{"id":"536d5eb3-28c7-43d5-a5f8-b75fffc67700","arxiv_id":"2412.19711","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"The authors extend the DR-learner and EP-learner to handle outcomes missing at random by adding inverse-probability-of-censoring weights, and show the resulting estimators are oracle efficient.","lead":"This paper proposes two new statistical estimators that measure how differently a treatment affects different patient subgroups when some patients' outcomes are missing. They are designed to correct the bias that missing data create in modern machine-learning-based treatment effect estimates.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"mEP-learner targeting step solves the wrong estimating equation: weight H(A,C,Z) lacks the (2A-1) factor in the EIF, so the drift term is not zeroed and oracle efficiency is unsupported.","rationale":"The reader correctly identified the mEP-learner's oracle-efficiency claim as the weakest assumption, but our concern is stronger: the algorithm as written does not satisfy even the stated mechanism (zero sample-average drift). The EIF coefficient is (2A-1)H, while the targeting regression uses weight H, so the score equation being solved differs from the required one. This is a concrete algebraic mismatch, not merely a missing proof. The mDR-learner's excess-risk bound (Theorem 1, Appendix E) appears internally consistent and is not affected by this issue. Since the paper's central claim covers both estimators, the mEP failure requires a major revision: either correct the targeting weight (e.g., use (2A-1)H or equivalently the full (A-pi)C/[pi(1-pi)G] coefficient), provide a proof with the corrected algorithm, or weaken the claim. Until then, the mEP oracle-efficiency result should not be accepted as stated. A CONDITIONAL verdict is appropriate because the mDR contribution is credible and the mEP issue is potentially fixable, but the current manuscript overstates its findings. The proposed simulation check would settle whether the drift-zero property fails empirically, corroborating the algebraic derivation.","tokens_in":36636,"tokens_out":13007,"duration_ms":120377,"concrete_test":"Use a simple DGP (n=2000, X~U(-1,1), pi=0.5, G=0.7 constant, mu_0=mu_1=0, Y~N(0,1), C~Bern(0.7)) and run Algorithm 2 exactly, with a cosine sieve phi_k(X)=cos(k arccos(X)) for k=1..5. After the targeting step, compute T_k = (1/n) sum_i (2A_i-1) H_i phi_k(X_i) (Y_i - mu_{A_i}*) for each k. If max|T_k| is not negligible (e.g., >0.05) while the same quantities from a corrected regression with weight (2A-1)H are approximately 0, the drift-zero property in Section 3.2 fails. Report the same check for the complete-data EP-learner in Appendix C.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The claim that the mEP-learner is oracle efficient rests on Section 3.2's assertion that the sieve-based iTMLE update with weight H(A,C,Z)=CA/(G pi)+C(1-A)/(G(1-pi)) 'ensures that the sample average of the second term in eq. (7) converges to 0.' This assertion is not merely unproved; it is inconsistent with Algorithm 2. The EIF drift term in eq. (7) is -2 theta(X)(A-pi)C/[pi(1-pi)G](Y-mu_A) = -2 theta(X)(2A-1)H(Y-mu_A). Algorithm 2 runs a weighted least squares regression of Y on phi(X) with offset mu_A and weight H, whose score equation is sum H_i phi_i (Y_i - mu_{A_i} - eps^T phi_i) = 0, i.e., sum theta_i H_i (Y_i - mu_{A_i}*) = 0 for theta in the sieve. The required drift-zero condition is sum theta_i (2A_i-1) H_i (Y_i - mu_{A_i}*) = 0. With the update mu_{A_i}* = mu_{A_i} + (2A_i-1) eps^T phi_i, the empirical drift becomes -2 sum theta_i(2A_i-1)H_i(Y_i-mu_{A_i}) + 2 sum theta_i H_i eps^T phi_i = 4 sum_{A_i=0} theta_i H_i (Y_i - mu_{0,i}), which is not zero in general. Thus the targeting regression as specified does not set the EIF drift to zero; the mEP-learner does not inherit the EP-learner's oracle-efficiency guarantee. The same apparent sign issue appears in the complete-data EP-learner in Appendix C. The mDR-learner proof (Appendix E) is a separate argument and appears internally consistent.