{"id":"f613f7cf-1057-4ed0-ab56-f45221267824","arxiv_id":"2505.13072","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"The paper proposes a general toolbox of orthogonal survival learners with custom weighting functions to estimate heterogeneous treatment effects robustly under treatment, censoring, and survival overlap violations.","lead":"This paper presents a general method for estimating how treatments affect survival differently across patients, designed to stay reliable when data have censoring and other overlap problems. Practitioners can plug in any machine learning model and choose a weighting function targeting the specific overlap issue in their study.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Weighting-function selection is the load-bearing gap: experiments choose f with oracle knowledge of the overlap violation, and mis-targeted weighting can be worse than no weighting (Table 3, T+S PEHE 3.55 vs ∅ 1.64).","rationale":"Agree with the reader that the weakest point is the oracle choice of the weighting function. I did not find a clear internal error in the derivation of Theorems 5.1 and 5.2; the algebra is long, but the key steps—conditional-zero martingale residuals in Lemma H.2 and the E[ρ|X]=f identity in Theorem 5.2—are standard, and the claimed results are plausible for a fixed f. The strongest real-world threat is therefore not the mathematical orthogonality result but the missing link from 'any f' to 'the right f.' The experimental protocol confirms this: targeted learners are highlighted per violation type, and the paper itself reports that unsuitable weighting can be worse than no weighting. A concrete, implementable check is to replace the oracle choice with the overlap diagnostics the paper recommends and measure the loss in PEHE. If diagnostics pick the wrong f in a material fraction of runs, the advertised robustness advantage disappears. This does not change the reader's conditional verdict; it sharpens the condition under which the paper's practical claims would hold.","tokens_in":28161,"tokens_out":23671,"duration_ms":255001,"concrete_test":"Run the Scenario 2 simulations (Section I.1) and, in each training fold, estimate π̂, Ŝ_{t−1}, and Ĝ_{t−1} with the same nuisance architectures. Apply the paper's recommendation by choosing the weighting f corresponding to the overlap type whose estimated minimum (or 10th percentile) is smallest, and no weighting if all exceed, say, 0.1. Evaluate the PEHE of this selected learner on the four violation settings over the 10 runs. If the selected learner is not within one standard error of the oracle gray-shaded learner in Table 3 and is not at least as good as the ∅-learner, then the robustness claim depends on oracle knowledge and the practical recommendation is not supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Theorems 5.1 and 5.2 are best read as claims about a fixed f: for a pre-specified positive weighting function, the population loss is Neyman-orthogonal and has τ_t as its minimizer. The advertised practical claim—robustness to treatment/censoring/survival overlap violations—requires an additional step that the paper does not supply: choosing f in a data-driven way. Section 5 tells practitioners to 'inspect overlap' from estimated nuisances, but no diagnostic or selection rule is given. In Section 7, the targeted learner is selected per simulated violation type (gray cells), i.e., oracle knowledge of the true overlap violation. The paper's own Table 3 shows that the wrong f can hurt: in the no-violation column, T+S weighting gives PEHE 3.55 vs 1.64 for ∅. Appendix J concedes that inappropriate weighting 'can significantly slow down convergence.' Since real analysts do not know which overlap type is violated, the reported gains may not transfer unless a reliable f-selection rule is established. This concern is about the practical central claim, not about the internal validity of the orthogonality proofs.