{"id":"a247bb8f-b8ab-4048-86fe-6f28ba07e09e","arxiv_id":"2411.17774","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A new LSTM-VAE method, TDCIV, learns time-varying conditional instrumental variables and their conditioning sets from proxy variables to debias causal effect estimates in time-series data.","lead":"This paper proposes TDCIV, a deep-learning method that learns time-varying conditional instrumental variables from time-series data to estimate causal effects when hidden confounders change over time. It combines LSTM memory with variational autoencoders and evaluates the approach on synthetic and climate data.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central claim hinges on a learned disentanglement of S_t and Z_t that Theorem 1 assumes rather than establishes, and the synthetic generator does not actually include a proxy of S_t in X_t.","rationale":"The reader's weakest assumption correctly identifies that Theorem 1 rests on an unproved disentanglement claim and that the synthetic data generator does not implement the proxy assumption. My stress-test confirms and sharpens this: the gap is not merely a missing formal detail but a structural mismatch between the theory and the only empirical validation. The paper's stated contribution is to learn time-varying CIV and conditioning-set representations from proxies without domain knowledge; for that claim to hold, the objective in Eq. (10) must be shown to recover the S_t and Z_t of Figure 3, or at least a representation that satisfies Definition 3. No such identifiability result is present. The experiments, which are the only evidence that the learned representations are valid, use a generator where S_t is not connected to X_t at all, so the 'at least one proxy variable' assumption is vacuously false. TDCIV's good performance in that setting suggests it is exploiting information from W_t (or from the outcome) rather than from a genuine proxy, which is precisely what an IV method must not do. This is load-bearing because if the representations can be learned only through the treatment, the estimated ACE is not identified. I therefore agree with the reader's conditional verdict: the paper should either provide an identifiability/disentanglement guarantee, or clearly reframe the contribution as heuristic; the synthetic evaluation should be revised to include proxy variables; and code should be released to verify the reported results. These are the same conditions the reader imposed, so no change to the verdict is needed.","tokens_in":19839,"tokens_out":5131,"duration_ms":52195,"concrete_test":"Regenerate the synthetic experiments with the proxy assumption explicitly imposed, e.g. replace the first component of Eq. (12) with X_t^1 = S_t + ε_S (or another noisy function of S_t) while keeping Eq. (13)-(16) unchanged; then rerun the Figure 5/Table II protocol and report |ACE_t| errors. If TDCIV's error remains low under this proxy-realizing generator, the current Eq. (12) is not a valid test of the method's assumptions; if the error degrades or becomes biased, the reported good results depend on leakage from W_t, which violates the IV exclusion restriction.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Theorem 1 (Section III-B) is conditional: it proves validity of S_t and Z_t only 'if the representations ... can be disentangled and learned' as in Figure 3. No identifiability or disentanglement result is provided for the LSTM-VAE objective in Eq. (10). The objective rewards predicting W_t from S_t and Y_{t+1} from Z_t, but a time-varying latent confounder U_t also predicts both; without an exclusion restriction or an auxiliary-variable identifiability argument, infinitely many latent decompositions can fit the same ELBO, and the 2SLS ratio in Eq. (11) may then be computed using a non-instrument. More concretely, the synthetic data generator violates the paper's own proxy assumption: Eq. (12) defines X_t as a function of X_{t-i} and W_{t-i} only, with no dependence on S_t or Z_t, while S_t is generated separately in Eq. (13). Thus no observed covariate proxies the latent CIV, contradicting Section III-A's assumption that at least one proxy variable is available at each step. The only way TDCIV could recover S_t in this setup is from W_t or other downstream effects, which would be circular: using a function of the treatment as its own instrument invalidates the IV exclusion restriction. Consequently, the empirical results in Section IV do not test the method under the assumptions of Theorem 1, and the central claim that TDCIV learns time-varying CIVs without domain knowledge remains unsupported.