{"id":"98cff830-f9a4-43bd-be5d-e162bc0fab72","arxiv_id":"2505.13092","paper_version":2,"verdict":"REJECT","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"The paper introduces a tunable, policy-targeted CATE estimator (PT-CATE) that trades prediction error against decision performance, though its central suboptimality proof is incomplete.","lead":"The paper argues that standard two-stage estimators of conditional average treatment effects, which are optimized for estimation accuracy, can lead to suboptimal treatment decisions when the model cannot exactly represent the true effect. It proposes a retargeted objective that balances CATE accuracy with expected decision outcome and a neural algorithm that implements it.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 4.1's proof is invalid: the constructed supremum M is infinite because S_G contains arbitrarily large constant CATEs, so the sequence argument breaks; the theorem as stated lacks a boundedness condition.","rationale":"The reader's weakest_assumption correctly identifies the load-bearing flaw in Theorem 4.1: the proof's supremum M over S_G is not finite, because S_G contains all positive constant functions for standard neural network classes, so the sequence converging to M and the step '||tau*||_inf > M' are not well-defined. This is not a minor technicality; it is the mechanism by which the proof claims existence of a CATE whose optimal L2 approximation gets the sign wrong on a set of positive measure. Without a boundedness assumption on the CATE or a different construction, the theorem has no valid proof. The central claim that two-stage CATE estimators can be suboptimal for decision-making therefore rests on an unsupported theoretical result, even though the empirical demonstrations are plausible and the proposed PT-CATE objective is reasonable. I also note the secondary issue in Theorem 4.4 that the bound leaves the optimization error R^m uncontrolled; this reinforces the conclusion that the current manuscript does not substantiate its theoretical guarantees. Because my concern matches the reader's and the reader's REJECT verdict is consistent with the unproved central theorem, I recommend no change to the verdict.","tokens_in":41734,"tokens_out":6368,"duration_ms":67483,"concrete_test":"Instantiate the proof with a concrete G and verify whether the construction can be repaired. Take X ~ Uniform[0,1], G = {g(x) = ax + b} (affine networks), and let S_G be defined as in Eq. (15). First check that the constant functions tau_c(x) = c for c > 0 are in S_G (their L2 projection is the constant c, sign correct), so M = inf and the current sequence argument is invalid. Then test the theorem's substance with a bounded alternative: for B = 1, define S_G^B = {tau in S_G : ||tau||_inf <= 1}, and check whether sup_{tau in S_G^B} ||tau||_inf < 1 or equals 1.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central theoretical claim is Theorem 4.1 (Sec. 4.1), proved in Appendix B.1. The proof defines S_G (Eq. 15) as step-function CATEs whose L2-optimal approximation in G has the correct sign everywhere, then sets M = sup_{tau in S_G} ||tau||_inf (Eq. 16). It then relies on a sequence tau_n in S_G with ||tau_n||_inf -> M (Eq. 17), on picking n_eps with ||tau_{n_eps}||_inf > M - eps (Eq. 18), and on constructing tau* with ||tau*||_inf > M (Eq. 24) to conclude tau* not in S_G. This argument requires M < inf. But S_G is unbounded: for any constant c > 0, the constant function tau_c(x) = c is a step function, and its L2 projection onto G (which contains constants for standard neural architectures) is itself, with correct sign everywhere; hence tau_c is in S_G for all c > 0, making M = inf. Then 'M - eps', '||tau*||_inf > M', and the definition of the sequence are meaningless. No boundedness assumption on tau or on the outcome appears in Theorem 4.1, and the proof does not restrict S_G. Thus the contradiction establishing tau* not in S_G is not derived; the theorem is unproved as stated. Since Theorem 4.1 is the paper's headline justification for why two-stage learners can be suboptimal for decisions, this is a load-bearing gap. A secondary issue: the Theorem 4.4 bound contains the uncontrolled optimization-dependent term R^m; even a repaired Theorem 4.1 would not by itself establish the error-rate claim.