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper addresses estimation of the conditional average treatment effect (CATE) when outcome data are missing at random. It proposes two estimators: the mDR-learner, which incorporates inverse probability of censoring weights into the DR-learner pseudo-outcomes, and the mEP-learner, which modifies the EP-learner's targeting step with a censoring-weighted clever covariate. The authors claim oracle efficiency for both estimators under reasonable nuisance convergence rates. The mDR-learner's excess-risk bound is derived in detail in Appendix E following Foster and Syrgkanis (2023), and simulation studies and a GBSG2 data application are used to illustrate performance. The mDR-learner derivation appears internally consistent, but the mEP-learner's oracle-efficiency claim is asserted without proof, and the targeting weight used in Algorithm 2 does not set the EIF drift term to zero as stated.","tokens_in":36956,"tokens_out":8375,"duration_ms":62711,"significance":"The mDR-learner is a solid contribution: its pseudo-outcomes follow from the efficient influence function of the MSE risk under MAR, and the excess-risk bound in Appendix E gives a concrete second-order robustness guarantee that includes the missingness model. The paper also provides reproducible code and a real-data illustration, which are strengths. If the mEP-learner's targeting step is corrected and its oracle-efficiency property is rigorously established, the paper would offer two practical debiased learners for a common missing-data problem. As it stands, the mDR half of the paper is well supported, while the mEP half lacks the theoretical grounding necessary for the paper's central claim.","major_comments":[{"comment":"The targeting weight H(A,C,Z)=CA/(G pi)+C(1-A)/(G(1-pi)) is not the correct clever covariate for the EIF in eq. (7). The EIF drift term is -2 theta(X)(A-pi)C/[pi(1-pi)G](Y-mu_A), whose treatment-arm coefficient is C/(G pi) and control-arm coefficient is -C/(G(1-pi)): the correct weight is CA/(G pi) - C(1-A)/(G(1-pi)). With the weight as defined, the weighted least squares regression in Algorithm 2 has score equation sum H_i phi_i (Y_i - mu_{A_i} - eps^T phi_i)=0. For any theta in the sieve, this gives sum theta_i H_i (Y_i - mu_{A_i} - eps^T phi_i)=0. The required drift-zero condition after the update mu_{A}* = mu_A + (2A-1)eps^T phi is sum theta_i (2A_i-1) H_i (Y_i - mu_{A_i,i}*)=0, which reduces, using the score equation, to a generally nonzero residual sum over the untreated complete cases (sum_{A_i=0} theta_i H_i (Y_i - mu_{0,i}) up to an overall sign). Therefore Algorithm 2 does not set the sample average of the second term in eq. (7) to zero. The same sign issue appears in the complete-data EP-learner in Appendix C. The manuscript's assertion in Section 3.2 that this weight 'ensures that the sample average of the second term in eq. (7) converges to 0' is unsupported.","section":"Section 3.2 and Algorithm 2"},{"comment":"The oracle-efficiency claim for the mEP-learner is stated without proof. Unlike the mDR-learner, which has the detailed excess-risk analysis in Appendix E and Theorem 1, the mEP-learner is only asserted to be oracle efficient on the basis of the targeting step. Even if the sign of the clever covariate were corrected, there is no theorem showing that the sieve-based iTMLE update makes the drift term vanish in the required metric, nor is there an excess-risk bound analogous to the mDR case. This missing proof is load-bearing because the abstract and introduction claim oracle efficiency for both estimators.","section":"Section 3.2"},{"comment":"The simulation evaluation emphasizes the root mean square median error (RMSME) because the mean RMSE is skewed by rare extreme estimates, but this metric obscures the instability that the paper itself identifies as a weakness of the mDR/mEP learners. Appendix K shows that the mean RMSE of mEP is often larger than that of mDR (e.g., DGP 3 at n=1600 and n=3200), which is not visible in the RMSME plots. The paper should present both metrics in the main text or clearly justify the choice of RMSME as the primary metric, and the stability claims in Section 4.2 should be qualified accordingly.","section":"Section 4 and Appendix K"}],"minor_comments":[{"comment":"The phrase 'Split the data randomly into K (e.g., 