{"id":"ee51724d-0697-4930-b382-7d17568c1104","arxiv_id":"2502.00606","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A cloglog-link unified BART framework for ordinal, survival, and density regression is introduced, with conjugate truncated-exponential augmentation and a minimax-optimal posterior contraction rate for the ordinal core.","lead":"This paper shows that the complementary log-log link, together with a simple truncated-exponential latent variable trick, can power a single Bayesian nonparametric framework for ordinal, survival, and density regression built on Bayesian additive regression trees. The resulting models are computationally simpler than probit or logit analogues, and the ordinal version is shown to concentrate at a minimax-optimal posterior rate.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 3's proof assumes Normal leaf priors (Condition P3 in Web Appendix S.5), while the proposed PHOBART/NPHOBART samplers use log-gamma leaf priors, so the stated contraction result does not cover the implemented model as written.","rationale":"The reader's weakest-assumption pick was Condition P5, the sup-norm approximation capability of the candidate split grid. That is a real condition, but it is explicitly stated as an assumption and is standard in BART contraction theory; a theorem may legitimately condition on it. A more load-bearing problem is internal: the proof's Condition P3 assumes Normal leaf priors, while the proposed PHOBART and NPHOBART samplers, and the conjugate backfitting algebra in Section 3 and Web Appendix S.1, rely on log-gamma leaf priors. Lemma 1, the prior-concentration input to the proof, is borrowed from results for Normal leaf priors and is not re-derived for log-gamma priors. Since the log-gamma density has exponential tails in mu and exp(mu), the prior mass assigned to an epsilon-neighborhood of a leaf value can be substantially smaller than for a Normal prior, and the proof supplies no bound. Thus Theorem 3, as stated, does not establish the claimed minimax rate for the actual model. This does not mean the claim is false; the log-gamma prior may well satisfy an analogous thickness condition with suitable parameter constraints. But the paper has not shown it, so the theory and implementation are not yet connected. The appropriate response is to require the authors to either prove Lemma 1 for the log-gamma prior or explicitly state Theorem 3 for the Normal-leaf variant and prove a separate result for the log-gamma sampler. The overall verdict remains CONDITIONAL: the methodology is promising and the gap is fixable, but the central theoretical guarantee currently applies to a different prior than the one used in practice. I therefore do not move the reader's verdict, but I identify a different weak point than the reader did.","tokens_in":25717,"tokens_out":13824,"duration_ms":152707,"concrete_test":"Replace Condition P3 in Web Appendix S.5 with the actual log-gamma leaf prior mu_tl ~ logGam(a,b), with a,b chosen so that E(mu_tl)=0 and Var(mu_tl)=sigma_mu^2 (i.e., psi(a)=log b and psi'(a)=sigma_mu^2). Re-derive the prior-thickness bound of Lemma 1 for this prior: verify whether Pi(||r-r0||_inf <= epsilon_n | S=S0) >= exp(-C n epsilon_n^2) for epsilon_n=(log n/n)^(alpha/(2alpha+D0)). If the log-gamma density at the leaf values needed to approximate r0 introduces a factor like exp(-c L |mu|) or exp(-c L ||r0||_inf), compute whether this factor can be absorbed into the target exponent n epsilon_n^2; if it cannot, the rate in Theorem 3 must be slowed accordingly. This single analytical check determines whether Theorem 3 applies to the model the paper actually fits.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central theoretical claim is that the cumulative-link cloglog ordinal model, fit with BART-based conjugate backfitting, attains the minimax-optimal posterior contraction rate of Theorem 3. However, the model actually fitted in Section 3 and Web Appendices S.1-S.2 uses leaf priors mu_tl ~ logGam(a,b): the text states 'provided that we use a log-gamma prior mu_tl ~ log Gam(a,b) rather than the usual normal distribution.' By contrast, Web Appendix S.5, Condition P3, states: 'The leaf node parameters mu_tl of T_t are given independent Normal(0, sigma_mu^2) priors.' Lemma 1, which supplies the prior-thickness bound needed for Condition B1, is imported from Orlandi et al. (2021) and Jeong and Rockova (2023) and is proved for Normal leaf priors; no analogous bound is derived for log-gamma