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript studies the common practice of thresholding CATE estimates to make binary treatment decisions. It argues that two-stage CATE learners, which minimize L2 error, may be suboptimal for decision-making when the second-stage model class is misspecified. It introduces the PT-CATE objective, a convex combination of L2 error and policy value, with an adaptive sigmoid approximation alpha(x), and proposes a three-step neural training algorithm. The paper claims three theoretical results: Theorem 4.1 (existence of CATEs for which L2-optimal CATE estimates are decision-suboptimal), Theorem 4.3 (consistency under oracle nuisances), and Theorem 4.4 (error rates under nuisance estimation error). Experiments on synthetic data and the Hillstrom dataset show improved policy loss for larger gamma at a modest PEHE cost.","tokens_in":42116,"tokens_out":12780,"duration_ms":132600,"significance":"The paper addresses a relevant and timely question: the disconnect between CATE accuracy and downstream decision performance. The PT-CATE objective is intuitive, the adaptive-sigmoid optimization trick is a reasonable practical proposal, and the empirical study is fairly thorough (four pseudo-outcomes, sample splitting, alternative nuisance models, and real-world data). The code is released, which is a concrete strength. If the theoretical claims are corrected, the paper would make a useful contribution. As it stands, however, the headline theoretical result is not established, so the theoretical contribution is currently limited.","major_comments":[{"comment":"The proof of Theorem 4.1 is invalid as stated because M is infinite. S_G in Eq. (15) contains, as the proof itself notes, every positive constant step function. For a fixed-architecture network that can represent constants, the constant tau_c(x)=c is in S_G with ||tau_c||_infty=c for every c>0, so M := sup_{tau in S_G} ||tau||_infty = infinity. Consequently, the sequence in Eq. (17), the selection of n_epsilon in Eq. (18), and the construction in Eq. (24) of tau* with ||tau*||_infty > M are not well-defined. No boundedness assumption on tau or Y appears in Theorem 4.1, so this is a genuine gap in the central argument, not merely a missing regularity line.","section":"Appendix B.1, Eq. (16)"},{"comment":"Even if M were finite, the proof's move from tau* not in S_G to a set E of positive measure on which sign(g_{tau*}) differs from sign(tau*) does not follow. By Eq. (15), S_G requires sign agreement at every x; a function whose optimal approximation agrees except on a null set is already outside S_G. The theorem needs a definition of S_G in terms of sign agreement up to null sets, or an additional argument that the mismatch set has positive measure, because the strict policy-value inequality in Eq. (29) depends on it.","section":"Theorem 4.1 / Eq. (15)"},{"comment":"The theorem is stated for an arbitrary fixed-architecture network class G, but no non-degeneracy condition is given. For the degenerate class consisting only of constant functions, every L2-optimal estimate is constant, the induced policy is either treat-all or treat-none, and pi_{g*_tau} is optimal within Pi_G; the claimed strict inequality cannot hold. The statement needs to be restricted (for example, to classes G that contain a sign-changing function, or to a concrete class such as linear functions), or replaced by an existence result for a specific class that is sufficient for the paper's qualitative claim.","section":"Theorem 4.1 statement"},{"comment":"The displayed bound contains the optimization-dependent term R^m. Since \\hat g is defined as the minimizer of the estimated loss and g* is in G, R^m is non-positive if optimization is exact; if optimization is inexact, no control on R^m is supplied. The sentence in Section 4.3 that good nuisance estimation suffices for a good second-stage learner therefore does not follow from Eq. (11) alone. The theorem should state explicitly how R^m is handled, for example by assuming a bound on optimization error or by proving that R^m is negligible under the stated conditions.","section":"Theorem 4.4, Eq. (11)"},{"comment":"Theorems 4.3 and 4.4 are stated for a fixed alpha, but Algorithm 1 alternately trains alpha (Step 2) and g (Step 3). The final g produced by the algorithm is a minimizer of a data-dependent loss in which alpha is no longer fixed. The paper does not provide a guarantee for this alternating procedure, so the theory and the algorithm are not directly connected. The authors should either extend the guarantees to the alternating scheme or explicitly state that the theory applies only to the fixed-alpha population objective and that the algorithm is a heuristic approximation.","section":"Section 4.4 / Algorithm 1"}],"minor_comments":[{"comment":"The symbol tau*_{n_epsilon} is not defined and appears to be a typo for tau_{n_epsilon}.","section":"Appendix B.1, Eq. (22)"},{"comment":"The binary cross-entropy term uses labels 1(tau(X)>0) and 1(tau(X)<0), but the