{"id":"d002d784-c139-4e8d-8ed5-9c00c5a283b9","arxiv_id":"2505.09113","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"DSIV-CFR learns instrumental variables from observed covariates and uses a generalized method of moments to estimate sequential treatment effects under unmeasured confounding, but the identification proof is incomplete and the experiments measure only factual prediction error.","lead":"This paper proposes a deep learning framework, DSIV-CFR, that aims to estimate sequential treatment effects from time series when some confounders are unobserved, by learning hidden instrumental variables from the data. The method combines transformers, negative control assumptions, and an adversarial moment condition, but the theoretical derivation and validation currently do not support the causal claims.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 3.8's proof is invalid: Eq. (3) is tautological and conditioning on Y_{t-1} cannot isolate ε(U_{t-1}) because Y_{t-1} also contains ε(U_{t-2}) per Eq. (2); the 'same function' assumption is unstated. IV identification is unproven.","rationale":"The reader's weakest_assumption pointed to Assumption 3.4 (E[U|X]=0, no autocorrelation) and the unstated same-function assumption in the proof of Theorem 3.8. I agree that the unstated same-function assumption is a genuine gap. However, the more direct and load-bearing problem is the inference step from Eq. (3) to Eq. (4): even if the same-function assumption is granted, conditioning on Y_{t-1} does not control ε(U_{t-1}), because Y_{t-1} also depends on ε(U_{t-2}) according to the paper's own Eq. (2). Eq. (3) is also tautological, so the proof does not establish the exclusion restriction that underlies IV recovery. This is an internal inconsistency, not merely a disagreement with common practice, so it is a correctness risk. The empirical MSE gains in Table 2 are suggestive but do not validate the counterfactual estimand; a model can fit factual outcomes well while failing to identify potential outcomes. The Impact Statement's admission that E[U|X]≠0 biases the estimand reinforces the fragility but is secondary to the invalid derivation of Theorem 3.8. Since the central claim rests on this theorem and the proof fails under the paper's own model class, the REJECT verdict should stand. The proposed SCM test would settle whether Eq. (4) can hold at all under the stated assumptions; if it fails, the theoretical contribution is unsupported.","tokens_in":15991,"tokens_out":12137,"duration_ms":102417,"concrete_test":"Construct the minimal SCM implied by Eq. (2) with t=2: independent standard normal U_0,U_1,U_2; observed X_1 = Z_1 + U_1 (so the recovered instrument is contaminated); Y_1 = h_1(A_1,C_0) + U_0 + U_1; Y_2 = h_2(A_2,A_1,C_1) + U_2 + U_1; A_2 ~ Bernoulli(σ(α Z_1 + β C_1 + γ U_1)). This satisfies Assumptions 3.1–3.5 and 3.7 with E[U|X]=0 and the same additive U_1 term in both outcomes. On a large synthetic sample, test whether E[Y_2|A_2,A_1,C_1,Z_1,Y_1] = E[Y_2|A_2,A_1,C_1,Y_1] for all relevant values. If the equality fails, Eq. (4) is false even within the paper's own model class, so Theorem 3.8 and the IV recovery collapse.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Theorem 3.8 is the theoretical foundation of DSIV-CFR: it asserts Z_{t-1} ⊥ Y_t | {A_t, \\bar{A}_{t-1}, \\bar{C}_{t-1}, Y_{t-1}}, justifying recovery of latent IVs from X and the moment condition in Eq. (5). The proof is not valid. First, it silently introduces a new assumption — \"we assume that the function of U_{t-1} on Y_{t-1} and Y_t is the same\" — which appears nowhere in Assumptions 3.1–3.5 or 3.7. Second, Eq. (3) is tautological: conditioning on ε(U_{t-1}) makes h(A_t,·)+ε(U_{t-1}) deterministic, so it cannot justify anything. Third, the leap from Eq. (3) to Eq. (4) replaces ε(U_{t-1}) by Y_{t-1} in the conditioning set. From Eq. (2), Y_{t-1} = h(A_{t-1}, \\bar{A}_{t-2}, \\bar{C}_{t-2}) + ε(U_{t-2}) + ε(U_{t-1}); even conditional on \\bar{A}_{t-1}, \\bar{C}_{t-1}, Y_{t-1} still contains the extra noise ε(U_{t-2}). Therefore conditioning on Y_{t-1} does not fix ε(U_{t-1}), and independence (4) does not follow. Consequently the recovered instrument may remain dependent on Y_t through U_{t-1}, and Eq. (5) — E[Y_t|\\bar{C}_{t-1},Z_{t-1}] = ∫ h dF(A_t|\\bar{A}_{t-1},\\bar{C}_{t-1},Z_{t-1}) — is unsupported. Since the experiments evaluate only factual MSE (Eq. 17), not the counterfactual estimand E[Y_t(a_t)|\\bar{H}_{t-1}], the central claim lacks a valid identification argument. The Impact Statement's E[U|X]≠0 caveat further exposes fragility, but the proof gap is more fundamental.