{"id":"72026a3c-7695-4442-877a-e0f73e13964c","arxiv_id":"2606.19892","paper_version":1,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":5.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":0,"one_line_summary":"Demonstrates estimation for Ghosh-Lin and Fine-Gray models under combined administrative and random censoring by modifying risk sets and IPCW with minimal modeling assumptions.","lead":"The paper shows how to adjust Ghosh-Lin recurrent events and Fine-Gray competing risks models for data that mixes administrative censoring (known study end) with random censoring. A smart generalist might read it to see practical ways to analyze registry or trial data without strong assumptions on the random part.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.3","headline":"Consistency requires administrative censoring times known exactly for every subject (including randomly censored ones); this is asserted but its operationalization is the least secure step.","rationale":"The reader's weakest_assumption directly identifies the same load-bearing precondition. With only the abstract available the concern cannot be checked inside the actual estimating equations, so the verdict remains CONDITIONAL pending verification that the paper's implementation truly uses the asserted knowledge without further assumptions.","tokens_in":1739,"tokens_out":358,"duration_ms":11602,"concrete_test":"Extract the precise definition of the modified risk set (or the indicator I(T_i > t, C_a,i > t)) used in the estimating equation for either model; recompute the estimator on a simulated data set in which C_a,i is set to missing whenever C_r,i < C_a,i and compare to the version that retains the known C_a,i values; if the two differ by more than Monte-Carlo error the method implicitly relies on C_a,i being recorded for all i.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim is that risk-set modification for administrative censoring can be combined with IPCW for random censoring under minimal assumptions once admin times are known for all. The abstract explicitly conditions on this knowledge. If the paper's construction (likely in the estimating equations for Ghosh-Lin or Fine-Gray) treats the admin time C_a,i as observed for every i even when random censoring time C_r,i < C_a,i, then any practical data set in which C_a,i is only recorded for subjects who reach it will break the risk-set adjustment and invalidate the claimed consistency without additional modeling of the joint (C_a, C_r) distribution.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.3","summary":"The manuscript addresses estimation in the Ghosh-Lin recurrent-events model (with terminal event) and the Fine-Gray competing-risks model when data are subject to a mixture of administrative censoring (known exactly for every subject) and random censoring. It proposes combining direct risk-set modification for the administrative component with inverse-probability-of-censoring weighting (IPCW) for the random component, claiming that this hybrid adjustment yields consistent estimators under minimal modeling assumptions.","tokens_in":1876,"tokens_out":481,"duration_ms":19350,"significance":"If the proposed hybrid estimators are consistent, the work would supply a practical, low-assumption route for registry or trial data in which administrative follow-up times are recorded for the entire cohort while a subset of subjects are lost to random censoring. This would be directly useful in biostatistical practice and would extend the existing IPCW and risk-set literatures without requiring a joint model for the two censoring mechanisms.","major_comments":[{"comment":"Abstract and introduction: the central claim that the combined estimator is consistent rests on the assertion that administrative censoring times are known exactly for every subject, including those whose observed time is determined by random censoring. The manuscript does not specify how this information is obtained or recorded when random censoring occurs first, nor does it demonstrate that the risk-set modification remains valid under the data-generating process in which C_a,i is unobserved for subjects with C_r,i < C_a,i.","section":"Abstract"},{"comment":"Estimating-equation section (presumably §3 or §4): the paper must show explicitly how the IPCW weights and the modified risk sets are combined inside the Ghosh-Lin and Fine-Gray estimating equations. Without the explicit form of the hybrid estimating function and a consistency argument (or at least a simulation study under the mixed-censoring regime), it is impossible to confirm that the two adjustments do not interfere.","section":"Methods"}],"minor_comments":[{"comment":"Notation for the two censoring times (C_a and C_r) should be introduced once and used consistently; the abstract uses “administrative censoring time” without a