{"id":"ca53f684-56c0-4933-bd8d-90bf4657bb16","arxiv_id":"2412.02986","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"TRADER is a source-guided horseshoe prior that shrinks target coefficients toward a weighted average of rescaled source estimates, improving posterior contraction and frequentist coverage in high-dimensional regression.","lead":"This paper introduces TRADER, a Bayesian transfer learning method that borrows information from multiple pre-trained source models to improve estimation and uncertainty quantification in a small target sample. A smart generalist might read it because it promises better statistical inference for underrepresented populations using only summary statistics from related studies, without sharing individual-level data.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Cosine-similarity Dirichlet parameters in Eq. (2) can be non-positive, making the proposed prior undefined for negatively correlated or orthogonal sources—exactly the heterogeneous regime the method claims to handle.","rationale":"The reader correctly identified a gap between the oracle theory and the data-dependent prior construction. That gap is real, but it presumes the data-dependent weights are at least well-defined. The failure of the Dirichlet prior for non-positive concentration parameters is more fundamental: the algorithm's own definition is invalid for a case it explicitly claims to cover. A quick fix is likely, such as using a positively transformed similarity (e.g., (1+θ_k)/2 or max(θ_k, ε)), but as written the method is internally inconsistent. This does not change the practical assessment that the paper is conditionally acceptable pending revisions that define and analyze the actual prior weight construction, so the verdict stays CONDITIONAL/UNCHANGED. Agreement is partial because both concerns involve the prior weight mechanism, but the reader did not identify the non-integrability of the Dirichlet density for negative similarities, which is a sharper correctness risk than the data-dependence gap alone.","tokens_in":18614,"tokens_out":10701,"duration_ms":102745,"concrete_test":"Run the TRADER algorithm as described on a minimal example with p=2, a target, and one source whose coefficient vector is exactly negatively correlated with the target (θ_1 = -0.5) and another with θ_2 = 0. The stated Dirichlet in Eq. (2) has concentration (-0.5, 0, 1) and should fail in Stan or produce a non-normalizable posterior. Alternatively, inspect the accompanying Stan code or supplement to see whether θ_k is transformed (e.g., max(θ_k, ε), (θ_k+1)/2, or |θ_k|) before entering the Dirichlet; if such a transformation exists, Eq. (2) and Section 2.2 must state it, and the theory must account for it. If no transformation exists and negative similarities appear in Setting II, the reported results cannot be generated by the stated method.","verdict_should_be":"UNCHANGED","load_bearing_attack":"In Section 2.2, Eq. (2) sets (η_1,...,η_K,η_{K+1}) ~ Dirichlet(θ_1,...,θ_K, ζ) with ζ=1, where θ_k = ŵ^{(k)T}β̂_val / (||ŵ^{(k)}|| ||β̂_val||) is the cosine similarity between the source estimate and the target validation estimate. The Dirichlet density requires concentration parameters α_i > 0 (or at most a degenerate boundary limit at α_i = 0 in some formalisms). A cosine similarity can be negative when the source coefficient vector is anti-correlated with the target, and zero when they are orthogonal. The paper never defines a transformation for non-positive similarities; it only states that sources with larger cosine similarity values receive greater weight. In Simulation Setting II, uninformative sources are generated as random vectors and are likely to have θ_k near zero or negative; Setting III explicitly uses ρ_k = 0.2, which can yield negative sample similarities. Under the stated prior, α_k ≤ 0 makes the Dirichlet density non-integrable (α_k < 0) or a boundary degenerate distribution (α_k = 0), so the prior in Eq. (2) is not a well-defined probability distribution. This is an internal inconsistency in the method as written, independent of the data-dependent-prior issue. If the implementation silently clips, shifts, or exponentiates θ_k, that modification is absent from the manuscript. Consequently, the claimed robustness to heterogeneous/uninformative sources—a core component of the central claim—rests on an undefined prior, and the simulations cannot be reproduced from the stated method.