{"id":"5a87bdc7-3dc1-48dd-93bc-d62a85303f7d","arxiv_id":"2507.10809","paper_version":1,"verdict":"REJECT","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":5,"one_line_summary":"The authors define an intervention-conditioned average treatment effect for event sequences and propose an IPW estimator with a Transformer-CNN intensity model, but the estimator's unbiasedness proof is flawed.","lead":"This paper proposes a causal framework for estimating how the effect of a cause event on an outcome event in a temporal point process changes under an out-of-domain intervention, together with a Transformer-CNN model and an inverse probability weighting estimator. It reports improved ATE estimation and process fitting over a Granger-causality baseline in simulations and two real-world datasets.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 2's IPW control arm is misspecified: 1 − e_t(v) is not P(c=0,v=v|x), so the estimator is biased whenever other (c,v) combinations occur.","rationale":"The paper's central contribution is the claim that Eq. (2), with propensity e_t(v) from Definition 2, gives an unbiased estimator of the intervention-conditioned ATE. That claim is load-bearing: it is what distinguishes the framework from a purely predictive model and what justifies the reported ATE numbers on simulated data. The control-arm denominator is not a stylistic choice; it is the inverse probability of the wrong event. The simulation protocol makes the problem concrete because v is activated independently of c, so observations with c=1,v=0 and c=0,v=1 are present. A biased estimator can still outperform a baseline in a specific simulation, but the theoretical guarantee, and with it the paper's main scientific claim, does not hold. The reader identified exactly this weakness, and the arithmetic check above confirms it. I also note that the proof's conditional-expectation step is invalid as written: the standard IPW tower argument uses an outer expectation over x, not a conditional expectation given the propensity score. Even if that step were fixed for the treated component, the control denominator remains misspecified. The absence of code and the narrow baseline comparison are secondary concerns that would matter even if the estimator were unbiased. Because the central proof fails, the REJECT verdict is appropriate, and my read does not change it.","tokens_in":13595,"tokens_out":8080,"duration_ms":100385,"concrete_test":"Run a no-confounder Monte Carlo check with c,v independent Bernoulli(1/2), potential outcomes λ(1,1)=2 and λ(0,1)=1, so τ(1)=1, and use the oracle joint propensity e(1)=P(c=1,v=1|x)=1/4 in Eq. (2). The treated component contributes (1/4)/(1/4)·2 = 2; the control component contributes (1/4)/(3/4)·1 = 1/3; Eq. (2) yields 5/3 instead of 1. Repeating with the corrected control weight 1/P(c=0,v=1|x)=4 recovers τ(1)=1. If a 10^6-window simulation reproduces this arithmetic, the misspecification in Theorem 2 is confirmed and is not an artifact of propensity-score estimation.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Definition 2 defines e_t(v) = P(c_t=1, v_t=v | x_t). Eq. (2) weights the control indicator 1{c_t=0, v_t=v} by 1/(1 − e_t(v)). Algebraically, 1 − e_t(v) = P(c=1,v≠v|x) + P(c=0,v≠v|x) + P(c=0,v=v|x), so the control arm is weighted by the inverse probability of the complement of {c=1,v=v}, not by 1/P(c=0,v=v|x). For the causal contrast τ(v) = E[λ(1,v) − λ(0,v)], the correct IPW denominators are P(c=1 | v=v, x) and P(c=0 | v=v, x); using the joint probability P(c=1,v=v|x) weights the wrong event. Unless P(c=1,v≠v|x) = P(c=0,v≠v|x) = 0 for all x, the control term contributes E[ P(c=0,v=v|x)/(1 − P(c=1,v=v|x)) · λ(0,v)(x) ] instead of E[λ(0,v)], producing a biased estimator. The simulations in Appendix B generate out-of-domain interventions independently of the cause state (Algorithm 2 activates v by a random state, separately from c), so the excluded combinations occur. The proof of Theorem 2 does not repair this: the step conditioning on e_t(v) asserts E[1{c=1,v=v} e^{-1} λ(1,v)|e] = E[λ(1,v)|e], which is not a valid IPW identity, and no correction there fixes the control denominator. Thus the paper's central unbiasedness claim fails as stated.