leaf priors. This matters because the log-gamma density is proportional to exp(a*mu - b*exp(mu)), which can be exponentially small in |mu| for large negative mu and in exp(mu) for large positive mu. If r0 takes values where that density is very small, the prior mass of the epsilon-neighborhood used in the proof may acquire an extra factor that is not controlled by the existing argument. As written, therefore, Theorem 3 establishes contraction for a different prior than the one used in the paper's algorithms, illustrations, and claimed computational advantages. This is a genuine internal inconsistency between the theoretical and computational sections, not merely a missing technical detail.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper develops a Bayesian nonparametric framework based on the complementary log-log (cloglog) link for ordinal regression, density regression via stick-breaking processes, and survival analysis. The key computational device is a truncated-exponential data augmentation scheme that, combined with BART and log-gamma leaf priors, yields conjugate Bayesian backfitting updates. The paper introduces the PHOBART and NPHOBART ordinal models, the PHSBP and NPHSBP stick-breaking priors, and discrete- and continuous-time proportional/non-proportional hazards survival models. The main theoretical result, Theorem 3, claims minimax-optimal posterior contraction for the ordinal model at rate (log n/n)^(alpha/(2alpha+D0)). The methods are illustrated on simulations, MEPS depression and BMI data, and the LeukSurv leukemia survival dataset.","tokens_in":26093,"tokens_out":17369,"duration_ms":161819,"significance":"If the results hold, the cloglog link provides a useful unification: it avoids cutpoint constraints in ordinal models, requires only one augmented latent variable per observation in stick-breaking mixtures, and yields simple Gibbs samplers. The augmentation algebra in Section 3 and the Web Appendices is coherent, and Theorem 1 is a clean verification that the PHSBP reduces to a Dirichlet process conditional on the regression function. The paper is strongest as a computational and methodological contribution. The theoretical claim, however, is currently mismatched with the implemented model: Theorem 3 is proved for Normal leaf priors, whereas all samplers use log-gamma leaf priors. Since the abstract and introduction explicitly advertise minimax-optimal contraction for the proposed ordinal models, this gap is load-bearing and must be resolved.","major_comments":[{"comment":"Theorem 3's contraction result is proved for a prior that does not match the model implemented in the paper. Condition P3 states that leaf node parameters mu_tl have independent Normal(0, sigma_mu^2) priors, and the prior-thickness bound in Lemma 1 is imported from Orlandi et al. (2021) and Jeong and Rockova (2023), which are proven for Normal leaf priors. However, Section 3 and Web Appendices S.1 through S.4 define all models with log-gamma leaf priors mu_tl ~ log Gam(a,b), with Section 3 explicitly stating \"provided that we use a log-gamma prior mu_tl ~ log Gam(a,b) rather than the usual normal distribution.\" No analogue of Lemma 1 is proven for log-gamma leaves, and the log-gamma density e^{a*mu - b*e^mu} has tail behavior substantially different from the Normal, so the existing prior-thickness argument does not transfer without new work. As written, Theorem 3 establishes contraction for a different ordinal BART model than the PHOBART and NPHOBART procedures whose samplers, illustrations, and computational claims are the subject of the paper. Please either extend the prior-concentration lemma to log-gamma leaf priors or state explicitly that the theoretical guarantee applies only to a normal-leaf variant.","section":"Section 4 and Web Appendix S.5, Condition P3"}],"minor_comments":[{"comment":"The definition of Z_ij contains an internal contradiction: the text says \"Z_ij = 1 if Yi > j, Z_ij = Zi if Yi = j and Z_ij = 0 if Yi > j\"; the last clause should presumably be \"Z_ij = 0 if Yi < j.\"","section":"Web Appendix S.2"},{"comment":"The lambda_b update contains a typographical double comma in the gamma distribution: \"Gam(a_lambda + sum delta_i, , b_lambda + ...)