{"id":"d0e6a19e-fd53-48c2-a7f7-aab476a34e70","arxiv_id":"2507.10774","paper_version":2,"verdict":"REJECT","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"high","formal_verification":"none","parameter_count":1,"one_line_summary":"A proposed longitudinal cross-world weighting estimand is not identified by the proof given, because the derivation silently assumes the two regimes' counterfactual covariate processes coincide.","lead":"This paper proposes a new 'cross-world' weighted effect for comparing two treatment schedules in longitudinal studies with positivity violations, and claims it can be identified and estimated without positivity assumptions. The identification proof is flawed: it swaps counterfactual covariate histories from the two schedules as if they were the same, which is only true under a stronger assumption the paper does not state.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 1 identifies a different estimand than (1): the g-formula evaluates both weights on one covariate history, while (1) uses propensities from two distinct counterfactual histories; Assumption 1 does not bridge this gap.","rationale":"The reader's weakest assumption is exactly the load-bearing gap. The paper's headline claim is identification of (1) as a g-formula; if (1) and (4)-(5) are different functionals, then the proposed estimator is not estimating the stated cross-world causal effect. The concern is not a disagreement with the cross-world conceptual framework or with the single-world g-formula, both of which are coherent; it is an internal inconsistency between the estimand and the identification theorem. The concrete DGP makes the discrepancy explicit and requires only standard normal and Bernoulli variables with full positivity, so it isolates the cross-world substitution from positivity complications. The paper itself contains the admission in Remark 1 that the needed equality fails in general when treatment affects intermediate covariates. A repair would be either to impose an explicit cross-world independence or equality assumption and check whether it is scientifically plausible, or to rename the identified functional as a single-world weighted g-formula estimand and adjust the interpretation accordingly.","tokens_in":23205,"tokens_out":12455,"duration_ms":131939,"concrete_test":"Simulate a T=2 NPSEM: X1~Bern(0.5), A1~Bern(0.5), X2=X1+A1+eps with eps~N(0,1), A2~Bern(expit(X2)), Y=A1+A2+X2+eta with eta~N(0,1). Let a_T=(1,1), a'_T=(0,0) and use overlap weights w_t(p)=p, w'_t(p')=p'. Compute the true estimand (1) by Monte Carlo over the structural exogenous variables, i.e., E[(Y(1,1)-Y(0,0)) * 0.25 * expit(X2(1)) * (1-expit(X2(0)))], and compute the g-formula (4)-(5) from the implied observed law, i.e., 0.25 * E[Y(1,1) * expit(X2(1)) * (1-expit(X2(1))) - Y(0,0) * expit(X2(0)) * (1-expit(X2(0)))]. If these disagree beyond Monte Carlo error, Theorem 1 fails even under full positivity, since all propensities lie in (0,1).","verdict_should_be":"UNCHANGED","load_bearing_attack":"Claimed identification of psi(a_T,a'_T) in Theorem 1 is invalid as stated. In (1), the time-t weight includes p'_t{X_t(a'_{t-1})} = P{A_t(a'_{t-1})=a'_t | X_t(a'_{t-1})}, the natural propensity under the a' regime evaluated at the covariate history generated by a'. The identified functional (4)-(5) instead evaluates both weights w_t{pi_t(x_t)} w'_t{pi'_t(x_t)} at a single covariate history x_t: the first integral integrates over dP(x_t | A_{t-1}=a_{t-1}, x_{t-1}) (history under a), and the second over the same-style history under a'. Lemma 1 identifies p'_t{X_t(a'_{t-1})} only as pi'_t(X'_t) with X'_t the history under a'_{t-1}; it does not license replacing X_t(a'_{t-1}) by X_t(a_{t-1}) inside the same expectation. The 'Timepoint 2' step of the proof silently writes w'_2(pi'_2) in an expectation over X_2 under A_1=a_1. Assumption 1 is a single-world conditional independence restriction and implies no equality in distribution between X_t(a_{t-1}) and X_t(a'_{t-1}). The paper's own Remark 1 concedes these histories are not equal in distribution when treatment affects intermediate covariates. Consequently (4)-(5) is the g-formula for a different, single-world weighted estimand, not for (1); the estimator in Section 5 and Theorem 2 target that different functional.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces a new longitudinal causal estimand, the cumulative cross-world weighted effect, which weights the difference in potential outcomes under two treatment regimes by a product of natural propensity scores evaluated under both counterfactual covariate histories. The authors claim that this estimand isolates the mechanistic contrast between regimes while adapting to positivity violations, is identifiable without a positivity assumption under strong sequential randomization, and admits doubly robust-style estimators with n^{-1/4} nuisance convergence rates. They derive an efficient influence function, propose a sample-split estimator, and illustrate the method with a union-membership wage analysis. The central identification theorem (Theorem 1) is, however, invalid as stated because