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a general toolbox for constructing Neyman-orthogonal meta-learners that estimate conditional survival treatment effects τ_t(x) from censored time-to-event data. The core theoretical contribution is a family of losses L_f(g, η_t) (Eq. 7) parametrized by an arbitrary positive weighting function f; the paper proves that this loss is Neyman-orthogonal with respect to the nuisance functions (Theorem 5.1) and that its population minimizer coincides with the true τ_t whenever τ_t lies in the function class G (Theorem 5.2). The toolbox is instantiated with several weighting functions targeting treatment, censoring, and survival overlap, recovering the Survival DR-learner and Survival R-learner as special cases and yielding new learners for survival-specific overlap violations. The authors evaluate the learners on synthetic data with simulated overlap violations and on the Twins dataset, reporting PEHE improvements when the weighting function is matched to the violation type.","tokens_in":28407,"tokens_out":5536,"duration_ms":55828,"significance":"If the theoretical claims hold, this is a valuable unifying framework: it subsumes existing survival meta-learners, extends orthogonal estimation to censoring- and survival-overlap regimes, is model-agnostic, and comes with concrete extensions to marginalized effects, continuous time, and other estimands in the appendices. The proof strategy is a direct derivative computation and correctly reproduces the known special cases f=1 (Survival DR) and f=π(1−π) (Survival R), which is reassuring. The authors also provide code and the derivations are sufficiently detailed to be checked. The main weakness is that the advertised practical robustness depends on choosing f to match the unknown overlap violation, and the paper does not yet supply a data-driven selection rule; the experiments select f with oracle knowledge of the simulated violation. This gap is acknowledged in the limitations but remains load-bearing for the practical claim.","major_comments":[{"comment":"The empirical evaluation of the toolbox's central practical advantage—robustness to unknown overlap violations—selects the targeted learner per simulated violation type (gray cells in Tables 2 and 3), i.e., with oracle knowledge of the true overlap problem. The paper's own Table 3 shows that a mismatched weighting can be substantially worse than no weighting (no-violation column: T+S PEHE 3.55±1.13 vs ∅ 1.64±0.19; survival-violation column: T 6.77±1.08 vs ∅ 4.54±0.36). Section 5's suggestion to 'inspect overlap' from estimated nuisances is not operationalized with a concrete diagnostic or a selection rule, and Appendix J concedes that inappropriate weighting 'can significantly slow down convergence.' Since real analysts do not know which overlap type is violated, the advertised robustness gain is not yet established; the manuscript should either supply a data-driven f-selection procedure with evidence that it recovers the oracle choice, or explicitly reframe the claim as robustness conditional on an appropriate choice of f.","section":"Section 7, Table 3, Appendix J"},{"comment":"The proof of Neyman orthogonality is the load-bearing theoretical result, but it is a long hand-derived computation that does not state the regularity conditions on f under which the required derivatives and the interchanges of differentiation and expectation are valid; notation such as ∂f/∂S_{t−1}(·,A) is also compressed. The computation appears to be correct—it reduces to the known Survival-DR (f=1) and Survival-R (f=π(1−π)) objective—but for a central theorem, the authors should either provide a more structural proof, state the smoothness/boundedness conditions on f and the hazards, or accompany the algebra with machine-checked verification.","section":"Section 5, Theorem 5.1, Appendix H"},{"comment":"The empirical support for the practical claims is limited by the small number of runs (10) and the large, often overlapping standard deviations; in several settings the targeted learner is not the best (e.g., Table 2, Survival column: S PEHE 2.72±1.60 vs ∅ 2.70±1.63). The conclusion that 'targeted weighting generally improves performance' should be backed by more seeds and a formal comparison, or the claims should be softened to reflect the variability in the reported results.","section":"Section 7, Tables 2 and 3"}],"minor_comments":[{"comment":"In the introduction, 'enable unbiased and robust estimation under both confounding end censoring' should read 'under both confounding and censoring.'","section":"Section 1"},{"comment":"The sentence 'we can use parametric models λS_j(x_i,a_i,θ) parametrized by θ ... to minimize the resulting the resulting log-likelihood loss' contains a duplicated phrase 'the resulting the resulting.'","section":"Appendix C"},{"comment":"The phrase 'This results in weighting results in' should be corrected to 'This results in the following weighting' or similar.","section":"Section 6, S-learner"},{"comment":"The text says 'we report the PEHE×10−4' but the table captions only say 'PEHE'; please clarify the scaling in both the text and the table