{"id":"b24f1d15-c963-42ef-873c-af9fb22803cc","arxiv_id":"2508.11622","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":8.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"ProTrans transfers profiled residuals from source datasets to the target to estimate model shift free of hidden-confounding bias, achieving optimal rates.","lead":"A new statistical method, ProTrans, removes bias from hidden confounders by borrowing confounding information from related source datasets. It estimates model shifts and treatment effects without instrumental or proxy variables, under assumptions about source-target similarity. It matters because it could make regression and causal estimates more trustworthy in genetics, economics, and medicine when extra datasets are available.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 1 relies on an unstated cross-moment identifiability condition; equation (6)'s '≈' is not implied by Assumption 1.","rationale":"The reader's weakest-assumption comment points in the right direction: the profiled-residual transfer step is the crux, and the approximation near equation (6) is not formalized. I agree that Assumption 1.d is not sufficient by itself. However, the more precise failure mode is broader: even with φ_s^(k)=φ_t for every source, the transfer step needs cross-moment equality between source and target designs, and it also needs the weighted source-heterogeneity term (1/n_s)Σ X_s^(k)^T X_s^(k)(β_s^(k)−β_s) to vanish or be controlled. This term is absent from the text and is not covered by Assumption 1.d, though it may be negligible under an implicit homogeneity of the source and target distributions. Because the proof is relegated to an unavailable supplement, the central claim that \\hat η achieves the confounding-free rate is conditional on a missing technical condition. The paper's core idea is plausible and the simulations are encouraging, but the theorem as stated is not self-contained. A re-derivation or a targeted counterexample can settle whether the condition is actually needed.","tokens_in":20175,"tokens_out":21152,"duration_ms":257293,"concrete_test":"Independently re-derive Theorem 1 from Assumption 1, tracking R = (1/n_t)X_t^T H_t φ_t − (1/n_s)Σ_k X_s^(k)^T H_s^(k) φ_s^(k) + (1/n_s)Σ_k X_s^(k)^T X_s^(k)(β_s^(k) − β_s). Show that ||R||_∞ = O(λ_t). If this cannot be shown, construct a valid instance of Assumption 1 with p=1000, n_t=100, n_s=600, ||β_s^(1)−β_s||_1 = ||β_s^(2)−β_s||_1 = 1, source covariance matrices differing by O(1), and φ_s^(k)=φ_t, and evaluate ||\\hat η−η_0||_2. If the error does not match √(s_η log p/n_t), the missing cross-moment condition is load-bearing.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim is the oracle-rate bound for \\hat\\eta in Theorem 1. The only mechanism that removes confounding is the informal approximation following equation (6): (1/n_s)Σ_k X_s^(k)^T \\hat Z_s^(k) ≈ (1/n_t)X_t^T H_t φ_t − (1/n_t)X_t^T X_t(\\hat β_init − β_s). For (7) to cancel the bias in (6), this approximation must hold in ℓ_∞ with error o(λ_t); otherwise the unresolved term is exactly the confounding bias the method claims to remove. Assumption 1.d only controls ||(1/n_s)Σ n_s^(k) φ_s^(k) − φ_t||_2. It does not state the two empirical cross-moment equalities that the approximation needs: (1/n_s)Σ X_s^(k)^T H_s^(k) φ_s^(k) ≈ X_t^T H_t φ_t/n_t, and (1/n_s)Σ X_s^(k)^T X_s^(k)(β_s^(k) − β_s) ≈ 0. The second term is not automatically small: the parameter space only bounds ||β_s^(k) − β_s||_1 by a constant, and if source covariances differ it can be Ω(1). Remark 3's doubly robust discussion concerns a different term and additionally requires ||\\hat β_init − β_s||_1 = O(√(log p/n_t)), which is stronger than Theorem 1's stated assumption that this norm is merely bounded. The promised supplementary proof is not available, so the optimal-rate conclusion is not established from the stated assumptions.