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes a causal framework for temporal point processes in which an exogenous \"out-of-domain\" intervention v can shift the causal effect of a cause event c on an outcome event o. The authors define an intervention-conditioned average treatment effect τ(v), introduce a propensity-score-based inverse probability weighting (IPW) estimator, and prove unbiasedness in Theorem 2 under SUTVA, unconfoundedness, and overlap. They then design a Transformer-CNN model to estimate the conditional intensity function and validate the approach on simulated Hawkes processes and on predictive-maintenance and diabetes datasets. The central theoretical result is Theorem 2, which claims E[τ̂(v)] = τ(v) for v ∈ {0,1} using the weight α_t(v) = 1{c=1,v=v}/e_t(v) − 1{c=0,v=v}/(1−e_t(v)), where e_t(v) = P(c_t=1, v_t=v | x_t).","tokens_in":14005,"tokens_out":7424,"duration_ms":92323,"significance":"If Theorem 2 were correct, the paper would address a genuine gap: existing causal point-process work focuses on causal effects within a fixed domain and does not model how exogenous interventions modify cause–outcome relations. The paper also contributes a concrete neural architecture and extensive simulation experiments with injected ground-truth effects, which is a strength. The diabetes and predictive-maintenance applications are practically motivated and the authors are explicit about several limitations. However, the central unbiasedness claim is invalid as stated because the control-arm weight is misspecified, and the only real-world causal validation is qualitative. The framework is potentially salvageable with a corrected propensity score, but the current manuscript cannot be accepted without substantial theoretical and experimental revision.","major_comments":[{"comment":"The control-arm weight is misspecified. Definition 2 defines e_t(v) = P(c_t=1, v_t=v | x_t), so 1 − e_t(v) = P(c_t=1, v_t=1−v | x_t) + P(c_t=0, v_t=1−v | x_t) + P(c_t=0, v_t=v | x_t). The indicator 1{c_t=0, v_t=v} is therefore weighted by the inverse probability of the complement of {c_t=1, v_t=v}, not by 1/P(c_t=0, v_t=v | x_t). Unless the first two terms vanish for all x_t, the control term contributes E[ P(c=0,v=v|x) / (1 − P(c=1,v=v|x)) · λ(0,v) ] instead of E[λ(0,v)], making the estimator biased. The overlap assumption in Assumption 3 cannot force the other cells to have zero probability; in fact the simulated data in Appendix B generate the out-of-domain intervention independently of the cause state (Algorithm 2), so the excluded combinations do occur. Theorem 2 is therefore false as stated.","section":"§3.4, Definition 2 and Eq. (2)"},{"comment":"The proof's key step is not a valid IPW identity. The manuscript asserts that conditioning on e_t(v) yields E[ 1{c=1,v=v}/e_t(v) · λ(1,v)(t) | e_t(v) ] = E[ λ(1,v)(t) | e_t(v) ], but inverse probability weighting requires conditioning on the full covariate vector x_t, and even then the treatment-arm identity does not transfer to the control arm because 1 − e_t(v) is not the probability of the observed control event. No subsequent step repairs the control denominator. A corrected estimator could define e_t(c,v) = P(c_t=c, v_t=v | x_t) and weight by 1{e_t=c, v_t=v}/e_t(c,v), which would restore unbiasedness under Assumptions 1–3, but the current proof and estimator must be rewritten.","section":"Appendix A.2, proof of Theorem 2"},{"comment":"The real-world causal validation is only qualitative. Table 4 reports no ATE estimates, confidence intervals, or effect sizes; it lists qualitative conclusions marked with a checkmark and cites medical literature. The text in Section 5.2 states that \"ground truth ATEs for the diabetes dataset are unavailable\" and that the method was \"validated\" against literature, but no numerical ATE values appear for either real-world dataset. This evidence does not support the claim that the estimator performs well outside simulation. The authors should report τ̂(1) and τ̂(0) with standard errors (or credible intervals) for the diabetes scenarios, and ideally a quantitative comparison against a meaningful baseline.","section":"§5.2, Table 4"}],"minor_comments":[{"comment":"The text says \"raining is performed on batches\" and should read \"Training is performed on batches.\"","section":"§4, optimization paragraph"},{"comment":"The column header contains the typo \"symptopm\"; it should be \"symptom.\"","section":"§5.2, Table 4"},{"comment":"The second enumerated item uses \"b)\" while the other items use numbers 1, 3, 4; it should be numbered consistently.","section":"§1, contribution list"},{"comment":"References [3] and [7] have \"null\" as the DOI and publication field; these citations need to be completed.","section":"Bibliography"},{"comment":"The proof is a single sentence and does not formally use the definition of process independence or the exclusion of trivial direct causes; a complete proof is needed.","section":"Appendix A.1, proof of Theorem 1"},{"comment":"The propensity-score estimator should clarify how the indicator 1{x_t^w} is defined over sliding windows and how the ratio is computed when x_t changes continuously in time.","section":"Eq. (3)"}],"recommendation":"major_revision","confidential_remarks":"The stress-test concern lands: the control-arm denominator in Eq. (2) is wrong, and Theorem 2 is false as stated. I nonetheless recommend major revision rather than immediate rejection because the error is local and a corrected propensity score e_t(c,v) = P(c_t=c, v_t=v | x_t) would restore the intended result. The authors must then rerun the simulations and, crucially, report quantitative ATE estimates in the real-world validation. The paper is not acceptable in its current form."