{"id":"c582aec1-21db-44c3-a4ea-d159e42b26dd","arxiv_id":"2411.12555","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A Bayesian method that jointly models multiple outcomes and sequentially updates across datasets with a learned weight that limits negative transfer.","lead":"This paper extends the RECaST Bayesian transfer learning method to handle multiple health outcomes at once and to update models sequentially as new data arrive. The method borrows strength from a well-represented group (White dental patients) to improve predictions for underrepresented groups while keeping patient data private.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The model treats β_i as independent of the source prediction f_i, but in the canonical linear case β_i|f_i has scale proportional to 1/|f_i|; the paper only reports marginal coverage, so the UQ claim rests on an untested conditional approximation.","rationale":"The reader's weakest assumption was that the beta_i distribution family (multivariate Cauchy or Gaussian copula with Cauchy marginals) is adequate outside the canonical linear setting. I agree that this is a real limitation, but the deeper issue is that the model uses the marginal distribution of beta_i while the likelihood conditions on f_i, and even in the univariate canonical case the conditional distribution of beta_i given f_i has scale 1/|f_i|. No choice of Cauchy versus copula can fix this dependence. The paper's simulations report only marginal average coverage, so the central claim about uncertainty quantification is not yet supported for feature-dependent scales. This concern does not change the reader's CONDITIONAL verdict: the multivariate Cauchy variant is still a plausible approximation and the simulations are encouraging, but the requested revision should include either a conditional-coverage analysis or a theoretical justification for ignoring the f_i-dependence. The proposed test would settle whether the concern actually degrades calibration in practice.","tokens_in":24511,"tokens_out":20341,"duration_ms":216581,"concrete_test":"Replicate the Section 6.2 simulation for Σ = I, (a, b) = (1, -1), and n_T = 50. After fitting MV Cauchy and MV Copula, compute the empirical 95% coverage and the average Mahalanobis distance separately within quartiles of the held-out test source predictions f_1(x_i) and f_2(x_i). If coverage remains near 95% in every quartile, the independence approximation is harmless for the claimed uncertainty quantification; if low-|f| quartiles undercover or high-|f| quartiles overcover, the credible sets are not conditionally calibrated. An analytic companion check is to compare the model-implied conditional scale of y_i given f_i, roughly |f_i| γ plus the outcome-noise scale, with the true conditional scale √(σ_W^2 / f_i^2 + σ_U^2); a ratio that depends on |f_i| confirms the misspecification.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The motivating derivation in Section 3 (Lemma 1) shows that β_i = x_i^T θ_T / x_i^T θ_S is marginally Cauchy when x_i is random Gaussian, but the likelihoods in Equations (2) and (3) marginalize β_i over that marginal law without conditioning on f_i = x_i^T θ_S. In the canonical linear-Gaussian case, (A_i, B_i) = (x_i^T θ_S, x_i^T θ_T) are jointly normal, so B_i | A_i has constant conditional variance σ_W^2 and conditional mean ρ A_i. Consequently β_i | A_i is N(ρ, σ_W^2 / A_i^2): its scale is 1/|A_i|, not a constant γ shared across observations. The model instead assigns every observation a common Cauchy scale, so observations with small |f_i| are modeled as much less dispersed than they truly are, and observations with large |f_i| are modeled as more dispersed. Section 4.1 concedes the multivariate Cauchy is only an approximation, but this conditional dependence is an additional, unacknowledged approximation that already appears in the univariate canonical case and affects both the multivariate Cauchy and the copula constructions, since both specify a joint law for β_i that is independent of f_i. The posterior predictive credible sets in Section 4.2 and Section 4.3 inherit this misspecification, so average near-nominal coverage in the simulations can hide meaningful conditional miscalibration. The paper only reports coverage averaged over test sets, so this is never probed.