14% of units the linear program was infeasible and states 'We exclude these units from the analysis.' This exclusion changes the estimand: the reported bounds and confidence intervals are for the subpopulation for which the estimated constraints are feasible, not for the Oregon sample described in the paper. Since feasibility depends on estimated nuisance functions, this is a form of selection on estimated values and should be addressed explicitly, for example by reporting sensitivity analyses, re-weighting the feasible units, or estimating bounds under a model for the infeasible units.","section":"Section 5, Medicaid enrollment analysis"}],"minor_comments":[{"comment":"In the proof of Lemma D.3, the Jacobians for the upper-bound dual problem are labeled with subscript L (∇_b λ^η_L and ∇_c λ^η_L) after introducing G_U; these should be subscript U. This is a typographical error, but it makes the proof harder to follow.","section":"Lemma D.3 proof"},{"comment":"The notation ∥f∥_∞ for vector-valued functions is defined as max_i sup_x |f_i(x)|, but this is not stated until after the first use; a one-sentence definition earlier would improve readability.","section":"Section 2.1, notation"},{"comment":"The paragraph following Theorem 2 says one can 'expect to find an η that is larger than this minimal value for a finite sample,' but this is not a mathematical guarantee and is precisely the point where the compatibility problem with Assumption 3 arises; the text should flag this as a condition that must be verified or imposed.","section":"Section 3.2"}],"recommendation":"major_revision","confidential_remarks":"The reader's report and the stress-test analysis correctly identify the entropic estimator's regularity condition as the load-bearing weakness of the paper. The BFS estimator and the policy-learning results are more solid, and the paper is a good fit for a statistical methodology journal. In revision, I would require either a corrected Theorem 2/Corollary D.1 with explicit restrictions on α and the tail of Δ_L, or a different limiting theory for the entropic estimator, plus a principled treatment of the infeasible units in the Oregon analysis."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know: this is a solid paper with a real contribution. The conditional LP framework is clean, and the de-biased BFS estimator—using simplex output rather than vertex enumeration—is a practical step forward. The policy learning extensions are nice and the proofs are detailed. But the entropic regularized estimator's normality claim is not as general as stated. The margin condition alone doesn't support the η growth that Theorem 2 needs; the authors acknowledge this, and their Corollary D.1 shows the problem: for α≤1 no η sequence satisfies its requirements. For 1<α≤4, Theorem 2's pointwise η condition won't be met with o(n^{1/4}) growth, so you'd have to rely on Corollary D.1, which for α just above 1 demands parametric nuisance rates. That is a real limitation, not a nitpick. The Oregon application picks η=100 without checking, so those Wald intervals are outside the theorem's verified regime.\n\nWhat's new: the BFS estimator with LP-solver-based debiasing avoids vertex enumeration, and the entropic Jacobian debiasing is new. The simulation study is careful and shows the entropic estimator is not sensitive to η scaling and outperforms log-sum-exp. The framework unifies several causal inference problems, and the paper is honest—Section 3.2 flags the η tail issue explicitly, and Appendix D gives the fallback result.\n\nThe empirical analysis drops 14% of units whose LPs are infeasible. That's a selection problem; it should be reported with sensitivity analysis or repaired. Minor relative to the theory, but needs addressing.\n\nI'd send this to a serious referee. The BFS estimator and framework deserve publication; the entropic claims need either a corrected theorem (with explicit α restrictions or a different regularization) or a downgrade to a heuristic with simulation support. The author is clearly capable and the paper is worth engaging.","headline":"A useful unification and a genuinely practical BFS estimator, but the entropic estimator's root-n theory has a real gap for small margin exponents.","tokens_in":40858,"tokens_out":2334,"would_cite":true,"duration_ms":26766,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62G05","62G20","90C05"],"pacs":[],"model":"deepseek-v4-flash","headline":"Two debiased estimators make partial-identification bounds asymptotically normal at root-n and support Wald confidence intervals without vertex enumeration.","keywords":["partial identification","conditional linear programs","debiased estimation","entropic regularization","policy learning","margin condition","asymptotic normality","causal bounds"],"falsifier":"Simulate a data-generating process with two nearly identical feasible vertices, so that $P(0<\\Delta_L(X)\\le t)$ decays like $t^{0.1}$; compute the plug-in BFS estimator's bias under nuisance error $\\varepsilon$ and check whether it decreases like $\\varepsilon^{1.1}$ and whether 95% Wald intervals cover the true lower bound. If coverage drops well below nominal as $n$ grows, the margin-dependent expansion in Theorem 1 is falsified.","tokens_in":39806,"feed_emoji":"📊","tokens_out":8471,"duration_ms":109156,"temperature":0.7,"pith_summary":"This