\".","section":"Web Appendix S.4"},{"comment":"The theorem's rate is conditional on Condition P5, but the paper does not state how the number of candidate split points b_n must scale with n for the approximation error bound to be satisfied for an alpha-Hoelder r0; please provide this guidance explicitly.","section":"Section 4, Condition P5"},{"comment":"The entropy bound is written as \"L log dT bT n LU/epsilon\", which is ambiguous; please clarify the intended product (presumably d*T*b_n*L*U/epsilon).","section":"Web Appendix S.5, Lemma 3"},{"comment":"The cross-validated deviance difference between the PH BART model and the Cox Linear model is small (-1.91); please provide standard errors or intervals for the deviance differences to support the claim that the PH BART model outperforms both competitors.","section":"Section 5.3, Table 2"}],"recommendation":"major_revision","confidential_remarks":"This is a solid methods paper with a coherent computational framework and a clean DP connection. The main obstacle to publication is the gap between Theorem 3's conditions and the implemented log-gamma leaf prior, which is fixable either by proving the needed prior-concentration bound for log-gamma leaves or by explicitly restricting the theoretical claim to a normal-leaf variant. I would also encourage the authors to connect the theoretical tree-prior conditions (P1, P2) to the default BART prior used in the illustrations, and to qualify the minimax-optimality claim by the approximation condition P5."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short take: this is a genuinely useful methods paper, but it has a real internal mismatch between its headline theory and what it actually fits. The stress-test note is correct. Section 3 and Web Appendices S.1-S.4 build the Gibbs samplers on log-gamma leaf priors (mu ~ logGam(a,b)), and that is what makes the conjugate backfitting work. Condition P3 in Web Appendix S.5, however, states that leaf node parameters have independent Normal(0, sigma_mu^2) priors, and Lemma 1, imported from Orlandi et al. and Jeong and Rockova, is a prior-thickness bound for normal leaves. No analogous bound is derived for log-gamma leaves. So Theorem 3 proves minimax contraction for a different prior than the one used in every implementation and experiment. That is a load-bearing gap, not a footnote: the minimax-optimal posterior contraction claim is the paper's central theoretical result. It may well be repairable, since the log-gamma prior might satisfy a similar bound, but as written the theorem does not cover the model. What is genuinely good: the cloglog link is exactly the right device to make cumulative-link and continuation-ratio ordinal models coincide, and the truncated-exponential augmentation with log-gamma priors is a neat trick that sidesteps the cutpoint ordering constraints that plague probit and logit ordinal BART. The PHSBP/NPHSBP construction is natural, and Theorem 1's DP special case is a clean Sethuraman calculation. The survival extensions are a reasonable byproduct, not a revolution, but they are coherent. Soft spots beyond the theory mismatch: no code is shipped, benchmark tables have no uncertainty, and there are no MCMC diagnostics. For a methods paper these are fixable, but they should be fixed before anyone relies on the empirical comparisons. The paper also leans on borrowed lemmas in a way that makes the proof hard to check independently, though that is common in this literature and not by itself a flaw. This paper deserves a serious referee. The core ideas are solid and the cloglog trick will be citable even if the theory needs work. If the authors close the prior mismatch, either by proving a log-gamma version of the prior-thickness bound or by honestly restricting the theorem to a normal-leaf variant, and add code and uncertainty-aware comparisons, this becomes a strong contribution. As is, the minimax claim should not be taken at face value. Send it to peer review, but flag the mismatch and require a fix or a clear caveat before acceptance.","headline":"A clever cloglog unification for BART-style ordinal, survival, and density models, but the headline contraction theorem is proved for normal leaf priors while the samplers and experiments all use log-gamma leaf priors.","tokens_in":716,"tokens_out":1479,"would_cite":true,"duration_ms":38626,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62F15","62G08","62N02","62G07"],"pacs":[],"model":"deepseek-v4-flash","headline":"The complementary log-log link makes ordinal, survival, and density regression tractable with a single latent-variable scheme, at minimax-optimal rates.","keywords":["complementary log-log link","Bayesian additive regression trees","ordinal regression","survival analysis","stick-breaking