{"id":"02f0e555-d5c9-4b40-8e35-35a8ed1ce2eb","arxiv_id":"2505.11351","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"A cross-validation-based choice of the response residual variance can push joint factor models to learn factors that predict the response, with small empirical gains.","lead":"The paper proposes tuning the response noise level in a joint factor model by cross-validation so the model focuses on predicting the outcome. The method shows small predictive gains in one simulation and on phthalate and BMI data, but the improvements are modest.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The reported gains from TEB-FAR are too small and too noisy to support the claim of substantial improvement; the actual procedure's largest edge over the joint Bayesian factor model is about 0.003 in MSE, with no uncertainty quantification.","rationale":"The strongest claim in the abstract is not a theorem about phase transitions; it is an empirical promise of 'substantial improvements' in simulation and application. For that promise to be true, the procedure with data-driven sigma_y^2 selection must beat the joint Bayesian factor model by a meaningful margin. The tables show the opposite: after implementing the actual estimation step, the margins are around 0.001-0.003 MSE, with no error bars, and in one training size TEB-FAR is worse. The one-factor phase transition motivates the method but is not load-bearing for the empirical headline; even a fully general theory would not make a 0.002 MSE difference 'substantial.' The reader's weakest_assumption flagged the k>1 postulate, which is a legitimate secondary concern, but the more pressing issue is that the reported effect sizes do not support the central claim. Therefore I agree with the conditional verdict: the paper should be accepted only if claims are tempered, uncertainty measures are added, and the actual estimation procedure's gains are demonstrated to be distinguishable from noise.","tokens_in":12079,"tokens_out":7596,"duration_ms":77887,"concrete_test":"Reanalyze the NHANES and Scenario 1 splits: for each of the 50 train/test splits, record the paired difference in test MSE between TEB-FAR (with CV-selected sigma_y^2) and the joint Bayesian factor model, and report the mean difference with a paired bootstrap 95% CI and the fraction of splits where TEB-FAR wins. If the CI includes zero at ntrain>=400, or if fewer than 60% of splits favor TEB-FAR, the claim of substantial improvement fails.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that targeted empirical Bayes estimation of sigma_y^2 produces substantial predictive gains. The reported evidence does not bear this out. In Table 1, when sigma_y^2 is selected by the proposed 10-fold CV, the best TEB-FAR edge over the joint Bayesian factor model is 0.003 (n=400: 0.973 vs 0.976), and at n=200 TEB-FAR is worse (0.993 vs 0.992). In Table 2, Scenario 1, the only scenario where TEB-FAR claims an advantage, the edge at n=1500 is 0.002 (0.983 vs 0.985) and at n=2000 it is 0.001 (0.981 vs 0.982), with all methods effectively tied. These differences are far smaller than the word 'substantial' in the abstract and are presented without standard errors, confidence intervals, or per-split distributions, so they could easily be sampling noise. Figure 3's visually substantial superiority relies on an oracle-style sweep over sigma_y^2 values, but the actual method must select sigma_y^2 from data, and when it does (Table 1), the gains largely disappear. Thus the load-bearing empirical claim is unsupported as reported.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Targeted Empirical Bayes Factor Regression (TEB-FAR), a modification of joint Bayesian Gaussian factor models in which the residual variance of the response, sigma_y^2, is estimated by an empirical Bayes procedure that targets out-of-sample prediction of y from x, rather than being inferred from the joint likelihood. The motivation is that joint factor models can miss weak-but-predictive latent factors when the predictors dominate the joint likelihood. The authors demonstrate with a one-factor example that fixing sigma_y^2 to a small value can reorder which factor is prioritized, and then postulate that this intuition extends to k>1. They evaluate TEB-FAR on an NHANES phthalate-BMI application and on three simulation scenarios, comparing against a joint Bayesian factor model, lasso, ridge, and OLS. The reported findings are that TEB-FAR can outperform competitors in some settings, but the gains are small in magnitude and are presented without uncertainty quantification.","tokens_in":12294,"tokens_out":2404,"duration_ms":24570,"significance":"The idea of targeting the residual variance of the response in a joint factor model is simple and potentially useful, and the paper is clearly written with reproducible code provided. If the claimed improvements were robust and substantial, this would be a useful