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes TRADER, a Bayesian transfer learning method for high-dimensional linear regression in which multiple source datasets are used only through pre-trained coefficient estimates. The prior centers each target coefficient on a rescaled, adaptively weighted average of source estimates, with Dirichlet weights informed by cosine similarities to a validation estimate and with horseshoe-type local-global shrinkage. The manuscript claims oracle posterior contraction rates faster than target-only horseshoe when sources are well aligned, protection against negative transfer from heterogeneous sources, and improved finite-sample frequentist coverage for moderate signal strengths. Theoretical results include an oracle contraction inequality (Theorems 1-2), a heuristic for the global shrinkage parameter (Proposition 1), and a finite-sample marginal posterior approximation (Theorem 3). The paper also presents simulation comparisons with target-only horseshoe and TransGLM, plus a MIMIC-III application estimating the association between HbA1C and insulin use in Hispanic patients.","tokens_in":18975,"tokens_out":9006,"duration_ms":91692,"significance":"If the central claims were fully established, TRADER would address a timely and practical gap: multi-source transfer learning using only summary-level source estimates, with communication efficiency and privacy preservation, together with uncertainty quantification. The paper's strengths are its clear problem framing, the explicit oracle-type contraction rates in terms of the effective sparsity of the source-target difference, the worst-case robustness statement in Corollary 2, and the broad simulation study covering scale and correlation heterogeneity. However, several load-bearing points are not yet resolved: the Dirichlet prior as written is not a valid distribution for non-positive cosine similarities, the estimator bβ_val used to build the prior is not defined in the high-dimensional regimes considered, the theory treats the source vector and weights as fixed and independent of target data while the method estimates them from a validation split, and Theorem 3's 'there exists F_j' construction does not directly establish a property of the actual posterior. These issues block acceptance of the core methodological and coverage claims in the current version.","major_comments":[{"comment":"The Dirichlet prior on the source weights is not a well-defined probability distribution when any cosine similarity θ_k is non-positive. The Dirichlet density requires strictly positive concentration parameters, but θ_k = bω^(k)T bβ_val / (||bω^(k)|| ||bβ_val||) can be negative for anti-correlated sources and zero for orthogonal sources. The paper explicitly targets heterogeneous and uninformative sources, and Simulation Settings II and III generate sources with near-zero or small positive correlations. If the implementation clips, shifts, or exponentiates θ_k, that transformation is absent from the manuscript. As written, the prior in Eq. (2) is internally inconsistent in exactly the regime the method claims to handle, and the simulations are not reproducible without an additional definition.","section":"Section 2.2, Eq. (2)"},{"comment":"The paper does not specify how bβ_val is computed, and in the settings considered it cannot be the usual OLS estimator. In the simulations n0 = 120 and p = 200, so a 1/3 validation split gives n_val = 40 < p, making (X_val^T X_val)^{-1} undefined. The same issue appears in the real-data application with p = 121 and a Hispanic target sample of n = 189, giving n_val ≈ 63 < p. Since bβ_val enters both the scaling factors ||bβ_val||/||bω^(k)|| and the cosine similarities θ_k, the proposed prior is not fully specified in the regimes where the method is evaluated. An explicit regularized estimator or a Bayesian procedure for bβ_val is needed.","section":"Section 2.1, Step 1"},{"comment":"The theoretical results treat the scaled source vector eω and the source weights η* as fixed and independent of the target data, while the method estimates them from a validation split of the target data. Theorem 2 explicitly assumes that eω^(1),...,eω^(K) are independent of the target data, but the scaling factors ||bβ_val||/||bω^(k)|| depend on target validation data, and the Dirichlet parameters are functions of bβ_val. No theorem or argument shows that the data-driven eω and η satisfy conditions such as (6) or the existence of the oracle weights η*. In addition, Theorems 1 and 2 require τ of order p^{-u} or n0^{-u} for an unspecified 'appropriate u', while Proposition 1 fixes a specific τ0 = (p - bψ)/(√n0 bψ); the paper does not verify that τ0 satisfies the theorem conditions. This gap separates the oracle statements from the actual TRADER procedure.","section":"Sections 2.1-2.2 and Theorems 1-2"},{"comment":"The statement 'there exists a distribution F_j with support contained in B' does not establish a property of the actual marginal posterior π(β_j | D0, σ²). The theorem constructs a marginal f_j from an arbitrary F_j and bounds its total variation distance to the actual posterior, but unless the actual posterior of β_{-j} is shown to concentrate on B, the bound is not informative about the real posterior. No lemma in the main text proves posterior mass of B. Part (ii) has the same problem, and its conditional density involving an unnormalized function