{"id":"ea007a8c-bf1d-4904-b502-3f2b226ee3d4","arxiv_id":"2501.03058","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":1.0,"correctness_risk":"low","formal_verification":"none","parameter_count":2,"one_line_summary":"Poisson regression with a constant event rate is a special case of the Cox proportional hazards model when the baseline hazard is constant, a standard equivalence presented here as a tutorial for fall-risk analysis.","lead":"This paper explains how classical survival analysis models, from logistic and Poisson regression to the Cox proportional hazards model, relate to each other, and shows that Poisson regression is a special case of the Cox model with a constant baseline hazard. It is a tutorial using elderly fall risk as a running example, with no new data, code, or experiments.","discovery_kind":"review","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The Poisson-Cox unification in §6 requires a constant baseline hazard h0(u)=1; the paper does not test or justify this for fall risk, so the central claim is conditional, not unconditional.","rationale":"The paper's derivations are mathematically transparent, and the special-case relationship is correct under the stated constant-baseline-hazard assumption. The reader's CONDITIONAL verdict is appropriate: the central claim is sound as a conditional statement, but the paper overstates its scope and novelty. I find no internal inconsistency in the derivation. The most load-bearing limitation is that the equality S(t|X)=exp(-t exp(βᵀX)) holds only when H0(t)=t; for fall risk, the hazard is commonly time-varying, and the paper offers no way to verify this condition. This does not warrant rejection, because the paper itself acknowledges the assumption in §6, but it does warrant keeping the verdict conditional on the applicability of that assumption. Censoring is a related but secondary gap: the model-level equality does not imply equivalence of the Poisson and Cox estimators, since the paper never derives the Poisson likelihood under censoring or connects it to the partial likelihood. I therefore agree with the reader and recommend keeping the verdict unchanged.","tokens_in":11516,"tokens_out":7647,"duration_ms":75972,"concrete_test":"Simulate n=500 survival times from a Weibull distribution with shape γ=1.5, scale α=5, and one binary covariate with true log-hazard ratio 0.5; generate independent right-censoring times. Fit (a) Poisson regression with a log-time offset and (b) a Cox PH model. Compare the estimated coefficients and predicted S(6|X=0). If the estimates and predictions differ beyond sampling error, the equality in §6 fails when H0(t)≠t, confirming the constant-hazard condition is load-bearing. A complementary check: plot the Nelson-Aalen cumulative hazard for a real fall-risk cohort on a log-log scale; if the curve is not linear with slope 1, constant hazard is rejected.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim rests on Section 6, which compares S_Poisson(t|X)=exp(-t exp(βᵀX)) with S_Cox(t|X)=exp(-H0(t) exp(βᵀX)). These are equal for all t only if H0(t)=t, i.e., h0(u)=1; any other constant can be absorbed into the intercept, so the substantive requirement is that the baseline hazard is constant. The paper states this in §6, but the abstract presents the claim without this qualifier. In fall risk, hazards typically increase with age, comorbidity, and time since baseline, so the constant-hazard condition is not innocuous; it excludes exactly the time-varying risk patterns that motivate survival analysis. The paper provides no empirical check on any fall dataset, such as a Nelson-Aalen plot or a test of the constant-hazard assumption. Thus the unification holds for the exponential submodel, but the paper does not establish its relevance to real fall-risk data.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper is a tutorial-style manuscript that revisits logistic regression, Poisson and exponential distributions, generalized linear models, and the Cox proportional hazards model, using fall risk as a motivating application. Its central technical claim is that Poisson regression in the survival context is a specific case of the Cox model when the baseline hazard is constant. Sections 3 through 6 contain step-by-step derivations of survival functions, GLM formulations, and the Cox model, and Section 7 illustrates coefficient estimation, baseline hazard estimation, and survival prediction with a hypothetical example. No real dataset is analyzed.","tokens_in":11697,"tokens_out":6577,"duration_ms":63188,"significance":"As a pedagogical contribution, the paper's main strength is its careful step-by-step derivations and the explicit identification of the constant-baseline-hazard condition in Section 6. The observation that a constant-hazard Poisson/exponential survival model is nested in the Cox model is mathematically sound but is a standard textbook result rather than a new methodological finding. The paper does not present new estimation procedures, new theory, or empirical evidence. Its value would be as an accessible tutorial, provided the scope is stated precisely and the unqualified claims in the abstract and conclusion are corrected. The fall-risk framing is currently