{"id":"5cca8b79-b1cf-4f00-af9e-5affd41df32f","arxiv_id":"2509.10354","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A Bayesian joint frailty model on the gap-time scale for multitype recurrent events and a terminal event, with gamma-process baselines and closed-form estimators.","lead":"This paper proposes a Bayesian statistical model for patients who experience several kinds of repeated medical events, such as heart attacks, heart failure episodes, and eventually death. The model estimates how treatments and patient characteristics affect the wait time between events, using a hidden patient-level risk factor shared across all event types.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Simulation resets all event clocks after every event, so data do not follow the type-specific renewal model in §2.1; the reported accuracy/prediction results do not validate the central claim.","rationale":"The reader's weakest_assumption concerns conditional independence and noninformative censoring, which are model assumptions acknowledged in the Discussion. The more load-bearing issue is that the simulation—the primary evidence for the paper's central claim—does not generate data from the model defined in §2.1. The reset-clock scheme in §3.1.1 is a competing-risks model where every event type's clock restarts after any event, whereas the likelihood in Eq. (2) and the model in §2.1 assume independent type-specific renewal processes. This means the simulation evaluates a misspecified model, so the reported accuracy and prediction results do not support the claim of superior performance. The coverage probabilities below nominal and the implausible ALLHAT hazard ratios are also concerning, but they are secondary to the simulation mismatch, which undermines the core empirical validation. The methodology might still be sound, and the simulation could be corrected, so the verdict remains CONDITIONAL rather than REJECT; however, the authors must re-run the simulation under the correct data-generating process before the claims can be accepted.","tokens_in":24199,"tokens_out":12671,"duration_ms":138094,"concrete_test":"Re-implement the simulation in §3.1.1 under the model in §2.1: for each subject, maintain a separate last-event time for each type q and generate the next type-q event from its residual gap, only resetting the clock for the type that actually occurs. Re-run Tables 1–2 for n=100, γ=1.1, ν=2 and compare Bayesian RMSE/coverage with the reported values. If results differ materially (e.g., RMSE changes by >10% or coverage changes by >0.05), the current simulation does not validate the model. Additionally, simulate a single dataset under both schemes and compare the empirical distribution of type-specific gap times to quantify the discrepancy.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The most load-bearing concern is an internal inconsistency between the simulation design and the model. In §3.1.1 the authors state: 'At each gap origin, for components q∈{1,2,0}, we draw independent candidate gaps using a Weibull baseline ... The next event time is min(T1,T2,T0); we realize the corresponding type, reset the gap-time clock, and repeat.' This resets the clock for all event types after every event of any type. However, the model in §2.1 defines T^q_ij as IID gap times between successive events of the same type q, and the likelihood in Eq. (2) uses type-specific backward recurrence times B_qi(u) = u − S_qi N†_{qi}(u−), i.e., time since the last type-q event. Under the simulation's reset scheme, B_qi(u) is the time since the last event of any type for every q, so the simulated data do not follow the assumed type-specific renewal structure. Consequently, the empirical accuracy and prediction results in Tables 1–2 and Figures 2–5 evaluate the method under a misspecified data-generating process and cannot support the claim that 'our model and MCMC algorithm demonstrated superior performance' in the proposed model. A correct simulation would maintain separate last-event times for each type, drawing a new gap only for the type that just occurred while carrying forward residual gaps for other types.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a Bayesian semiparametric joint frailty model for multitype recurrent events and a terminal event on the gap-time scale. A shared gamma frailty links all event types; each baseline cumulative hazard is assigned an independent gamma-process prior. The authors derive closed-form conditional posterior forms for the cumulative hazard and frailty components, propose a Metropolis-within-Gibbs sampler, and claim that the cumulative hazard estimators recover Breslow-Aalen-type estimators as the prior precision tends to zero. The empirical evaluation compares the Bayesian approach with a frequentist EM-based Breslow-type estimator in simulations and applies the model to ALLHAT data, with the abstract asserting 'superior performance in each criterion.'","tokens_in":1753,"tokens_out":2153,"duration_ms":130237,"significance":"If the model and computational claims are correct, the paper offers a useful and relatively simple Bayesian alternative for joint analysis of multitype recurrent events and a terminal event. The gamma-process prior yields closed-form posterior updates, and the near-linear scaling MCMC sampler is attractive for large biomedical datasets. The ALLHAT application addresses a clinically relevant question. However, the simulation design does not generate data from the model specified in Section 2.1, and several internal notation errors affect the likelihood and posterior derivations. The current evidence is therefore not sufficient to support the central claims; the paper would need corrected derivations and a redesigned simulation study.","major_comments":[{"comment":"The simulation DGP resets the gap-time clock for all event types after every event: 'At each gap origin, ... draw independent candidate gaps ... The next event time is min(T1,T2,T0); ... reset the gap-time clock, and repeat.' This does not generate data from the type-specific renewal model of Section 2.1, where T^q_ij are IID gaps between successive type-q events and the likelihood uses