case tau(X)=0 is not specified; this should be clarified, even if it is a measure-zero event under the assumed data-generating process.","section":"Section 4.2, Eq. (8)"},{"comment":"There is a typo in 'overregluarizing' in the discussion of Bonvini et al.; the text should read 'overregularizing'.","section":"Section 2"},{"comment":"The sentence 'We also compare the as compared to the behavioral policy' is ungrammatical and should be rewritten.","section":"Section 5"},{"comment":"The step marked with 'proportional to' hides the constant E[(Y^m_eta - tau(X))^2] and the cancellation of the cross term via E[Y^m_eta - tau(X) | X]=0; this is correct, but the notation is sloppy and should be made explicit.","section":"Theorem 4.3 proof, Eq. (35)"}],"recommendation":"major_revision","confidential_remarks":"The stress-test concern about the proof of Theorem 4.1 is correct and lands: M is infinite because S_G contains all positive constant functions. I would not, however, go as far as rejection. The qualitative suboptimality claim is likely repairable by restricting the theorem to a concrete class such as linear functions (as in the authors' own toy example) and adding boundedness assumptions, and the empirical part is solid. Please require the authors to restate Theorem 4.1, repair or replace its proof, and clarify the role of R^m and the alternating alpha-updates in the theory. The current version cannot be accepted as is."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Let me give you my take on arXiv:2505.13092. The paper asks a good question: if you're going to threshold a CATE estimate to make a policy, maybe you shouldn't train the CATE estimator purely for L2 accuracy. The PT-CATE objective (convex combination of L2 error and policy value) and the adaptive sigmoid alpha(X) trick are sensible, and the three-step algorithm is clearly described. The experiments are more thorough than most: four pseudo-outcomes, sample splitting, TARNet robustness, and a real dataset. The empirical pattern is consistent: increasing gamma trades a little PEHE for a noticeably better policy loss.\n\nBut the theory does not hold up as stated. Theorem 4.1 is the paper's headline justification for why two-stage learners are suboptimal for decisions. The proof in Appendix B.1 defines S_G as step functions whose L2-optimal approximation has the right sign everywhere, then sets M = sup_{tau in S_G} ||tau||_inf. The problem: S_G contains all positive constant functions, so M = infinity. You can't choose the sequence tau_n with ||tau_n|| -> M, and 'M - epsilon' is meaningless. No boundedness assumption appears in the theorem. The stress-test note is right: this is a load-bearing gap. The theorem as stated is unproved.\n\nA secondary issue: the consistency theorem (Thm 4.3) assumes perfectly estimated nuisances, so it's a tower-property exercise, and the error-rate theorem (Thm 4.4) carries an uncontrolled optimization-dependent term R^m, so it doesn't deliver a real rate. These are minor compared to Thm 4.1, but they mean the paper's theoretical contributions are currently just the objective and algorithm, not the claimed optimality gap.\n\nThat said, the gap looks fixable. If you add a boundedness condition on tau (say ||tau||_inf <= B), M becomes finite and the argument might go through, though you'd also need to handle the non-uniqueness of argmin when claiming the result for every optimal L2 approximation. I'd send this to peer review rather than desk-reject, because the question is important, the method is new, and the empirical section is honest and reproducible (code is linked). The authors need to repair the theorem or demote it to a conjecture supported by examples. For a reader, the paper is worth engaging with, but only with the proof gap in mind.","headline":"The paper's core theorem has a load-bearing proof gap (unbounded M), but the proposed retargeting objective and empirical evaluation make it worth a serious revision.","tokens_in":42647,"tokens_out":3231,"would_cite":true,"duration_ms":33093,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Accurate CATE estimates can still yield wrong treatment decisions when the model class is misspecified, and a retargeted objective fixes the mismatch.","keywords":["conditional average treatment effects","two-stage meta-learners","policy learning","thresholding policies","doubly robust estimation","neural networks","treatment effect heterogeneity","off-policy learning"],"falsifier":"Recompute the quantity $M$ in the proof of Theorem 4.1 for a specific function class such as linear functions on $[0,1]$; since every positive constant belongs to the set $S_G$, the supremum $M$ is