{"id":"e4b40951-be18-4d00-b2b5-8ef17ef149d8","arxiv_id":"2505.17283","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"DWTS debiases and selects features from observational data, then warm-starts Thompson sampling with those estimates, achieving lower cumulative regret than LinTS in simulations.","lead":"The paper proposes Deconfounded Warm-Start Thompson Sampling (DWTS), which uses a doubly debiased lasso on confounded observational data to select important covariates and then initializes a Thompson sampling bandit with those estimates. The authors report lower cumulative regret than standard LinTS in synthetic and cardiovascular-risk simulations, suggesting a way to speed up adaptive clinical trials.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The deconfounding step is not isolated: DWTS is never compared to a warm-started LinTS initialized with non-debiased offline estimates, so the reported regret gains cannot be attributed to DDL.","rationale":"The reader's verdict is CONDITIONAL, and I agree that the paper should not be accepted without changes. However, the most load-bearing concern is not the reader's labeled weakest assumption (that hidden confounders are measured online), which the authors explicitly acknowledge as a limitation and which is a scope restriction rather than an internal flaw. The more consequential issue is that the empirical comparison does not isolate the DDL debiasing step. Without a warm-started baseline using non-debiased estimates, the observed regret reduction could come from the generic benefit of using any offline estimates, from the reduced dimension of the online model, or from the oracle threshold used in the synthetic experiments. This directly undermines the abstract's causal attribution of the improvement to 'deconfounded warm-start.' The proposed concrete test would settle the question by holding the warm-start mechanism fixed and varying only the debiasing method. If the non-debiased control performs equally well, the central contribution should be reframed as an empirical study of warm-starting rather than a demonstration that deconfounding helps. If DWTS dominates the control, the current framing is supported. I therefore recommend keeping the reader's CONDITIONAL verdict unchanged, with the ablation as a required condition.","tokens_in":12271,"tokens_out":8538,"duration_ms":114518,"concrete_test":"Run the identical synthetic (p=20, 40, 100) and NHANES experiments with a control warm-start baseline that uses the same mask and prior-variance construction but replaces DDL estimates with non-debiased estimates: (a) ordinary least squares on the selected submodel when n > p_eff, or (b) plain LASSO coefficient estimates without the DDL correction, using the same kappa_o selection rule. Keep all online LinTS details identical. If the non-debiased control matches DWTS's cumulative regret to within noise, the DDL deconfounding step is not the source of the reported improvement and the central claim should be weakened; if DWTS is clearly better across settings, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"DWTS's central claim is that initializing LinTS with DDL-debiased estimates of selected measured coefficients reduces regret relative to standard LinTS. The reported experiments compare DWTS against LinTS from scratch and LinTS on the true support, but never against LinTS warm-started with non-debiased estimates (e.g., plain LASSO, ridge, or OLS on the same offline data). The regret gap in Figures 1 and 2 can therefore be explained by the generic benefit of any offline warm-start, by the reduced dimension p_eff+q, or by the oracle threshold kappa_o used in the synthetic experiments (Remark 3.1 requires beta_Theta, the minimum true coefficient, which is infeasible in practice). The abstract's phrase 'deconfounded warm-start' attributes the improvement to the DDL step, but the experimental design does not isolate that step. This is load-bearing because if a naive warm-start achieves the same or better regret, the paper's contribution reduces to a known warm-start heuristic with an extra debiasing device that is not shown to help.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Deconfounded Warm-Start Thompson Sampling (DWTS), a two-stage algorithm for linear contextual bandits with access to a confounded offline dataset. In the offline stage, a Doubly Debiased LASSO (DDL) procedure estimates per-coordinate coefficients and standard errors for the measured features, and a thresholding rule selects a sparse support. In the online stage, LinTS is warm-started with these estimates on the selected measured coordinates while keeping uninformative priors on the q hidden-confounder coordinates, and the online learner operates in the reduced p_eff+q dimensional space. The authors evaluate DWTS in synthetic experiments and in a virtual environment built from NHANES cardiovascular-risk data, reporting lower cumulative regret than standard LinTS, LinTS restricted to the true support, and an OFUL method for partially observable offline data. The paper positions the contribution as a practical way to transfer