unsupported because all applications are illustrative and no real fall data are examined.","major_comments":[{"comment":"The paper's central claim is established only under a constant baseline hazard. Section 6 explicitly sets h0(u)=λ0 and then shows that the Poisson and Cox survival functions coincide when λ0=1 (or, more generally, when the constant is absorbed into the intercept). However, the abstract and Section 8 state that 'Poisson regression in the survival context is a specific case of the Cox model' without this qualifier. Since fall hazards are generally time-dependent, the unqualified statement overstates the result. The abstract and conclusion should be revised to include 'under a constant baseline hazard,' or the paper should state at the outset that the unification is conditional.","section":"§6 and Abstract"},{"comment":"The fall-risk application is not supported by any data. Section 7.2 uses a hypothetical table of H0(t) values for months 1 through 6 with no source, and no real fall dataset is analyzed anywhere in the paper. Consequently, the manuscript does not demonstrate that the constant-hazard assumption is reasonable for fall risk, nor does it show that the proposed framework yields useful predictions in practice. If the manuscript is intended as a tutorial, the applied claims in the abstract and introduction should be removed or explicitly labeled as illustrative; if applied relevance is claimed, an empirical section with at least a descriptive check of the hazard shape, such as a Nelson-Aalen plot, and a model fitting exercise on real data is needed.","section":"§7.2 and overall paper"},{"comment":"The estimation section omits censoring, which is central to survival analysis. The partial likelihood L(β)=∏_{i=1}^n exp(β^T X_i)/∑_{j∈R(t_i)} exp(β^T X_j) and the Breslow-type estimator for H0(t) are written without a censoring indicator, implicitly assuming that every subject's event time is observed and that there are no tied event times. In fall-risk follow-up, right censoring is the norm rather than the exception, so the tutorial's estimation recipe is incomplete as a practical guide. The authors should either state the complete-data assumption explicitly and point to standard censored-data formulas, or include censoring indicators and a brief note on tied event times.","section":"§7.1"}],"minor_comments":[{"comment":"The exponential CDF is misprinted: F(T)=P(T≤t)=1−exp(−λT) should read 1−exp(−λt), and the notation should use t consistently as the time argument.","section":"§3"},{"comment":"The sentence 'If λ0=1 in the Cox model, its formula for the survival function becomes identical to that of Poisson regression' could be misinterpreted as making the result depend on a specific numerical value. Because any constant λ0 can be absorbed into the intercept exp(β0), the substantive requirement is the constancy of h0, not λ0=1; this could be stated explicitly.","section":"§6"},{"comment":"The exponential regression mean hβ(X)=1/(−β^T X) is correct, but the derivation would be clearer if the authors explicitly recalled that E[T]=1/λ for T∼Exp(λ) and that the canonical parameterization gives η=β^T X=−λ.","section":"§4"},{"comment":"There is a typo in the final paragraph: 'classical statistical survivla frameworks' should be 'classical statistical survival frameworks.'","section":"§8"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is a readable tutorial, but the central contribution is a textbook-level observation and no empirical analysis is included. For a research venue, the lack of data and the unqualified central claim are significant concerns. A major revision that adds an empirical demonstration or clearly restricts the paper to a pedagogical scope, and that fixes the abstract/conclusion wording, could make it acceptable for a teaching-oriented outlet. The fall-risk framing is currently more motivational than substantive."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"I read this as a tutorial-style preprint, not a research contribution. The algebra is correct and the exposition is mostly clear, but the headline claim—that Poisson regression is a special case of the Cox model—is standard survival analysis, and it is conditional on a constant baseline hazard. The paper itself says this in Section 6, but the abstract drops the qualifier, which overstates the point.\n\nWhat the paper does well: the step-by-step derivations from the Poisson/exponential survival function to GLM-based Poisson regression and then to Cox are transparent and would be useful in a classroom. The comparison of survival functions in Section 6 is honest: it sets h0(u)=λ0, absorbs λ0 into the intercept, and shows the two forms coincide. The worked example for estimating median survival time from a baseline cumulative hazard table is a nice touch. The math checks out; I did not find a load-bearing error.