infinite and the sequence $\\{\\tau_n\\}$ with norms approaching $M$ cannot be chosen, so the proof as written fails unless a boundedness assumption is added.","tokens_in":41482,"feed_emoji":"🎯","tokens_out":5588,"duration_ms":53822,"temperature":0.7,"pith_summary":"Medical and other high-stakes decisions are often made by estimating the conditional average treatment effect (CATE) and treating everyone whose estimated CATE is positive. This paper proves that a two-stage CATE estimator (such as the DR-learner) that is optimal for estimating the CATE can still produce a strictly suboptimal thresholded policy when the true CATE falls outside the model class used in the second stage. The reason is that squared-error estimation spends effort on regions far from the decision boundary, where mistakes do not change the decision. To fix this, the paper introduces the policy-targeted CATE (PT-CATE), a retargeted estimand that blends CATE accuracy with policy value, and a neural procedure with an adaptively smoothed decision indicator. If correct, the work gives practitioners a principled way to keep interpretable CATE-based decisions while approaching the decision quality of direct policy optimization.","feed_headline":"Accurate CATE estimates can still yield wrong treatment decisions","feed_subtitle":"A retargeted learning objective fixes the mismatch by balancing estimation error against decision value.","key_machinery":"The load-bearing object is the $\\gamma$-policy-targeted CATE ($\\gamma$-PT-CATE), defined as the minimizer of $L_\\gamma(g) = (1-\\gamma)\\mathbb{E}[(\\tau(X)-g(X))^2] - \\gamma\\,\\mathbb{E}[\\mathbf{1}(g(X)>0)\\tau(X)]$. To make this objective differentiable, the paper replaces the indicator with a sigmoid $\\sigma(\\alpha(X)g(X))$, where $\\alpha(X)>0$ is a learned covariate-dependent sharpness parameter. The function $\\alpha$ is trained through a weighted binary-cross-entropy surrogate so that it becomes large where the current estimate already has the correct sign and small where the sign is wrong, thereby giving gradient signal exactly in decision-relevant regions. The second stage can use plug-in, RA, IPW, or DR pseudo-outcomes, and the DR version inherits a doubly robust error rate: the PT-CATE error is controlled as long as either the propensity score or the response functions are estimated well.","core_discovery":"The central claim is Theorem 4.1: for any fixed neural-network class $G$, there exists a true CATE $\\tau^*$ outside $G$ such that every $L^2$-optimal approximation $g^*_{\\tau^*}$ in $G$, when thresholded at zero, yields a policy with strictly lower value than the best thresholded policy available in $G$. Equivalently, minimizing mean-squared error of the CATE and maximizing the value of the induced thresholded policy are different objectives, and the former does not dominate the latter under model misspecification. The proposed remedy is the $\\gamma$-PT-CATE loss $L_\\gamma(g) = (1-\\gamma)\\mathbb{E}[(\\tau(X)-g(X))^2] - \\gamma\\,\\mathbb{E}[\\mathbf{1}(g(X)>0)\\tau(X)]$, whose minimizer interpolates between pure CATE estimation ($\\gamma=0$) and pure policy-value maximization ($\\gamma=1$). The paper shows consistency when pseudo-outcomes are used and an error bound with doubly robust nuisance rates, and it demonstrates empirically that increasing $\\gamma$ reduces policy loss while only mildly worsening CATE error.","pith_inferences":["Editorial inference: The same mismatch likely appears for other decision rules built from CATE estimates, such as top-k treatment allocation or cost-sensitive thresholds, wherever squared-error loss weights regions by outcome variance rather than by decision consequence.","Editorial inference: The learned sharpness parameter $\\alpha(X)$ can be read as a soft confidence measure for the decision, suggesting a testable extension where $\\alpha$ is reported alongside the thresholded policy to quantify decision uncertainty.","Editorial inference: A natural benchmark is to compare the PT-CATE curve of policy value versus PEHE against direct off-policy learning on datasets with strong misspecification; the paper's experiments suggest the trade-off curve dominates standard learners, but a systematic comparison across public benchmarks is not included."],"forward_implications":["Two-stage CATE learners that target squared error can be suboptimal for thresholded decisions whenever the second-stage model class is misspecified, even if the estimators are otherwise state-of-the-art.","The $\\gamma$-PT-CATE loss provides a continuous trade-off: $\\gamma=0$ recovers standard CATE estimation, $\\gamma=1$ becomes direct policy optimization, and intermediate