symbol.","section":"Abstract"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":{"model":"grok-4.3","summary":"We thank the referee for the constructive comments. We address each major comment below and will revise the manuscript accordingly to improve clarity.","responses":[{"response":"In the registry and trial settings considered, the administrative censoring time C_a,i is the fixed study end date, which is known and recorded for every subject independently of the random censoring time. The observed data structure includes this known C_a,i for all i, with the observed time being the minimum of the event time, random censoring, and C_a,i. The risk-set modification therefore uses the known C_a,i values and remains valid under this data-generating process; we do not consider the case where C_a,i is unobserved. We will revise the abstract and introduction to explicitly describe this data structure and recording practice.","revision_made":"yes","referee_comment":"[Abstract] Abstract and introduction: the central claim that the combined estimator is consistent rests on the assertion that administrative censoring times are known exactly for every subject, including those whose observed time is determined by random censoring. The manuscript does not specify how this information is obtained or recorded when random censoring occurs first, nor does it demonstrate that the risk-set modification remains valid under the data-generating process in which C_a,i is unobserved for subjects with C_r,i < C_a,i."},{"response":"We agree that the explicit combination of the two adjustments should be presented more clearly. The hybrid estimating equations are obtained by inserting the administratively modified risk-set indicators into the standard Ghosh-Lin and Fine-Gray estimating functions and then weighting the contributions by the IPCW terms for random censoring. We will add the precise mathematical form of these hybrid estimating functions, a brief consistency outline under independent censoring, and simulation results under the mixed-censoring regime to the methods section.","revision_made":"yes","referee_comment":"[Methods] Estimating-equation section (presumably §3 or §4): the paper must show explicitly how the IPCW weights and the modified risk sets are combined inside the Ghosh-Lin and Fine-Gray estimating equations. Without the explicit form of the hybrid estimating function and a consistency argument (or at least a simulation study under the mixed-censoring regime), it is impossible to confirm that the two adjustments do not interfere."}],"tokens_in":1412,"tokens_out":497,"duration_ms":25364,"standing_objections":[]},"desk_editor":{"model":"grok-4.3","letter":"The main takeaway is that this paper gives an explicit way to handle the mix of administrative and random censoring in the Ghosh-Lin recurrent-events model and the Fine-Gray competing-risks model. They keep the simple risk-set change that works for pure administrative censoring and layer IPCW on top for the random part, all while claiming consistency under minimal assumptions once the administrative times are known for everyone.\n\nWhat the work does well is target a setup that actually shows up in registry data, where the study end date is fixed for all but some subjects still drop out randomly. The abstract is clear that this particular combination for these two models had not been laid out before, so the contribution is the direct construction that avoids extra modeling of the joint censoring distribution.\n\nThe soft spot is the standing assumption that administrative censoring times are observed exactly even for subjects who are randomly censored first. The abstract conditions on this knowledge, and the stress-test note is right to flag it: if real data sets only record the administrative time when it is reached, the risk-set adjustment will not go through without additional work. The paper does not appear to add simulations or derivation steps in the abstract, so the full text needs to show that the estimating equations stay consistent under their stated conditions.\n\nThe approach reads as a direct extension of existing IPCW and risk-set ideas rather than anything circular. Citations follow the standard references for these models.