captions.","section":"Section 7, Tables 2 and 3"},{"comment":"The caption 'Ratios of PEHE of the targeted learner wrt. the learner without the correct target' is confusing; please define the numerator and denominator explicitly and explain the color coding for the reader.","section":"Section 7, Figure 3"}],"recommendation":"major_revision","confidential_remarks":"The theoretical core appears sound, and the toolbox is a useful unifying contribution. The main risk is overclaiming robustness without a selection rule for f; the experiments use oracle knowledge of the violation type, and the paper's own results show that a wrong f can hurt. A revision that adds a data-driven selection procedure or substantially tempers the practical claims would resolve my main concern. The orthogonality proof, while long, seems correct, but it would benefit from stated regularity conditions."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a serious paper worth refereeing carefully. The theoretical core is a general weighted orthogonal loss for heterogeneous treatment effects from censored time-to-event data, and it checks out as far as I can see. The new thing is not orthogonality per se—Morzywolek et al. and Xu et al. were already in that neighborhood—but the unifying weighting function that lets you target treatment, censoring, or survival overlap, and the concrete C- and S-learners. The proof in Appendix H is long algebra, not machine-checked, but it reduces to known special cases and uses the right martingale-style lemmas. The code is available, and the simulations keep architectures fixed so the comparisons are about the loss.\n\nThe soft spot is exactly where the stress-test note lands. Theorems 5.1 and 5.2 are statements about a pre-specified f. The practical claim—that these learners give robustness to whatever overlap violation is present—requires knowing which f to use. Section 5 suggests inspecting estimated overlap, but there's no diagnostic, no selection rule, no sensitivity analysis. In Section 7 the targeted learner is chosen with oracle knowledge of the simulated violation. Table 3 shows that the wrong f can be worse than no weighting: T+S gets 3.55 PEHE in the no-violation column versus 1.64 for ∅, and Appendix J concedes that inappropriate weighting can slow convergence. This is not a hole in the theory, but it is a load-bearing gap in the practical story.\n\nMinor quibbles: the PEHE differences in the experiments are modest and overlapping; the Twins case study is illustrative, not a strong benchmark. The citations line up; self-citations are contextual, not load-bearing.\n\nWho this is for: researchers working on causal meta-learners for survival outcomes, and applied statisticians who know their overlap regime. It deserves a serious referee. I'd recommend engaging with it and pressing for a data-driven f-selection rule or, failing that, a sensitivity analysis that doesn't assume oracle knowledge.","headline":"A genuinely useful generalization of orthogonal learners to survival data, with a real practical gap: the advertised overlap robustness depends on an f-selection rule the paper doesn't provide.","tokens_in":28953,"tokens_out":2586,"would_cite":true,"duration_ms":26160,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that one weighted loss family yields orthogonal, model-agnostic survival treatment effect learners whose minimizer equals the true heterogeneous treatment effect.","keywords":["heterogeneous treatment effects","survival analysis","time-to-event data","censoring","Neyman orthogonality","overlap violations","meta-learners","causal inference"],"falsifier":"Take a synthetic dataset with known $\\tau_t$, deliberately misspecify one nuisance function (for example, replace the true propensity score by a constant $0.5$) while keeping the others accurate, and evaluate Eq. (12) with several positive weighting functions $f$. If the PEHE changes at first order with the misspecification for some $f$, or if the empirical minimizer departs systematically from $\\tau_t$ even though $\\tau_t\\in G$, the orthogonality claim fails; equivalently, compute a finite-sample estimate of the cross-derivative $D_{\\eta}D_g L_f$ and check whether it is zero.","tokens_in":27949,"feed_emoji":"⏳","tokens_out":12068,"duration_ms":112205,"temperature":0.7,"pith_summary":"This paper tries to establish that heterogeneous treatment effects (HTEs) from censored time-to-event data can be estimated by minimizing one orthogonal weighted loss, no matter which positive weighting function is used. The loss