the proof substitutes counterfactual covariate histories across regimes without a justifying assumption, so the identified functional in equations (4)-(5) does not correspond to the estimand in equation (1).","tokens_in":23546,"tokens_out":3633,"duration_ms":41359,"significance":"If the identification claim were correct, the paper would contribute a novel class of estimands for longitudinal positivity violations, along with a concrete estimation strategy and a useful conceptual discussion of the tradeoff between mechanistic and policy relevance. The development of an efficient influence function and the reformulation of density-ratio estimation as binary regression are methodologically interesting, and the included code and data analysis are positive features. However, because Theorem 1 is the foundation for the estimator and the data analysis, the paper's main contribution is not currently supported. The manuscript is candid about the cross-world nature of the estimand, but that candor makes the missing cross-world assumption in the proof more consequential rather than less.","major_comments":[{"comment":"The proof of Theorem 1 identifies a different estimand than the one defined in equation (1). In equation (1), the time-t weight includes p'_t{X_t(a'_{t-1})}, the natural propensity under the a' regime evaluated at the covariate history generated by a'. The identified functional in equations (4)-(5) instead evaluates both weights w_t{pi_t(x_t)}w'_t{pi'_t(x_t)} at a single covariate history x_t that is generated under the a-regime in the first integral and under the a'-regime in the second integral. The 'Timepoint 2' step of the proof replaces X_t(a'_{t-1}) by X_t inside an expectation conditional on A_1=a_1, which is only justified if the counterfactual covariate histories under the two regimes coincide in law. Assumption 1 is a single-world conditional independence restriction and implies no such equality in distribution. Remark 1 explicitly concedes that X_t(a_{t-1}) and X_t(a'_{t-1}) are generally not equal in distribution when treatment affects intermediate covariates. Consequently, the functional in (4)-(5) is the g-formula for a single-world weighted estimand, not for psi(a_T,a'_T) in (1). This invalidates Theorem 1 as a statement about the proposed estimand.","section":"Section 4, Theorem 1 and Appendix A.2"},{"comment":"Lemma 1 identifies the natural propensity P{A_t(a'_{t-1})=a'_t | X_t(a'_{t-1})} under positivity of the a'-regime propensity scores and conditioning on A_{t-1}=a'_{t-1}. In the proof of Theorem 1, however, this lemma is invoked inside an expectation in which the conditioning event involves the a-regime history (e.g., after 'iterated expectations on X_2 | A_1 = a_1, X_1'). The lemma does not license replacing the counterfactual covariate history X_t(a'_{t-1}) with the observed history X_t under the target regime A_{t-1}=a_{t-1}. The manuscript's own Example in Section 4.1 shows that the support of X_t under the two regimes can be disjoint even without positivity violations, which makes the substitution particularly problematic. The proof therefore relies on an unstated cross-world equivalence that is neither implied by Assumption 1 nor otherwise justified.","section":"Section 4, Lemma 1 and its use in Theorem 1"},{"comment":"The efficient influence function in Lemma 2 and the estimator in Algorithm 1 are derived for the functional in equations (4)-(5), not for the estimand in equation (1). Since Theorem 1 is invalid, the estimator's consistency claim for psi(a_T,a'_T) in Theorem 2 is unsupported. If the authors intend to estimate the single-world weighted functional that appears in (4)-(5), they should redefine the target estimand accordingly and re-derive the influence function under that target. The current presentation conflates the two functionals, and the data analysis in Section 6 therefore does not provide evidence about the estimand advertised in the abstract and introduction.","section":"Section 5, Lemma 2 and Theorem 2"}],"minor_comments":[{"comment":"The proof of Lemma 1 contains typographical errors, such as 'A1 = a1X1' where a comma is missing, and notation is used inconsistently (for example, X_3(a_2) versus X_3). These should be corrected.","section":"Appendix A.1"},{"comment":"The abstract states that the proposed effect 'circumvents the limitations of existing longitudinal methods' and 'isolates mechanistic differences,' but Section 3.1 and Remark 1 appropriately emphasize that the effect is not policy-relevant because it is cross-world. The abstract could mislead readers about the applicability of the estimand, and it should be qualified accordingly.","section":"Abstract and Section 3"},{"comment":"The identity for the density ratio is stated as holding 'when rho_t(X_t) < infinity almost surely,' but the convention that propensity scores are set to zero when the conditioning event has probability zero is not enough to ensure the ratio is well-defined for the intermediate expressions. The conditions under which the four probabilities in the ratio are obtained from observed data should be stated more carefully.","section":"Section 