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes ProTrans, a transfer-learning method for estimating target regression coefficients and model shifts when both source and target regressions are affected by hidden confounders. The method constructs source 'profiled residuals' using an arbitrary initial estimator, transfers their covariate-moment information to the target via a projection-type step, and then fits a penalized regression for the model shift. The authors claim that the resulting model-shift estimator is asymptotically as accurate as if no confounding were present, that the target parameter estimator attains minimax rates, and that a source-selection extension improves robustness to noninformative sources. The paper includes simulation studies and two real-data applications. The theoretical results are stated as Theorems 1-4 and Propositions 1-4, with proofs deferred to a supplementary document that is not included in the reviewed manuscript.","tokens_in":20631,"tokens_out":5641,"duration_ms":67644,"significance":"If the central claim is correct, ProTrans would be a genuinely useful contribution: it would allow transfer learning to remove hidden confounding without instrumental variables, proxy variables, or strong parametric assumptions on the confounding mechanism. The algorithmic idea is constructive and the paper includes extensive simulations and real-data illustrations. However, the main theorem's proof depends on a key approximation that is not formally stated or derived from Assumption 1. Since the optimal-rate conclusion is the paper's principal contribution, this gap is load-bearing. The paper is therefore best seen as promising but not yet established; with a substantially added set of explicit conditions and a complete, self-contained proof, the work could become publishable.","major_comments":[{"comment":"The derivation of Theorem 1 requires that the bracketed term in inequality (6) be o_∞(λ_t). Tracing the construction of Z_t through (7) and the definition of Z_s^(k), this requires at least three empirical approximations: (a) (1/n_s)Σ_k X_s^(k)ᵀ X_s^(k)(β_s^(k)-β_s) ≈ 0; (b) (1/n_s)Σ_k X_s^(k)ᵀ H_s^(k)φ_s^(k) ≈ (1/n_t)X_tᵀ H_t φ_t; and (c) (X_sᵀX_s/n_s - X_tᵀX_t/n_t)(β_s - βhat_init) ≈ 0. None of these is stated as an assumption or proven from Assumption 1. Assumption 1.d only controls the average of φ_s^(k) versus φ_t in ℓ_2; it does not control the covariance-weighted cross moments. In particular, (a) is not implied by ‖β_s^(k)-β_s‖_1 ≤ s_δ when source covariate distributions differ; it can be Ω(1). Thus the oracle-rate bound in Theorem 1 is not established from the stated assumptions.","section":"§3.1, Eq. (6)-(7) and Theorem 1"},{"comment":"Remark 3 attempts to address heterogeneity by invoking a 'doubly robust' property of (X_sᵀX_s/n_s - X_tᵀX_t/n_t)(βhat_init-β_s). This concerns only approximation (c) above, and in its heterogeneous-case discussion it requires ‖βhat_init-β_s‖_1 ≲ sqrt(log p / n_t), which is strictly stronger than Theorem 1's stated assumption that this norm is merely bounded by a constant. The remaining approximations (a) and (b) are not discussed. Therefore the remark does not fill the gap in the proof of Theorem 1.","section":"§3.1, Remark 3"},{"comment":"The paper repeatedly cites 'the online Supplementary Material' for proofs of Theorems 1-4 and Propositions 1-4. The provided manuscript contains no proofs of these statements. Given that the central claim is a non-asymptotic oracle rate for ηhat, the absence of a verifiable proof is a load-bearing omission. If the supplement was inadvertently omitted from the review packet, it should be provided; otherwise, the paper does not currently support its main theoretical conclusion.","section":"Supplementary Material / overall"}],"minor_comments":[{"comment":"Typo: 'confounding effectss' should be 'confounding effects'.","section":"§2.2, Remark 1"},{"comment":"In the multi-source Scenario 2, p=2000 row, the ProTrans standard errors (0.1853 and 0.1732) are identical to the TransFarm standard errors in neighboring cells; this looks like a copy-paste error and should be corrected.","section":"§5, Table 1"},{"comment":"The text in Algorithm 1 takes τ_s as the ⌊n_s/2⌋-th ordered singular value, while Assumption 2.a says τ_s ≍ p^{1/2} n_s^{-1/2}. These two descriptions are not obviously consistent; please clarify.","section":"§2.1, Assumption 2.a"},{"comment":"The reference list contains minor inconsistencies, e.g., 'Gholizade' vs. 