10) equal sized folds of n observations' is imprecise; each fold contains approximately n/K observations, not n observations.","section":"Algorithm 1, step 1"},{"comment":"In the complete-data EP-learner, the outcome models are written as mu_0(Z)=E[Y|A=0, C=1, Z] and mu_1(Z)=E[Y|A=1, C=1, Z], but the variable C is not defined in the no-missing-data setting; this appears to be a copy-paste error from the mEP-learner algorithm.","section":"Appendix C, Algorithm 2"},{"comment":"The statement that the mEP-learner and EP-learner are more stable than the mDR-learner should be tied more explicitly to the RMSME metric, since the mean RMSE results in Appendix K show a different ordering in some settings.","section":"Section 4.2"}],"recommendation":"major_revision","confidential_remarks":"The mEP-learner is the weaker half of the paper. The sign error in the clever covariate is easy to fix, but the absence of any proof of oracle efficiency for the targeted learner is a substantive gap that will require a new theorem and a re-run of the simulations for the corrected estimator. The mDR contribution, by contrast, is solid and could stand on its own. I would not reject the paper on the strength of the mDR alone, but the revision must address the mEP targeting issue rigorously."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Read the paper. The mDR-learner is a genuine, useful extension of the DR-learner to MAR outcomes: IPCW-weighted pseudo-outcomes from the EIF of the MSE risk, with an excess-risk bound in Appendix E that follows Foster-Syrgkanis and appears correct. The orthogonality checks in Appendix F go through. That part deserves serious attention.\n\nThe mEP-learner is another story. Algorithm 2's targeting step regresses Y on the sieve φ(X) with offset µ_A and weight H(A,C,Z) = CA/(Gπ) + C(1-A)/(G(1-π)), then updates µ_1* = µ_1 + ε·φ and µ_0* = µ_0 - ε·φ. The EIF drift in eq. (7) is -2θ(A-π)C/[π(1-π)G](Y-µ_A). But (A-π)C/[π(1-π)G] = (2A-1)H, so zeroing the drift requires sum (2A-1)H φ (Y - µ_A - (2A-1)εφ) = 0, i.e. sum (2A-1)H φ (Y-µ_A) = sum H φ φ^T ε. The regression as written solves sum H φ (Y-µ_A - εφ) = 0, which is a different equation. The claim that the weight \"ensures\" the drift goes to zero is not supported; the same sign issue appears in the complete-data EP-learner in Appendix C. So the mEP's oracle-efficiency claim is not merely unproved, it's inconsistent with the algorithm.\n\nThe simulations and example are helpful but need caveats. The RMSME metric is median-based and hides the DR-learner's heavy tails; mean RMSE in Appendix K shows more instability. The half-sample bootstrap CIs have very poor conditional coverage in some subgroups (min in single digits), which the authors honestly flag. The GBSG2 analysis is illustrative, not confirmatory.\n\nNet: the mDR-learner is a solid incremental contribution with a formal excess-risk bound and reproducible code. The mEP-learner either needs a corrected targeting step or a proof that the current step works; as is, the claim doesn't hold. The paper deserves a serious referee, but the referee should push on the mEP and on uncertainty reporting. If the mEP can't be fixed, the paper can still stand on the mDR with an explicit caveat.","headline":"The mDR-learner is a solid IPCW extension with a real excess-risk bound, but the mEP-learner's targeting step as written does not zero the EIF drift, so its oracle-efficiency claim is unsupported.","tokens_in":54,"tokens_out":6195,"would_cite":true,"duration_ms":473464,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62D20","62G05","62G20"],"pacs":[],"model":"deepseek-v4-flash","headline":"Two modified learners estimate treatment effects without bias when outcomes are missing at random.","keywords":["causal machine learning","heterogeneous treatment effects","missing outcome data","missing at random","oracle efficiency","doubly robust estimation","targeted learning","inverse probability of censoring weights"],"falsifier":"Run the mEP-learner on data where the true CATE is a step function that the trigonometric sieve basis cannot approximate, with fast nuisance estimators; if the sample average of the drift term after the targeting step is not zero, or