5.4"},{"comment":"The table lists 'Weighting towards a_T only' with w_t = p_t and w'_t = 1. This weight is unbounded when p_t is near zero and does not satisfy Condition 1 of Theorem 1, since the product w_t w'_t does not necessarily vanish when pi_t pi'_t = 0. The authors should clarify which weights in the table are actually covered by their main results.","section":"Table 1"}],"recommendation":"reject","confidential_remarks":"The paper's core identification result is invalid as stated, and the error is load-bearing: Theorem 1, Lemma 2, Theorem 2, and the data analysis all depend on it. The missing assumption is a form of cross-world independence between the counterfactual covariate processes under the two regimes, which is generally untestable and would substantially change the contributions if added explicitly. The authors' Remark 1 undermines their own identification argument, so this is not merely a presentational issue. The paper may contain useful ideas for a future revision, but the current manuscript does not establish its central claims."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this paper has a genuinely novel estimand and a useful conceptual frame, but Theorem 1 does not identify the object defined in equation (1). The proof replaces the counterfactual covariate history under the a' regime, X_t(a'_{t-1}), with the history under the a regime inside the same expectation. That substitution needs an equality in distribution or a cross-world independence that Assumption 1 doesn't provide and that the paper's own Remark 1 concedes fails when treatment affects intermediate covariates. So the identified functional (4)-(5) is a single-world weighted g-formula, not the cross-world estimand. The stress-test note is correct.\n\nWhat's good: the null preservation property (Proposition 1) is genuinely nice; the mechanism-relevance vs policy-relevance distinction is worth publishing; the partial common support analysis (Section 4.1) is useful and clearly separates covariate overlap from positivity; the density-ratio-as-regression identity in Section 5.4 is practical; and the efficiency theory in Section 5 appears correct for the single-world functional (4)-(5) if you take that as the target. The writing is clear and the authors are honest about the tradeoffs they see.\n\nThe soft spot is exactly the gap above. Lemma 1 identifies the natural propensity scores only when the conditioning event is from the same regime. Using Lemma 1 inside an expectation over the other regime's covariate history is the error. Condition 1 of Theorem 1 (weight zero when either propensity is zero) doesn't fix it. The data analysis estimates the single-world functional, so the empirical illustration doesn't actually demonstrate the cross-world estimand.\n\nWho does this help? People working on longitudinal positivity violations will get a good discussion, but they should not rely on the main theorem as stated. This is a major-revision paper, not a desk-reject; the idea is important enough and the fix is concrete (either state a cross-world independence assumption, or rebrand the estimand as the single-world weighted g-formula and adjust the claims). I'd send it to review, with the clear expectation that the authors must address the gap.","headline":"The identification theorem is broken in a way that is both central and repairable; the paper's best ideas survive a redefinition of the target.","tokens_in":24061,"tokens_out":3416,"would_cite":false,"duration_ms":37169,"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":"A cross-world weighted estimand can identify longitudinal treatment-regime contrasts without positivity assumptions, but only under a partial common-support condition.","keywords":["longitudinal causal inference","positivity violations","propensity score weighting","cross-world estimands","doubly robust estimation","efficient influence function","common support","time-varying treatments"],"falsifier":"Set T=2 with X2 = A1 so that the conditional law of X2 given A1=1 and given A1=0 have disjoint support, choose treatments A_t and outcome Y so that Y(1,1) - Y(0,0) is nonzero, and compute both sides of the identification display in Theorem 1 under Assumption 1; the weighted g-formula side collapses to zero while the estimand's defining expectation is nonzero, which would show the stated identification formula does not follow from the paper's stated assumptions.","tokens_in":22925,"feed_emoji":"⚖️","tokens_out":8048,"duration_ms":84880,"temperature":0.7,"pith_summary":"The paper introduces a new causal estimand for longitudinal studies where one or both treatment regimes are unobservable for some subjects: the cumulative cross-world weighted effect. The estimand multiplies the individual potential-outcome difference Y(a_T) - Y(a'_T) by a product of natural propensity scores under both regimes, so that subjects with near-zero chance of following either regime receive little weight. The authors claim this effect isolates the mechanistic difference between two treatment regimes, is identifiable without a positivity assumption under strong sequential randomization, and