{"id":"bec567fb-3b21-4740-b74c-d42554c2ffee","arxiv_id":"1908.08444","paper_version":5,"verdict":"CONDITIONAL","confidence":"LOW","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"The proposed hierarchical Beta (Polya tree) posterior mean achieves lower mean squared error than MLE, Lasso, ridge, and adjusted MLE in three simulated high-dimensional logistic regression settings.","lead":"This statistics paper proposes a Bayesian method that shrinks noisy regression estimates using a flexible prior over the unknown distribution of the true coefficients. It reports lower average error than common regularized regression methods in simulated high-dimensional logistic regression, so it is relevant to anyone doing statistical estimation with many predictors.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The practical superiority claim for hBeta in GLMs rests on single-realization simulations in Section 6.2; without repeated experiments and sensitivity analysis, the claimed uniform 'best' performance is not established.","rationale":"The paper's Section 2 compound-decision optimality result is internally reasonable, but it is explicitly restricted to oracle separable rules in the symmetric sequence model. The hBeta method in the logistic GLM is therefore an empirical proposal, and its central claim rests on Table 3. That table reports single-realization comparisons, with no error bars and no sensitivity analysis. Although the differences shown are large, a single favorable draw of X and Y is insufficient to support the paper's abstract-level claim of uniformly better estimation and prediction accuracy. The reader's weakest assumption also flags the grid and the oracle-to-hBeta gap; my emphasis is on the single-realization basis of the empirical claim and the absence of any repeated simulation or hold-out evaluation. The proposed check directly tests whether the claimed dominance survives independent replication and whether it depends on the prior grid. This is a correctness-risk issue, not an ad hominem point, and it is addressable. I therefore keep the verdict CONDITIONAL rather than REJECT: the method may be genuinely good, but the current evidence is not strong enough to accept the general statement as stated.","tokens_in":23443,"tokens_out":4770,"duration_ms":54324,"concrete_test":"Re-run the three Section 6.2 examples with 100 independent replications of X and Y (same n, m, and beta configurations), reporting mean and standard error of the normalized MSE for beta, mu, and q for all five methods. If hBeta is not the best in a clear majority of replications, or is within one standard error of the best in any example or target, the headline claim should be weakened. In the same replications, vary the grid endpoints (e.g., ±12 and ±48) and L in {4, 6, 8}; if performance collapses when the grid does not contain the coefficient support, the practical claim is conditional on tuning choices that the paper does not address.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The load-bearing assertion is Section 6.2's claim that hBeta 'performs best in each of the examples and for each choice of the model parameters.' The evidence in Table 3 is each number based on a single realization of X and Y. There are no standard errors, no repeated independent replications, no code, and no sensitivity analysis for the choices amin = -24, amax = 24, L = 6, or for the i.i.d. N(0,1/n) design. If this particular realization is favorable to hBeta, the abstract's general claim of 'better estimation and prediction accuracy' in GLMs does not follow. The theoretical optimality in Section 2 is for the oracle separable Bayes rule in the symmetric sequence model, not for the hBeta posterior mean in a logistic GLM, so the theory does not compensate for the missing replication. The 'prediction' part is also in-sample MSE on q_i, not hold-out prediction.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a hierarchical Bayes approach to large-scale inference in which the unknown coefficient vector is modeled as i.i.d. draws from a common mixing distribution, and the mixing distribution is assigned a hierarchical Beta (Polya-tree-type) prior supported on a fixed grid. Section 2 gives a compound-decision justification for the oracle Bayes rule in the symmetric sequence model, showing that the oracle separable rule minimizes the compound risk for every fixed parameter vector. Section 3 defines the L-level hierarchical Beta model, derives the no-noise posterior, and provides Gibbs samplers for the sequence model and for logistic regression. Sections 4 and 5 apply the method to normal-means multiple testing and to the accident data of Simar et al. (1976). Section 6 applies the method to three high-dimensional logistic