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper studies the estimation of sequential treatment effects E[Y_t(a_t) | H_{t-1}] in the presence of unmeasured confounders. The proposed method, DSIV-CFR, uses a transformer to embed the observed history, separates the covariates into instrument and confounder representations via mutual information losses, and then estimates the outcome with a GMM-style two-stage regression that includes an adversarial loss on the residuals. The identification argument relies on Theorem 3.8, which claims that a latent instrument Z_{t-1} can be recovered from observed covariates using the previous outcome as a negative control, and on the moment condition in Eq. (5). Experiments are reported on one synthetic dataset, Tumor Growth, Cryptocurrency, and MIMIC-III, with one-step-ahead MSE and a five-step decision-making experiment.","tokens_in":16570,"tokens_out":8827,"duration_ms":77548,"significance":"The problem addressed is important, and a valid method for sequential treatment effects with unmeasured confounders would be a significant contribution. The paper contains concrete algorithmic details, a hyperparameter study, running-time measurements, and an extended decision-making experiment, which are useful elements of an experimental study. However, the theoretical foundation of the method is not established. The proof of the central identification theorem contains unstated assumptions and invalid steps, the moment condition is not derived correctly, and the empirical evaluation uses factual prediction error rather than counterfactual outcomes. As a result, the paper's central claim is currently unsupported. If the identification result can be repaired and the experiments re-focused on the target estimand, the approach could become promising; in its present form the contribution does not meet the standard for publication.","major_comments":[{"comment":"The proof of Theorem 3.8 is not valid as written. It introduces an unstated assumption in the proof paragraph: 'we assume that the function of U_{t-1} on Y_{t-1} and Y_t is the same,' which appears neither in Assumptions 3.1-3.5 nor in Assumption 3.7. Eq. (3) is tautological: once epsilon(U_{t-1}) is conditioned on, the expression h(A_t, A_{t-1}, C_{t-1}) + epsilon(U_{t-1}) is deterministic given A_t, A_{t-1}, C_{t-1}, so its independence from Z_{t-1} is vacuous and cannot serve as a characterization of a valid instrument. The step from Eq. (3) to Eq. (4) replaces the conditioning on epsilon(U_{t-1}) with conditioning on Y_{t-1}; by the paper's own Eq. (2), Y_{t-1} also contains epsilon(U_{t-2}), so conditioning on Y_{t-1} does not fix epsilon(U_{t-1}). Consequently the claimed conditional independence Z_{t-1} independent of Y_t given {A_t, A_{t-1}, C_{t-1}, Y_{t-1}} is unsupported, and the recovered representation cannot be claimed to be a valid instrument.","section":"Section 3, Theorem 3.8 and proof"},{"comment":"Eq. (5) is not a valid derivation of the moment condition. The left-hand side conditions on C_{t-1} and Z_{t-1}, while the right-hand side integrates h(A_t, A_{t-1}, C_{t-1}) over the treatment distribution conditional on A_{t-1}, C_{t-1}, and Z_{t-1}; these conditioning sets must match, and a term conditioning on A_{t-1} is missing on the left. In addition, the term E[epsilon(U_{t-1}) | C_{t-1}, Z_{t-1}] is dropped using Assumption 3.4's E[U|X]=0, but epsilon is an arbitrary function, so E[epsilon(U)|X]=0 does not follow from E[U|X]=0. The expectation also omits E[epsilon(U_t) | C_{t-1}, Z_{t-1}] that appears in Eq. (2). The Impact Statement admits that E[U|X] not equal to 0 biases the potential outcome, but even under E[U|X]=0 the displayed moment condition is not established.","section":"Section 3, Eq. (5)"},{"comment":"The