h(β_j) and an indicator I(β_j ∈ R(β_{-j})) is not shown to be a valid probability kernel. Since Theorem 3 is the basis for the paper's finite-sample coverage claims, this is a load-bearing gap.","section":"Theorem 3, Eqs. (12)-(13)"}],"minor_comments":[{"comment":"The figure reference appears as 'Figure ??' and should be corrected to the appropriate figure number.","section":"Section 4.1"},{"comment":"The Caucasian-only 95% confidence interval [-1.349, -0.132] is not centered at the reported point estimate -1.091; please verify the reported interval or point estimate.","section":"Table 1"},{"comment":"The description of ψ as 'counting' the number of informative elements is informal because κ_j are continuous shrinkage factors; the formula E(ψ | τ) = p/(1 + τ√n0) should be derived explicitly using the half-Cauchy marginal calculation and the approximation X^(0)T X^(0) ≈ n0 I.","section":"Section 2.4, Proposition 1"},{"comment":"The symbol p is used both for the dimension and for a separate threshold integer (e.g., 'there exists p > s, p ≍ s' and 'p > max{e_s(r_sup), s}'); a different symbol such as \\bar p would remove the ambiguity.","section":"Assumption 3 and Theorem 3"},{"comment":"In the definition MSE = (1/p) Σ_{j=1}^p (bβ_mean_j - β_i)^2, the subscript i in β_i should be j.","section":"Section 4, MSE definition"},{"comment":"No sensitivity analysis is reported for the user-chosen hyperparameters: the 1/3 validation split fraction, bψ = p/2, and the Dirichlet concentration ζ = 1; given the role of τ and the validation estimate in source weighting, some sensitivity checks would strengthen the empirical claims.","section":"Sections 2.1 and 2.4"}],"recommendation":"major_revision","confidential_remarks":"The manuscript addresses a relevant problem and the general idea is plausible, but the current version is not reproducible because the prior is undefined for non-positive similarities and bβ_val is unspecified in high-dimensional settings. The theoretical connection between the oracle results and the actual data-driven procedure also needs to be made explicit. I see no indication of misconduct; the issues are technical completeness and rigor. If the authors can define the prior on a valid support, specify and analyze bβ_val, and repair Theorem 3 so that it concerns the actual posterior, the paper could become a useful contribution."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nTRADER is a neat idea that deserves a proper look, but the version I just read has a load-bearing flaw in the prior specification. The prior in Eq. (2) sets (η_1,...,η_K,η_{K+1}) ~ Dirichlet(θ_1,...,θ_K,ζ) with θ_k the cosine similarity between the k-th source estimate and the target validation estimate. Cosine similarity can be negative or zero. The Dirichlet density requires positive concentration parameters; with α ≤ 0 it is not a probability distribution. The paper never defines a transformation for non-positive θ_k. This matters because the method's headline claim is robustness to heterogeneous/uninformative sources. Simulation Setting II generates uninformative sources as random vectors, which will often have θ_k near zero or negative; Setting III explicitly uses ρ_k=0.2, which can produce negative sample similarities. So the prior is undefined in exactly the regime the method claims to handle. If the Stan implementation silently clips or shifts these values, that's not in the manuscript.\n\nThat said, there is real substance here. The two-step rescaling idea is sensible—normalizing source estimates by the ratio of validation norms is a practical way to handle scale differences—and the source-guided horseshoe prior is a genuinely new construction. The posterior contraction oracle inequality (Theorems 1 and 2) is a solid contribution to the transfer-learning theory literature, and the finite-sample analysis of marginal coverage (Theorem 3) adds value, though it is an existence result rather than a direct guarantee for the actual posterior. The simulations are extensive and the real-data analysis is a good, relevant use case.\n\nThe softer spots: the theorems treat the rescaling factors and Dirichlet weights as fixed or oracle, while the method estimates them from a validation split; no result covers the data-dependent prior. Proposition 1 is asserted without derivation. The real-data credible interval for the Hispanic analysis is likely overconfident because source estimation uncertainty is ignored.\n\nWho will get value: anyone working on Bayesian transfer learning or high-dimensional shrinkage priors. The idea is worth discussing, but the current version is not reproducible.