can be estimated by a doubly robust estimator with root-n normality when nuisance estimators converge at $n^{{-1/4}}$ rates. They also show a tradeoff: the estimand corresponds to a non-implementable intervention, and it collapses to zero when covariate distributions under the two regimes have no common support, even if the underlying causal effect is nonzero. If right, this supplies a principled alternative to flip interventions for longitudinal positivity violations.","feed_headline":"Cross-world weights identify regime effects without positivity","feed_subtitle":"Products of natural propensity scores across both regimes yield a doubly robust estimator that isolates mechanistic contrasts.","key_machinery":"The central object is the cumulative cross-world weighted effect \\psi(a_T, a'_T) = E[(Y(a_T) - Y(a'_T)) \\prod_{t=1}^T w_t{p_t(X_t(a_{t-1}))} w'_t{p'_t(X_t(a'_{t-1}))}], where p_t and p'_t are natural propensity scores under the two intervention histories. The argument works by identifying those natural propensity scores as ordinary observed propensity scores \\pi_t and \\pi'_t conditional on regime-consistent histories, then re-expressing \\psi as a difference of weighted g-formula integrals. The efficiency analysis is carried by the efficient influence function \\varphi = \\varphi_m + \\varphi_w, where \\varphi_m debiases the sequential regressions and \\varphi_w accounts for estimating the propensity-score weights; the covariate density ratio \\rho_t = dP(X_t | A_{t-1}=a_{t-1}, X_{t-1})/dP(X_t | A_{t-1}=a'_{t-1}, X_{t-1}) is handled by writing it as a ratio of four binary-regression probabilities.","core_discovery":"On the paper's own terms, the central discovery is that the cross-world weighted contrast \\psi(a_T, a'_T) is identifiable from observed data without positivity, provided the weights are zero whenever either natural propensity score is zero and strong sequential randomization holds. The proof rewrites \\psi as a difference of two weighted g-formula functionals, one per regime, with the same cross-world weight product applied to each. The same analysis reveals that the estimand's informativeness requires a partial common support assumption on time-varying covariate distributions; when the conditional laws of covariates under the two regimes have disjoint support, the identified functional vanishes identically, so the estimand is only meaningful as a mechanistic contrast when some overlap remains. The paper also derives an efficient influence function and a sample-split doubly robust estimator that converges to a normal distribution at root-n rate, recasting the challenging covariate density ratio as a ratio of four binary-regression probabilities.","pith_inferences":["If the unstated cross-world covariate equivalence fails, as it will whenever treatment changes intermediate covariates, the identification proof's replacement of one counterfactual covariate history by the other may not hold; a simple two-timepoint simulation with X2 = A1 could test this directly.","Because the estimand depends on natural propensity scores under both regimes, the same weighting construction should extend to continuous treatments or multi-valued actions by replacing propensity scores with dose-response or generalized propensity functions, though the density-ratio conditions would need reworking.","The collapse-to-zero behavior suggests that any applied report of this effect should also report the empirical distribution of the covariate density ratios, as the paper's data analysis does; otherwise a null result could reflect support failure rather than absence of mechanism.","A natural next comparison is against flip interventions on the same dataset: differences between the two estimates would quantify the bias flip effects incur from their additional effects on intermediate treatments and covariates."],"forward_implications":["Researchers can estimate contrasts between two treatment regimes when some subjects have near-zero probability of following one or both regimes, without assuming full positivity.","The doubly robust estimator achieves \\sqrt{n}-consistent, asymptotically normal inference when nuisance models converge at n^{-1/4} rates, so standard machine learning can be used for the nuisance steps.","The estimand's null-preservation property distinguishes it from flip interventions: if the two potential outcomes are almost surely equal, the weighted effect is exactly zero.","In practice, estimates should be accompanied by checks of the covariate density ratio \\rho_t, because absence of overlap makes the identified functional collapse to zero.","The interpretability-implementability tradeoff is made explicit: mechanism-relevant effects need not correspond to interventions anyone could perform."],"supporting_citations":[{"why":"Supplies the single-world intervention graph framework and the natural treatment value concept used to define counterfactual covariates and the propensity scores in the estimand.","marker":"Richardson and Robins [2013]"},{"why":"Provides the g-formula factorization that the identification theorem rewrites the cross-world effect as a difference