regression examples from Sur and Candès (2019), reporting in Table 3 that the hierarchical Beta method achieves the smallest normalized MSE for beta, mu, and q in each example. The abstract claims 'better estimation and prediction accuracy' compared with parametric and nonparametric alternatives.","tokens_in":23623,"tokens_out":3694,"duration_ms":41456,"significance":"If the empirical claims are taken at face value, the paper offers a practically useful nonparametric shrinkage method for high-dimensional GLMs, going beyond sequence-model deconvolution, and the Section 2 oracle optimality result is a clean contribution to compound decision theory. The accident-data analysis is a strength: it reports risk estimates with standard errors based on 400 replications. However, the central new claim about GLM performance rests on single-realization simulations without standard errors, and the theory in Section 2 does not cover the implemented GLM procedure. The paper also provides no code or sensitivity analysis for the key tuning inputs. Thus the broad claims in the abstract are not currently supported, although the underlying methodology is plausible and the issues appear addressable.","major_comments":[{"comment":"The load-bearing empirical claim that 'The hierarchical Beta method performs best in each of the examples and for each choice of the model parameters' is based entirely on a single realization of X and Y for each example. No standard errors, no repeated replications, and no code are provided, so the reader cannot assess whether the reported ordering is stable or is an artifact of one favorable dataset. The authors should report results over many independent replications (with standard errors or boxplots) for all three examples, and should temper the abstract's general claim unless the replicated results support it.","section":"Section 6.2, Table 3"},{"comment":"The optimality result in Section 2 is established only for separable decision rules in the symmetric sequence model, as made explicit in equations (7)-(9). The hierarchical Beta posterior mean for the logistic regression coefficients in Section 6 is not a separable rule: the posterior for each beta_i depends on the entire vector Y and the full design matrix X through the conditional distribution in equation (23). Consequently, the Section 2 theory does not provide theoretical support for the GLM implementation, and the paper should either extend the theory or explicitly state that the GLM method is heuristic, with the sequence-model theory serving only as motivation.","section":"Section 2 versus Section 6"},{"comment":"The abstract claims 'better estimation and prediction accuracy,' but Table 3 measures only in-sample estimation error: normalized MSE for beta, mu, and q on the same data used to fit the model. There is no held-out prediction experiment, no cross-validation, and no evaluation on independent test data. The term 'prediction' is therefore unsupported, and the claims should be restricted to estimation accuracy unless genuine out-of-sample experiments are added.","section":"Abstract and Section 6.2"},{"comment":"The method depends on user-specified inputs: the grid endpoints amin = -24, amax = 24, number of levels L = 6, the dense grid of K = 1280 values used in Algorithm 2, and the MCMC chain lengths. No sensitivity analysis is given for any of these choices, and Section 7 states only that an R package and vignette 'will provide guidance' in the future. Without a sensitivity study or a data-driven rule, the claim that the method is 'free of tuning' (Section 3 intro and Section 7) is overstated, and the possibility remains that performance degrades substantially for other reasonable grid choices, especially if the true coefficients lie outside [amin, amax].","section":"Section 6.2 and Section 7"},{"comment":"Algorithm 2 as printed is incomplete: the loop over i in lines 3-6 computes values of f(b_k|pi) and f(y|b_k, beta_(i)) but never specifies how beta_i^(g) is actually sampled from its conditional distribution. This makes the main simulation results impossible to reproduce from the manuscript text alone. The authors should provide a complete, unambiguous pseudocode and, ideally, working code.","section":"Algorithm 2"}],"minor_comments":[{"comment":"The abstract says the prior 'assigns equal mass to every permutation of the fixed coefficient vector,' which is a useful intuition, but the body of the paper formalizes this only in the symmetric sequence model; the phrasing should be aligned with the formal scope.","section":"Abstract"},{"comment":"There is a typo: 'accodring' should be 'according.'","section":"Introduction, page 2"},{"comment":"The text says Examples 1 and 2 used a single MCMC simulation of 1000 