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper extends the RECaST Bayesian transfer learning framework in two directions: it replaces the univariate Cauchy random effect with a multivariate Cauchy or a Gaussian-copula-with-Cauchy-marginals distribution to model multiple outcomes jointly, and it introduces an online variant that combines the source-model prior with the posterior from previous target data sets through a learned convex weight alpha. The authors claim that the multivariate methods outperform univariate RECaST and target-only ridge regression in simulated and real periodontal data, that the online method mitigates negative transfer between sequentially arriving target sets, and that the posterior predictive credible sets achieve near-nominal coverage. The main methodological content is in Sections 4 and 5; simulation and data analyses are in Sections 6 and 7.","tokens_in":24873,"tokens_out":6607,"duration_ms":67269,"significance":"If the results hold, the paper offers a practically useful extension of a Bayesian transfer learning framework with attractive properties: privacy preservation (only models and posterior summaries are shared), scalability to small target samples, and built-in uncertainty quantification. The simulation study is reasonably extensive (100 replications, several covariance and similarity settings), and the dental data application addresses a socially relevant problem of demographic underrepresentation. The multivariate and online constructions are natural and the exposition is generally clear. However, the central uncertainty-quantification claim rests on an unexamined conditional approximation, and one appendix presents a sampler for a different model, so the current form needs substantive revision.","major_comments":[{"comment":"The model assumes the random effect beta_i is independent of the source prediction f_i = f(Theta_S, x_i), but in the motivating canonical linear case this is false. With (A_i, B_i) = (x_i^T Theta_S, x_i^T Theta_T) jointly Gaussian, the conditional law of beta_i = B_i / A_i given A_i = f_i is N(rho, sigma_W^2 / f_i^2), whose scale depends on the realized source prediction. Equations (2) and (3) instead assign every observation a common Cauchy scale gamma (or gamma_j), independent of f_i. Section 4.1 only acknowledges that the marginal multivariate law of beta_i is not generally multivariate Cauchy; it does not mention this conditional misspecification, which already occurs in the univariate setting and affects both the multivariate Cauchy and copula models. The paper reports only test-set-averaged coverage (e.g., 96-97% in Tables 1-4), so conditional miscalibration for observations with unusually small or large |f_i| is never probed. The authors should either demonstrate conditional calibration empirically (e.g., coverage stratified by |f_i|) or modify the prior to depend on f_i.","section":"Sections 3-4, Eqs. (2)-(3)"},{"comment":"The Gibbs sampler presented in Appendix A does not target the model in Equation (2). Its full conditional distributions are for the hierarchical model Y_i | mu, Sigma ~ N_m(mu, Sigma), mu | u, delta, Gamma ~ N_m(delta, u^{-1} Gamma), u ~ Gamma(nu/2, nu/2), i.e., a common multivariate t prior on a shared outcome mean mu. In contrast, the proposed RECaST model has observation-specific random effects via y_i = diag(beta_i) f(Theta_S, x_i) + U_i with beta_i ~ Cauchym. These are not equivalent unless f_i is constant and one interprets the t prior on a single common mean, which is not the setting of the paper. The sentence in Section 4.2 that a Gibbs sampler 'could also be used; the details ... are outlined in Appendix A' is therefore misleading. The appendix should be corrected to describe a sampler for Equation (2), or removed if it is not used in the numerical results.","section":"Appendix A and Section 4.2"},{"comment":"The claim that alpha is 'learned' is overstated because the Uniform(0,1) prior restricts the posterior mean to the interval [1/3, 2/3]. The derivation in Section 5 shows E(alpha | data) = (2/3 k1 + 1/3 k2)/(k1 + k2), so alpha = 0.67 is the maximum possible value, not evidence that the posterior for T1 is 'more useful' than the uninformative prior. Tables 11-13 show posterior means concentrated at 0.67 (close setting) and around 0.34-0.47 (far setting), indicating that the procedure is often pushing against the prior boundary. The authors should present this as a deliberate bounded-influence design and discuss its implications: with a uniform prior the method cannot fully reject T1 information (lower bound 1/3) nor fully embrace it (upper bound 2/3). A sensitivity analysis with, e.g., Beta(1/a,1/a) priors would clarify how much the conclusions depend on this choice.","section":"Section 5, online alpha prior; Tables 11-13"},{"comment":"The comparison between multivariate and univariate RECaST is not fully specified. The paper computes a joint Mahalanobis distance using the test-set empirical covariance S, but for univariate RECaST only marginal posterior predictive distributions are provided (the paper reports marginal coverages of 100 