adversarial objective in Eq. (18) is unbounded. The bridge function f is unconstrained, so M_{t-1}=f(A_{t-1}, C_{t-1}, Z_{t-1}) can take arbitrarily large values; for a fixed h the maximum over f of sum_i M_i (Y_hat_i - Y_i) is either +infinity or -infinity. No normalization or boundedness condition is imposed on the weights. The text describes this as 'a special form of the generalized method of moments,' but no argument connects Eq. (18) to the moment condition in Eq. (5). Without a well-posed minimax objective, the overall optimization in Eq. (19) is not well defined.","section":"Section 4.2, Eq. (18)"},{"comment":"The reported one-step-ahead results evaluate factual predictive accuracy, not the counterfactual estimand in Eq. (1). The loss in Eq. (17) compares h(A_t^i, A_{t-1}^i, C_{t-1}^i) to the factual outcome Y_t^i observed under the assigned treatment A_t^i. This does not measure E[Y_t(a_t)|H_{t-1}] for alternative treatments, so the large improvement over baselines in Table 2 is not evidence that unmeasured confounding bias is corrected. The five-step decision-making experiment does compare against oracle counterfactual outcomes, but it is limited to one synthetic setting; no counterfactual evaluation is reported on the real-world datasets.","section":"Section 5, Table 2 and Eq. (17)"},{"comment":"The representation-learning losses impose the conditions that define the instrument: L_ZY enforces Z_{t-1} independent of Y_t given the conditioning set, and L_ZC enforces Z_{t-1} independent of C_{t-1}. Since the representation is optimized to satisfy these conditions, the good performance in Table 2 cannot be read as empirical evidence that valid instruments are discovered from observed covariates; the conditions are built into the objective rather than tested. This does not, by itself, invalidate the estimation strategy, but it means the experiments provide no independent support for the identification assumption.","section":"Section 4.1, Eqs. (8)-(15)"}],"minor_comments":[{"comment":"The line 'Calculate overall loss L <- L_MSE by Equation (19)' appears to be a typo; it should read L <- L_MSE + alpha*L_MI + beta*L_adv as in Eq. (19).","section":"Algorithm 1, near Eq. (19)"},{"comment":"The notation E[epsilon(U)t-1 | C_{t-1}, Z_{t-1}] contains a misplaced parenthesis; it should presumably be E[epsilon(U_{t-1}) | C_{t-1}, Z_{t-1}], and the surrounding derivation should be corrected accordingly.","section":"Section 3, Eq. (5)"},{"comment":"The softmax normalization of the RBF kernel weights is unclear as written because softmax is applied to a single scalar exp(-||v_i-v_j||^2/2 sigma^2) rather than to a vector; please specify the domain over which the softmax is taken and how this yields the claimed conditional independence weighting.","section":"Section 4.1, Eq. (11)"},{"comment":"Assumption 3.4 informally states Y_{t-1} does not cause Y_t and U_{t-1} does not cause U_t, but Eq. (2) places the same epsilon(U_{t-1}) in both Y_{t-1} and Y_t; the formal relationship between this assumption and the structural equations should be clarified.","section":"Section 3, Assumption 3.4"},{"comment":"The abstract and Section 1 claim the method identifies optimal treatments for dynamic systems, but the decision-making evaluation is conducted only on the synthetic simulator over a horizon of 5 steps; the wording should be tempered to match the evidence.","section":"Abstract and Section 1"}],"recommendation":"reject","confidential_remarks":"To the editor: the identification theorem is the foundation of the paper and its proof is invalid; the experimental metric does not match the claimed estimand. These are not presentation issues, and I do not see how they can be fixed within the scope of a revision. I recommend rejection, though a substantially rewritten paper with a correct identification argument and counterfactual evaluation could be reconsidered."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nThe central claim of this paper is that DSIV-CFR can estimate sequential treatment effects under unmeasured confounding by recovering latent instrumental variables from observed covariates using negative controls and GMM. That claim is not yet supported. The identification proof in Theorem 3.8 is invalid as written, and the experiments validate factual outcome prediction, not the counterfactual estimand. Still, this is a serious attempt at a hard problem, and the machinery is worth a careful look.