of weighted g-formula integrals.","marker":"Robins [1986]"},{"why":"Describes flip interventions, the main alternative the paper contrasts with, and the source of the null-preservation failure motivating the new estimand.","marker":"McClean et al. [2025]"},{"why":"Gives the stochastic-intervention g-formula and weighting identities that the cross-world weights generalize.","marker":"Kennedy [2019]"},{"why":"Establishes the doubly robust longitudinal estimating strategy that the efficient influence function estimator extends to cross-world weights.","marker":"Bang and Robins [2005]"},{"why":"Develops longitudinal modified treatment policies, the class of interventions whose estimand is compared for implementability.","marker":"Díaz et al. [2023]"},{"why":"Provides identification results for interventions depending on natural treatment value, used for the propensity-score identification in Lemma 1.","marker":"Young et al. [2014]"},{"why":"Contributes the weighted sequential regression recursion that the paper's estimation algorithm adapts to weight two regimes simultaneously.","marker":"Schomaker et al. [2024]"},{"why":"Supplies the von Mises expansion lemma used to turn the efficient influence function into a root-n convergence result.","marker":"Kennedy et al. [2023]"}],"fun_headline_variants":["Cross-world weights identify effects without positivity","Identifiable effects without positivity assumptions","Doubly robust estimator for cross-world effects","New estimand isolates mechanistic contrasts under violations","Causal effects when positivity fails: new weighting"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The proof's key identification step assumes, without stating it, that the counterfactual covariate history under the comparison regime can be treated as equivalent to the covariate history under the target regime inside the same expectation, a cross-world equivalence that strong sequential randomization alone does not imply and that generally fails when treatment affects intermediate covariates.","fun_headline_variants_meta":{"raw":{"variants":["Cross-world weights identify effects without positivity","Identifiable effects without positivity assumptions","Doubly robust estimator for cross-world effects","New estimand isolates mechanistic contrasts under violations","Causal effects when positivity fails: new weighting"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000898,"raw_usage":{"total_tokens":3878,"prompt_tokens":964,"completion_tokens":2914,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":580,"completion_tokens_details":{"reasoning_tokens":2850}},"tokens_in":580,"tokens_out":2914,"duration_ms":25293,"temperature":1.0,"reasoning_tokens":2850,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T17:27:21.830102+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Set T=2 with X2 = A1 so that the conditional law of X2 given A1=1 and given A1=0 have disjoint support, choose treatments A_t and outcome Y so that Y(1,1) - Y(0,0) is nonzero, and compute both sides of the identification display in Theorem 1 under Assumption 1; the weighted g-formula side collapses to zero while the estimand's defining expectation is nonzero, which would show the stated identification formula does not follow from the paper's stated assumptions.","supporting_citations":[{"cited_title":"Single world intervention graphs (swigs): A unification of the counterfactual and graphical approaches to causality","cited_arxiv_id":null,"evidence_quote":"Supplies the single-world intervention graph framework and the natural treatment value concept used to define counterfactual covariates and the propensity scores in the estimand."},{"cited_title":"A new approach to causal inference in mortality studies with a sustained exposure period—application to control of the healthy worker survivor effect","cited_arxiv_id":null,"evidence_quote":"Provides the g-formula factorization that the identification theorem rewrites the cross-world effect as a difference of weighted g-formula integrals."},{"cited_title":"Nonparametric causal effects based on incremental propensity score interventions","cited_arxiv_id":null,"evidence_quote":"Gives the stochastic-intervention g-formula and weighting identities that the cross-world weights generalize."},{"cited_title":"Identification, estimation and approximation of risk under interventions that depend on the natural value of treatment using observational data","cited_arxiv_id":null,"evidence_quote":"Provides identification results for interventions depending on natural treatment value, used for the propensity-score identification in Lemma 1."},{"cited_title":"Causal inference for continuous multiple time point interventions","cited_arxiv_id":null,"evidence_quote":"Contributes the weighted sequential regression recursion that the paper's estimation algorithm adapts to weight two regimes simultaneously."},{"cited_title":"Semiparametric counterfactual density estimation","cited_arxiv_id":null,"evidence_quote":"Supplies the von Mises expansion lemma used to turn the efficient influence function into a root-n convergence result."}],"review_version":1}