practical contribution to supervised factor analysis and to environmental health applications. However, as reported, the empirical support for the central claim is weak: the largest improvements over the joint Bayesian factor model are on the order of 0.001--0.003 in MSE, with no standard errors or per-split distributions, and the more favorable results in Figure 3 rely on an oracle-style choice of sigma_y^2 rather than the data-driven CV procedure that the method actually uses. The paper also explicitly postulates, rather than proves, that the one-factor phase-transition intuition generalizes to k>1, with only simulation evidence for that generalization.","major_comments":[{"comment":"The central claim of the abstract, that the method leads to 'substantial improvements in simulation performance,' is not supported by the reported numbers. In Scenario 1, the largest TEB-FAR advantage over the joint Bayesian factor model is 0.002 (n=1500) and 0.001 at n=2000; at n=200 the joint model is better (1.010 vs 1.007). In Scenarios 2 and 3, TEB-FAR is essentially tied with or slightly worse than the joint model. These differences are far smaller than the word 'substantial' implies, and no standard errors, confidence intervals, or per-split results are provided, so the differences could easily be sampling noise. The authors should either report uncertainty quantification (e.g., standard errors across the 50 replications) or soften the claim.","section":"Section 3.2.2, Table 2"},{"comment":"Figure 3, which visually shows a substantial advantage for TEB-FAR, is based on sweeping over a grid of fixed sigma_y^2 values, which is an oracle-style comparison. The actual procedure selects sigma_y^2 by 10-fold cross-validation within the training set, and Table 1 shows that when this is done, the gains over the joint Bayesian factor model shrink to at most 0.003 (n=400: 0.973 vs 0.976) and are negative at n=200 (0.993 vs 0.992). The main evidence for the method's practical value therefore appears to be the oracle curve, not the data-driven procedure. The authors should present the CV-selected results as the primary evidence and clarify that Figure 3 is diagnostic rather than a fair comparison.","section":"Section 3.1, Figure 3 vs Table 1"},{"comment":"In Figure 3, lasso and ridge are evaluated at 'the optimal choice of their tuning parameters over a grid,' apparently selected using the test data, while TEB-FAR is evaluated at a fixed grid of sigma_y^2 values. This is not a fair comparison, because it gives lasso/ridge oracle tuning on the test set. The comparison in Table 1, where lasso and ridge use internal cross-validation, is more appropriate; the text should emphasize that and not rely on Figure 3 for the performance claims.","section":"Section 3.1, Figure 3 (lasso/ridge comparison)"},{"comment":"The paper states, 'we postulate that this intuition generalizes for k>1, and we support this hypothesis with empirical results in Section 3.' This is a key load-bearing assumption: the entire method's mechanism for improving prediction relies on shrinking sigma_y^2 causing the model to reallocate factors toward those predictive of y. The one-factor analytic example is instructive, but it does not establish that the same reordering occurs in multi-factor models with shrinkage priors. The simulation evidence in Section 3 is not connected to the phase-transition mechanism—the authors do not show, for example, that the estimated loadings in Scenario 1 shift to favor the weak factor when sigma_y^2 is reduced. The manuscript would be strengthened by either a formal argument for k>1 or a direct simulation demonstration of the reordering mechanism.","section":"Section 2.3"}],"minor_comments":[{"comment":"Typo: 'emprical Bayes' should be 'empirical Bayes'.","section":"Section 3.2.1"},{"comment":"Typo: 'targetting' should be 'targeting'.","section":"Section 4"},{"comment":"The reference list entry for Friedman et al. has 'Repositary' instead of 'Repository'.","section":"References"},{"comment":"The grid for sigma_y^2 is described as 'values from 0 to 1' in Section 2.2 but '0.01 to 1' in Section 3.1; please make the description consistent.","section":"Section 2.2"},{"comment":"In the sentence about the induced covariance matrix, the sum of squared differences is reported in the text as dropping from 5.6 x 10^-2 to 4.8 x 10^-3, but the supplement reports mean squared differences of 1.1 x 10^-3 and 1.1 x 10^-4. Please clarify which metric is being reported in each place.","section":"Section 3.1.1"}],"recommendation":"major_revision","confidential_remarks":"The paper's idea is interesting and the writing is clear, but the empirical evidence is currently too weak to justify the abstract's claim of 'substantial improvements.' The most favorable results come from an oracle-style comparison in Figure 3, and the actual CV-based procedure shows differences on the order of 0.001--0.003 in MSE with no uncertainty quantification. I would be willing to reconsider after the authors provide standard errors or interval estimates, reframe the claims to match the effect sizes, and make the simulation analysis directly address the k>1 generalization. The paper may also need a stronger connection to the phase-transition mechanism in the multi-factor setting. A reject is not warranted because the core idea is sound and the weaknesses are fixable within the paper's scope."