\n\nWhat is genuinely new is the specific combination of existing building blocks — negative control proxies, IV decomposition, generalized method of moments, and transformer backbones — into a single sequential deep learning pipeline. That combination is new, and the paper does a reasonable job situating itself in the literature.\n\nWhat the paper does well: the method is fully specified, the authors run experiments on four datasets, and the factual one-step-ahead MSE is much lower than the baselines. I also credit the honest Impact Statement, which admits that E[U|X]=0 limits unbiasedness of the potential outcome estimator.\n\nNow the soft spots. They are not minor.\n\nFirst, the proof of Theorem 3.8 introduces an unstated assumption — \"we assume that the function of U_{t-1} on Y_{t-1} and Y_t is the same\" — which appears only in the proof paragraph. More seriously, Eq. (3) is tautological: conditioning on ε(U_{t-1}) makes the expression deterministic, so it cannot justify anything. The step from Eq. (3) to Eq. (4) replaces ε(U_{t-1}) with Y_{t-1} in the conditioning set, but Y_{t-1} also contains ε(U_{t-2}), so it does not isolate the unmeasured confounder. The claimed independence Z_{t-1} ⊥ Y_t | {A_t, ¯A_{t-1}, ¯C_{t-1}, Y_{t-1}} is therefore not established. Eq. (5) also omits ¯A_{t-1} from the left-hand conditioning set.\n\nSecond, the adversarial loss in Eq. (18) is unconstrained. The max over the bridge function f is unbounded, so the GMM interpretation does not follow as written. The authors need to restrict the function class or justify the min-max game more carefully.\n\nThird, the experiments report factual MSE (Eq. 17), which is a prediction metric, not a causal one. No comparison against ground-truth potential outcomes is presented for the one-step-ahead task, so the headline empirical result does not actually test the causal estimand.\n\nThe paper is repairable, but the central identification argument must be redone or restated with the missing assumptions made explicit, and the experiments need to evaluate counterfactual accuracy. I would send it to a serious referee — the problem matters and the negative-control-plus-IV idea is worth exploring — but I would expect major revision.\n\nBest,\n[You]","headline":"A serious attempt at a hard problem, but the identification proof is invalid and the experiments validate only factual prediction, so the causal claims are unsupported.","tokens_in":17030,"tokens_out":4742,"would_cite":false,"duration_ms":39991,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper proposes DSIV-CFR, a framework that estimates sequential treatment effects under unmeasured confounders by recovering latent instrumental variables from observed covariates using negative controls.","keywords":["sequential treatment effects","unmeasured confounding","instrumental variables","negative controls","counterfactual regression","generalized method of moments","time series causal inference"],"falsifier":"Simulate the paper's data-generating process, then break one load-bearing premise at a time: shift $E[U\\mid X]$ away from zero, add a direct edge $Y_{t-1}\\rightarrow Y_t$, or let $U_{t-1}$ affect $Y_{t-1}$ and $Y_t$ with different coefficients. Comparing DSIV-CFR's estimated treatment-response function $h$ to the true one in these variants would show whether the identification holds under the assumed graph and fails when the graph changes.","tokens_in":15821,"feed_emoji":"🎯","tokens_out":9532,"duration_ms":86597,"temperature":0.7,"pith_summary":"The paper tackles a standard failure mode in sequential decision-making: treatments are chosen partly on hidden factors that also affect future outcomes, so estimators that ignore those factors are biased. It claims that the causal effect of a sequence of treatments can still be recovered by treating the previous outcome as a negative control outcome and a hidden instrumental variable as a negative control exposure. The