Cox Proportional Hazards — Regression Model for Time-to-Event Data
- Introduced by
- Sir David Roxbee Cox
- Year
- 1972
- Type
- Semi-parametric regression model
- Family
- Clinical / epidemiology
What does it do?
The Cox proportional hazards model is a semi-parametric regression method that estimates the effect of one or more covariates on the hazard — the instantaneous rate of an event such as death, relapse, or failure — while making no assumption about the shape of the baseline hazard function.
Introduced by David Cox in 1972, it is the dominant tool for multivariable survival analysis in clinical and epidemiological research.
The intuition behind the method
Suppose you want to know whether age and treatment type jointly affect how quickly patients reach remission. The Cox model answers this without requiring you to specify whether the underlying risk rises, falls, or stays flat over time. It works by ranking when events occur and estimating which covariate patterns are associated with earlier events. The 'proportional hazards' assumption means that the ratio of hazard between two individuals with different covariate values stays constant over time — if a treated patient is twice as likely to recover at month 3, they remain twice as likely at month 12. Each covariate produces a hazard ratio: an HR of 1.5 means 50 % higher instantaneous event rate per unit increase in that covariate.
History
David Cox published the proportional hazards model in a landmark 1972 paper in the Journal of the Royal Statistical Society, Series B. The model was revolutionary because it allowed covariate adjustment in survival analysis without specifying the baseline hazard distribution. Cox introduced the partial likelihood concept — which he formalised more completely in a 1975 follow-up — enabling estimation from the event ordering alone. The paper became one of the most cited in all of statistics and secured Cox a knighthood. Extensions including time-varying covariates, stratified baselines, and frailty (random effects) terms have since made the framework applicable to clustered, recurrent-event, and competing-risks settings.
When to use it
Use the Cox model when your outcome is time-to-event, censoring is present (not all participants experience the event), and you need to adjust for covariates. It is the default multivariable method for cohort studies, clinical trials, and registry analyses with survival endpoints. Do not use it when the proportional hazards assumption is severely violated across all covariates — consider an accelerated failure time model or flexible parametric alternatives instead. Avoid it when events are very rare relative to the number of covariates (fewer than 10 events per predictor is a recognised rule of thumb for instability). For a single exposure without covariates, Kaplan-Meier curves with log-rank tests are simpler and sufficient.
Limitations and considerations
- The proportional hazards assumption must hold — if hazard ratios change over time, estimates are averaged and potentially misleading.
- Does not produce absolute survival time predictions without additional estimation of the baseline hazard.
- Performance degrades when events are sparse relative to covariates; overfitting risk rises with fewer than 10 events per predictor.
- Censoring must be non-informative; if patients drop out because they are deteriorating, estimates are biased.
Source conditions do not replace independent scientific review or validation by an analysis engine.
Strengths
- Requires no assumption about the shape of the baseline hazard, making it robust across a wide range of survival time distributions.
- Handles right-censoring naturally, which is ubiquitous in clinical follow-up data.
- Produces interpretable hazard ratios that quantify the relative effect of each covariate on the instantaneous event rate.
- Easily extended to stratified analyses, time-varying covariates, and clustered data (frailty models).
- Widely implemented in all major statistical software (R, Stata, SAS, SPSS) with well-established diagnostic tools.
Applications
- Assessing the effect of treatment allocation on overall survival or progression-free survival in randomised clinical trials.
- Identifying prognostic factors for disease recurrence or mortality in cancer registry and hospital cohort studies.
- Estimating the association between occupational exposures and time to chronic disease onset in occupational epidemiology.
- Modelling time to hospital readmission after adjusting for age, comorbidities, and discharge status in health services research.
- Evaluating device or implant failure rates adjusted for patient-level risk factors in biomedical engineering studies.
How it works
How it works
Before modelling, verify the proportional hazards assumption for each covariate using Schoenfeld residual plots or a formal test (correlation of scaled residuals with time). Also check for near-zero event rates in covariate strata, extreme outliers in follow-up time, and the adequacy of the censoring mechanism — censoring must be non-informative for valid inference.