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes TDCIV, a method that combines LSTM and variational autoencoders to learn latent representations of a time-varying conditional instrumental variable (CIV) S_t and its conditioning set Z_t from observed time-series data, and then applies two-stage least squares (2SLS) to estimate the average causal effect of a treatment W_t on an outcome Y_{t+1} in the presence of time-varying latent confounders. The authors define a time-varying CIV in a full-time DAG, state Theorem 1 asserting that if the representations of S_t and Z_t can be disentangled and learned as in their Figure 3, then these representations, together with historical data, instrumentalise S_t relative to W_t -> Y_{t+1}. The paper reports synthetic and real-world climate experiments that appear to show that TDCIV outperforms several causal-effect estimators. However, the central claim that TDCIV actually learns valid time-varying CIV representations without domain knowledge is not established: Theorem 1 is conditional on a disentanglement assumption that is never proved, and the synthetic data generator does not satisfy the paper's own proxy-variable assumption.","tokens_in":20189,"tokens_out":6917,"duration_ms":62869,"significance":"The problem addressed is important: estimating causal effects from time-series data with time-varying latent confounders, while avoiding the need for predefined instrumental variables, would be a valuable contribution to causal inference. The formalization of a time-varying CIV in a full-time DAG (Definition 3) is a useful conceptual step, and the idea of using deep generative models for this task is interesting. If the central claim were rigorously supported, the method would likely be of significant interest to the machine learning and causal inference communities. As the manuscript stands, however, the theoretical result is conditional and the experiments do not test the method under its stated assumptions, so the significance is potential rather than demonstrated.","major_comments":[{"comment":"Theorem 1 is conditional in a way that leaves the paper's central claim unproved. The theorem states that if the representations of S_t and Z_t 'can be disentangled and learned' as in Figure 3, then they instrumentalise S_t; the proof is only a d-separation check within an assumed graph. No identifiability or disentanglement result is provided for the LSTM-VAE objective in Eq. (10), so the antecedent of the theorem is not established. The abstract and introduction claim that the paper 'theoretically establish[es] the validity of these learned representations,' which is stronger than what is actually shown. This is load-bearing because the 2SLS estimate in Eq. (11) is valid only if the learned S_t satisfies the CIV conditions.","section":"Section III-B (Theorem 1)"},{"comment":"The synthetic data generator violates the proxy-variable assumption of Section III-A. In Eq. (12), X_t is generated only from past values X_{t-i} and W_{t-i} plus noise, with no dependence on the current S_t or Z_t, while S_t and Z_t are generated separately in Eqs. (13)-(14). Thus no observed covariate at time t is a proxy of the latent CIV S_t, contradicting the assumption that at least one proxy variable is available at each step. The experiments therefore do not test TDCIV under the assumptions of Theorem 1, and the low reported errors in Figures 5-6 and Tables II-III cannot independently validate the central claim that the learned S_t is a valid time-varying CIV.","section":"Section IV-A, Eqs. (12)-(14)"},{"comment":"There is an inconsistency in what S_t is. Definition 3 defines a time-varying CIV as a variable S_t in X_t, while Section III-A states that S_t is a latent factor approximated through measurement errors in the observed covariates X_t, with at least one proxy variable available. These are different data-generating assumptions: if S_t is observed, no latent representation needs to be learned; if S_t is latent and only proxied, the conditions in Definition 3 should be stated for the latent variable and the proxy structure should be made explicit. This ambiguity matters because the validity conditions for 2SLS depend on which variable is the instrument.","section":"Section III-A and Definition 3"},{"comment":"The learning objective does not enforce the exclusion restriction or instrument exogeneity for the learned S_t. The term beta * E_q log p(Y_{t+1} | Z_t, H_t) predicts Y_{t+1} without S_t, but nothing prevents S_t from encoding information about Y_{t+1} through H_t or Z_t, and no constraint ensures that S_t is independent of the latent confounders U_t. Without additional identifiability assumptions, infinitely many latent decompositions can fit the ELBO, and the representation extracted as S_t may be correlated with U_t, making it an invalid instrument. The paper's claim that TDCIV learns a valid CIV without domain knowledge is therefore not supported by the objective as stated.","section":"Section III-C, Eq. (10)"}],"minor_comments":[{"comment":"The title contains a typo: 'Leaning' should be 'Learning.'","section":"Title"},{"comment":"The noise term for Z_t is written as epsilon_U, the same symbol used for the latent confounder noise in Eq. (12); this appears to be a typo and should be a distinct noise term.","section":"Section IV-A, Eq. (14)"},{"comment":"The equality E[Y_{t+1}(w_t) | W_{t-1}, X_t] = E[Y_{t+1} | w_t, W_{t-1}, X_t] is not true without a no-unmeasured-confounding assumption; as presented it conflates the