for many settings, indicating that the univariate credible sets are uninformative). It is unclear how a joint mean vector and joint covariance matrix are formed from the univariate predictions to compute the reported Mahalanobis distances. Without this detail, the claim 'in all cases, the multivariate RECaST methods have better predictive performance than the univariate method' (Section 6.2) cannot be independently verified, and the comparison may be unfair. Please clarify the construction of the univariate joint predictive distribution or use a comparison metric that is well-defined for both methods.","section":"Section 6.1 and Tables 1-4"}],"minor_comments":[{"comment":"The phrase 'even when Θ_T2 if far from Θ_T1' appears several times; 'if' should be 'is'.","section":"Section 6.3"},{"comment":"The last column header reads 'MV Copulah' with a stray 'h'; this formatting error should be corrected.","section":"Table 4 header"},{"comment":"The sentence describing Maddox et al. (2021) ends with '...large covariance matrix in .' and is incomplete; please rephrase.","section":"Section 2"},{"comment":"The sentence 'considering The relationship between the outcomes...' has a capitalization/grammar error; 'The' should be 'the'.","section":"Section 4.1"},{"comment":"The paper does not include a data or code availability statement. Given that the simulations are a key part of the claims, providing code or at least detailed random seeds would improve reproducibility.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The paper is an early arXiv manuscript with a promising idea and relevant application, but the conditional calibration issue strikes at the heart of the uncertainty-quantification claim, and the Appendix A mismatch suggests the manuscript needs careful technical editing. I would encourage the editor to send the paper back for major revision rather than reject, since the core extension is sound in spirit and the problems appear fixable within the scope of a revision. I also note that the alpha weighting scheme, as presented, is essentially a bounded shrinkage method; the authors should be honest about that in the revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe genuinely new pieces here are the two multivariate random-effect priors for beta (multivariate Cauchy, and Gaussian copula with Cauchy marginals) and the online convex-prior weighting with learned alpha. All are natural extensions, clearly presented, and the simulation design is thorough. The privacy-preserving property — sharing only posterior summaries, not data — is a real practical strength, and the dental application is relevant. This is a workmanlike paper, not a breakthrough.\n\nThe main soft spot is conditional calibration. In the motivating canonical linear-Gaussian case, beta_i given f_i is normal with variance sigma^2 / f_i^2, so the scale depends on the source prediction. The model instead assigns one constant scale gamma, independent of f_i. Section 4.1 concedes only that the multivariate Cauchy is not the exact marginal; it does not acknowledge the conditional misspecification, which affects both the multivariate Cauchy and the copula constructions. Coverage is reported only averaged over test sets, so near-nominal mean coverage can hide miscalibration for observations with small or large |f_i|. The authors should either defend the approximation for conditional UQ or report coverage stratified by |f|.\n\nSecond, Appendix A's Gibbs sampler targets a different model: it places a multivariate t prior on the outcome mean mu, not the multiplicative random effect diag(beta) f(x) used in Sections 4.2 and 4.3. That appendix does not support the paper's main posterior and should be corrected or removed.\n\nThird, the learned alpha is bounded by the Uniform(0,1) prior: its posterior mean lies in [1/3, 2/3]. The paper states this, but the data analysis reads values of 0.67 as \"significant weight\" and 0.33 as \"minimized use,\" which overstates what the data can say. Also, the claim that no other transfer learning methods provide uncertainty quantification is too strong; the cited Bayesian transfer literature itself provides UQ in various forms.\n\nMinor: no code is released; the Mahalanobis differences in the dental data are within one standard error; and univariate RECaST's frequent 100% coverage makes it a weak comparison.