is built so that small errors in the nuisance estimates (propensity score, survival hazards, censoring hazards) do not shift the second-stage estimate, and its population minimizer equals the true conditional survival treatment effect whenever the model class contains it. The weighting function is the design choice: it can down-weight regions with low treatment, censoring, or survival overlap, yielding both existing survival DR- and R-learners as special cases and new learners targeted at overlap problems specific to survival data. If the claim is right, practitioners get a model-agnostic way to build survival HTE learners whose robustness to overlap violations is controlled by a single weighting function.","feed_headline":"One weighted loss handles every overlap type for survival effects","feed_subtitle":"Weighted orthogonal losses recover the true treatment effect under censoring and low overlap, whatever the weighting.","key_machinery":"The central object is the orthogonal loss $L_f(g,\\eta_t)$ of Eq. (7), assembled from a positive weighting function $f(\\tilde\\eta_t(X))$ and two pseudo-outcome terms $\\rho(Z,\\eta_t)$ and $\\varphi(Z,\\eta_t)$ in Eqs. (8)-(9). The terms $\\xi^S$ and $\\xi^G$ in Eq. (10) are residual processes whose conditional means given $(X,A)$ are zero (Lemma H.2); those cancellations make all cross-derivatives $D_{\\eta}D_g L_f$ vanish, which is exactly Neyman orthogonality. For the targeting claim, the $g$-dependent part of $L_f$ reduces to the weighted squared loss $\\mathbb{E}[f(\\tilde\\eta_t(X))(\\tau_t(X)-g(X))^2]$ up to terms independent of $g$, so the minimizer is $\\tau_t$ whenever the model class contains it. The user controls $f$, choosing products of propensity, survival, or censoring functions to retarget the loss toward well-supported regions.","core_discovery":"The central claim is that the population loss $L_f(g,\\eta_t)$ defined in Eq. (7), with $\\rho$ and $\\varphi$ given by Eqs. (8)-(10), is Neyman-orthogonal (that is, first-order insensitive to nuisance estimation errors) with respect to the nuisance functions $\\eta_t=(\\pi,\\lambda^S_i(\\cdot,1),\\lambda^S_i(\\cdot,0),\\lambda^G_i(\\cdot,1),\\lambda^G_i(\\cdot,0))_{i=0}^t$ for any positive weighting function $f$, and that its minimizer over a class $G$ coincides with the true conditional survival treatment effect $\\tau_t(x)=P(T(1)>t\\mid X=x)-P(T(0)>t\\mid X=x)$ whenever $\\tau_t\\in G$ (Theorems 5.1 and 5.2). This means weighting changes only which population the loss emphasizes, not the target estimand. The paper instantiates the toolbox with $f=1$ (recovering the survival DR-learner), $f=\\pi(1-\\pi)$ (recovering the survival R-learner), $f=G_{t-1}(\\cdot,1)G_{t-1}(\\cdot,0)$ for censoring overlap, $f=S_{t-1}(\\cdot,1)S_{t-1}(\\cdot,0)$ for survival overlap, and combinations of these. In synthetic experiments, the learner whose weighting matches the induced overlap violation attains the lowest PEHE, and a twins birth-weight case study shows faster convergence for censoring-weighted learners.","pith_inferences":["A data-driven rule for choosing $f$ from estimated overlap diagnostics would remove the main practical gap, since the paper's experiments select the weighting that matches the simulated violation and mismatched weighting is shown to hurt.","The same orthogonal-loss construction should transfer to related survival estimands such as restricted mean survival times and treatment-specific survival curves using the appendix derivations; the main test would be whether the weighting choices remain equally beneficial there.","Because $f$ indexes a continuum, practitioners could interpolate between DR-type and R-type behavior or tune the retargeting strength by cross-validation on a PEHE-style criterion rather than committing to one overlap type.","For clinical use, pairing the orthogonal loss with an interpretable second-stage class (trees, linear models) would give directly readable HTE estimates that still carry the orthogonality guarantee."],"forward_implications":["Any learner instantiated from the toolbox inherits orthogonality, so first-order errors in the propensity, survival-hazard, or censoring-hazard estimates do not bias the HTE estimate.","The loss minimizer is the true conditional survival treatment effect $\\tau_t(x)$ for every positive $f$ whenever the second-stage model class is rich enough, so weighting affects robustness and variance but not the estimand.","The survival DR- and R-learners are recovered as the special cases $f=1$ and $f=\\pi(1-\\pi)$; the new C- and