definition of the ACE with its identification.","section":"Section II-A, Eq. (1)"},{"comment":"The proof of condition (i) states that 'S_t and W_t are dependent given any set of covariates,' which is not generally true in a DAG; conditioning on particular sets can create or remove dependence. The specific conditioning set in Definition 3 may be sufficient, but the statement as written is too strong.","section":"Section III-B, proof of Theorem 1"},{"comment":"The line 'Update parameters of LT DCIV' contains a typo (LT DCIV should be L_TDCIV or TDCIV).","section":"Algorithm 1"},{"comment":"The paper states that source code will be made publicly accessible upon publication; no code or data repository is provided at submission, which limits reproducibility verification.","section":"Section IV-A, evaluation"}],"recommendation":"major_revision","confidential_remarks":"The manuscript addresses an interesting problem, but the central claim that TDCIV learns valid time-varying CIV representations is currently not established. The main issues are substantive: Theorem 1 is conditional on a disentanglement assumption that is not proved, the learning objective has no identifiability guarantee, and the synthetic experiments do not implement the paper's own proxy-variable assumption. These are load-bearing problems that would require substantial additional theory and revised experiments to address. If the authors can provide an identifiability result under clearly stated assumptions (e.g., temporal structure, multiple proxies, or additional constraints) and then evaluate the method in a setting that satisfies those assumptions, a resubmission could be of interest. As it stands, the paper is not ready for publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The headline is that this paper is a legitimate but incomplete extension: it takes the static CIV.VAE idea, adds LSTM history, and calls the result time-varying CIV. The specific object is new, and the formal definition of a time-varying conditional IV in a full-time DAG is a useful contribution. What the paper does well: the d-separation proof in Theorem 1 is internally consistent as a conditional claim — if S_t and Z_t were learned to match Figure 3, then Z_t plus history does instrumentalise S_t. The synthetic experiments show TDCIV with low absolute error against reasonable baselines, and the climate case study is at least consistent with domain knowledge.\n\nThe soft spots are real and load-bearing. Theorem 1 assumes the disentanglement it needs to establish. The objective in Eq. (10) rewards predicting W_t and Y_{t+1} from the learned latents, but that is exactly what a latent confounder U_t also does. Without an exclusion restriction or an identifiability argument, infinitely many latent decompositions can fit the same ELBO, and the 2SLS ratio may be computed from a non-instrument. The paper acknowledges this in the Limitations section, but does not supply the missing proof.\n\nThe synthetic data are more troublesome. In Eq. (12), X_t is a function of past X and W only; S_t is generated separately in Eq. (13). So no observed covariate proxies S_t, contradicting the paper's own assumption in Section III-A that at least one proxy is available. The only way TDCIV can recover S_t in this setup is through W_t, which makes the \"instrument\" a function of the treatment — circular and invalid. As a result, the experiments do not test the method under the conditions of Theorem 1. Code is promised but not released, and key hyperparameters (alpha, beta, latent dimensions) are unreported, so the empirical claims are hard to reproduce.\n\nWho is this for: readers working on IV-based causal inference for longitudinal data will find the time-varying CIV definition and the conditional theorem worth thinking with, but nobody should take the empirical results as evidence that the method learns valid instruments from proxies. It deserves a serious referee, not a desk reject, because the formalization is useful and the gap is clearly fixable in principle: either prove identifiability under explicit assumptions or reframe the paper as a heuristic with a sensitivity analysis, rebuild the synthetic benchmark to actually include proxies of S_t, and release code. I would not cite the empirical claims in my own work, but I would bring it to reading group to discuss where representation learning for IVs gets circular.","headline":"A useful formalization of time-varying conditional instruments, but the central identifiability claim is assumed, not proved, and the synthetic experiments don't satisfy the paper's own proxy assumption.","tokens_in":20781,"tokens_out":2584,"would_cite":false,"duration_ms":23883,"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":"A deep generative method learns time-varying conditional instrumental variables directly from proxy variables, enabling unbiased causal effect estimates under time-varying latent confounders.","keywords":["causal effect estimation","time-series data","time-varying latent confounders","conditional instrumental variables","instrumental