\n\nMy recommendation: send it to referees—this is not a desk reject—but the referee report should first ask for a well-defined prior for non-positive similarities and theoretical treatment of the data-dependent weights. If those are fixed, the paper could be a solid contribution. As it stands, I would not cite it until the prior is fixed.","headline":"A promising transfer-learning prior with a load-bearing bug: the Dirichlet weights are undefined for negative or zero cosine similarities, which is the very regime the paper claims to handle.","tokens_in":19436,"tokens_out":5393,"would_cite":false,"duration_ms":50439,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62F15","62J07","62G20"],"pacs":[],"model":"deepseek-v4-flash","headline":"TRADER, a source-guided horseshoe prior, lets Bayesian transfer learning borrow from multiple pre-trained models and provably improves estimation and coverage in high-dimensional regression when sources align with the target.","keywords":["Bayesian transfer learning","horseshoe prior","high-dimensional linear regression","posterior contraction","uncertainty quantification","multi-source transfer","coverage probability","summary-level data"],"falsifier":"Run a simulation where the validation split is small (e.g., $n_0/10$) and the source estimates are noisy or biased, so that the estimated scaling factors and cosine-similarity weights make the combined source estimate violate the sparse-support condition (6); if the posterior contraction rate then matches the target-only rate or the credible intervals undercover for moderate signals, the claim that the adaptive construction preserves the theoretical guarantees is refuted.","tokens_in":18425,"feed_emoji":"📉","tokens_out":5253,"duration_ms":42997,"temperature":0.7,"pith_summary":"This paper introduces TRADER, a Bayesian transfer learning method that shrinks the target regression coefficients toward a weighted average of rescaled estimates from several pre-trained source models. The authors aim to show that when at least one source is informative, the posterior contracts around the true coefficients at a faster rate than the standard horseshoe prior achieves using target data alone, and that this also repairs the well-known undercoverage of moderate-strength signals. If correct, this would give practitioners a communication-efficient and privacy-preserving way to borrow information across studies while still producing credible intervals with frequentist validity. The method needs only the source parameter estimates, not the underlying individual-level data.","feed_headline":"Guided horseshoe prior sharpens Bayesian transfer learning","feed_subtitle":"Pre-trained source summaries narrow credible intervals and restore 95% coverage for moderate signals.","key_machinery":"The source-guided horseshoe prior. Each target coefficient $\\beta_j$ is given a normal prior with mean $\\eta_{K+1}\\cdot 0$ plus the weighted sum of rescaled source estimates, where source $k$ is scaled by the ratio $\\|\\hat\\beta_{\\text{val}}\\|/\\|\\hat\\omega^{(k)}\\|$ to match the target's scale and the weights $(\\eta_1,\\ldots,\\eta_{K+1})$ follow a Dirichlet distribution whose concentration parameters are cosine similarities between each source estimate and the target validation estimate. The global shrinkage parameter $\\tau$ is set not by a half-Cauchy hyperprior but by the informative-element count $\\psi=\\sum_j\\kappa_j$, which determines how much the posterior leans on the source; this avoids the instability of $\\tau$ collapsing to zero when sources are near the truth.","core_discovery":"The central claim is that a horseshoe prior whose mean is a data-adaptive combination of rescaled source estimates enjoys posterior contraction rates up to $\\sqrt{\\tilde s(r)\\log p/n_0}$ with $\\tilde s(r)$ potentially much smaller than the true sparsity $s$ when the combined source estimate is well aligned, while degenerating gracefully to the target-only rate $\\sqrt{s\\log p/n_0}$ when no source is helpful. Theorems 1 and 2 establish this as an oracle inequality with the effective sparsity determined by how much the combined source estimate deviates from the truth, and Theorem 3 shows that the marginal posterior for each coefficient can be approximated by a normal centered near the MLE with a bias term that shrinks at the improved rate, which is what restores nominal coverage for coefficients of moderate signal strength. The paper argues, with simulations and a Hispanic subgroup analysis of insulin and HbA1C, that this yields real gains in estimation and inference over using the target data alone and over an individual-level-data transfer method that is sensitive to scale mismatches.","pith_inferences":["The validation-split estimation of the scaling factors and weights is not covered by the theory, so in practice the guarantees require the validation estimate to be sufficiently accurate; users should check the stability of the constructed source estimate.","The mechanism suggests a recipe for other shrinkage priors: any global-local prior can be 'guided' by replacing the zero mean with an adaptively weighted external estimate, potentially extending the result to Dirichlet–Laplace or horseshoe+ priors.","Because