\n\nSoft spots, in order: (1) Novelty is minimal. The relationship is covered in the very references the author cites, e.g., Kalbfleisch and Schaubel's 'Fifty years of the Cox model', and in any decent survival analysis text. (2) The unification is definitional. Setting the baseline hazard constant is exactly the exponential/Poisson assumption, so the paper is showing that a special case is a special case. (3) The fall-risk framing is not backed by evidence. There is no data, no Nelson-Aalen plot, no test of the constant-hazard assumption. Real fall hazards often increase with age and time since baseline, so the claimed relevance is asserted, not shown. (4) Censoring is absent. The partial likelihood in Section 7.1 is written for uncensored data, which would mislead a newcomer. (5) Minor: the exponential CDF has a typo (T vs t); the deep-learning comparison is rhetorical rather than empirical.\n\nWho is this for? A graduate student or practitioner who wants a gentle derivation-based review of why Poisson/exponential models sit inside Cox. For that audience, it is serviceable. As a research paper in cs.LG, it does not clear the bar: the contribution is pedagogical, not methodological, and there is no empirical component.\n\nOn peer review: if the venue has a tutorial or review track, it could go out after revision; as a regular research submission, I would desk reject rather than spend referee time. Repositioning as a teaching note would be the honest move.","headline":"Correct and clear tutorial, but the Poisson–Cox unification is standard and the abstract overstates it.","tokens_in":12228,"tokens_out":2325,"would_cite":false,"duration_ms":21770,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62N01","62J12"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper argues that Poisson regression, when applied to survival data, is a special case of the Cox proportional hazards model, with the two survival functions identical when the Cox baseline hazard is constant and normalized to one.","keywords":["survival analysis","Cox proportional hazards model","Poisson regression","exponential regression","fall risk prediction","generalized linear models","time-to-event analysis","model unification"],"falsifier":"Take a survival dataset with a visibly time-varying baseline hazard, fit both Poisson regression and Cox regression, and compare their estimated survival curves; the unification would be contradicted if the two curves agree despite a non-constant cumulative hazard, and it would be confirmed if they diverge exactly where $H_0(t)$ departs from linearity.","tokens_in":11284,"feed_emoji":"⏱️","tokens_out":15525,"duration_ms":129331,"temperature":0.7,"pith_summary":"The paper is an expository re-derivation of classical time-to-event models, organized around fall-risk prediction. Its central thesis is that Poisson regression applied to survival data is not a separate model but a special case of the Cox proportional hazards model: with a constant event rate $\\lambda = \\exp(\\beta^T X)$, the Poisson survival function $S(t|X) = \\exp(-t \\exp(\\beta^T X))$ is exactly the Cox survival function when the baseline hazard is constant and set to one. The paper builds a ladder from logistic regression through Poisson and exponential distributions to generalized linear models and Cox regression, showing what each step adds. The motivation is practical: a single survival model can simultaneously predict risk at fixed horizons, interpret covariate effects, and estimate time-to-event, which is particularly useful in healthcare settings compared with deep-learning pipelines that need separate training and post-hoc explanation.","feed_headline":"Poisson regression is Cox regression when the hazard is constant","feed_subtitle":"Using fall-risk data, the paper shows Poisson and Cox survival functions coincide when the baseline hazard is constant.","key_machinery":"The central object is the survival-function identity $S(t|X) = \\exp(-H(t|X))$ with cumulative hazard $H(t|X) = H_0(t) \\exp(\\beta^T X)$. It carries the unification because setting the baseline cumulative hazard to $H_0(t) = \\lambda_0 t$ turns the Cox survival function into $\\exp(-t \\lambda_0 \\exp(\\beta^T X))$, which coincides with the Poisson regression survival function $\\exp(-t \\exp(\\beta^T X))$ when $\\lambda_0 = 1$. The other load-bearing pieces are the GLM link $\\lambda = \\exp(\\beta^T X)$ for Poisson regression and Cox's partial likelihood for estimating $\\beta$ without specifying the baseline hazard.","core_discovery":"The central claim is that Poisson regression in the survival context is a specific case of the Cox model. The paper derives this by setting the event rate as $\\lambda = \\exp(\\beta^T X)$, giving survival $S(t|X) = \\exp(-t \\exp(\\beta^T X))$. In the Cox model, $h(t|X) = h_0(t) \\exp(\\beta^T X)$, and if the baseline hazard is constant, $h_0(u) = \\lambda_0$, the survival function becomes $S(t|X) = \\exp(-t \\lambda_0 \\exp(\\beta^T X))$. With $\\lambda_0 = 1$, the two survival functions are identical. The paper therefore frames the Cox model as the more general framework, with Poisson and exponential survival models as its constant-hazard special case, and identifies a time-varying baseline hazard as the feature that distinguishes them.","pith_inferences":["Beyond the paper, the same nesting argument suggests a specification test: fit a Cox model, plot the estimated baseline cumulative hazard $H_0(t)$ against $t$, and check for linearity through the origin; curvature indicates the Poisson simplification is not appropriate for that dataset.","Beyond the paper, the