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's the short version: the paper identifies a genuine gap—how causal effects between events change when an out-of-domain intervention is active—and defines a sensible estimand (Definition 1). But Theorem 2, the load-bearing claim, is not proven and the proposed IPW estimator is actually biased under their own simulation setup. I would not send this out for review as it stands.\n\nWhat is genuinely new: conditioning the ATE on an intervention state in a temporal point process is a natural and useful extension of [9], and they are careful to define the potential-outcome framework with proximal history and binary treatment/intervention. The hybrid Transformer-CNN architecture is a reasonable engineering contribution, and the simulation study with injected ground truth shows the model can beat CAUSE in several settings. That part is credible.\n\nThe soft spots are not minor. Definition 2 sets e_t(v) = P(c_t=1, v_t=v | x_t). Equation (2) weights the control term 1{c_t=0, v_t=v} by 1/(1 - e_t(v)). But 1 - e_t(v) is the probability of everything except {c=1, v=v}; it is not the probability of {c=0, v=v}. So the control arm gets the wrong denominator, and the estimator is biased whenever other (c,v) combinations have positive probability. Their own simulation toggles v independently of c, so that happens all the time. The proof in Appendix A.2 does not repair this; it just asserts the conditional expectation identity that would hold for the correct weight. This is a foundational problem, not a cosmetic one.\n\nThe empirical side also has gaps. Real-data validation reports only qualitative agreement with medical literature and no ATE estimates, so the main claim is not actually checked there. The only baseline is CAUSE; the closest methodological relative, [9], is discussed but not compared. No code is provided, which limits reproducibility.\n\nWho is this for? Researchers working on causal inference for event sequences will find the framing useful and the error instructive. But the paper as written overclaims the central results, and the theory does not support the estimator. My recommendation: do not send to peer review in current form. Ask the authors to either fix the IPW estimator (e.g., use separate propensity scores for cause and intervention, or weight by P(c=0,v=v|x) directly) or drop the unbiasedness claim and reposition the paper as a heuristic framework. A revised version with correct theory and a comparison to [9] could be worth a serious look.","headline":"The paper names a real gap in defining intervention-conditioned ATE for event sequences, but the central IPW estimator is biased as stated and the proof of unbiasedness does not hold.","tokens_in":14491,"tokens_out":2412,"would_cite":false,"duration_ms":28914,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper extends average treatment effect estimation to temporal event sequences under out-of-domain interventions and claims an inverse-probability-weighted estimator that is unbiased for the intervention-conditioned effect.","keywords":["causal inference","temporal point processes","out-of-domain interventions","average treatment effect","inverse probability weighting","Transformer","event sequences","Hawkes process"],"falsifier":"Simulate a point process with a known positive effect of cause $c$ on outcome $o$, let intervention $v$ toggle independently of $c$ so all four combinations $(c,v)\\in\\{0,1\\}^2$ appear, and estimate $\\hat{\\tau}(v)$ with the paper's plug-in estimator; if the control-weight misspecification matters, the estimate's bias will fail to shrink to zero as the number of sequences grows, which would contradict Theorem 2's unbiasedness claim.","tokens_in":13413,"feed_emoji":"🎯","tokens_out":7994,"duration_ms":89743,"temperature":0.7,"pith_summary":"This paper tackles a gap in causal analysis of event sequences: an external \"out-of-domain\" intervention can change which event types cause which, and most existing models assume the causal structure is fixed. The authors propose a causal framework that defines the average treatment effect of a cause event on an outcome's intensity separately for windows where the intervention is present or absent. They define a propensity score for the cause-and-intervention pair and an inverse-probability-weighted estimator, and their Theorem 2 asserts this estimator is unbiased