identification rests on an additive-noise model plus a generalized moment condition, which turns counterfactual regression into a two-stage instrumental-variable problem. If the claim holds, optimal treatment plans can be selected from observational records without ever measuring the confounders, which matters for medicine, finance, and policy.","feed_headline":"Hidden confounders no longer block sequential treatment effects","feed_subtitle":"Negative-control decomposition recovers latent instruments from history, debiasing counterfactual forecasts.","key_machinery":"The carrying object is the IV/confounder decomposition of the observed pre-treatment covariates, enforced by mutual-information losses, together with the generalized method of moments (GMM), a standard way to estimate parameters from moment conditions. The moment condition is $$E[Y_t\\mid\\bar{C}_{t-1},Z_{t-1}]=\\int h(A_t,\\bar{A}_{t-1},\\bar{C}_{t-1})\\,dF(A_t\\mid\\bar{A}_{t-1},\\bar{C}_{t-1},Z_{t-1}).$$ The previous outcome $Y_{t-1}$ serves as a negative control outcome because, under the additive-noise assumption, the unmeasured confounder $U_{t-1}$ is the only common cause it shares with $Y_t$. The bridge function $f$ learns weights $M_{t-1}$ in an adversarial loss that enforces the moment condition while the outcome network $h$ minimizes prediction error.","core_discovery":"The central claim is Theorem 3.8's identification: from observed history one can decompose the covariates $\\bar{X}_{t-1}$ into an instrument $Z_{t-1}$ and confounders $\\bar{C}_{t-1}$ such that $Z_{t-1}\\perp\\perp \\bar{C}_{t-1}$ and $Z_{t-1}\\perp\\perp Y_t\\mid\\{A_t,\\bar{A}_{t-1},\\bar{C}_{t-1},Y_{t-1}\\}$. Under this decomposition, the conditional expectation $E[Y_t\\mid\\bar{C}_{t-1},Z_{t-1}]$ becomes the integral of the treatment-response function $h$ against the conditional treatment distribution, so $h$ is identified by two-stage regression. The paper implements this by learning $\\phi_Z$ and $\\phi_C$ representations with a transformer, enforcing the independence conditions through mutual-information losses, and adding a bridge function that supplies adversarial moment weights. In experiments on four datasets it reports lower one-step-ahead prediction error than six baselines, and on a simulator it selects five-step treatment plans close to the oracle.","pith_inferences":["Inference: The same negative-control construction could transfer to off-policy evaluation in reinforcement learning, where a lagged action and a lagged reward often satisfy the required exclusion restrictions; testing the moment condition on logged bandit or RL data would be a natural next step.","Inference: The paper's own Impact Statement concedes that when $E[U\\mid X]\\neq 0$ the outcome prediction is biased but the treatment-effect estimate may remain consistent, so a follow-up could report effect-error separately from prediction error.","Inference: Because the additive-noise assumption forbids autoregressive outcomes and persistent hidden confounders, real-world use would hinge on preprocessing such as detrending or differencing to make residuals approximately independent, a burden the paper does not discuss."],"forward_implications":["If the identification holds, one can estimate personalized counterfactual outcomes for the next treatment without measuring latent confounders, using only observed history and the previous outcome as a negative control.","The learned IV and confounder representations plug directly into a GMM objective, giving a principled estimation target that goes beyond plain supervised loss and extends two-stage least squares to high-dimensional sequential settings.","Multi-step decision making follows by enumerating treatment sequences and scoring each with the estimated response function, which the paper demonstrates on a simulator with a five-step horizon.","If the framework transfers to real clinical, financial, or policy logs, treatment policies can be selected from observational records where unmeasured lifestyle, socioeconomic, or market factors would otherwise confound decisions."],"supporting_citations":[{"why":"Supplies the negative-control formulation that motivates using an instrument