causal estimates from biased observational data into an adaptive clinical trial.","tokens_in":12420,"tokens_out":4966,"duration_ms":36315,"significance":"If the reported gains are real and attributable to the debiasing step, DWTS would be a practically useful contribution to precision medicine and sequential decision-making under confounding. The paper is honest about its limitations: Section 6 explicitly acknowledges the assumption that all hidden confounders become observable online, the restriction to linear models, the absence of formal regret guarantees, and the lack of a treatment for non-stationary or multiple offline datasets. These admissions are a strength. However, the empirical evidence does not yet isolate the mechanism behind the regret improvement, and two experiment-design choices (the missing naive warm-start baseline and the oracle threshold used in the synthetic evaluation) weaken the central claim as it stands.","major_comments":[{"comment":"The experimental design lacks a key ablation: DWTS is compared only against LinTS from scratch and LinTS on the true support, but never against LinTS warm-started with non-debiased offline estimates (e.g., OLS, ridge, or plain LASSO on the same offline data, with the same or a comparable selection rule). Without this baseline, the regret reduction in Figure 1 cannot be attributed to the DDL debiasing step; it may reflect the generic benefit of any offline warm-start or the reduced dimension p_eff+q. Adding this comparison is load-bearing for the paper's central claim.","section":"§4 and Figure 1"},{"comment":"In the synthetic experiments, the threshold κ_o is chosen using the true minimum nonzero coefficient β_Θ, which is infeasible in practice, as the remark itself acknowledges. This introduces oracle information into the evaluation and makes the reported regret gaps potentially optimistic. The authors should report results with a feasible threshold selection procedure (e.g., cross-validation, a quantile of |θ̂_a,i|, or a sensitivity analysis over κ_o) to support the practical claim that the method can be deployed without knowledge of the true parameters.","section":"Remark 3.1 and §4"},{"comment":"The prior precision is initialized as B̂_{0,a}^{-1} = diag([σ̂_a^{eff}, 1_q]), but σ̂_a,i is defined in step 3 as the standard error of θ̂_a,i. Precision should be the inverse variance, namely diag([1/(σ̂_a^{eff})^2, 1_q]), if σ̂ denotes a standard error. As written, the algorithm does not implement the claimed warm-start with DDL-estimated variances, and the pseudocode may not be reproducible. This needs to be corrected and the resulting experiments re-run or justified.","section":"§3 and Algorithm 1, step 7"},{"comment":"The virtual NHANES environment generates rewards as r_i(a) = ρ_a · CVDRisk_i + ε, where CVDRisk_i is the output of nonlinear cardiovascular risk calculators. This reward model does not satisfy the linear response model in Eq. (2) used by all LinTS variants and by DDL. The real-data evaluation therefore tests the algorithms under misspecification without discussing this mismatch. To support the claim that DWTS works under the paper's linear-model assumptions, the authors should either construct a linearized version of the virtual environment, or explicitly analyze and justify the effect of this misspecification on the comparison.","section":"§5.2"}],"minor_comments":[{"comment":"The score is computed as s_{t,a} = μ_{t,a}^⊺ X_t^h, but μ_{t,a} lies in R^{p_eff+q} while X_t^h lies in R^{p+q}; the context must be restricted to the effective and hidden coordinates, e.g., X_t^{h,eff}, or the dimensions should be specified consistently.","section":"Algorithm 1, line 15"},{"comment":"The second baseline is labeled 'LinTS (Important dimensions)' in the captions but is not defined in the main text; the paper should state explicitly whether this baseline is warm-started or initialized from scratch, and whether it uses the true support of θ*.","section":"Figures 1 and 2"},{"comment":"The text says 'In LinTS, we learn all=p+q dimensions from scratch' but does not describe the arm-specific structure of θ*_a (e.g., θ*_a = [(a+1)1_5; 0_5] is only given for one arm); the construction for both arms should be stated completely.","section":"§4"},{"comment":"The noise level σ² for the virtual environment is not specified; report the value used in the experiments so the results are reproducible.","section":"§5.2"},{"comment":"There are several typos and grammatical issues (e.g., 'we chose to do not overburden the notations', inconsistent use of Y_t^h(a) vs. Y_t^h); a careful proofread would improve clarity.","section":"Throughout"},{"comment":"No code or data availability statement is provided; for an empirical paper, releasing the code and the data-processing pipeline would substantially strengthen the contribution.","section":"Reproducibility"}],"recommendation":"major_revision","confidential_remarks":"The paper addresses a timely and practically important problem, and the overall idea of combining DDL with a warm-started LinTS is reasonable. However, the central empirical claim is not yet isolated: the missing naive warm-start baseline and the oracle-informed threshold in the synthetic experiments mean the reported gains cannot be cleanly attributed to the deconfounding step. The prior-precision inconsistency and the misspecified NHANES reward model are additional correctness concerns. These issues are fixable within the paper's scope, so I recommend a major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nThis is a decent practical paper on warm-starting linear Thompson sampling with confounded observational data, but the headline claim overreaches the evidence. The idea is simple and sensible: apply DDL to the offline data to get debiased coefficient estimates, select a sparse support, and use those estimates to initialize LinTS priors. The paper is well written, the limitations section is candid, and the NHANES-based virtual environment is a nice touch. I believe the combination of DDL with a warm-started bandit is new in the literature; the literature review covers the relevant warm-starting work.\n\nThe problem is experimental isolation. DWTS is compared to LinTS from scratch, LinTS on the true support, and OFUL with partially observable data. It is never compared to LinTS warm-started with the same offline data but without the debiasing step — plain LASSO, ridge, or OLS on the observational data. Without that baseline, the regret gains cannot be attributed to DDL. They could come from the generic benefit of any warm-start or from dimension reduction. The stress-test note gets this right, and I don't think the paper currently rebuts it.\n\nThere is also a smaller issue with the synthetic experiments: the threshold κ_o is chosen using the true minimum nonzero coefficient (Remark 3.1 acknowledges this is infeasible). That means the synthetic evaluation injects oracle information. The real-data section uses a fixed κ_o=0.01, which is better, but the main figures in Figure 1 are the ones making the load-bearing claim.\n\nOn the positive side, the paper does not oversell theory it doesn't have. It explicitly says there are no regret guarantees. The assumption that all hidden confounders become observable online is stated and then flagged in the conclusion as likely unrealistic. That honesty counts for something.\n\nSo who gets value here? Practitioners building adaptive trials with rich EHR data will find the recipe useful, and researchers working on warm-starting will want to see it as related work. But as a contribution, the central causal claim — that deconfounding is what makes warm-starting work — is not yet demonstrated. I would send this to a serious referee, but the referee should require the missing ablation and a data-driven threshold before acceptance.\n\nRecommendation: engage with it as a conditionally acceptable paper.","headline":"A practically motivated warm-start heuristic with an honest limitations section, but the missing naive warm-start baseline and the oracle threshold leave the deconfounding claim unproven.","tokens_in":13015,"tokens_out":2055,"would_cite":false,"duration_ms":22377,"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":"Deconfounded warm-start Thompson sampling claims lower regret than standard LinTS by seeding priors with offline causal estimates.","keywords":["Thompson sampling","contextual bandits","warm-start","hidden confounding","doubly debiased lasso","observational data","precision medicine","cumulative regret"],"falsifier":"Construct a semi-synthetic environment from NHANES where one of the offline hidden confounders (e.g., demographic factors) is excluded from the online context, and compare DWTS against standard LinTS; if DWTS no longer shows lower cumulative regret, the claim that it harnesses confounded observational data is undercut.","tokens_in":12024,"feed_emoji":"💊","tokens_out":4473,"duration_ms":32139,"temperature":0.7,"pith_summary":"The paper tries to show that biased observational data, usually set aside, can be safely used to start an online learning algorithm for treatment choice. It proposes Deconfounded Warm-Start Thompson Sampling (DWTS), which first applies a debiasing procedure to the offline data to estimate which measured patient features matter and with what accuracy, then feeds those estimates into the priors of Thompson sampling while leaving hidden features uninformative. Across synthetic and NHANES-based experiments, DWTS reports consistently lower cumulative regret than standard LinTS and a state-of-the-art baseline. If true, this would let adaptive clinical trials reach good treatment policies with fewer enrolled patients.","feed_headline":"Warm-started Thompson sampling cuts regret in adaptive trials","feed_subtitle":"Debiased offline estimates seed the bandit's priors, so treatment choices improve early without extra patient cohorts.","key_machinery":"The load-bearing object is the prior initialization scheme: Doubly Debiased LASSO (DDL), a high-dimensional debiasing method that adjusts for hidden confounding, provides per-coordinate deconfounded estimates (mean and standard error) for the regression coefficients of the measured features; a threshold converts