iterations while Example 3 used 10 simulations of 150 iterations. No convergence diagnostics are reported, and the discrepancy across examples is not explained. A brief justification or trace plots would help.","section":"Section 6.2"},{"comment":"Table 3 reports MSE values only as fractions of the MLE MSE. Absolute MSE values and the scale of the target quantities would make the comparisons more interpretable, especially since beta and q are on very different scales.","section":"Table 3"},{"comment":"The accident-data analysis is one of the more compelling parts of the paper because it includes risk estimates with standard errors from 400 replications. Consider presenting this type of replication detail in the logistic regression section as a model for the missing analysis.","section":"Section 5"}],"recommendation":"major_revision","confidential_remarks":"The paper's central claim about GLM performance is not yet established because Table 3 is based on a single realization and the theoretical result does not cover the GLM estimator. However, the methodology is promising and the accident-data results are carefully evaluated. The manuscript would become publishable after a substantive revision that adds replicated simulations, sensitivity analyses, a complete algorithm, and code, and that recalibrates the abstract's claims to the actual evidence. The authors should also address the relationship between the Polya-tree/hierarchical-Beta terminology and the actual prior specification, since this may affect how the contribution is positioned."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a real methodological contribution—a hierarchical Beta (Polya tree) prior on the mixing distribution, with a working Gibbs sampler that can move beyond the sequence model to logistic regression. The parts that are standard, like the compound decision oracle optimality in Section 2, are clearly presented and do what they claim. The accident data section even does a proper simulation with 400 replicates and reports standard errors, which is more than I can say for the section that made the paper famous.\n\nThe soft spots are concentrated in Section 6. Table 3 is the load-bearing evidence for the claim that hBeta 'performs best in each of the examples and for each choice of the model parameters,' and every number in it is one realization. No standard errors, no independent replications, no code. The grid endpoints [-24, 24] comfortably contain all nonzero coefficients in all three examples, and the design is i.i.d. N(0,1/n), so the favorable result could be partly due to favorable settings. The 'prediction' comparison is in-sample MSE for q_i, not out-of-sample prediction. And the theoretical support in Section 2 applies to the oracle separable rule in the symmetric sequence model; the hBeta posterior mean in a logistic GLM has no theorem attached to it. So the abstract's talk of 'universally optimal' and 'better prediction accuracy' goes beyond what is proven.\n\nNone of this is fatal. The method is plausible, the modeling is coherent, and the accident-data risk simulation suggests the shrinkage can genuinely help. But the paper needs to either weaken the claims or strengthen the evidence. Re-running the logistic experiments with multiple realizations and SEs, adding a sensitivity analysis for L and the grid, and releasing code would address most of my concerns. Also worth fixing: the supplied full text is titled 'Hierarchical Bayes Modeling for Large-Scale Inference,' not the arXiv title; make sure the version under review matches.\n\nWho should read this: anyone working on nonparametric empirical Bayes or shrinkage for high-dimensional regression, especially logistic. It deserves a proper peer review, but a referee should demand the replication and code before publication. I would send it out, with the expectation of major revision.","headline":"A promising Polya-tree shrinkage method for large-scale inference, but the flagship logistic regression claim rests on single-realization simulations and needs replication and sensitivity analysis before it can be taken as stated.","tokens_in":24168,"tokens_out":2598,"would_cite":true,"duration_ms":26281,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62J07","62C12","62G05","62F15"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that a single Polya-tree shrinkage prior, applied through a hierarchical Bayes model, yields more accurate estimates of high-dimensional generalized linear model coefficients than standard penalized and empirical-Bayes…","keywords":["Polya tree","hierarchical Beta model","shrinkage estimation","generalized linear models","empirical Bayes","compound decision theory","nonparametric deconvolution","logistic regression"],"falsifier":"Repeat