\n\nWho this is for: statisticians working on Bayesian transfer learning with small, multi-outcome target samples, especially in medical and demographic applications. It deserves a serious referee. With the conditional calibration issue addressed and the appendix fixed, it would be publishable. Send it out.","headline":"A competent multivariate/online extension of RECaST with a real conditional-calibration soft spot; worth refereeing if the Appendix A mismatch and alpha interpretation are fixed.","tokens_in":25380,"tokens_out":2185,"would_cite":true,"duration_ms":23539,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62F15","62H12","62J07"],"pacs":[],"model":"deepseek-v4-flash","headline":"Joint transfer learning cuts prediction error and keeps 96% coverage","keywords":["Bayesian transfer learning","Dental records","Informative Bayesian prior","Online learning","Racial bias","Multivariate outcomes","Uncertainty quantification","Negative transfer"],"falsifier":"Simulate target data from a mechanism whose ratio to the source is far from Cauchy—for example, bounded, skewed outcomes such as the dental averages, or a nonlinear source-to-target map—and check whether the claimed 96–97% joint coverage falls well below the 95% nominal level. A specific test is to generate targets from a log-normal or beta regression and compute the empirical joint coverage of the multivariate Cauchy RECaST credible sets; if coverage drops below roughly 90%, the ratio-distribution assumption is doing the work.","tokens_in":24314,"feed_emoji":"🦷","tokens_out":5333,"duration_ms":48193,"temperature":0.7,"pith_summary":"This paper extends the RECaST Bayesian transfer-learning framework in two directions: it models several correlated outcomes at once, and it lets target datasets arrive one after another. The authors argue that modeling the outcomes jointly through a vector of source-to-target ratio random effects improves predictive accuracy relative to fitting each outcome separately, while the online extension borrows information from previous target datasets only when it helps. If the claims hold, medical and other settings with small, underrepresented target groups could get both sharper predictions and trustworthy credible intervals without ever sharing raw data.","feed_headline":"Joint transfer learning cuts prediction error, keeps 96% coverage","feed_subtitle":"Extending RECaST to multivariate and sequential targets sharpens periodontal predictions without sharing patient data.","key_machinery":"The central object is the source-to-target ratio random effect $\\beta_{i,j}=g(\\Theta_{T,j},x_i)/f(\\Theta_{S,j},x_i)$, which replaces the unestimable target mapping with a scaled version of the fitted source model. Univariate RECaST models each ratio as Cauchy; this paper supplies two multivariate priors: a full multivariate Cauchy with location vector $\\delta$ and scale $\\Gamma$, and a Gaussian copula that keeps the univariate Cauchy marginals $\\beta_j\\sim\\text{Cauchy}(\\delta_j,\\gamma_j)$ while coupling them through a correlation matrix $R$. The online extension stacks a learned posterior from a previous target into the prior for the next target as $\\pi^\\star(\\Omega,\\alpha)=\\pi(\\alpha)\\{\\alpha\\,p(\\Omega\\mid T_1)+(1-\\alpha)\\pi(\\Omega)\\}$, with $\\alpha\\in[0,1]$ estimated from the new target's data so that uninformative transfer can be downweighted. MCMC sampling of the posterior and posterior predictive draws yields elliptical credible sets via Mahalanobis distances.","core_discovery":"The paper's central claim is that transferring knowledge through the random ratio $\\beta_{i,j} = g(\\Theta_{T,j},x_i)/f(\\Theta_{S,j},x_i)$—the factor by which the source model must be scaled to match the target—works for multivariate outcomes if $\\boldsymbol\\beta_i$ is given either a multivariate Cauchy distribution or a multivariate Gaussian copula with Cauchy marginals. In simulation, both multivariate RECaST variants beat univariate RECaST in every setting and beat a target-only ridge baseline whenever the target sample is small; joint coverage sits at 96–97% for a 95% nominal level, whereas univariate credible sets are often the whole outcome space. The online extension writes the prior for a new target as a convex combination $\\alpha$ times the posterior from a previous target plus $(1-\\alpha)$ times an uninformative prior, learning $\\alpha$ from data; the paper finds no negative transfer from a distant previous target, with the online methods performing as well as or better than their offline counterparts. On dental outcomes (clinical attachment level and pocket depth), the multivariate and online methods improve Mahalanobis prediction error over ridge regression for the smallest underrepresented groups while maintaining slightly conservative joint coverage.","pith_inferences":["The learned weight $\\alpha$ could be repurposed as a diagnostic for negative transfer: a posterior mean near the prior floor suggests the previous target is uninformative for the new one.","The copula variant under-covers slightly (93–94%) in the online setting at larger sample sizes; a natural next step is to test whether a heavier-tailed copula or a wider prior on $R$ restores nominal coverage.","The dental outcomes are bounded and skewed whole-mouth averages, so the Cauchy-ratio assumption is untested there; applying the method to synthetic data with bounded, skewed outcomes would show how much coverage degrades.","The 96–97% coverage in linear simulations suggests the credible sets are slightly conservative; calibrating them to exactly 95% could trade a small amount of over-coverage for tighter regions."],"forward_implications":["Practitioners can transfer a source model to a small target group with multiple correlated outcomes and get joint prediction regions that are near-nominally calibrated.","The online prior mixture means that when a new target dataset is very different from a previous one, the method can shrink back to the uninformative prior and avoid negative transfer.","Data privacy is preserved: only the source model and the learned posterior from prior targets are shared, never raw patient records.","Because the framework is agnostic to the source model, the same ratio-random-effect machinery could be applied atop neural-network or other flexible source fits.","The posterior mean of $\\alpha$ gives a built-in, data-driven measure of how much information a previous target contributes."],"supporting_citations":[{"why":"Defines the RECaST framework and its univariate Cauchy ratio lemma that this paper extends to multivariate and online settings.","marker":"Hickey et al. (2024)"},{"why":"Provides the result that ratios of Gaussian random variables are Cauchy, motivating the heavy-tailed random-effect distribution and explaining why the multivariate ratio is not jointly Cauchy.","marker":"Pillai and Meng (2016)"},{"why":"Describes the dental-records data set used for the periodontal case study.","marker":"Guan et al. (2020)"},{"why":"Supplies the ridge-regression implementation used as the target-only baseline and as the source model.","marker":"Friedman et al. (2010)"},{"why":"Provides the MCMC engine used to sample the posterior distributions.","marker":"de Valpine et al. (2017)"},{"why":"Supplies the copula theory used to build the multivariate dependence from Cauchy marginals.","marker":"Nelsen (2006)"}],"fun_headline_variants":["Multivariate RECaST cuts error, keeps 96% coverage in dental data","Online transfer learning matches offline, avoids negative transfer","Better periodontal predictions via multivariate Bayesian transfer","Joint modeling of outcomes improves small-group estimates","Transfer learning with UQ: accurate even for underrepresented groups"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the source-to-target ratio $\\beta_i$ is adequately described by a multivariate Cauchy distribution or a Gaussian copula with Cauchy marginals; the paper itself notes this is only approximate even in the canonical linear setting, and for the real dental outcomes the approximation is untested.","fun_headline_variants_meta":{"raw":{"variants":["Multivariate RECaST cuts error, keeps 96% coverage in dental data","Online transfer learning matches offline, avoids negative transfer","Better periodontal predictions via multivariate Bayesian transfer","Joint modeling of outcomes improves small-group estimates","Transfer learning with UQ: accurate even for underrepresented groups"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000311,"raw_usage":{"total_tokens":1788,"prompt_tokens":975,"completion_tokens":813,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":591,"completion_tokens_details":{"reasoning_tokens":736}},"tokens_in":591,"tokens_out":813,"duration_ms":8888,"temperature":1.0,"reasoning_tokens":736,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T17:23:41.931738+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Simulate target data from a mechanism whose ratio to the source is far from Cauchy—for example, bounded, skewed outcomes such as the dental averages, or a nonlinear source-to-target map—and check whether the claimed 96–97% joint coverage falls well below the 95% nominal level. A specific test is to generate targets from a log-normal or beta regression and compute the empirical joint coverage of the multivariate Cauchy RECaST credible sets; if coverage drops below roughly 90%, the ratio-distribution assumption is doing the work.","supporting_citations":[{"cited_title":"Multivariate and Online Transfer Learning with Uncertainty Quantification","cited_arxiv_id":"2411.12555","evidence_quote":"Defines the RECaST framework and its univariate Cauchy ratio lemma that this paper extends to multivariate and online settings."}],"review_version":1}