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the idea is simple and legitimately new in this form, but the abstract overclaims. The reported MSE advantages over the joint Bayesian factor model are on the order of 0.001 to 0.003, with no standard errors or per-split distributions, so \"substantial\" is doing a lot of work. I'd send it to a serious referee, but the referee should ask for a more honest summary and a stronger empirical section.\n\nWhat is genuinely nice: the targeted EB selection of the outcome residual variance sigma_y^2 is a clean, practical modification of joint factor analysis. The one-factor algebra in Section 2.3 is correct as far as I can tell, and the phase-transition demonstration is an instructive toy. The authors are also honest that the generalization to k>1 is a postulate—they don't dress it up as a theorem. The simulation design covers a useful range of scenarios, and the code is promised.\n\nThe soft spots are mostly about evidence. In Table 1, the best TEB-FAR edge is 0.003, and in Table 2 the best edge is 0.002; all methods are effectively tied at larger n. Without error bars or split-level distributions, those differences could easily be noise. The Figure 3 gains are visually striking but rely on an oracle-style sweep over sigma_y^2; when sigma_y^2 is actually selected by CV in Table 1, the gains mostly vanish. The motivation phase transition is computed from true parameters, not fitted models, so it is a heuristic rather than proof of the mechanism. And the most relevant baselines—supervised PCA, partial factor modeling, PLS—are not compared, so it is hard to tell whether the modest win over JBFM transfers to a win over methods already designed for supervised dimensionality reduction.\n\nOn the other side of the ledger, the method is not broken. The empirical work is reproducible in principle, the authors flag the limits of their own scenarios (Scenarios 2 and 3 show no meaningful gain), and the inferential differences they highlight are plausible. This is a useful contribution in search of sharper evidence.\n\nFor peer review: yes, send it out. The idea is worth discussing and the execution is careful enough to engage with seriously. But the revision should temper the claims, add uncertainty measures, and add at least the supervised PCA and partial factor baselines. If the evidence stays this thin, the abstract needs to say so.","headline":"A sensible and clearly written paper that promises more than its own tables deliver; the gains over the joint Bayesian factor model are tiny and come without uncertainty bounds, so it needs a tempering of claims and a real comparison set before it earns its abstract.","tokens_in":12846,"tokens_out":2646,"would_cite":true,"duration_ms":26421,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62H25","62F15","62J07"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that a joint Gaussian factor model can be made more supervised by estimating the response's residual variance with an empirical Bayes step aimed at predicting y from x, and that this improves predictive performance and…","keywords":["joint factor analysis","targeted empirical Bayes","response residual variance","supervised factor models","principal component regression","high-dimensional predictors","mixture exposures","phthalates"],"falsifier":"Compute the KL-optimal k-factor approximation to a known multi-factor Gaussian model as a function of fixed $\\sigma_y^2$, with the outcome loading only on the weakest factor; if no $\\sigma_y^2$ below the marginal variance causes the loadings to align with the outcome factor, the mechanism fails. Alternatively, run the paper's Scenario 1 at n_train = 1500 many times: if TEB-FAR's cross-validated $\\sigma_y^2$ does not yield lower test MSE than the joint Bayesian factor model, the central predictive claim is contradicted.","tokens_in":11812,"feed_emoji":"🎯","tokens_out":12452,"duration_ms":114356,"temperature":0.7,"pith_summary":"Joint Bayesian factor models infer latent structure from the joint distribution of predictors and response, so when the number of predictors p is large, the predictor covariance can dominate and the fitted factors can miss signal that is weak in x but decisive for y. The paper claims that one hyperparameter—the response's residual variance, $\\sigma_y^2$—can re-target the model without changing the model family: instead of sampling it with the joint likelihood, estimate it by cross-validation to minimize prediction