paper develops a way to estimate, test, and learn policies for parameters that are only partially identified, in the common case where bounds on the parameter can be written as expectations of solutions to linear programs whose objectives and constraints depend on covariates. It proposes two debiased estimators: a plug-in basic-feasible-solution estimator that reads de-biasing information out of an ordinary LP solver, and an entropic-regularized estimator that smooths the program so that solutions are differentiable in the nuisance functions. The paper's central theoretical claim is that both estimators are asymptotically normal at the $n^{1/2}$ rate, and therefore support Wald confidence intervals for the identified set, whenever nuisance estimates converge fast enough and a margin condition on sub-optimality gaps holds. It also shows that optimizing these estimated bounds yields policies whose excess regret is controlled by the policy class's Rademacher complexity plus the same debiased error rates. The methods are applied to the Oregon Medicaid experiment to bound counterfactual and regret estimands for emergency-department visits.","feed_headline":"Two estimators turn partial-ID bounds into root-n normal CIs","feed_subtitle":"Debiased LP plug-in and entropic smoothing yield Wald intervals without vertex enumeration.","key_machinery":"The central object is the conditional linear program $\\theta_L=\\mathbb{E}[\\min_{p\\in P(X)}\\langle c(X),p\\rangle]$ with feasible set $P(x)=\\{p\\in\\mathbb{R}_+^K:Ap=b(x)\\}$; the paper rewrites the program over basic feasible solutions $p=A_B^{-1}b(x)$. The plug-in estimator selects bases $B$ by optimizing with estimated $\\hat b,\\hat c$ and debiases using Riesz-representer corrections $\\hat\\varphi^{(b)},\\hat\\varphi^{(c)}$---weighted residual functions that make debiased estimates first-order robust to nuisance error. The entropic estimator replaces the objective with $\\langle c,p\\rangle+(1/\\eta)\\sum_k p_k(\\log p_k-1)$, whose dual variables give an explicit smooth solution and Jacobians $\\nabla_b p^\\eta,\\nabla_c p^\\eta$ used in the debiasing Taylor expansion. The argument is carried by the sub-optimality gap $\\Delta_L(x)$ and the margin condition $P(0<\\Delta_L(X)\\le t)\\lesssim t^\\alpha$, which controls the probability that plug-in bases are misclassified, and by the Rademacher complexity of the policy class for the policy-learning results.","core_discovery":"The central claim is that plug-in and entropic-regularized estimates of bounds defined by conditional linear programs can be debiased so that inference and policy learning proceed as if the bounds were smooth point-identified functionals. Theorem 1 shows the plug-in BFS estimator satisfies $\\hat\\theta_L-\\theta_L$ equals a mean-zero expansion plus $O_p((\\|\\hat b-b\\|_\\infty+\\|\\hat c-c\\|_\\infty)^{1+\\alpha}+r_n+\\|\\hat b-b\\|_2\\|\\hat c-c\\|_2)+o_p(n^{-1/2})$; Theorem 2 shows the entropic estimator satisfies an analogous expansion with an added $e^{-\\eta}$ approximation-error term. Corollaries give conditions---unique optimal bases, non-degenerate solutions, or zero conditional variance of the debiasing functions---under which the asymptotic variance depends only on true optimal bases, and the empirical variance estimators yield Wald intervals. Theorems 3 and 4 bound excess regret for estimated policies by $R_n(\\Pi)+\\|\\hat b-b\\|_\\infty^{1+\\alpha}+r_n$ and, for entropic policies, by $\\eta R_n(\\Pi)+e^{-\\eta}+r_n+\\eta^2\\|\\hat b-b\\|_2^2$. The paper applies the estimators to the Oregon health insurance experiment, estimating bounds on extra emergency-department visits under waitlist randomization versus an oracle ED-minimizing rule, and on regret under power-law collective utility functions.","pith_inferences":["The margin exponent $\\alpha$ is in practice unknown; a natural sensitivity analysis would re-estimate the bounds under worst-case small $\\alpha$ or replace normality-based intervals with bounds that do not rely on the margin condition.","Because Theorem 2's condition $\\eta\\ge (R_1(X_i)+R_H(X_i))/\\Delta_L(X_i)$ depends on unobserved sub-optimality gaps, a testable extension would estimate per-unit gaps and choose $\\eta$ from their empirical distribution, then compare coverage across choices.","The dual formulation of the entropic program suggests an immediate extension to continuous outcomes: discretize the outcome space, compute the dual variables, and check whether the resulting bounds remain valid as the grid refines, which the paper flags as an open limitation.","The excess-regret rates imply an explicit trade-off: shrinking the policy class reduces the Rademacher term but may increase the approximation error from not containing the true optimum; a data-driven rule for model class selection could be derived from Theorem 4's bound."],"forward_implications":["Analysts can build Wald confidence intervals for partially identified parameters using standard LP solver output, without enumerating the possibly combinatorial set of feasible vertices.","Flexible machine-learning estimates of constraint and objective functions are allowed: only their debiased rate and margin-adjusted errors need to vanish faster than $n^{-1/2}$, not the raw nuisance errors themselves.","The entropic regularization level $\\eta$ acts as a sensitivity parameter moving continuously between the no-assumption bounds ($\\eta\\to\\infty$) and a maximum-entropy point identification ($\\eta\\to 0$), with approximation error decaying like $e^{-\\eta}$.","For policy learning, excess regret of the estimated optimal policy is controlled by the Rademacher complexity of the policy class plus the debiased estimation error, so choosing a simpler policy class can compensate for slower nuisance convergence.","In point-identified cases such as the average treatment effect on the treated, the estimators reduce to standard debiased estimators and are insensitive to the regularization choice."],"supporting_citations":[{"why":"Establishes the LP-bounds formulation for treatment effects with imperfect compliance, which the paper generalizes to conditional linear programs.","marker":"Balke and Pearl (1997)"},{"why":"Supplies the margin-condition framework used in Assumption 3 to control misclassification of optimal bases.","marker":"Audibert and Tsybakov (2007)"},{"why":"Provides the double/debiased machine-learning rate conditions used in Assumption 2's de-biased rate $r_n$ and in the Riesz-representer construction.","marker":"Chernozhukov et al. (2018)"},{"why":"Defines the covariate-assisted intersection bounds estimator of which the plug-in BFS estimator is a special case.","marker":"Semenova (2024)"},{"why":"Develops covariate-assisted debiased bounds for instrumental variables, providing the comparison and log-sum-exp approximation used in simulations.","marker":"Levis et al. (2023)"},{"why":"Gives the exponential decay of the entropic regularization approximation error that Theorem 2 relies on.","marker":"Weed (2018)"},{"why":"Provides the method for combining one-sided intervals for lower and upper bounds into a confidence interval for the partially identified parameter.","marker":"Imbens and Manski (2004)"},{"why":"Supplies Lemma 1, which controls the empirical-process remainder terms in the asymptotic expansions.","marker":"Kennedy (2024)"},{"why":"Proposes the dual estimator discussed as a potential path to valid bounds under approximation or continuous outcomes.","marker":"Ji et al. (2024)"}],"fun_headline_variants":["Debiased LP bounds yield root-n normal CIs without vertex enumeration","Partial-ID bounds get Wald intervals via debiased conditional LPs","Entropic smoothing debiases partial-ID bound estimators for inference","Policy learning with partial ID via debiased conditional linear programs","Two debiased estimators turn partial-ID bounds into normal confidence intervals"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The central claim collapses if the margin condition fails---that is, if the gap between the best and next-best feasible solutions is frequently tiny---because then the basis-misclassification bias decays slowly; the entropic result additionally requires an unverifiable unit-level lower bound on the regularization parameter.","fun_headline_variants_meta":{"raw":{"variants":["Debiased LP bounds yield root-n normal CIs without vertex enumeration","Partial-ID bounds get Wald intervals via debiased conditional LPs","Entropic smoothing debiases partial-ID bound estimators for inference","Policy learning with partial ID via debiased conditional linear programs","Two debiased estimators turn partial-ID bounds into normal confidence intervals"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000196,"raw_usage":{"total_tokens":1447,"prompt_tokens":1117,"completion_tokens":330,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":733,"completion_tokens_details":{"reasoning_tokens":240}},"tokens_in":733,"tokens_out":330,"duration_ms":78238,"temperature":1.0,"reasoning_tokens":240,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T00:55:12.998724+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Simulate a data-generating process with two nearly identical feasible vertices, so that $P(0<\\Delta_L(X)\\le t)$ decays like $t^{0.1}$; compute the plug-in BFS estimator's bias under nuisance error $\\varepsilon$ and check whether it decreases like $\\varepsilon^{1.1}$ and whether 95% Wald intervals cover the true lower bound. If coverage drops well below nominal as $n$ grows, the margin-dependent expansion in Theorem 1 is falsified.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the margin-condition framework used in Assumption 3 to control misclassification of optimal bases."},{"cited_title":"Debiased Machine Learning of Aggregated Intersection Bounds and Other Causal Parameters","cited_arxiv_id":"2303.00982","evidence_quote":"Defines the covariate-assisted intersection bounds estimator of which the plug-in BFS estimator is a special case."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Gives the exponential decay of the entropic regularization approximation error that Theorem 2 relies on."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the method for combining one-sided intervals for lower and upper bounds into a confidence interval for the partially identified parameter."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies Lemma 1, which controls the empirical-process remainder terms in the asymptotic expansions."}],"review_version":1}