variable","variational autoencoder","LSTM","causal inference"],"falsifier":"Generate synthetic time series with a known latent time-varying confounder and no proxy variable for $S_t$, so the proxy assumption fails; if TDCIV still removes the confounding bias, the stated premise is not doing the work, and if it fails, the premise is confirmed necessary. Alternatively, train on data generated from a graph that violates Figure 3 by adding a direct edge $S_t \\to Y_{t+1}$; the 2SLS estimate should show bias proportional to that edge.","tokens_in":19624,"feed_emoji":"📈","tokens_out":5786,"duration_ms":49174,"temperature":0.7,"pith_summary":"The paper sets out to show that the average causal effect of a time-varying treatment on a later outcome can be estimated without bias even when unmeasured confounders change over time, by learning a time-varying conditional instrumental variable from observed data. Existing instrumental-variable methods for time series either require an expert to supply the instrument or rely on the stronger assumption of a plain time-varying IV; the proposed TDCIV instead disentangles a latent instrument representation $S_t$ and its conditioning set $Z_t$ from proxy variables using LSTM, VAE, and CVAE components. The central theoretical claim is that, if the learned representations match the assumed full-time causal graph, then $Z_t$ together with the historical record instrumentalises $S_t$ for the effect of $W_t$ on $Y_{t+1}$, so two-stage least squares gives unbiased estimates. A sympathetic reader would care because this would remove a key obstacle to causal effect estimation in healthcare, economics, climate science, and epidemiology when latent confounding is present and no domain-specific instrument is known.","feed_headline":"Hidden instruments learned from data debias time-series effects","feed_subtitle":"Learns time-varying instruments from proxy variables to remove latent-confounder bias, no expert input needed.","key_machinery":"The central object is the time-varying conditional instrumental variable $S_t$: a variable that behaves as an instrument only after conditioning on a set $Z_t$ and on the historical record. The carrying mechanism is the TDCIV architecture, which uses an LSTM to encode history $\\bar{H}_t$, a VAE to disentangle $S_t$ and $Z_t$ from proxy variables in $\\bar{X}_t$, and a CVAE to generate the conditioning set; the objective (Equation 10) augments the ELBO with predictors of $W_t$ and $Y_{t+1}$. Theorem 1 is what connects these learned representations to identification: it certifies the three CIV conditions in the full-time DAG, so the extracted $S_t$ and $Z_t$ can be fed to 2SLS for an unbiased effect estimate.","core_discovery":"The paper claims that time-varying conditional instrumental variables can be learned rather than predefined. In the full-time DAG it proposes, $S_t$ is a latent cause of treatment $W_t$, $Z_t$ is a conditioning set generated from observed covariates, and $Y_{t+1}$ is a sink outcome. Theorem 1 states that if the representations $S_t$ and $Z_t$ shown in Figure 3 can be disentangled and learned from the time-series data, then $Z_t$ together with the historical data $\\bar{S}_{t-1}, \\bar{W}_{t-1}, \\bar{Y}_t, \\bar{Z}_{t-1}$ makes $S_t$ a valid conditional instrument for $W_t \\to Y_{t+1}$ over time. The proof verifies the three CIV conditions: $S_t$ is associated with $W_t$ given the conditioning set; $S_t$ is independent of $Y_{t+1}$ in the graph with the $W_t \\to Y_{t+1}$ edge removed; and $Z_t$ contains no descendants of $Y_{t+1}$. Given that validity, the paper's 2SLS estimator (its Equation 11) recovers the average causal effect $ACE_t(W_t, Y_{t+1})$ at each time step despite time-varying latent confounders.","pith_inferences":["An implication the authors leave implicit is that the practical value of TDCIV hinges on whether the VAE objective identifies $S_t$ and $Z_t$; the paper proves validity of the representations only after they are correctly disentangled, and no identifiability theorem for Equation 10 is provided.","A direct testable extension would be to degrade the proxy assumption by adding increasing measurement noise to $\\bar{X}_t$; the paper's assumptions predict the bias of the 2SLS estimate should grow, which would isolate the mechanism.","The same conditioning-on-the-past logic could be carried over to multiple treatments or survival outcomes, settings where time-varying CIVs are currently not available.","If latent confounders also affect $S_t$ directly, the assumed graph in Figure 3 is violated and the d-separation argument breaks; detecting such violations is an open problem."],"forward_implications":["A researcher no longer needs to name an instrument ahead of time: the conditioning set and instrument are read from data.","Average causal effects can be estimated at each time step with time-varying latent confounders present, a setting where standard sequential-randomisation methods are biased.","The method applies beyond simulated data: its climate case study claims to recover known drivers of rainfall as