'Gholizadeh' and the formatting of the GTEx and NLS URLs. These should be cleaned up.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The paper's core idea is interesting and the simulations are suggestive, but the main theorem cannot be verified from the submitted text because the crucial cross-moment approximation is unstated and the proofs are deferred. I would be willing to reconsider after the authors add explicit assumptions under which the approximation in (6)-(7) holds with error o(λ_t), provide a rigorous proof, and either include the supplement or make it available to the referees. If the supplement cannot supply the missing control of the source-parameter heterogeneity term, the central claim would need to be substantially weakened."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The core idea is good and the paper is worth a serious referee. ProTrans—transferring profiled residuals across datasets to cancel hidden confounding—is new relative to the cited transfer and deconfounding literature, and the simulations back it up in both linear and nonlinear confounding settings. The paper is also honest about what it does not require: no instruments, no proxies, no linear confounding structure. I see no circular fitting or invented entities.\n\nThat said, the central theorem has a load-bearing gap. The only mechanism that removes confounding is the informal approximation near equation (6): (1/n_s)Σ X_s^(k)ᵀH_s^(k)φ_s^(k) ≈ X_tᵀH_tφ_t/n_t, plus covariances aligning. Assumption 1.d only controls the average φ's in ℓ₂; it does not imply either cross-moment equality. The stress-test note is right that the second term involving (β_s^(k) − β_s) can be Ω(1) if source covariances differ, and the parameter space allows that. So Theorem 1's oracle-rate conclusion is not derived from the stated assumptions. Remark 3's doubly robust rescue requires ||β̂_init − β_s||₁ = O(√(log p/n_t)), which is stronger than Theorem 1's boundedness assumption. The supplementary proofs are promised, not present, so I can't check whether a missing condition is hiding there.\n\nMinor points: the trimming threshold and ρ for source selection are heuristic, with only numerical support; the real-data gold standards rely on proxies, so the evaluation is suggestive but not independent. None of these are fatal by themselves, but the unproven cross-moment condition is.\n\nBottom line: the method is promising and the paper should be sent to peer review, but the referee should be told to focus on whether Theorem 1's proof supplies the missing identifiability condition or whether Assumption 1 needs to be strengthened. I'd bring it to my reading group and would likely cite it once the theory is tightened.","headline":"A genuinely new transfer-learning deconfounding idea with a real gap between its theorem and proof: worth refereeing, but Theorem 1 as stated isn't established from the stated assumptions.","tokens_in":21097,"tokens_out":534,"would_cite":true,"duration_ms":7997,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62J07","62C20","62F12"],"pacs":[],"model":"deepseek-v4-flash","headline":"ProTrans transfers profiled residuals from source regressions to the target, canceling hidden-confounder bias in model shift estimation at the optimal no-confounding rate.","keywords":["unmeasured confounding","transfer learning","deconfounding","model shift","profiled residuals","high-dimensional regression","treatment effect","minimax optimality"],"falsifier":"Simulate the Section 5 model with a single source, fix $\\phi_s=\\phi_t$, then increase $\\|\\phi_s-\\phi_t\\|_2$ while holding $n_s,n_t,p$ fixed. The theorem's bound for $\\|\\hat\\eta-\\eta_0\\|_2$ should stay near the no-confounding rate until Assumption 1.d fails; beyond the threshold $\\|\\phi_s-\\phi_t\\|_2^2>C\\log p/n_t$, the error should grow with the discrepancy instead of following $\\sqrt{s_\\eta\\log p/n_t}$. A clear error floor at that threshold would falsify the deconfounding claim.","tokens_in":20088,"feed_emoji":"🎯","tokens_out":16395,"duration_ms":169372,"temperature":0.7,"pith_summary":"ProTrans is a transfer-learning estimator that targets a setting where the target regression is contaminated by unmeasured confounders but additional source datasets share a similar confounding structure. The paper's central claim is that the model shift between the average source coefficient and the target coefficient can be estimated at the minimax-optimal rate $\\sqrt{s_\\eta\\log p/n_t}$—the same rate as if no confounders existed—by transferring 'profiled residuals' (source residuals that carry the hidden confounding) into the target estimating equation. This matters because standard deconfounding requires instruments, proxy variables, or a linear hidden-factor structure, all of which are hard to justify; ProTrans needs only a large source sample and average confounding similarity. A byproduct is direct estimation of conditional average treatment effects as the model shift between control and treatment groups. The paper proves the rate with non-asymptotic bounds and minimax lower bounds, and shows by comparison that conventional transfer learning retains a confounding bias that no source size can remove.","feed_headline":"Profiled residuals kill hidden-confounder bias in transfer learning","feed_subtitle":"Only need source data whose confounding is close to the target's - no instruments or proxies.","key_machinery":"Profiled residual transfer. A profiled source residual is $\\hat Z_s^{(k)}=Y_s^{(k)}-X_s^{(k)}\\hat\\beta_{\\rm init}$, whose projection onto the covariate space encodes the source confounding moment $X_s^{(k)\\top}H_s^{(k)}\\phi_s^{(k)}/n_s^{(k)}$. The method then solves $\\hat Z_t=\\arg\\min_{Z_t}\\|X_t^\\top Z_t/n_t-\\sum_k X_s^{(k)\\top}\\hat Z_s^{(k)}/n_s\\|_\\infty$, aligning the source and target confounding-induced moments so that, when $\\hat Z_t$ is subtracted from $Y_t$ in the target LASSO, the confounding bias cancels. The construction carries the argument: it converts unmodeled confounders into a transferable residual moment and lets the model shift be estimated as in an unconfounded regression.","core_discovery":"The core discovery is that hidden confounders can be cancelled out of the target's shift-estimation problem without ever estimating them. Define the source profiled residual $\\hat Z_s^{(k)}=Y_s^{(k)}-X_s^{(k)}\\hat\\beta_{\\rm init}$; it contains $X_s^{(k)}(\\beta_s^{(k)}-\\hat\\beta_{\\rm init})+H_s^{(k)}\\phi_s^{(k)}+\\varepsilon_s^{(k)}$, so it carries the source confounding effect. ProTrans chooses the target residual $\\hat Z_t$ so that $X_t^\\top \\hat Z_t/n_t$ matches the average $X_s^{(k)\\top}\\hat Z_s^{(k)}/n_s$ componentwise. Under Assumption 1.d—the average source confounding vector $\\sum_k n_s^{(k)}\\phi_s^{(k)}/n_s$ is within $C\\log p/n_t$ (squared) of $\\phi_t$—this alignment makes the target","pith_inferences":["Continuous weighting is a natural extension the paper does not study: replacing the hard threshold in $I(\\rho)$ with weights proportional to $1/v^{(k)}$ could interpolate between selection and pooling, and plausibly inherit the Theorem 4 rates.","The moment-matching construction is loss-agnostic; the same $X^\\top Z$ alignment should transfer to generalized linear or survival models, but that extension is not proved here.","The treatment-effect application implicitly assumes the control group's unmeasured confounding represents the treatment group's; if confounding is treatment-specific, residual transfer could add bias instead of removing it. A practical diagnostic is to estimate $\\hat\\eta$ with different candidate source populations and check stability."],"forward_implications":["With $\\lambda_t \\asymp \\sqrt{\\log p/n_t}$, the shift estimator satisfies $\\|\\hat\\eta-\\eta_0\\|_2=O(\\sqrt{s_\\eta\\log p/n_t})$ under hidden