if the excess risk fails to shrink at the oracle rate as n grows, the mEP oracle-efficiency claim is refuted. The same check applies to the mDR-learner, where the bound is proven and should hold.","tokens_in":36343,"feed_emoji":"📊","tokens_out":6293,"duration_ms":52969,"temperature":0.7,"pith_summary":"This paper tackles a common but underappreciated problem: when outcome data is missing at random (MAR), standard machine-learning estimators of the conditional average treatment effect (CATE) can give biased estimates for subgroups with high dropout. The authors propose two modified learners, the mDR-learner and the mEP-learner, that multiply the usual correction term by the inverse probability of censoring, so that complete cases are reweighted to represent the full population. They prove that the mDR-learner is oracle efficient under rate conditions, meaning its estimation error behaves as if the nuisance functions were known, and they argue the same holds for the mEP-learner. In simulations the new learners outperform available-case and imputation-based versions of standard learners, and in the GBSG2 breast cancer example they yield smaller treatment-effect estimates than available-case analyses, consistent with correction for selective dropout.","feed_headline":"Treatment-effect ML fixed for missing outcome data","feed_subtitle":"New mDR and mEP estimators reweight by censoring odds to keep subgroup estimates unbiased.","key_machinery":"The efficient influence function of the mean-squared-error risk for the CATE under missing outcomes, given in the paper as eq. (7), whose correction term is (A - pi(Z)) C / (pi(Z)(1 - pi(Z)) G(A,Z)) times (Y - mu_A(Z)). The mDR-learner builds its pseudo-outcomes directly from this EIF; the mEP-learner uses the same expression to define the weight H(A,C,Z) = C A/(G pi) + C(1-A)/(G(1-pi)) in a weighted linear regression onto a sieve basis $\\varphi$(X), updating the outcome predictions so the sample average of the drift term converges to zero. The excess-risk bound of Theorem 1, obtained via the orthogonal statistical learning framework, carries the oracle-efficiency argument.","core_discovery":"The paper's central claim is that incorporating inverse probability of censoring weights into the DR-learner and EP-learner pseudo-outcomes produces estimators whose excess risk is governed by products of nuisance-function errors, not by the errors themselves. Theorem 1 bounds the mDR-learner's excess risk by the oracle regression rate plus terms such as $Rate_pi^{4}$, $Rate_pi^{2}$ times $Rate_mu^{2}$, $Rate_pi^{2}$ times $Rate_G^{2}$, $Rate_mu^{2}$ times $Rate_G^{2}$, and $Rate_G^{4}$; when those product rates beat the oracle rate, the learner is oracle efficient. The same oracle-efficiency claim is made for the mEP-learner, whose sieve-based iTMLE update uses the weight H(A,C,Z) = C A/(G pi) + C(1-A)/(G(1-pi)) to zero out the drift term. The authors support the theoretical claim with simulations in which the mDR and mEP learners match or beat available-case and imputation-based versions of the standard learners, and with a GBSG2 analysis showing that correcting for selective dropout reduces the estimated benefit of hormonal therapy.","pith_inferences":["If the mEP target step is verified, the mEP-learner's stability advantage (fewer extreme pseudo-outcomes) would make it the safer default in finite samples; until then the mDR-learner rests on a proven bound.","The paper's Appendix N sketch suggests the mDR construction generalizes to longitudinal settings by multiplying the correction term by cumulative censoring weights, so the same EIF logic may transfer to post-baseline missingness.","The half-sample bootstrap intervals, as applied here, had poor conditional coverage for some individuals; users should treat the interval estimates as less mature than the point-estimate theory.","A natural testable extension is to replace the fixed trigonometric sieve in mEP with a data-adaptive basis chosen by cross-validation; if the targeting step still zeros the drift term, oracle efficiency may hold for a wider class of CATEs."],"forward_implications":["Under MAR outcomes, mDR and mEP provide CATE estimates that do not require imputation