S-learners target censoring and survival overlap, with combined variants for multiple violations.","Empirically, the learner whose weighting matches the overlap violation present in the data achieves the lowest PEHE and faster convergence; mismatched weighting can increase variance.","Because the loss is model-agnostic, the same second-stage loss can be paired with any nuisance estimator or regression model for $g$, including neural networks, trees, or linear models."],"supporting_citations":[{"why":"Defines Neyman orthogonality and the double/debiased estimation principle the loss is designed to satisfy.","marker":"[5]"},{"why":"Supplies the survival HTE problem setting and the Scenario 2 data-generation mechanism used in the experiments.","marker":"[8]"},{"why":"Provides the directional-derivative calculus used to define and verify orthogonality of the loss.","marker":"[11]"},{"why":"Provides the efficient influence function and von Mises expansion background underlying the orthogonal-loss construction.","marker":"[22]"},{"why":"The DR-learner whose survival version is recovered as the no-weighting special case.","marker":"[23]"},{"why":"Provides the twins birth-weight dataset used in the real-world case study.","marker":"[29]"},{"why":"The weighted orthogonal learner formulation for uncensored CATE that the toolbox extends to censoring and survival overlap.","marker":"[30]"},{"why":"The R-learner whose survival version is recovered as the treatment-overlap weighting special case.","marker":"[31]"},{"why":"Underlies the DR-learner and the semiparametric efficiency framing that the no-weighting learner recovers.","marker":"[44]"},{"why":"Provides the censoring-unbiased-transformation baselines and the survival DR/R learner special cases.","marker":"[50]"}],"fun_headline_variants":["Orthogonal survival learners unaffected by weighting choice","Any weighting keeps survival effect estimators orthogonal","Weighting never breaks orthogonality in survival learners","Robust survival treatment effects via orthogonal learners"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the analyst can tell which overlap type (treatment, censoring, or survival) is actually violated and choose the matching weighting function $f$, because the paper's simulations pick $f$ with oracle knowledge of the violation and report that unsuitable weighting can harm performance, e.g., in the no-violation Scenario 2, the T+S learner gives PEHE 3.55 versus 1.64 for the unweighted learner.","fun_headline_variants_meta":{"raw":{"variants":["Orthogonal survival learners unaffected by weighting choice","Any weighting keeps survival effect estimators orthogonal","Weighting never breaks orthogonality in survival learners","Robust survival treatment effects via orthogonal learners"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000337,"raw_usage":{"total_tokens":1958,"prompt_tokens":1131,"completion_tokens":827,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":747,"completion_tokens_details":{"reasoning_tokens":771}},"tokens_in":747,"tokens_out":827,"duration_ms":8753,"temperature":1.0,"reasoning_tokens":771,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T20:20:47.517052+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a synthetic dataset with known $\\tau_t$, deliberately misspecify one nuisance function (for example, replace the true propensity score by a constant $0.5$) while keeping the others accurate, and evaluate Eq. (12) with several positive weighting functions $f$. If the PEHE changes at first order with the misspecification for some $f$, or if the empirical minimizer departs systematically from $\\tau_t$ even though $\\tau_t\\in G$, the orthogonality claim fails; equivalently, compute a finite-sample estimate of the cross-derivative $D_{\\eta}D_g L_f$ and check whether it is zero.","supporting_citations":[{"cited_title":"SurvITE: Learning heterogeneous treatment effects from time-to-event data","cited_arxiv_id":null,"evidence_quote":"Supplies the survival HTE problem setting and the Scenario 2 data-generation mechanism used in the experiments."},{"cited_title":"Semiparametric doubly robust targeted double machine learning: A review","cited_arxiv_id":null,"evidence_quote":"Provides the efficient influence function and von Mises expansion background underlying the orthogonal-loss construction."},{"cited_title":"Causal effect inference with deep latent-variable models","cited_arxiv_id":null,"evidence_quote":"Provides the twins birth-weight dataset used in the real-world case study."}],"review_version":1}