In an extensive evaluation of clinical prediction methodology, researchers Peter C. Austin and Daniele Giardiello examined how violating the proportional hazards assumption impacts the discriminatory performance of the widely used Cox proportional hazards regression model. The study, published in Diagnostic and Prognostic Research, utilized a series of large-scale Monte Carlo simulation experiments to evaluate predictive discrimination and overall accuracy across diverse scenarios. The authors found that departures from proportional hazards directly altered the discriminatory capacity of the Cox regression model, demonstrating that the direction and magnitude of the time-dependent hazard ratio systematically govern whether model performance degrades or paradoxically improves relative to scenarios where model assumptions strictly hold.
Clinical prediction models based on survival analysis are routinely deployed across medical disciplines to estimate an individual patient’s absolute probability of experiencing a critical health event within a prespecified prediction horizon, such as mortality over a five-year or ten-year follow-up window. The Cox regression framework relates baseline patient characteristics and covariates to the hazard function, which represents the instantaneous rate of event occurrence conditional on survival up to that moment. A foundational requirement of standard Cox modeling is the proportional hazards assumption, which dictates that the relative hazard comparing any two individuals remains constant across time. When estimating individual survival probabilities, investigators frequently couple estimated regression coefficients with Breslow’s estimator of the cumulative baseline hazard function, yet the empirical consequences of assumption violations on predictive discrimination have historically received minimal formal inquiry.
Validating predictive instruments requires careful attention to both calibration and discrimination, two fundamental dimensions of model performance. Calibration assesses the numerical agreement between predicted probabilities and observed empirical event rates, whereas discrimination reflects the statistical ability of predicted risk estimates to systematically separate individuals who develop the event from those who remain event-free. Prior literature noted that proportional hazards violations generally exert negligible effects on model calibration, leading earlier methodological commentaries to hypothesize that standard Cox models would maintain robust discriminatory power even when non-proportionality is present. However, the simulation experiments conducted by Austin and Giardiello reveal that non-proportional hazards can meaningfully reshape discriminatory metrics, contradicting assumptions that discrimination remains uniformly preserved.
To rigorously quantify these dynamics, the investigators constructed thirty-four distinct Monte Carlo simulation scenarios using samples of two thousand subjects, simulating one thousand distinct datasets within each defined scenario. For each individual subject, the data-generating mechanism established two correlated baseline covariates consisting of a continuous predictor variable and a binary predictor variable derived from a underlying bivariate normal distribution. Simulated time-to-event outcomes were generated from a parametric hazards model incorporating time-varying hazard ratios over a maximal follow-up window of 1,825 days, representing five complete years. Non-proportional hazards were introduced by specifying a regression coefficient that changed linearly over time, allowing the parameter beta one to span seventeen discrete increments between negative 0.0008 and positive 0.0008, while setting average hazard ratios across the five-year duration to two.
The experimental setup systematically isolated two core analytical dimensions: whether the non-proportional relationship occurred on a continuous variable or on a binary variable, and the exact slope and direction of the hazard ratio’s temporal evolution. Within each generated dataset, researchers partitioned observations evenly into a derivation cohort and a validation cohort. In the derivation cohorts, investigators fitted an intentionally misspecified Cox model that omitted time-dependent covariates, directly mimicking typical observational analytical practice where proportional hazards violations are frequently ignored. In validation sets, investigators evaluated discrimination across one-, two-, three-, four-, and five-year horizons using time-varying area under the receiver operating characteristic curve, Harrell’s concordance index, Uno’s concordance statistic, and the time-varying index of prediction accuracy.