these into a mask selecting an effective dimension; LinTS then runs on the reduced dimension, with prior means and variances seeded from the offline estimates on the selected coordinates and uninformative priors on the hidden features. This transfers the causal information from the offline structural equation model into the online linear bandit, cutting both bias and dimension.","core_discovery":"The central claim is that a contextual bandit can be warm-started from confounded observational data without inheriting its bias, provided the offline hidden confounders are measured in the online phase. The paper models the offline data with a linear structural equation where hidden confounders affect both treatment and outcome, and the online phase with a linear response model sharing the same regression coefficients. DWTS uses Doubly Debiased LASSO to estimate deconfounded coefficients and standard errors on measured covariates, selects a sparse support by thresholding, and initializes LinTS priors on those coordinates while leaving priors on the hidden features uninformative. The paper reports that this reduces regret relative to standard LinTS and to an OFUL baseline in synthetic settings and in a virtual environment built from NHANES cardiovascular risk data.","pith_inferences":["If the online phase cannot measure all offline-hidden confounders, the same warm-start trick should fail; a natural test is to run DWTS with one confounder left unmeasured online and check whether the regret advantage disappears.","The threshold is a tuning knob: the paper's theoretical choice requires knowing the smallest true coefficient, so in practice one would need data-driven selection; the sensitivity of DWTS to this choice is not established.","The same prior-seeding idea could be transplanted to other Bayesian bandit algorithms, where DDL estimates would have to be replaced by corresponding debiased estimators."],"forward_implications":["Adaptive clinical trials can begin with informative priors on measured features, improving early treatment decisions without enrolling additional patients.","The effective dimension of the online learning problem is reduced, which should lower sample complexity under sparsity.","The approach outperforms OFUL with partially observable confounded data in the paper's experiments, suggesting that full online measurement of confounders combined with offline debiasing is a stronger design.","The combined offline-online procedure gives a practical route to incorporate EHR or registry data into sequential decision-making, supporting personalized treatment."],"supporting_citations":[{"why":"Supplies the DDL estimator that produces deconfounded means and standard errors on measured covariates.","marker":"[Guo et al., 2022]"},{"why":"The LinTS algorithm that DWTS warm-starts and compares against.","marker":"[Agrawal and Goyal, 2013]"},{"why":"The OFUL baseline for partially observable confounded data, benchmarked in the experiments.","marker":"[Tennenholtz et al., 2021]"},{"why":"CVD risk equation used to build the virtual environment's reward signal.","marker":"[Basu et al., 2017]"},{"why":"AHA/ACC pooled cohort risk equations also used for the virtual environment.","marker":"[Goff et al., 2014]"}],"fun_headline_variants":["Warm-start Thompson sampling from observational data cuts regret","Debiased offline estimates seed bandit priors for lower regret","Kick-start adaptive trials with deconfounded Thompson sampling","Lower regret in adaptive trials using debiased offline data","Observational data warm-starts Thompson sampling for better trials"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The online phase must measure every confounder that was hidden in the offline data, because the whole warm-start relies on the online linear model being correctly specified; the authors themselves call this potentially unrealistic for real clinical settings.","fun_headline_variants_meta":{"raw":{"variants":["Warm-start Thompson sampling from observational data cuts regret","Debiased offline estimates seed bandit priors for lower regret","Kick-start adaptive trials with deconfounded Thompson sampling","Lower regret in adaptive trials using debiased offline data","Observational data warm-starts Thompson sampling for better trials"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000883,"raw_usage":{"total_tokens":3776,"prompt_tokens":865,"completion_tokens":2911,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":481,"completion_tokens_details":{"reasoning_tokens":2829}},"tokens_in":481,"tokens_out":2911,"duration_ms":16655,"temperature":1.0,"reasoning_tokens":2829,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T14:49:32.923210+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Construct a semi-synthetic environment from NHANES where one of the offline hidden confounders (e.g., demographic factors) is excluded from the online context, and compare DWTS against standard LinTS; if DWTS no longer shows lower cumulative regret, the claim that it harnesses confounded observational data is undercut.","supporting_citations":[],"review_version":1}