identity extends naturally to piecewise-constant hazards, where a Poisson regression on time-stratified data would correspond to a stratified Cox model, making the paper's result the limiting case of a more general nesting.","Beyond the paper, nothing in the derivation restricts the outcome to falls, so the constant-hazard nesting should transfer to other time-to-event settings such as disease progression, hospital readmission, or equipment failure.","Beyond the paper, the paper leaves censoring unmodeled; incorporating independent censoring is the natural next step, and the counting-process formulation of the Cox partial likelihood is the standard bridge for checking whether the identity survives in that setting."],"forward_implications":["Any dataset well described by a Poisson or exponential survival model can be fit by a Cox model with a constant baseline hazard, and the coefficients retain the same hazard-ratio interpretation.","The practical choice between Poisson regression and Cox regression reduces to one assumption: whether the baseline hazard is constant, with the simpler model sufficient only when it is.","For fall-risk monitoring, a single fitted survival model can output 3-, 6-, and 12-month risk probabilities, covariate hazard ratios, and median time-to-fall without separate training runs.","The paper's equivalence between $P(N_t \\ge 1)$ and $P(T \\le t)$ means that event-count questions and time-to-first-event questions are answered by the same exponential survival formula only under the memoryless, constant-rate assumption."],"supporting_citations":[{"why":"Supplies the Poisson and exponential process relationship $P(N_t=0)=\\exp(-\\lambda t)$ that becomes the survival function.","marker":"[16]"},{"why":"Provides the statistical regression background and GLM framework used to write $\\lambda=\\exp(\\beta^T X)$.","marker":"[17]"},{"why":"Supplies the generalized linear models introduction that underlies the Poisson regression derivation.","marker":"[18]"},{"why":"Defines the Cox proportional hazards model $h(t|X)=h_0(t)\\exp(\\beta^T X)$, the general model that Poisson regression is claimed to specialize.","marker":"[19]"},{"why":"Provides the partial-likelihood justification used in the paper for estimating Cox coefficients without specifying the baseline hazard.","marker":"[20]"}],"fun_headline_variants":["Poisson is Cox with a constant hazard","Constant hazard unifies Poisson and Cox","Survival analysis: Poisson is Cox's special case","Poisson regression as Cox with constant baseline","One model family: Poisson, Exponential, Cox unified"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The claimed unification holds only if the event rate is constant over time, meaning the underlying process is memoryless; if the hazard actually varies with time, the Poisson survival curve $\\exp(-\\lambda t)$ no longer equals the general Cox survival curve $\\exp(-H_0(t) \\exp(\\beta^T X))$.","fun_headline_variants_meta":{"raw":{"variants":["Poisson is Cox with a constant hazard","Constant hazard unifies Poisson and Cox","Survival analysis: Poisson is Cox's special case","Poisson regression as Cox with constant baseline","One model family: Poisson, Exponential, Cox unified"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00145,"raw_usage":{"total_tokens":5853,"prompt_tokens":970,"completion_tokens":4883,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":586,"completion_tokens_details":{"reasoning_tokens":4814}},"tokens_in":586,"tokens_out":4883,"duration_ms":31351,"temperature":1.0,"reasoning_tokens":4814,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T21:57:37.320764+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a survival dataset with a visibly time-varying baseline hazard, fit both Poisson regression and Cox regression, and compare their estimated survival curves; the unification would be contradicted if the two curves agree despite a non-constant cumulative hazard, and it would be confirmed if they diverge exactly where $H_0(t)$ departs from linearity.","supporting_citations":[{"cited_title":"Introduction to probability models","cited_arxiv_id":null,"evidence_quote":"Supplies the Poisson and exponential process relationship $P(N_t=0)=\\exp(-\\lambda t)$ that becomes the survival function."},{"cited_title":"Statistical regression and classification: from linear mod els to machine learning","cited_arxiv_id":null,"evidence_quote":"Provides the statistical regression background and GLM framework used to write $\\lambda=\\exp(\\beta^T X)$."},{"cited_title":"An introduction to generalized linear models","cited_arxiv_id":null,"evidence_quote":"Supplies the generalized linear models introduction that underlies the Poisson regression derivation."},{"cited_title":"Regression models and life-tables","cited_arxiv_id":null,"evidence_quote":"Defines the Cox proportional hazards model $h(t|X)=h_0(t)\\exp(\\beta^T X)$, the general model that Poisson regression is claimed to specialize."},{"cited_title":"A bayesian justification of cox’s partial likelihood","cited_arxiv_id":null,"evidence_quote":"Provides the partial-likelihood justification used in the paper for estimating Cox coefficients without specifying the baseline hazard."}],"review_version":1}