TRADER needs only summary statistics, it could be combined with federated learning pipelines where each site runs its own horseshoe regression and shares only the fitted coefficients.","A testable extension would be to let the Dirichlet concentration parameters themselves be estimated from the data rather than fixed as the raw cosine similarities, which could improve robustness in small-validation settings."],"forward_implications":["If at least one source estimate is sparse and close to the target, the effective sparsity $\\tilde s(r)$ is smaller than $s$, so credible intervals shrink faster than target-only horseshoe and still cover the truth.","With multiple sources that individually capture only part of the support, the adaptive Dirichlet weighting recovers an oracle-weighted combination, improving the contraction rate without knowing which sources are informative.","In the worst case of uninformative sources, the prior collapses to a standard zero-mean horseshoe, so TRADER cannot perform worse than target-only analysis.","The real-data example suggests that borrowing from larger racial/ethnic groups can turn a non-significant association in a small underrepresented subgroup into a significant, appropriately uncertain estimate."],"supporting_citations":[{"why":"Introduces the horseshoe prior whose global-local shrinkage structure TRADER extends.","marker":"Carvalho et al., 2009"},{"why":"Documents the undercoverage of moderate signals in horseshoe priors, the target problem TRADER addresses.","marker":"van der Pas et al., 2017"},{"why":"Provides the bias-term analysis of conditional posteriors used to explain and fix the undercoverage.","marker":"Wu et al., 2023"},{"why":"Supplies the target-only posterior contraction rate $\\sqrt{s\\log p/n_0}$ that serves as the baseline for comparison.","marker":"Song and Liang, 2023"},{"why":"Motivates the informative-element-count choice of the global shrinkage parameter $\\tau$.","marker":"Piironen and Vehtari, 2017a"},{"why":"Provides TransGLM, the individual-level-data transfer learning method used as the state-of-the-art comparison.","marker":"Tian and Feng, 2023"},{"why":"Establishes minimax multi-source transfer learning theory that motivates the oracle-weighted combination of multiple sources.","marker":"Li et al., 2022"}],"fun_headline_variants":["Bayesian transfer via guided horseshoe improves coverage","Multi-source Bayesian transfer shrinks intervals and restores coverage","TRADER: Bayesian transfer learning that beats target-only inference","Horseshoe-guided transfer yields faster posterior contraction","Bayesian transfer from source summaries restores 95% coverage"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The theory treats the combined source estimate and oracle weights as fixed and independent of the target data, but the method constructs them from a validation split of the target data; the load-bearing premise is that this data-driven construction still satisfies the support-size and alignment conditions (such as condition (6)) that the theorems require.","fun_headline_variants_meta":{"raw":{"variants":["Bayesian transfer via guided horseshoe improves coverage","Multi-source Bayesian transfer shrinks intervals and restores coverage","TRADER: Bayesian transfer learning that beats target-only inference","Horseshoe-guided transfer yields faster posterior contraction","Bayesian transfer from source summaries restores 95% coverage"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000188,"raw_usage":{"total_tokens":1347,"prompt_tokens":977,"completion_tokens":370,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":593,"completion_tokens_details":{"reasoning_tokens":291}},"tokens_in":593,"tokens_out":370,"duration_ms":34619,"temperature":1.0,"reasoning_tokens":291,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T22:53:28.109494+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run a simulation where the validation split is small (e.g., $n_0/10$) and the source estimates are noisy or biased, so that the estimated scaling factors and cosine-similarity weights make the combined source estimate violate the sparse-support condition (6); if the posterior contraction rate then matches the target-only rate or the credible intervals undercover for moderate signals, the claim that the adaptive construction preserves the theoretical guarantees is refuted.","supporting_citations":[{"cited_title":"M., Polson, N","cited_arxiv_id":null,"evidence_quote":"Introduces the horseshoe prior whose global-local shrinkage structure TRADER extends."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Documents the undercoverage of moderate signals in horseshoe priors, the target problem TRADER addresses."},{"cited_title":"and Liang, F","cited_arxiv_id":null,"evidence_quote":"Supplies the target-only posterior contraction rate $\\sqrt{s\\log p/n_0}$ that serves as the baseline for comparison."}],"review_version":1}