{"id":"5e785209-d53f-492c-aafd-a988acaa73f7","arxiv_id":"2606.19982","paper_version":1,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":3.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":0,"one_line_summary":"Simulations demonstrate built-in selection bias in Cox PH hazard ratios from omitted covariates, with comparisons to frailty models, AFT models, and nonparametric survival differences.","lead":"This preprint uses simulations to show that Cox proportional hazards models produce biased hazard ratio estimates for treatment effects when important covariates are omitted, even in randomized trials due to non-collapsibility. Smart generalists might read it to see why common survival metrics can mislead and what alternative measures like survival differences or frailty models are suggested.","discovery_kind":"extension","skeptic_critique":{"model":"grok-4.3","headline":"No significant objection identified","rationale":"Reader correctly flags the independence premise as central, but that premise is the deliberate design choice that makes the non-collapsibility demonstration valid rather than a vulnerability. The claim matches textbook results on HR non-collapsibility; simulations and real-data illustration are the natural next verification steps but do not alter the logical structure of the argument itself.","tokens_in":1783,"tokens_out":272,"duration_ms":17930,"concrete_test":"Implement a minimal gamma-frailty simulation (variance=1, treatment effect log(HR)=0.5, n=500 per arm) and fit both Cox and AFT models; confirm Cox marginal HR attenuates toward 1 while AFT log-time ratio recovers the conditional effect and KM survival difference at t=median remains consistent.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim—that Cox PH hazard ratios exhibit built-in attenuation due to non-collapsibility when important covariates are omitted, even under baseline independence from treatment—is a standard, theoretically derived result in survival analysis. The independence condition isolates the non-collapsibility mechanism exactly as required; the described alternatives (frailty HR, AFT coefficients, nonparametric survival differences) are established remedies. No internal inconsistency or unsupported premise appears in the stated argument.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.3","summary":"The paper claims that hazard ratios from the Cox proportional hazards model exhibit built-in selection bias due to non-collapsibility when important covariates are omitted, even if those covariates are independent of treatment at baseline (as holds in RCTs). It reviews relevant literature on unobserved heterogeneity, conducts simulations to quantify bias in the semi-parametric Cox PH and parametric PH models across scenarios, compares these to alternatives including frailty-model HRs, AFT regression coefficients, and nonparametric survival differences (Kaplan-Meier or Cox with time-dependent effects), and demonstrates the alternatives on the RTOG 9202 randomized trial data.","tokens_in":1856,"tokens_out":430,"duration_ms":30729,"significance":"If the simulations confirm substantial and systematic bias under the stated conditions, the work would usefully consolidate known theoretical results on non-collapsibility, supply concrete numerical evidence of its practical magnitude, and evaluate established remedies, thereby informing reporting practices for treatment effects in survival analyses of randomized trials.","major_comments":[{"comment":"Abstract: the simulation plan and alternatives are described but no quantitative results, model specifications, or data-generating details are supplied, preventing assessment of whether the evidence supports the central claim of built-in bias.","section":"Abstract"},{"comment":"Simulation study: without explicit statements of the data-generating process (baseline hazard, distribution and effect sizes of omitted covariates, censoring mechanism, sample sizes, and number of replications), it is impossible to verify that the design isolates non-collapsibility from other sources of bias.","section":"Simulation study"}],"minor_comments":[{"comment":"The abstract would benefit from one or two key quantitative findings (e.g., range of bias observed) to convey the practical importance of the results.","section":"Abstract"},{"comment":"Notation for the frailty distribution and the precise definition of the AFT acceleration factor should be stated explicitly when first introduced.","section":"Methods"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":{"model":"grok-4.3","summary":"We thank the referee for their constructive comments, which highlight opportunities to improve the clarity and reproducibility of our manuscript. We address each major comment below and will incorporate revisions to strengthen the presentation of our simulation evidence on non-collapsibility in proportional hazards models.","responses":[{"response":"We agree that the abstract would be strengthened by including key quantitative results and a concise statement of the simulation design. In the revised manuscript, we will add specific findings (e.g., the percentage attenuation in the Cox HR under moderate omitted covariate effects) along with brief mentions of the baseline hazard form, omitted covariate distribution, and sample size. This will allow readers to immediately gauge the magnitude of the reported bias while remaining within abstract length constraints.","revision_made":"yes","referee_comment":"[Abstract] Abstract: the simulation plan and alternatives are described but no quantitative results, model specifications, or data-generating details are supplied, preventing assessment of whether the evidence supports the central claim of built-in bias."},{"response":"We acknowledge the need for greater explicitness. Although the simulation methods are described in the main text, we will add a new summary table (or expanded subsection) that lists every parameter: baseline hazard (Weibull shape/scale), omitted covariate distribution and coefficients, treatment effect size, independent censoring mechanism and rate, sample sizes per arm, and number of Monte Carlo replications. This will make transparent that the design holds treatment independent of the omitted covariate at baseline, thereby isolating the non-collapsibility mechanism from confounding or other biases.","revision_made":"yes","referee_comment":"[Simulation study] Simulation study: without explicit statements of the data-generating process (baseline hazard, distribution and effect sizes of omitted covariates, censoring mechanism, sample sizes, and number of replications), it is impossible to verify that the design isolates non-collapsibility from other sources of bias."