B_qi(u)=u minus S_qi at the last type-q event, i.e., time since the last type-q event. Under the simulation, the hazard for a type-q event at time u depends on the time since the last event of any type, not on the time since the last type-q event. The terminal event is also re-drawn after each recurrent event, so its hazard depends on time since the last recurrent event rather than on time since study entry. Consequently, Tables 1-2 and Figures 2-5 evaluate the method under a misspecified DGP and cannot s","section":"3.1.1"}],"minor_comments":[{"comment":"'We proposed' should be 'We propose'.","section":"Abstract"},{"comment":"The product-integral notation is used heavily but not defined for readers unfamiliar with Jacod's formulation; a brief explanation or reference to Andersen et al. [5] would help.","section":"2.2, Eq. (2)"},{"comment":"The 'misspecified' baseline-prior analysis (Figure 3) replaces the reference mean functions with exponential forms and also changes the precision c_q, c_0 from 0.1 to 0.01. It would be clearer to separate misspecification of the baseline mean from the change in precision.","section":"3.1.1, 'Bayesian priors'"},{"comment":"The columns labeled 'Lower CI' and 'Upper CI' appear to be on the log-hazard scale, while the HR column is exponentiated. Please label the columns explicitly (e.g., log-scale 95% CI) to avoid confusion.","section":"Table 7"},{"comment":"Reference [21] and reference [45] are the same work (Ripatti and Palmgren, 2000) and should be merged or cited once.","section":"References"},{"comment":"The three panels are referred to as (a), (b), and (c) in the text; the figure caption should indicate the correspondence between the panels and the event types.","section":"Figure 6"}],"recommendation":"major_revision","confidential_remarks":"The simulation DGP issue is the most serious problem and, in my view, is load-bearing: the current Monte Carlo results cannot be used to validate the model. The notation errors in the likelihood and terminal-event posterior blocks also need careful correction. I would ask the authors to re-derive the relevant equations and rerun the simulation study with a correct type-specific renewal generator before the manuscript can be reconsidered. The application to ALLHAT is interesting, but its interpretation also depends on the corrected model."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The kernel of this paper is worth attention: a gamma-process prior on baseline hazards, a shared frailty linking multitype recurrent events and a terminal event, all on the gap-time scale, with closed-form posterior estimators for cumulative hazards and frailties and a clean Breslow-Aalen limit as the precision goes to zero. The MCMC scheme is simple, nearly linear in sample size, and the robustness checks (prior sensitivity, initialization) are thoughtful. That is a genuinely useful extension of existing gap-time frailty work.\n\nThe soft spots are serious, though. The simulation design in §3.1.1 resets the gap-time clock for every event type after each event, then re-draws all candidate gaps. That does not generate data from the type-specific renewal model in §2.1, where each type's gaps are IID and the at-risk process uses backward recurrence time since the last event of that same type. As a result, the accuracy and prediction results in Tables 1–2 and Figures 2–5 are evaluating the method under a misspecified data-generating process. The reported \"superior performance\" claim is not supported.\n\nThere are also internal notation errors in the likelihood and posterior blocks for the terminal event (Y†_{qi} appears where Y†_{0i} is needed, e.g., Eqs. (2), (5)–(8), (11)–(12)). These errors are load-bearing because the estimators are defined by those terms.\n\nThen there is the ALLHAT application. The hazard ratios in Table 7, such as amlodipine 1.77 and lisinopril 2.30 for ACE, Black race 4.15 for ACE and 6.78 for death, are far outside what the original JAMA publications report. The Discussion claims these findings align with the clinical results, but they do not. That mismatch needs an explanation or a substantial caveat.\n\nOn the positive side, the posterior derivations are otherwise coherent, the notation errors are fixable, and the core idea is a reasonable contribution. But the evidence for the central claims currently fails because of the simulation error and the application discrepancy. Minor issues: no code or data released.\n\nThis paper deserves a serious referee—the idea is salvageable and useful to biostatisticians working on joint frailty models—but it needs a corrected simulation, fixed notation, and a reconciled application before it is defensible.","headline":"Solid Bayesian machinery, but the simulation does not generate data from the proposed model, and the ALLHAT numbers contradict the published trial; needs major revision before the claims can be trusted.","tokens_in":25030,"tokens_out":4120,"would_cite":false,"duration_ms":46917,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62N01","62F15","62P10"],"pacs":[],"model":"deepseek-v4-flash","headline":"A Bayesian joint model with gamma-process priors ties multiple recurrent event types to death and beats a frequentist EM fit.","keywords":["recurrent events","terminal event","joint frailty model","gap time","gamma process prior","Bayesian semiparametric","cumulative hazard","ALLHAT"],"falsifier":"Simulate data with two event types whose dependence is not captured by a single shared gamma frailty—for example, with type-specific frailties or with an acute event raising the subsequent chronic-event hazard—and check whether the model's posterior intervals cover the true regression and frailty parameters and whether the closed-form cumulative hazard estimates deviate from the truth. A direct comparison on such data would settle the scope of the claim.","tokens_in":24036,"feed_emoji":"📊","tokens_out":4113,"duration_ms":45167,"temperature":0.7,"pith_summary":"This paper proposes a Bayesian semiparametric joint model for data in which patients experience several types of recurrent events—such as heart attacks and strokes—and eventually a terminal event, death. The