\n\nThis is for analysts who work with competing risks or recurrent events in large bio-medical registries and need to deal with this exact censoring pattern. A reader who runs into mixed censoring will get a usable recipe. It deserves a serious referee because the practical problem is common and the proposed fix is straightforward enough to check in detail.","headline":"The paper shows how to blend risk-set adjustment and IPCW for Ghosh-Lin and Fine-Gray under mixed censoring, provided administrative times are known for every subject.","tokens_in":2355,"tokens_out":434,"would_cite":false,"duration_ms":27349,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.3","headline":"When administrative censoring times are known for all subjects, Ghosh-Lin and Fine-Gray models yield consistent estimates under mixed censoring by risk-set modification plus targeted inverse-probability weighting.","keywords":["recurrent events","competing risks","administrative censoring","random censoring","Ghosh-Lin model","Fine-Gray model","IPCW","marginal models"],"falsifier":"A simulation or real dataset in which the proposed estimators remain inconsistent after the administrative times are correctly inserted into the risk sets would falsify the claim.","tokens_in":2638,"feed_emoji":"","tokens_out":673,"duration_ms":25543,"temperature":0.7,"pith_summary":"The paper examines estimation for two marginal models commonly used in biomedical data: the Ghosh-Lin model for recurrent events ending in a terminal event and the Fine-Gray model for competing risks. It shows that the usual difficulty of random censoring can be isolated when administrative censoring times are recorded exactly for every individual. In that case the risk set is adjusted directly for the administrative part while inverse-probability-of-censoring weights are applied only to the random part, without needing a joint model for the combined censoring mechanism. Readers in registry studies or trials would care because this mixed censoring pattern is routine yet previously required stronger parametric assumptions on the entire censoring process.","feed_headline":"Ghosh-Lin and Fine-Gray models fixed under mixed censoring","feed_subtitle":"Known administrative times for every subject allow risk-set changes plus partial weighting to deliver consistent estimates without a full ce","key_machinery":"The hybrid adjustment that performs direct risk-set modification for known administrative censoring times while restricting inverse-probability-of-censoring weighting to the random-censoring component alone.","core_discovery":"For the Ghosh-Lin model for recurrent events with terminal events and the Fine-Gray model for competing events, when both administrative and random censoring are present and administrative censoring times are known for all subjects, correct estimation is obtained by modifying the risk-set for administrative censoring and using IPCW adjustments only for random censoring, relying on a minimum of modeling assumptions.","pith_inferences":["The same separation of censoring types may apply to other marginal survival models that currently rely on full IPCW.","In clinical-trial settings the method could be implemented by simply flagging the known administrative date for each participant and weighting only the observed random losses.","Empirical checks could compare the hybrid estimator against naive full-IPCW estimators when the proportion of random censoring varies."],"forward_implications":["Consistent parameter estimates become available for the Ghosh-Lin recurrent-event model under mixed censoring.","Consistent parameter estimates become available for the Fine-Gray competing-risks model under the same mixed censoring.","A full parametric model for the combined censoring distribution is no longer required.","The approach applies directly to large registry data where administrative end-of-study dates are recorded for the entire cohort."],"fun_headline_variants":["Ghosh-Lin and Fine-Gray adjusted for mixed censoring","Risk-set changes fix Ghosh-Lin and Fine-Gray models","Mixed censoring handled via risk sets in Ghosh-Lin Fine-Gray","Ghosh-Lin Fine-Gray consistent estimates with mixed censoring","Combining risk-set and IPCW for Ghosh-Lin and Fine-Gray models"],"cache_read_input_tokens":2112,"weakest_assumption_plain":"Administrative censoring times are known exactly for every subject.","fun_headline_variants_meta":{"raw":{"variants":["Ghosh-Lin and Fine-Gray adjusted for mixed censoring","Risk-set changes fix Ghosh-Lin and Fine-Gray models","Mixed censoring handled via risk sets in Ghosh-Lin Fine-Gray","Ghosh-Lin Fine-Gray consistent estimates with mixed censoring","Combining risk-set and IPCW for Ghosh-Lin and Fine-Gray models"]},"model":"grok-4.3","cost_usd":0.005995,"raw_usage":{"total_tokens":2839,"prompt_tokens":667,"num_sources_used":0,"completion_tokens":91,"cost_in_usd_ticks":59949500,"prompt_tokens_details":{"text_tokens":667,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":2081,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":667,"tokens_out":91,"duration_ms":16279,"temperature":1.0,"reasoning_tokens":2081,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-06-26T16:35:38.616891+00:00","model_set":{"reader":"grok-4.3"},"falsifier":"A simulation or real dataset in which the proposed estimators remain inconsistent after the administrative times are correctly inserted into the risk sets would falsify the claim.","supporting_citations":[],"review_version":1}