under standard causal assumptions adapted to point processes. A Transformer-CNN model with intervention embeddings estimates the outcome intensity that feeds the estimator. If the unbiasedness claim is correct, practitioners could quantify causal relation shifts from observational logs rather than from randomized experiments.","feed_headline":"New estimator recovers causal shifts under out-of-domain interventions","feed_subtitle":"A weighted estimator promises unbiased treatment effects in temporal event data, tested on diabetes and machine-maintenance logs.","key_machinery":"The central object is the intervention-conditioned average treatment effect $\\tau(v^w_t)$, defined as the expected time-average difference between the counterfactual outcome intensity with cause present, $\\lambda(1,v^w_t)(t)$, and with cause absent, $\\lambda(0,v^w_t)(t)$, within a proximal time window. Carrying the argument is the propensity score $e^w_t(v)=P(c^w_t=1, v^w_t=v \\mid x^w_t)$, which the paper estimates by a duration ratio of observed window states. The weight $\\alpha^w_t(v)=\\frac{\\mathbf{1}\\{c^w_t=1,v^w_t=v\\}}{e^w_t(v)}-\\frac{\\mathbf{1}\\{c^w_t=0,v^w_t=v\\}}{1-e^w_t(v)}$ converts observed intensities into inverse-probability weighting, and Theorem 2 claims this adjustment cancels confounding. The neural component is a Transformer with event embeddings, trigonometric time encodings, and binary intervention encodings combined through weighted sums, followed by a CNN over basis-function weights that outputs the conditional intensity.","core_discovery":"On its own terms, the paper's central claim is that causal relation shifts induced by out-of-domain interventions are estimable from observed event sequences. It defines the intervention-conditioned average treatment effect as $\\tau(v^w_t)$, comparing the conditional intensity of an outcome under cause present versus absent, within proximal windows where the out-of-domain intervention variable $v^w_t$ is fixed. The estimator reweights observed intensities by $\\alpha^w_t(v)$, built from the propensity score $e^w_t(v)=P(c^w_t=1,v^w_t=v\\mid x^w_t)$, and Theorem 2 asserts that under SUTVA, unconfoundedness, and overlap, $\\mathbb{E}[\\hat{\\tau}(v)]=\\tau(v)$ for both $v=0$ and $v=1$. The paper also introduces a hybrid Transformer-CNN point-process model that encodes intervention activity as part of the input and estimates the latent intensity $\\lambda(t)$; experiments on simulated interventions and two real datasets, diabetes and predictive maintenance, compare favorably to a Granger-causality baseline.","pith_inferences":["An unstated condition for Theorem 2 appears to be that the only possible joint states are $(c=1,v=v)$ and $(c=0,v=v)$; if other cause-intervention combinations occur, $1-e^w_t(v)$ is not the control-arm probability, and the control component of the IPW weight is mis-specified.","Replacing the scalar propensity score with a joint propensity score over all four cells would preserve inverse-probability weighting when $c$ and $v$ toggle independently, and would be directly testable on the paper's simulated data.","Because the estimator is a linear weighting of estimated intensities, the main practical bottleneck is accurate propensity-score estimation; the paper's duration-ratio estimator assumes covariates and windows are stable enough for that ratio to be a good probability.","The reported real-data gains over a Granger-causality baseline are likely to reflect both the causal framing and the intervention-aware architecture; a controlled ablation that removes only the intervention embedding would separate the two sources."],"forward_implications":["Causal relation shifts under external interventions can be estimated from observational event logs without randomized assignment, provided the binary window model is adequate.","The estimated $\\tau(1)$ versus $\\tau(0)$ comparison gives a quantitative signal for when an intervention creates, removes, or reverses a cause-outcome relationship.","Predictive systems that fit the outcome intensity can absorb intervention information directly through the embedding, which the paper reports improves prediction error on maintenance logs and diabetes records.","The framework offers a point-process analogue of the classical treatment-effect decomposition, with the intervention modeled as a second binary treatment rather than as effect modification."],"supporting_citations":[{"why":"Supplies the IPW treatment-effect estimator for event pairs in multivariate point processes that this work extends to out-of-domain interventions.","marker":"[9]"},{"why":"Supplies the potential-outcomes and propensity-score framework from which the paper's assumptions and estimator are adapted.","marker":"[11]"},{"why":"Supplies the process-independence