process","density regression","posterior contraction","truncated exponential augmentation"],"falsifier":"Simulate data from a sparse Hölder function with smoothness $\\alpha=0.5$ and fit the ordinal model with a very coarse candidate split grid (e.g., two candidate split points per predictor). If the posterior's integrated total-variation error does not decay at the claimed rate for $\\alpha=0.5$, namely $(\\log n/n)^{1/(2+2D_0)}$, then the approximation condition behind the contraction theorem is violated.","tokens_in":25506,"feed_emoji":"📊","tokens_out":7354,"duration_ms":61678,"temperature":0.7,"pith_summary":"The paper argues that the complementary log-log (cloglog) link, long overshadowed by probit and logit links, is a natural backbone for Bayesian nonparametric regression. Because a binary cloglog model is equivalent to thresholding a latent exponential variable, fitting it reduces to an exponential-regression problem that is conditionally conjugate under BART with a log-gamma leaf prior. The same mechanism makes cumulative-link and continuation-ratio ordinal models coincide, frees inference from monotone cutpoint constraints, and extends to new stick-breaking priors and to discrete- and continuous-time survival models with simple Gibbs samplers. For the ordinal model the posterior is shown to contract at the minimax-optimal rate $(\\log n/n)^{\\alpha/(2\\alpha+D_0)}$ up to logarithmic factors, confirming that the computational convenience does not come at the cost of statistical efficiency.","feed_headline":"One simple link powers ordinal, survival, density Bayesian models","feed_subtitle":"Complementary log-log link gives simple Gibbs samplers and minimax-optimal rates.","key_machinery":"The load-bearing identity is that a cloglog Bernoulli model is exactly a latent exponential thresholding model: $Y=\\mathbf{1}\\{E\\le 1\\}$ with $E\\sim\\mathrm{Exp}(e^{r(x)})$, so truncated exponential augmentation reduces inference to exponential regression. Under a $\\log\\text{Gam}(a,b)$ prior on BART leaves, each tree's marginal likelihood collapses to $\\Gamma(a+A_\\ell)/(b+B_\\ell)^{a+A_\\ell}$, preserving the conjugate backfitting algorithm of BART. The second pillar is the equality of the continuation-ratio and cumulative-link parameterizations under cloglog, with cutpoints $c_k=\\log\\sum_{j\\le k} e^{\\gamma_j}$, which removes ordering constraints on the cutpoints and yields log-gamma full conditionals for the $\\gamma_j$'s.","core_discovery":"The central discovery is that the cloglog link sits at a unique intersection: under the cloglog link, the cumulative-link and continuation-ratio formulations of ordinal regression coincide, and the resulting likelihood has a Poisson-gamma form that admits conjugate updates for BART leaf parameters when leaves carry a log-gamma prior. Augmenting one truncated exponential per observation turns inference into Bayesian backfitting of a Gumbel regression, avoiding the constrained cutpoint sampling that plagues probit and logit cumulative models. The paper uses this to construct the proportional hazards stick-breaking process (PHSBP) and its nonproportional extension (NPHSBP), shows that a special case of the PHSBP is a Dirichlet process with covariate-dependent concentration, and builds proportional and non-proportional hazards survival models. It further proves that the ordinal regression posterior concentrates around the truth at the minimax-optimal rate under a Hölder smoothness and sparsity assumption.","pith_inferences":["Beyond the paper: the same truncated-exponential machinery should transfer to other asymmetric links or to multivariate ordinal outcomes; nothing in the augmentation restricts it to the cloglog link.","Because the contraction-rate theorem is proved only for the ordinal model, the natural next check is whether the stick-breaking and survival models also reach the minimax bounds; the paper conjectures but does not prove this.","The NPHSBP's prior shrinks the nonparametric function toward the proportional-hazards form by controlling splits on the category index, so the posterior probability of such splits could serve as a model-selection test for the proportional-hazards assumption; the paper does not yet run that test."],"forward_implications":["A BART-based ordinal model can be fit with essentially the same ease as a normal regression, with no cutpoint-ordering constraint step.","The new stick-breaking priors need only one latent variable per observation, unlike probit and