values sacrifice a little CATE accuracy for better decisions.","The DR pseudo-outcome version of the PT-CATE loss carries a doubly robust guarantee, so the retargeted estimand can be learned reliably even if only one of the nuisance components is accurately estimated.","In the real-data experiment, the retargeted policy improves response probability by about 24.45% over a standard CATE-based policy at $\\gamma=0.98$, with modest increases in estimated PEHE.","If the second-stage class is unrestricted and the true CATE lies in it, the method reduces to ordinary CATE estimation and introduces no additional bias."],"supporting_citations":[{"why":"Supplies the DR-learner and the two-stage estimation principle that the paper shows can be suboptimal for decision-making.","marker":"[32]"},{"why":"Provides the early doubly robust learner formulation that the paper builds on as a state-of-the-art CATE estimator.","marker":"[53]"},{"why":"Establishes the pseudo-outcome framework and the identity that each pseudo-outcome is conditionally unbiased for the CATE, which underlies the consistency theorem.","marker":"[8]"},{"why":"Formalizes thresholding a CATE estimate as a decision rule and connects CATE sign to policy value.","marker":"[12]"},{"why":"Gives minimax-optimality results for thresholded two-stage meta-learners under well-specified models, which the paper contrasts with its misspecification setting.","marker":"[6]"},{"why":"Provides the doubly robust policy evaluation surrogate used to derive the loss for learning the adaptive sharpness parameter.","marker":"[10]"}],"fun_headline_variants":["CATE accuracy ≠ decision accuracy: retargeting fixes it","Retarget CATE to optimize decisions, not just estimation","When perfect CATE still gives suboptimal policies","Balancing CATE error and policy value improves decisions","Two-stage CATE: estimation-optimal, decision-suboptimal"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The proof of the suboptimality theorem assumes a finite upper bound on the size of a certain family of step-function treatment effects, yet that family includes arbitrarily large constant functions, so the bound is not finite and the constructed counterexample does not go through.","fun_headline_variants_meta":{"raw":{"variants":["CATE accuracy ≠ decision accuracy: retargeting fixes it","Retarget CATE to optimize decisions, not just estimation","When perfect CATE still gives suboptimal policies","Balancing CATE error and policy value improves decisions","Two-stage CATE: estimation-optimal, decision-suboptimal"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000266,"raw_usage":{"total_tokens":1631,"prompt_tokens":983,"completion_tokens":648,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":599,"completion_tokens_details":{"reasoning_tokens":568}},"tokens_in":599,"tokens_out":648,"duration_ms":6095,"temperature":1.0,"reasoning_tokens":568,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T20:20:26.958602+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Recompute the quantity $M$ in the proof of Theorem 4.1 for a specific function class such as linear functions on $[0,1]$; since every positive constant belongs to the set $S_G$, the supremum $M$ is infinite and the sequence $\\{\\tau_n\\}$ with norms approaching $M$ cannot be chosen, so the proof as written fails unless a boundedness assumption is added.","supporting_citations":[{"cited_title":"Towards optimal doubly robust estimation of heterogeneous causal effects","cited_arxiv_id":null,"evidence_quote":"Supplies the DR-learner and the two-stage estimation principle that the paper shows can be suboptimal for decision-making."},{"cited_title":"Statistical inference for variable importance","cited_arxiv_id":null,"evidence_quote":"Provides the early doubly robust learner formulation that the paper builds on as a state-of-the-art CATE estimator."},{"cited_title":"Nonparametric estimation of heterogeneous treat- ment effects: From theory to learning algorithms","cited_arxiv_id":null,"evidence_quote":"Establishes the pseudo-outcome framework and the identity that each pseudo-outcome is conditionally unbiased for the CATE, which underlies the consistency theorem."},{"cited_title":"Causal classification: Treatment effect estimation vs. outcome prediction","cited_arxiv_id":null,"evidence_quote":"Formalizes thresholding a CATE estimate as a decision rule and connects CATE sign to policy value."},{"cited_title":"Doubly robust policy evaluation and learning","cited_arxiv_id":null,"evidence_quote":"Provides the doubly robust policy evaluation surrogate used to derive the loss for learning the adaptive sharpness parameter."}],"review_version":1}