definition used to establish direct cause-outcome relations in point processes.","marker":"[7]"},{"why":"Supplies the proximal-history window simplification that defines the binary cause and intervention variables.","marker":"[3]"},{"why":"Provides the CAUSE baseline, a Granger-causality model that estimates intensity without representing out-of-domain interventions.","marker":"[29]"},{"why":"Provides a neural point-process intensity model whose log-intensity formulation informs the paper's conditional intensity estimation.","marker":"[18]"},{"why":"Supplies the Transformer Hawkes process and temporal positional encoding used as the architectural backbone.","marker":"[32]"},{"why":"Provides the Hawkes-process simulation library that the paper modifies to inject out-of-domain interventions in the numerical study.","marker":"[2]"}],"fun_headline_variants":["Causal shifts in event sequences under out-of-domain interventions","New estimator captures causal relation shifts in events","Unbiased ATE estimation under out-of-domain interventions","How out-of-domain interventions alter event causalities","Transformer method unveils causal shifts from outside-domain effects"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that $1-e^w_t(v)$ correctly gives the probability of observing no cause with the same intervention, but because $e^w_t(v)$ is defined to include only the case where the cause is present, its complement also includes other combinations of cause and intervention, so the estimator's control weight is correct only if those other combinations do not occur.","fun_headline_variants_meta":{"raw":{"variants":["Causal shifts in event sequences under out-of-domain interventions","New estimator captures causal relation shifts in events","Unbiased ATE estimation under out-of-domain interventions","How out-of-domain interventions alter event causalities","Transformer method unveils causal shifts from outside-domain effects"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000218,"raw_usage":{"total_tokens":1434,"prompt_tokens":933,"completion_tokens":501,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":549,"completion_tokens_details":{"reasoning_tokens":427}},"tokens_in":549,"tokens_out":501,"duration_ms":6127,"temperature":1.0,"reasoning_tokens":427,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T17:25:12.576729+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Simulate a point process with a known positive effect of cause $c$ on outcome $o$, let intervention $v$ toggle independently of $c$ so all four combinations $(c,v)\\in\\{0,1\\}^2$ appear, and estimate $\\hat{\\tau}(v)$ with the paper's plug-in estimator; if the control-weight misspecification matters, the estimate's bias will fail to shrink to zero as the number of sequences grows, which would contradict Theorem 2's unbiasedness claim.","supporting_citations":[{"cited_title":"AdvancesinNeuralInformationProcessingSystems 34,17311–17324(2021)","cited_arxiv_id":null,"evidence_quote":"Supplies the IPW treatment-effect estimator for event pairs in multivariate point processes that this work extends to out-of-domain interventions."},{"cited_title":"Cambridge University Press (2015)","cited_arxiv_id":null,"evidence_quote":"Supplies the potential-outcomes and propensity-score framework from which the paper's assumptions and estimator are adapted."},{"cited_title":"null (2005).https://doi.org/null","cited_arxiv_id":null,"evidence_quote":"Supplies the process-independence definition used to establish direct cause-outcome relations in point processes."},{"cited_title":"null (2018).https://doi.org/null","cited_arxiv_id":null,"evidence_quote":"Supplies the proximal-history window simplification that defines the binary cause and intervention variables."},{"cited_title":"In: Inter- national Conference on Machine Learning","cited_arxiv_id":null,"evidence_quote":"Provides the CAUSE baseline, a Granger-causality model that estimates intensity without representing out-of-domain interventions."},{"cited_title":"Advances in neural information pro- cessing systems30 (2017)","cited_arxiv_id":null,"evidence_quote":"Provides a neural point-process intensity model whose log-intensity formulation informs the paper's conditional intensity estimation."},{"cited_title":"In: International conference on machine learning","cited_arxiv_id":null,"evidence_quote":"Supplies the Transformer Hawkes process and temporal positional encoding used as the architectural backbone."},{"cited_title":"Tick: a Python library for statistical learning, with a particular emphasis on time-dependent modelling","cited_arxiv_id":"1707.03003","evidence_quote":"Provides the Hawkes-process simulation library that the paper modifies to inject out-of-domain interventions in the numerical study."}],"review_version":1}