Write the hazard function as h(t|X) = h_0(t) * exp(b_1*X_1 + ... + b_p*X_p), where h_0(t) is the unspecified baseline hazard and the b_i are log-hazard coefficients to be estimated. Select covariates based on subject-matter theory or prior literature, not purely on univariable p-values; include known confounders regardless of significance.
Formula
h(t|X) = h_0(t) * exp(b_1*X_1 + ... + b_p*X_p)
Coefficients are estimated by maximising the partial likelihood, a function that conditions on the observed event order and cancels out the baseline hazard. Because the baseline hazard drops out, no distributional assumption about the survival time is required — hence 'semi-parametric'. The method handles tied event times via Breslow or Efron approximations.
Evaluate model fit with log-likelihood ratio, Wald, and score tests for overall significance. Check individual covariate influence with dfbeta residuals, and assess linearity of continuous predictors using martingale residuals. If the proportional hazards assumption is violated for a covariate, consider stratification by that variable or adding a time-interaction term.
Exponentiate each coefficient to obtain a hazard ratio: HR = exp(b_i). An HR > 1 indicates higher event rate (shorter survival); HR < 1 indicates lower event rate (longer survival). Report HRs with 95 % confidence intervals and two-sided p-values. The HR represents an instantaneous relative rate, not a risk ratio — communicate this distinction clearly to non-specialist audiences.
Formula
HR = exp(b_i)
Common pitfalls
- Ignoring the proportional hazards assumption — failing to test and report Schoenfeld residuals leads to undetected violations that distort hazard ratio estimates across the entire follow-up period.
- Treating the hazard ratio as a risk ratio — the HR is an instantaneous rate ratio, not a cumulative risk ratio; over-translating it into absolute risk language misleads clinical audiences.
- Over-fitting with too many covariates relative to events — entering all candidate variables without considering the events-per-variable ratio inflates type I errors and produces unreliable CIs.
- Ignoring tied event times — the default Breslow approximation is adequate for few ties, but substantial ties require the Efron approximation or an exact method.
Frequently asked questions
What is the proportional hazards assumption and how do I test it?
The assumption states that the hazard ratio between any two covariate patterns is constant over time. Test it by plotting scaled Schoenfeld residuals against time for each covariate — a flat line suggests the assumption holds. The cox.zph() function in R's survival package provides both the plot and a formal correlation test; a significant p-value indicates a violation that should be addressed by stratification or a time-interaction term.
When should I prefer a parametric survival model over the Cox model?
If subject-matter knowledge strongly supports a specific distribution for survival times (e.g., Weibull for device failure), a fully parametric model is more efficient and produces direct predictions of median or mean survival time. The Cox model is preferable when you are agnostic about the hazard shape or when model robustness matters more than parametric efficiency — which is most of the time in clinical research.
Can I use the Cox model with time-varying covariates?
Yes. The counting-process formulation of the Cox model accommodates covariates whose values change over the follow-up period (e.g., updated lab values, treatment switches). The dataset must be restructured into a start–stop interval format for each observation window. This is fully supported in R (survival package), Stata, and SAS.
How many events do I need?
A widely used heuristic is at least 10 events per predictor variable (EPV) in the model. Below this threshold, coefficient estimates become unstable, confidence intervals widen, and the model may not converge reliably. For models with many candidate covariates, penalised regression (ridge, LASSO) or formal shrinkage methods can improve stability when EPV is low.
What is the difference between the Cox model and competing risks analysis?
When only one type of event is possible, the standard Cox model applies. Competing risks arise when a subject can experience one of several mutually exclusive events (e.g., cancer-specific death vs. death from other causes). In that setting, cause-specific Cox models or the Fine-Gray subdistribution hazard model are appropriate; the standard Cox model on the event of interest alone overestimates cause-specific cumulative incidence.
Related methods
Sources
Cox, D. R. (1972). Regression models and life-tables. Journal of the Royal Statistical Society: Series B (Methodological), 34(2), 187–202.
Open DOI referenceCopied.Collett, D. (2015). Modelling Survival Data in Medical Research (3rd ed.). CRC Press.
ISBN: 978-1439856789Copied.