The simulation results demonstrated that when the log-hazard ratio was positively associated with time—meaning the covariate effect started lower and expanded as time progressed—both model discrimination and overall predictive accuracy declined relative to the baseline scenario where proportional hazards held perfectly. Conversely, when the log-hazard ratio displayed a negative association with time—characterized by a hazard ratio that was initially elevated and subsequently diminished across the follow-up horizon—discrimination and predictive accuracy metrics increased relative to the proportional hazards reference. In extreme scenarios involving a continuous predictor evaluated at a one-year prediction horizon, time-varying area under the ROC curve plummeted from approximately 0.70 under proportional hazards down to approximately 0.55 under strong positive non-proportionality, representing a shift from moderate discriminatory ability to near-random prediction.
Beyond standard metrics of concordance, the research tracked the index of prediction accuracy, an index equivalent to the scaled Brier score that integrates discrimination and calibration against a null prediction baseline. Under the strongest conditions of positive time association for continuous predictors, the index of prediction accuracy fell below zero at the one-year horizon. A negative index of prediction accuracy demonstrates that the fitted, misspecified survival model generated risk forecasts that were demonstrably less accurate than a trivial null model assigning uniform baseline probabilities to all subjects. Concordance measures such as Harrell’s and Uno’s c-statistics similarly reflected monotonic declines as beta one rose, showing that rank-based concordance is inherently sensitive to temporal departures from constant proportional hazards.
The authors also evaluated alternative survival modeling strategies to test whether flexible non-Cox frameworks could overcome performance degradations in the presence of proportional hazards violations. These comparator frameworks included parametric accelerated failure time models with generalized gamma error distributions, Royston and Parmar’s flexible spline-based survival models utilizing four degrees of freedom, and generalized linear models constructed on non-parametric pseudo-observations employing bounded logit and complementary log-log links. Additionally, for scenarios where non-proportionality involved the binary covariate, a stratified Cox regression allowing distinct baseline hazards was implemented. Across these comprehensive comparisons, neither accelerated failure time formulations, spline-based flexible models, nor pseudo-observation regression yielded superior discrimination or predictive accuracy over the standard misspecified Cox model when non-proportionality occurred.
A secondary sensitivity analysis evaluated whether changing the prevalence of the binary covariate altered the vulnerability of the Cox model to proportional hazards violations. Fixing the non-proportional slope at a moderate value of 0.0004, the authors varied binary covariate prevalence across ten steps ranging from 0.05 to 0.50. Across these variations, prevalence shifts produced minimal impact on Harrell’s concordance index and only minor to modest fluctuations on time-varying area under the ROC curve and the index of prediction accuracy, indicating that baseline categorical prevalence does not substantially modify the underlying mathematical influence of non-proportionality on model discrimination.
Discussing the practical reach and design constraints of their investigation, the authors highlighted that the findings stem from synthetic datasets generated via Weibull-based survival models and linear non-proportional hazards structures. While empirical patient registries present complex non-linear hazards and confounding patterns, empirical cohorts cannot provide a counterfactual reference to observe how identical data would behave had proportional hazards remained unviolated. Consequently, the simulation framework provides unique mathematical clarity, showing that while mild departures from proportional hazards maintain stable performance, substantial violations meaningfully alter predictive discrimination in survival analysis, reinforcing the necessity of evaluating temporal hazard trajectories when developing prognostic clinical instruments.
Subject of Research: Impact of proportional hazards assumption violations on Cox model discrimination
Article Title: The impact of violation of the proportional hazards assumption on the discrimination of the Cox proportional hazards model
Article References: Austin, P. C., & Giardiello, D. (2026). The impact of violation of the proportional hazards assumption on the discrimination of the Cox proportional hazards model. Diagnostic and Prognostic Research, 10(1), Article 7. https://doi.org/10.1186/s41512-026-00223-0
Image Credits: AI Generated
DOI: 10.1186/s41512-026-00223-0
Keywords: Cox proportional hazards, survival analysis, clinical prediction models, statistical discrimination, Monte Carlo simulation, proportional hazards assumption, biostatistics, receiver operating characteristic, impact, violation, proportional, hazards
News Source: Ophelia Keating. (November 7, 2026). Simulation Study Evaluates Cox Model Discrimination Under Proportional Hazards Violations. Scienmag.