Example 1 with the true nonzero coefficients moved to ±30, outside the [-24, 24] grid, while keeping all other settings fixed; if the hierarchical Beta estimates fail to shrink toward the true support or their MSE rises above LASSO's, the claimed universality would be falsified.","tokens_in":23194,"feed_emoji":"🎯","tokens_out":12456,"duration_ms":119140,"temperature":0.7,"pith_summary":"In high-dimensional regression with fixed (nonrandom) coefficients, the paper asks for one shrinkage estimator that automatically adapts to the coefficient distribution—whether sparse, Gaussian, or mixed. It first shows that if the empirical distribution of the true coefficients were known, the Bayes rule against that distribution would be the optimal separable rule for the whole vector, in both frequentist and Bayesian senses. To approach this oracle without knowing the coefficients, the paper models the coefficients as i.i.d. draws from a common distribution and assigns that distribution a Polya-tree (hierarchical Beta) prior. The posterior mean of this mixing distribution estimates the empirical CDF of the coefficients nonparametrically, and the posterior means of the coefficients serve as shrinkage estimates. In three simulated high-dimensional logistic regressions, this method attains lower mean squared error for coefficients, linear predictor, and success probabilities than the MLE, the debiased MLE, LASSO, and Ridge.","feed_headline":"Bayesian shrinkage beats LASSO and Ridge in logistic regression","feed_subtitle":"A Polya-tree prior learns the coefficients' true spread, cutting mean squared error across three simulated settings.","key_machinery":"The central object is the L-level hierarchical Beta prior, which is a Polya tree on the space of coefficient distributions: it generates a piecewise-constant density on a fixed grid by recursively splitting intervals, where each split multiplies a parent probability by an independent Beta(1,1) variable. This construction gives conjugacy in the no-noise case, so posterior conditional probabilities are simple Beta updates, and it permits a Gibbs sampler for noisy and general likelihoods. In the logistic regression setting, each coefficient is updated conditionally on the others with a density proportional to the likelihood times the step-function prior, evaluated on a dense grid. The machinery does two jobs at once: the posterior mean of the mixing distribution performs nonparametric deconvolution, estimating the empirical CDF of the true coefficients, and the posterior means of the coefficients provide the shrinkage estimates that mimic the oracle rule.","core_discovery":"The paper's central claim is that one nonparametric hierarchical Bayes construction—the L-level hierarchical Beta model, a Polya tree with Beta(1,1) splits—can serve as a universal shrinkage prior for high-dimensional regression. The theoretical anchor is an oracle result: under a symmetric sequence model with nonrandom parameters, the Bayes estimator against the empirical distribution of the parameters is instance-optimal among separable decision rules, and its dependence on the unknowns is only through that empirical distribution. The hierarchical Beta prior is proposed as a fully Bayes way to estimate the mixing distribution and hence approximate the oracle. In the logistic regression application, the posterior mean of the coefficient vector, computed by a Gibbs sampler that conditions each coefficient on the others, shrinks the MLE toward the estimated mixing distribution, and in the three reported simulation examples it achieves the lowest mean squared error among the five methods compared, for the coefficients, for the linear predictor, and for the success probabilities.","pith_inferences":["The paper's fixed-grid choice (e.g., endpoints -24 and 24 in the logistic examples) is untested for sensitivity; a natural extension is an adaptive or expanding grid that tracks the support of the estimated mixing distribution, which would make the method genuinely tuning-free.","The oracle equivalence suggests a benchmark for other nonparametric Bayesian priors on the mixing distribution, such as Dirichlet-process mixtures or log-spline priors, since any prior that consistently estimates the empirical CDF should approach the same oracle.","Because the method estimates the full mixing distribution, it could be lifted to tasks the paper does not pursue, such as variable selection, ranking, or treatment-effect heterogeneity, by using posterior functions of the coefficients rather than just their means.","The reported MSE gains in Section 6.2 are based on single simulation realizations; averaging over repeated draws of the design and outcomes would give standard errors and clarify