logit stick-breaking processes, so density regression is more scalable.","The same augmentation yields discrete-time and continuous-time proportional hazards survival models with conjugate Gibbs updates; the proportional-hazards sampler processes each observation in constant time per tree.","The ordinal model attains minimax-optimal posterior contraction rates, so the computational convenience is not bought at the price of slower learning.","A special case of the PHSBP is a Dirichlet process whose concentration parameter varies with covariates, placing the new construction inside the classical dependent Dirichlet process family."],"supporting_citations":[{"why":"Provides the BART prior and Bayesian backfitting algorithm that the paper extends with a log-gamma leaf prior.","marker":"Chipman et al. (2010)"},{"why":"Supplies the latent-variable data augmentation paradigm for cumulative-link models that the cloglog approach replaces.","marker":"Albert and Chib (1993)"},{"why":"Defines the probit stick-breaking process that the new PHSBP and NPHSBP are compared against and improve upon.","marker":"Rodríguez and Dunson (2011)"},{"why":"Provides the discrete-time BART survival model that the new non-proportional hazards model offers a computationally cheaper alternative to.","marker":"Sparapani et al. (2016)"},{"why":"Supplies the stick-breaking construction used to identify the PHSBP as a Dirichlet process in Theorem 1.","marker":"Sethuraman (1994)"},{"why":"Establishes the sufficient conditions for posterior contraction that the proof of Theorem 3 verifies.","marker":"Ghosal et al. (2000)"},{"why":"Gives the prior thickness and entropy conditions (Condition P and Lemmas) used to derive the BART minimax rate.","marker":"Jeong and Rockova (2023)"},{"why":"Identifies the Poisson-gamma likelihood form that makes the Gibbs updates for the ordinal and survival models efficient.","marker":"Hill et al. (2020)"}],"fun_headline_variants":["Cloglog link unifies ordinal, survival, and density regression","Cloglog link simplifies Bayesian ordinal regression and more","One link: cloglog gives simple Gibbs samplers and optimal rates","Cloglog link makes Bayesian nonparametrics computationally simple","Cloglog: the elegant link for ordinal and survival Bayesian models"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The claimed minimax-optimal rate rests on the assumption that the true regression function can be approximated, everywhere on the predictor space and to within the target error, by a tree built on a fixed grid of candidate split points; if that grid is too coarse, the rate proof fails even though the algorithm still runs.","fun_headline_variants_meta":{"raw":{"variants":["Cloglog link unifies ordinal, survival, and density regression","Cloglog link simplifies Bayesian ordinal regression and more","One link: cloglog gives simple Gibbs samplers and optimal rates","Cloglog link makes Bayesian nonparametrics computationally simple","Cloglog: the elegant link for ordinal and survival Bayesian models"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000633,"raw_usage":{"total_tokens":2918,"prompt_tokens":939,"completion_tokens":1979,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":555,"completion_tokens_details":{"reasoning_tokens":1892}},"tokens_in":555,"tokens_out":1979,"duration_ms":14145,"temperature":1.0,"reasoning_tokens":1892,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-09T18:21:03.544865+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Simulate data from a sparse Hölder function with smoothness $\\alpha=0.5$ and fit the ordinal model with a very coarse candidate split grid (e.g., two candidate split points per predictor). If the posterior's integrated total-variation error does not decay at the claimed rate for $\\alpha=0.5$, namely $(\\log n/n)^{1/(2+2D_0)}$, then the approximation condition behind the contraction theorem is violated.","supporting_citations":[{"cited_title":"A., Logan, B","cited_arxiv_id":null,"evidence_quote":"Provides the discrete-time BART survival model that the new non-proportional hazards model offers a computationally cheaper alternative to."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the stick-breaking construction used to identify the PHSBP as a Dirichlet process in Theorem 1."},{"cited_title":"and Rockova, V","cited_arxiv_id":null,"evidence_quote":"Gives the prior thickness and entropy conditions (Condition P and Lemmas) used to derive the BART minimax rate."}],"review_version":1}