model works on the gap time between events, links all event types through a single shared frailty, and places gamma-process priors on the baseline hazards, giving smooth nonparametric baselines. Its main technical claim is that the cumulative hazard and frailty estimators have closed forms and reduce to Breslow–Aalen-type estimators as the prior precision goes to zero. In simulations the Bayesian fit is reported to be more accurate, more robust to initialization and prior choices, and 2–4 times faster than a comparable frequentist EM fit. Applied to the ALLHAT trial, the model estimates treatment and race effects on acute and chronic cardiovascular events and death while quantifying unobserved dependence.","feed_headline":"Bayesian gap-time joint model beats EM on multitype recurrent events","feed_subtitle":"Gamma-process priors yield closed-form cumulative hazards and a faster, more accurate fit in simulations and the ALLHAT trial.","key_machinery":"The central object is the shared gamma frailty W_i (unit mean, variance 1/nu) that multiplies every type-specific hazard, inducing dependence among recurrent event types and the terminal event. On the gap-time scale, each event type has its own baseline hazard assigned a gamma-process prior, which makes the increments Gamma-distributed and leads to conjugate conditional posteriors for the baseline cumulative hazards and for W_i. The key identities are the closed-form posterior mean estimators for the baseline cumulative hazards (Eqs. 11–12) and for the frailty (Eq. 10), which are sums over the grid of event-count increments divided by weighted at-risk sums; these are what connect the Bayesia","core_discovery":"On the author's terms, the paper establishes that a joint shared-frailty model on the gap-time scale with independent gamma-process priors on each baseline cumulative hazard yields tractable full conditional posteriors: the baseline hazards and frailties are updated in closed form, and the Bayes estimators in Eqs. (11)–(12) are explicit sums over the time grid. Taking the gamma-process precision parameters to zero recovers the Breslow–Aalen cumulative hazard estimator, so the Bayesian procedure contains the classical nonparametric estimator as a special case. The sampler avoids large matrix factorizations and scales nearly linearly with sample size. The authors argue that this combination—cl","pith_inferences":["The precision-to-zero limit suggests a broader bridge: other nonparametric Bayesian priors with a precision parameter may similarly recover classical nonparametric estimators, giving a principled Bayesian justification for running the classical analysis.","The single shared frailty is a strong structural assumption; a natural testable extension is to allow type-specific frailties or a factor model with more than one latent dimension and compare predictive performance on data with heterogeneous dependence.","The gap-time formulation invites modeling the effect of past events on future hazards, e.g., a prior acute event changing the subsequent chronic-event hazard; the current model assumes that dynamics are captured entirely by the frailty and covariates.","The authors' stated limitation of noninformative censoring suggests a concrete extension: a joint model with dependent censoring, which would be needed in studies where dropout is related to disease burden."],"forward_implications":["Clinicians and epidemiologists can fit a joint model of multiple recurrent event types and death without choosing arbitrary parametric baseline hazards, since gamma-process priors adapt to the hazard shape.","Because the cumulative hazard estimator has a closed form and reduces to the classical Breslow–Aalen estimator, results from this Bayesian model can be compared directly with standard frequentist nonparametric estimates.","The near-linear scaling and reported 2–4x speed gain over EM in the simulations make routine sensitivity analyses and larger datasets feasible.","The ALLHAT analysis illustrates the substantive payoff: treatment comparisons (Amlodipine and Lisinopril vs Chlorthalidone), race, and age effects on acute and chronic cardiovascular events and death, with a frailty estimate indicating strong within-patient dependence."],"fun_headline_variants":["Bayesian gap-time joint model outdoes EM on multitype recurrences","Closed-form gamma priors speed up joint recurrent-death modeling","Scalable Bayesian joint model for recurrences and death beats EM","Bayesian joint model for multitype events and death outperforms EM"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The model assumes that, once the shared frailty and covariates are known, all gap times of every event type and the terminal event are independent, and censoring is noninformative—so if one event type directly changes another type's future hazard, or dropout depends on outcome, estimates will be biased.","fun_headline_variants_meta":{"raw":{"variants":["Bayesian gap-time joint model outdoes EM on multitype recurrences","Closed-form gamma priors speed up joint recurrent-death modeling","Scalable Bayesian joint model for recurrences and death beats EM","Bayesian joint model for multitype events and death outperforms EM"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00083,"raw_usage":{"total_tokens":3493,"prompt_tokens":805,"completion_tokens":2688,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":549,"completion_tokens_details":{"reasoning_tokens":2621}},"tokens_in":549,"tokens_out":2688,"duration_ms":22173,"temperature":1.0,"reasoning_tokens":2621,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T17:53:48.088658+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Simulate data with two event types whose dependence is not captured by a single shared gamma frailty—for example, with type-specific frailties or with an acute event raising the subsequent chronic-event hazard—and check whether the model's posterior intervals cover the true regression and frailty parameters and whether the closed-form cumulative hazard estimates deviate from the truth. A direct comparison on such data would settle the scope of the claim.","supporting_citations":[],"review_version":1}