confounding; in the treatment-effect application this gives root-$n$ conditional average treatment effect estimates when covariates are bounded.","The target estimator $\\hat\\beta_t=\\hat\\beta_s+\\hat\\eta$ has error governed by the source-size term $\\sqrt{p\\|\\phi_t\\|_2^2/(n_s\\lambda_\\Psi^2)}$ plus the optimal shift error, so increasing source data shrinks the confounding penalty.","Classical transfer learning applied to confounded targets carries a lower bound with $n_t$ in place of $n_s$ in the confounding term, so additional source data cannot remove its bias; ProTrans moves that bias to the source scale.","The informative-source selection procedure keeps only sources with small quality score $v^{(k)}$; under Assumption 3, the selected estimator retains the minimax rate up to a factor $\\rho$, so filtering out noninformative sources costs little."],"supporting_citations":[{"why":"Supplies the spectral trim deconfounding transform used to build the deconfounded source estimator and the SingleTrim baseline.","marker":"´Cevid et al., 2020"},{"why":"Supplies the transfer-learning decomposition of the target coefficient into average source effect plus model shift, and the conventional two-step estimator compared against.","marker":"Li et al., 2022"},{"why":"Provides the doubly debiased lasso theory used to state the error bound for conventional transfer estimators under linear confounding.","marker":"Guo et al., 2022"},{"why":"Provides the factor-model deconfounding baseline used in simulations.","marker":"Fan et al., 2021"},{"why":"Supplies the synthetic instrument method used as a baseline and as an alternative source deconfounding estimator in fixed-dimensional settings.","marker":"Tang et al., 2023"},{"why":"Supports the trim-transform-based deconfounding comparison for high-dimensional confounded models.","marker":"Sun et al., 2024"}],"fun_headline_variants":["ProTrans: cancel hidden confounders with profiled residuals","Deconfound transfer learning without instruments or proxies","Kill confounder bias using only similar source data","Transfer learning that erases hidden confounder bias","Profiled residuals: no proxies needed to remove confounders"],"cache_read_input_tokens":2816,"weakest_assumption_plain":"The load-bearing premise is the near-equality of confounding moments around equation (6): the average source cross-moment $\\sum_k X_s^{(k)\\top}H_s^{(k)}\\phi_s^{(k)}/n_s$ must track $X_t^\\top H_t\\phi_t/n_t$, with Assumption 1.d bounding the average source confounding vector within $C\\log p/n_t$ of $\\phi_t$; if the source confounders are far from the target's, the transferred residuals do not cancel the bias and the optimal-rate conclusion collapses.","fun_headline_variants_meta":{"raw":{"variants":["ProTrans: cancel hidden confounders with profiled residuals","Deconfound transfer learning without instruments or proxies","Kill confounder bias using only similar source data","Transfer learning that erases hidden confounder bias","Profiled residuals: no proxies needed to remove confounders"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000151,"raw_usage":{"total_tokens":1058,"prompt_tokens":785,"completion_tokens":273,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":529,"completion_tokens_details":{"reasoning_tokens":195}},"tokens_in":529,"tokens_out":273,"duration_ms":3758,"temperature":1.0,"reasoning_tokens":195,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T19:47:17.064515+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Simulate the Section 5 model with a single source, fix $\\phi_s=\\phi_t$, then increase $\\|\\phi_s-\\phi_t\\|_2$ while holding $n_s,n_t,p$ fixed. The theorem's bound for $\\|\\hat\\eta-\\eta_0\\|_2$ should stay near the no-confounding rate until Assumption 1.d fails; beyond the threshold $\\|\\phi_s-\\phi_t\\|_2^2>C\\log p/n_t$, the error should grow with the discrepancy instead of following $\\sqrt{s_\\eta\\log p/n_t}$. A clear error floor at that threshold would falsify the deconfounding claim.","supporting_citations":[],"review_version":1}