error of y given x. In a one-factor calculation this produces a phase transition: as $\\sigma_y^2$ is forced down, the optimal factor loadings switch from the dominant predictor factor to the outcome-relevant factor. The paper states this mechanism as a postulate for multi-factor shrinkage-prior models and supports it empirically: in simulations where the outcome loads only on the weakest predictor factor, its Targeted Empirical Bayes Factor Regression (TEB-FAR) becomes the best predictor once training sets are large enough, and in NHANES phthalate–BMI data it outperforms joint Bayesian factor models, lasso, ridge, and OLS for training sets of 400 or more.","feed_headline":"Tuning one variance term unlocks weak response signals","feed_subtitle":"Selecting the response variance by prediction error beats joint-likelihood factor fitting and lasso/ridge on phthalate–BMI data.","key_machinery":"The load-bearing object is the response residual variance $\\sigma_y^2$ in the Gaussian joint factor model, estimated by cross-validated predictive accuracy of y|x rather than by the joint likelihood—the paper names the resulting procedure TEB-FAR (Targeted Empirical Bayes Factor Regression). A named identity carries the argument: the one-factor induced regression variance $\\sigma^2 = \\sigma_y^2 + \\gamma^2/(1 + \\sum_j \\lambda_j^2/\\sigma_j^2)$ and the coefficient formula $\\beta \\propto \\gamma \\lambda_j/\\sigma_j^2$ show that a small fixed $\\sigma_y^2$ forces the outcome loading $\\gamma$ away from zero and makes the likelihood penalize loadings that are poor for y|x. This produces the phase transition in the KL-optimal loadings, and the same $\\sigma_y^2$ then serves as a tuning parameter for how supervised the factor analysis is. The remaining machinery is an overfitted factor model with an increasing shrinkage prior that starts with a large upper bound on the number of factors and shrinks unnecessary columns of the loadings matrix to zero.","core_discovery":"The central discovery is that the response residual variance $\\sigma_y^2$ is not merely a noise parameter but a dial that controls which latent factors the joint likelihood rewards. The paper works out the one-factor special case, where the induced regression of y on x has coefficients $\\beta = \\left(\\gamma / \\left(1 + \\sum_j \\lambda_j^2/\\sigma_j^2\\right)\\right)(\\lambda_1/\\sigma_1^2,\\dots,\\lambda_p/\\sigma_p^2)^T$ and residual variance $\\sigma^2 = \\sigma_y^2 + \\gamma^2/\\left(1 + \\sum_j \\lambda_j^2/\\sigma_j^2\\right)$; forcing $\\sigma_y^2$ small forces $\\gamma^2$ up, which magnifies the penalty for choosing loadings $\\lambda$ that predict y poorly. In the motivating two-factor example, the KL-optimal one-factor approximation aligns with the first factor and models y as noise, but as $\\sigma_y^2$ is decreased along a grid, the optimal loadings undergo a sharp phase transition near $\\sigma_y^2 \\approx 0.06$ and align instead with the second factor—even though the resulting joint model is worse for x alone. The empirical Bayes estimator selects $\\sigma_y^2$ by cross-validation on a grid from 0 to 1 for standardized data, then the remaining parameters and the number of factors are learned with an overfitted shrinkage-prior factor model. The paper's claim is that this simple change makes the factor model \"more supervised\" without relaxing the conditional-independence structure.","pith_inferences":["If the phase transition is general, then for any two-factor Gaussian model there should be a critical $\\sigma_y^2$ below which the KL-optimal factor assignment flips, expressible in terms of loadings and idiosyncratic variances; deriving it in the multi-factor case would turn the paper's postulate into a theorem and could predict when TEB-FAR helps.","The same targeted-empirical-Bayes logic should transfer to any joint model with a nuisance likelihood component that threatens to dominate inference—high-dimensional covariates in mixed models, for example—by cross-validating only the target component's variance.","The NHANES coefficient shifts should not be read as causal exposure effects; they are predictions conditional on the fitted factor model, and the paper's own comparison is about predictive accuracy rather than causal identification.","A direct replication test is to apply TEB-FAR to other environmental-mixture datasets where the outcome is suspected to track a low-variance exposure pattern; predictive gains appearing exactly in that regime would support the mechanism, while gains everywhere would suggest generic shrinkage."],"forward_implications":["In settings where the response depends on a minor factor of the predictors, TEB-FAR can recover predictive signal that a fully Bayesian joint factor model misses; in the paper's Scenario 1, TEB-FAR achieves the lowest test MSE among all methods once the training set has at least 1,500 observations.","The selected $\\sigma_y^2$ value is stable and