the strongest time-varying causes.","Because it handles conditional rather than plain instruments, TDCIV is applicable where the instrument is only valid after adjusting for observed past states.","The learned representations feed any CIV-based estimator, so the 2SLS stage could be replaced by a nonlinear estimator without changing the identification argument."],"supporting_citations":[{"why":"Provides the static conditional-IV representation-learning approach that TDCIV extends to time series.","marker":"[22]"},{"why":"Supplies the framework for time-series instrumental variables with correction for the past, which the CIV conditions adapt.","marker":"[34]"},{"why":"Defines the two-stage least squares estimator used to convert learned $S_t$ and $Z_t$ into the causal effect estimate.","marker":"[23]"},{"why":"Gives the variational autoencoder machinery used to disentangle the latent representations.","marker":"[28]"},{"why":"Gives the conditional VAE machinery used to generate the conditioning set $Z_t$ from history.","marker":"[29]"},{"why":"Supports recovering latent causes from proxy variables with deep latent-variable models.","marker":"[37]"},{"why":"The closest prior time-varying IV method (TIFM), which learns a standard IV but not a conditional IV and is the main baseline to beat.","marker":"[17]"},{"why":"Provides the time-series deconfounder approach and the synthetic data-generation style used to evaluate performance under latent time-varying confounders.","marker":"[11]"}],"fun_headline_variants":["Learned time-varying instruments debias causal estimates","Auto-learned instruments remove hidden confounder bias in time series","Time-varying IVs learned from proxies debias causal effects","Model discovers conditional instruments for unbiased time-series effects","Learned instruments debias time series without expert knowledge"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing assumption is that at every time step at least one observed covariate is a proxy for the hidden time-varying conditional instrument, and that the learned representations $S_t$ and $Z_t$ match the assumed full-time causal graph.","fun_headline_variants_meta":{"raw":{"variants":["Learned time-varying instruments debias causal estimates","Auto-learned instruments remove hidden confounder bias in time series","Time-varying IVs learned from proxies debias causal effects","Model discovers conditional instruments for unbiased time-series effects","Learned instruments debias time series without expert knowledge"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001068,"raw_usage":{"total_tokens":4520,"prompt_tokens":1036,"completion_tokens":3484,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":652,"completion_tokens_details":{"reasoning_tokens":3407}},"tokens_in":652,"tokens_out":3484,"duration_ms":23094,"temperature":1.0,"reasoning_tokens":3407,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T12:19:25.621844+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Generate synthetic time series with a known latent time-varying confounder and no proxy variable for $S_t$, so the proxy assumption fails; if TDCIV still removes the confounding bias, the stated premise is not doing the work, and if it fails, the premise is confirmed necessary. Alternatively, train on data generated from a graph that violates Figure 3 by adding a direct edge $S_t \\to Y_{t+1}$; the 2SLS estimate should show bias proportional to that edge.","supporting_citations":[{"cited_title":"Causal inference with conditional instruments using deep generative models,","cited_arxiv_id":null,"evidence_quote":"Provides the static conditional-IV representation-learning approach that TDCIV extends to time series."},{"cited_title":"Two-stage least squares estimation of average causal effects in models with variable treatment intensity,","cited_arxiv_id":null,"evidence_quote":"Defines the two-stage least squares estimator used to convert learned $S_t$ and $Z_t$ into the causal effect estimate."},{"cited_title":"Learning structured output representation using deep conditional generative models,","cited_arxiv_id":null,"evidence_quote":"Gives the conditional VAE machinery used to generate the conditioning set $Z_t$ from history."},{"cited_title":"Causal effect inference with deep latent- variable models,","cited_arxiv_id":null,"evidence_quote":"Supports recovering latent causes from proxy variables with deep latent-variable models."},{"cited_title":"Instrumental variable estimation for causal inference in longitudinal data with time-dependent latent confounders,","cited_arxiv_id":null,"evidence_quote":"The closest prior time-varying IV method (TIFM), which learns a standard IV but not a conditional IV and is the main baseline to beat."},{"cited_title":"Time series deconfounder: Estimating treatment effects over time in the presence of hidden confounders,","cited_arxiv_id":null,"evidence_quote":"Provides the time-series deconfounder approach and the synthetic data-generation style used to evaluate performance under latent time-varying confounders."}],"review_version":1}