as a negative control exposure and the previous outcome as a negative control outcome.","marker":"[Miao et al., 2018]"},{"why":"Extends negative controls to time series and provides the confounding bridge interpretation used for the generalized moment condition.","marker":"[Miao et al., 2024]"},{"why":"Provides the nonparametric instrumental-variable estimation result that makes Equation (5) an inverse problem for the response function h.","marker":"[Newey & Powell, 2003]"},{"why":"Supplies the generalized method of moments framework that the paper's two-stage regression is presented as a special case of.","marker":"[Hall, 2003]"},{"why":"Motivates the automatic decomposition of covariates into instrument and confounder representations via independence conditions.","marker":"[Yuan et al., 2022]"},{"why":"Provides the contrastive log-ratio upper bound used as the mutual-information objective for disentangling instruments from confounders.","marker":"[Cheng et al., 2020]"}],"fun_headline_variants":["Debias sequential treatment effects with negative control decomposition","Decomposing history recovers instruments for unbiased sequential effects","Latent instruments from negative controls fix confounded treatment effects","Unmeasured confounders tamed via decomposition of observation history","New framework identifies sequential treatment effects despite hidden confounders"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the unmeasured confounder makes no net contribution after conditioning on observed covariates ($E[U\\mid X]=0$), that neither the hidden confounder nor the outcome carries over from one time step to the next, and that the hidden confounder enters the previous and current outcomes through the same functional form.","fun_headline_variants_meta":{"raw":{"variants":["Debias sequential treatment effects with negative control decomposition","Decomposing history recovers instruments for unbiased sequential effects","Latent instruments from negative controls fix confounded treatment effects","Unmeasured confounders tamed via decomposition of observation history","New framework identifies sequential treatment effects despite hidden confounders"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000411,"raw_usage":{"total_tokens":2139,"prompt_tokens":969,"completion_tokens":1170,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":585,"completion_tokens_details":{"reasoning_tokens":1090}},"tokens_in":585,"tokens_out":1170,"duration_ms":9942,"temperature":1.0,"reasoning_tokens":1090,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T21:40:11.339150+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Simulate the paper's data-generating process, then break one load-bearing premise at a time: shift $E[U\\mid X]$ away from zero, add a direct edge $Y_{t-1}\\rightarrow Y_t$, or let $U_{t-1}$ affect $Y_{t-1}$ and $Y_t$ with different coefficients. Comparing DSIV-CFR's estimated treatment-response function $h$ to the true one in these variants would show whether the identification holds under the assumed graph and fails when the graph changes.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the negative-control formulation that motivates using an instrument as a negative control exposure and the previous outcome as a negative control outcome."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the nonparametric instrumental-variable estimation result that makes Equation (5) an inverse problem for the response function h."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the generalized method of moments framework that the paper's two-stage regression is presented as a special case of."},{"cited_title":"Auto IV: counterfactual prediction via automatic instrumental variable decomposition","cited_arxiv_id":null,"evidence_quote":"Motivates the automatic decomposition of covariates into instrument and confounder representations via independence conditions."},{"cited_title":"CLUB: A contrastive log-ratio upper bound of mutual information","cited_arxiv_id":null,"evidence_quote":"Provides the contrastive log-ratio upper bound used as the mutual-information objective for disentangling instruments from confounders."}],"review_version":1}