interpretable: it sits on a fixed grid [0,1] for standardized data, so cross-validation within each training set does not suffer the tuning-instability that the paper observes for lasso and ridge in the NHANES comparison.","TEB-FAR can change substantive inferences, not just prediction: in the NHANES application the induced regression coefficient for MC1 moves from strictly positive to include zero, while MCOH moves from including zero to strictly positive, relative to the joint Bayesian factor model.","When the true data-generating process is a high-signal factor model or a sparse linear regression, TEB-FAR performs about as well as the joint Bayesian factor model, so the targeted modification does not appear to harm performance in settings that already work.","By shrinking $\\sigma_y^2$, TEB-FAR reduces regularization on the joint covariance, moving the estimated joint covariance substantially closer to the Pearson sample covariance (sum of squared differences drops from $5.6\\times 10^{-2}$ to $4.8\\times 10^{-3}$)."],"supporting_citations":[{"why":"Supplies the two-factor motivating example in which the joint-likelihood one-factor approximation ignores the outcome, and motivates predictor-dependent shrinkage.","marker":"Hahn et al. [2013]"},{"why":"Provides the increasing shrinkage prior that TEB-FAR uses to handle an overfitted number of factors.","marker":"Bhattacharya and Dunson [2011]"},{"why":"Articulates the principle that the response may be closely related to the least important principal component, motivating targeted supervision.","marker":"Cox [1968]"},{"why":"Gives real-world examples where low-variance components matter for the outcome, motivating reweighting toward y.","marker":"Jolliffe [1982]"},{"why":"Gives an extreme data example where the final principal component alone relates to the outcome, showing why variance-based selection fails.","marker":"Hadi and Ling [1998]"},{"why":"Defines the lasso competitor that TEB-FAR is compared against in the empirical sections.","marker":"Tibshirani [1996]"}],"fun_headline_variants":["A single variance knob sharpens factor analysis for weak signals","Response variance tuning reveals hidden factors in joint models","Shrinking one variance term surfaces predictive factors","Empirical Bayes twist targets response variance for better factors","Phase transition in factor loadings via targeted empirical Bayes"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The argument assumes that the phase transition proved for one factor—forcing $\\sigma_y^2$ down makes the optimal loadings switch from the dominant factor to the outcome-relevant factor—continues to hold when there are multiple factors and shrinkage priors; Section 2.3 labels this a postulate, and the only support offered is simulations.","fun_headline_variants_meta":{"raw":{"variants":["A single variance knob sharpens factor analysis for weak signals","Response variance tuning reveals hidden factors in joint models","Shrinking one variance term surfaces predictive factors","Empirical Bayes twist targets response variance for better factors","Phase transition in factor loadings via targeted empirical Bayes"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000534,"raw_usage":{"total_tokens":2605,"prompt_tokens":1019,"completion_tokens":1586,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":635,"completion_tokens_details":{"reasoning_tokens":1511}},"tokens_in":635,"tokens_out":1586,"duration_ms":11694,"temperature":1.0,"reasoning_tokens":1511,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T20:53:16.561857+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute the KL-optimal k-factor approximation to a known multi-factor Gaussian model as a function of fixed $\\sigma_y^2$, with the outcome loading only on the weakest factor; if no $\\sigma_y^2$ below the marginal variance causes the loadings to align with the outcome factor, the mechanism fails. Alternatively, run the paper's Scenario 1 at n_train = 1500 many times: if TEB-FAR's cross-validated $\\sigma_y^2$ does not yield lower test MSE than the joint Bayesian factor model, the central predictive claim is contradicted.","supporting_citations":[{"cited_title":"Partial factor modeling: predictor-dependent shrinkage for linear regression","cited_arxiv_id":null,"evidence_quote":"Supplies the two-factor motivating example in which the joint-likelihood one-factor approximation ignores the outcome, and motivates predictor-dependent shrinkage."},{"cited_title":"Sparse Bayesian infinite factor models","cited_arxiv_id":null,"evidence_quote":"Provides the increasing shrinkage prior that TEB-FAR uses to handle an overfitted number of factors."},{"cited_title":"Notes on some aspects of regression analysis","cited_arxiv_id":null,"evidence_quote":"Articulates the principle that the response may be closely related to the least important principal component, motivating targeted supervision."}],"review_version":1}