}],"tokens_in":1390,"tokens_out":413,"duration_ms":17354,"standing_objections":[]},"desk_editor":{"model":"grok-4.3","letter":"The core takeaway is that Cox hazard ratios can show built-in attenuation from omitted covariates that affect the outcome, even when those covariates are independent of treatment at baseline. The authors simulate this under varying degrees of heterogeneity, compare the usual Cox and parametric PH fits against frailty models, AFT coefficients, and nonparametric survival differences, and then show the numbers on the RTOG 9202 trial data.\n\nWhat the paper does well is give practitioners a side-by-side look at how large the bias can get in realistic scenarios and how the alternatives behave on the same data. The real-trial illustration is straightforward and helps translate the simulation results.\n\nThe limitation is that the central mechanism is not new; the literature already derives the non-collapsibility result, so the contribution is mainly the simulation grid and the head-to-head comparison. If the simulations stick to standard settings without heavy censoring or strong time dependence, the added value stays modest. The alternatives they highlight are also familiar, so the paper functions more as a synthesis than a fresh theoretical step.\n\nThis is aimed at applied statisticians and trial analysts who routinely report Cox HRs and want concrete numbers on when the bias matters. Readers already comfortable with frailty or AFT models will not learn much new. It still deserves a serious referee because the practical stakes in clinical reporting are real and the simulations can be checked for robustness.","headline":"The paper runs simulations to quantify how much omitted covariates bias Cox HRs even in RCTs and compares a few standard fixes, but the underlying non-collapsibility point is already established.","tokens_in":2323,"tokens_out":359,"would_cite":false,"duration_ms":18321,"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":"Hazard ratios from Cox models are biased by omitted covariates even in randomized trials.","keywords":["hazard ratio","proportional hazards","non-collapsibility","omitted covariates","selection bias","frailty models","accelerated failure time","randomized trials"],"falsifier":"A simulation or dataset in which all relevant covariates are measured and the estimated hazard ratio equals the known marginal effect without adjustment for frailty would falsify the built-in bias claim.","tokens_in":2696,"feed_emoji":"","tokens_out":546,"duration_ms":27069,"temperature":0.7,"pith_summary":"The paper shows that hazard ratios from proportional hazards models are non-collapsible, producing selection bias when important covariates are omitted from the analysis. This bias occurs even when the omitted covariates are independent of treatment assignment at baseline, as holds in randomized trials. Simulations quantify the bias magnitude in semi-parametric Cox models and parametric proportional hazards models under varying unmeasured heterogeneity. Alternative approaches such as frailty models, accelerated failure time models, and direct estimates of survival differences are compared for robustness. The methods are applied to data from a randomized radiation therapy trial to show practical differences.","feed_headline":"Omitted covariates bias Cox hazard ratios even in RCTs","feed_subtitle":"Non-collapsibility of the hazard ratio induces selection bias from unmeasured factors balanced at baseline.","key_machinery":"Non-collapsibility of the hazard ratio due to its conditioning on survival up to each time point.","core_discovery":"Hazard ratios derived from the Cox proportional hazards model are subject to built-in selection bias in the presence of unmeasured heterogeneity arising from omitted important covariates, even when these covariates are independent of the main exposure at baseline.","pith_inferences":["Researchers may need to report survival probabilities rather than hazard ratios to obtain collapsible effect measures in trials with potential unmeasured factors.","Meta-analyses that pool hazard ratios from different studies could systematically distort summary effects if covariate sets differ across trials.","The mechanism implies that adding measured covariates after randomization can change the hazard ratio estimate in ways not explained by confounding alone."],"forward_implications":["The estimated treatment hazard ratio will not equal the true marginal effect when covariates are omitted.","Frailty models recover an adjusted hazard ratio that accounts for unobserved heterogeneity.","Accelerated failure time model parameters remain collapsible and unaffected by the same selection mechanism.","Nonparametric survival differences or time-dependent effect models provide treatment effect measures free of the non-collapsibility bias."],"fun_headline_variants":["Cox HRs biased by omitted covariates even in RCTs","Unmeasured factors cause selection bias in Cox models","Built-in bias affects Cox HR estimates despite randomization","Non-collapsibility induces HR bias from missing covariates"],"cache_read_input_tokens":2112,"weakest_assumption_plain":"Omitted covariates affect the outcome hazard while remaining independent of treatment assignment at baseline.","fun_headline_variants_meta":{"raw":{"variants":["Cox HRs biased by omitted covariates even in RCTs","Unmeasured factors cause selection bias in Cox models","Built-in bias affects Cox HR estimates despite randomization","Non-collapsibility induces HR bias from missing covariates"]},"model":"grok-4.3","cost_usd":0.003694,"raw_usage":{"total_tokens":1929,"prompt_tokens":688,"num_sources_used":0,"completion_tokens":59,"cost_in_usd_ticks":36937000,"prompt_tokens_details":{"text_tokens":688,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":1182,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":688,"tokens_out":59,"duration_ms":11867,"temperature":1.0,"reasoning_tokens":1182,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-06-26T16:32:34.074537+00:00","model_set":{"reader":"grok-4.3"},"falsifier":"A simulation or dataset in which all relevant covariates are measured and the estimated hazard ratio equals the known marginal effect without adjustment for frailty would falsify the built-in bias claim.","supporting_citations":[],"review_version":1}