how much of the apparent margin is sampling noise, particularly in the example where the gap over Ridge is small."],"forward_implications":["Across the three simulated logistic regressions, the hierarchical Beta posterior means achieve the lowest normalized MSE among the MLE, adjusted MLE, LASSO, Ridge, and the proposed method, for the coefficient vector, the linear predictor, and the success probabilities.","The same Gibbs-sampler recipe, with no per-example tuning beyond the grid and level count, yields marginal 95% credible intervals with observed coverage between 92.5% and 99% in the three examples.","In the compound-decision theory section, the optimal separable rule for fixed nonrandom coefficients depends on the unknowns only through their empirical CDF, which is exactly what the posterior mean of the Polya-tree mixing distribution estimates.","In the accident-data application, the hierarchical Beta posterior mean for Poisson rates has lower risk than the NPMLE plug-in estimate and nearly matches two oracle estimators, while all four Bayesian estimators beat the MLE by roughly a factor of four.","Because the same prior and sampler are applied verbatim across very different logistic examples, the paper suggests the method is a single recipe rather than a tuned-per-problem procedure."],"supporting_citations":[{"why":"Supplies the three high-dimensional logistic regression simulation settings and the debiased-MLE baseline that the proposed method must outperform.","marker":"Sur and Cand` es (2019)"},{"why":"Provides the glmnet software used to compute the LASSO and Ridge competitors in the comparison.","marker":"Friedman et al. (2010)"},{"why":"Introduces the compound decision framework that defines the oracle separable rule the paper tries to approximate.","marker":"Robbins (1951)"},{"why":"Underlies the nonparametric maximum-likelihood deconvolution approach that the hierarchical Bayes method is designed to regularize.","marker":"Kiefer and Wolfowitz (1956)"},{"why":"Provides a nonparametric empirical Bayes deconvolution competitor in the normal-means example.","marker":"Efron (2016)"},{"why":"Gives the optimal multiple-testing rule for composite-null hypotheses that the paper extends to fixed nonrandom parameters.","marker":"Sun and McLain (2012)"},{"why":"Provides the selection-adjusted Bayesian framework used to justify that the oracle rule is unchanged after selection.","marker":"Yekutieli (2012)"},{"why":"Supplies the accident data and the NPMLE estimator compared in the Poisson example.","marker":"Simar et al. (1976)"}],"fun_headline_variants":["Polya tree prior shrinks coefficients adaptively, beating LASSO and Ridge","Nonparametric Bayesian shrinkage: Polya tree prior beats LASSO and Ridge in high-dim GLMs","Bayesian adaptive shrinkage via Polya trees improves GLM estimation over LASSO and Ridge","Polya tree prior shrinks toward oracle, outdoes LASSO and Ridge in GLMs","Nonparametric Bayes beats LASSO/Ridge: Polya tree prior adapts to coefficient spread"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method's practical success depends on the user choosing a fixed grid wide and fine enough to contain the true coefficients—the simulations do this by setting endpoints such as -24 and 24—and the paper provides no data-driven rule or sensitivity analysis for this choice, while the proved optimality applies to the oracle rule that already knows the coefficient distribution rather than to the actual posterior-mean estimator.","fun_headline_variants_meta":{"raw":{"variants":["Polya tree prior shrinks coefficients adaptively, beating LASSO and Ridge","Nonparametric Bayesian shrinkage: Polya tree prior beats LASSO and Ridge in high-dim GLMs","Bayesian adaptive shrinkage via Polya trees improves GLM estimation over LASSO and Ridge","Polya tree prior shrinks toward oracle, outdoes LASSO and Ridge in GLMs","Nonparametric Bayes beats LASSO/Ridge: Polya tree prior adapts to coefficient spread"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000575,"raw_usage":{"total_tokens":2733,"prompt_tokens":981,"completion_tokens":1752,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":597,"completion_tokens_details":{"reasoning_tokens":1639}},"tokens_in":597,"tokens_out":1752,"duration_ms":14014,"temperature":1.0,"reasoning_tokens":1639,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T11:39:26.646839+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Repeat Example 1 with the true nonzero coefficients moved to ±30, outside the [-24, 24] grid, while keeping all other settings fixed; if the hierarchical Beta estimates fail to shrink toward the true support or their MSE rises above LASSO's, the claimed universality would be falsified.","supporting_citations":[],"review_version":1}