and remain valid when subgroups with high dropout are under-represented in the complete cases.","When propensity and missingness models converge fast, slow outcome learners can still be used without breaking oracle efficiency, because nuisance errors enter only as products.","In the GBSG2 example, correcting for selective missingness lowers the estimated hormonal-therapy benefit at high progesterone-receptor levels relative to available-case analyses, showing the direction of the bias the new learners remove.","The new learners extend the comparison set for practitioners choosing among metalearners, giving a reason to prefer them over available-case or imputed versions of the DR and EP learners in MAR settings."],"supporting_citations":[{"why":"Introduces the DR-learner and its oracle-efficiency theory, which the mDR-learner extends to missing outcomes.","marker":"Kennedy, 2023"},{"why":"Introduces the EP-learner and iTMLE framework that the mEP-learner extends.","marker":"van der Laan et al., 2024"},{"why":"Supplies the orthogonal statistical learning framework and the excess-risk bound template used in Theorem 1.","marker":"Foster and Syrgkanis, 2023"},{"why":"Source of inverse probability of censoring weights used to reweight observed cases under MAR.","marker":"Robins et al., 1994"},{"why":"Provides the half-sample bootstrap method used to construct confidence intervals for the new learners.","marker":"Ritzwoller and Syrgkanis, 2024"},{"why":"Guides the choice of the trigonometric sieve basis and its dimension in the iTMLE step.","marker":"Zhang and Simon, 2023"}],"fun_headline_variants":["mDR and mEP learners fix missing outcome bias in CATE","Causal ML estimators handle missing outcomes with censoring weights","Debiased learners recover CATE under missing outcome data","CATE learners correct for missing outcome subgroups","Censoring-weighted learners robust to missing outcomes"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"For the mEP-learner, the claim of oracle efficiency depends on an unproved premise: that the sieve-based iTMLE update, reweighted by H(A,C,Z), actually drives the sample average of the drift term to zero and delivers the same guarantees as the EP-learner; if that targeting step fails, the mEP-learner may not be oracle efficient.","fun_headline_variants_meta":{"raw":{"variants":["mDR and mEP learners fix missing outcome bias in CATE","Causal ML estimators handle missing outcomes with censoring weights","Debiased learners recover CATE under missing outcome data","CATE learners correct for missing outcome subgroups","Censoring-weighted learners robust to missing outcomes"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000815,"raw_usage":{"total_tokens":3581,"prompt_tokens":962,"completion_tokens":2619,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":578,"completion_tokens_details":{"reasoning_tokens":2541}},"tokens_in":578,"tokens_out":2619,"duration_ms":24178,"temperature":1.0,"reasoning_tokens":2541,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T23:57:01.363859+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the mEP-learner on data where the true CATE is a step function that the trigonometric sieve basis cannot approximate, with fast nuisance estimators; if the sample average of the drift term after the targeting step is not zero, or if the excess risk fails to shrink at the oracle rate as n grows, the mEP oracle-efficiency claim is refuted. The same check applies to the mDR-learner, where the bound is proven and should hold.","supporting_citations":[{"cited_title":"Orthogonal statistical learning","cited_arxiv_id":null,"evidence_quote":"Supplies the orthogonal statistical learning framework and the excess-risk bound template used in Theorem 1."},{"cited_title":"Estimation of regression coefficients when some regressors are not always observed","cited_arxiv_id":null,"evidence_quote":"Source of inverse probability of censoring weights used to reweight observed cases under MAR."},{"cited_title":"Regression in tensor product spaces by the method of sieves","cited_arxiv_id":null,"evidence_quote":"Guides the choice of the trigonometric sieve basis and its dimension in the iTMLE step."}],"review_version":1}