Cox Regression

Cox regression, also known as the Cox proportional hazards model, is a semiparametric statistical method used in survival analysis to examine the relationship between one or more predictor variables and the time to an event of interest (such as death, disease recurrence, or equipment failure). Proposed by Sir David Cox in 1972, the model estimates hazard ratios that compare the instantaneous rate of the event (the hazard) at any given time, conditional on the predictor variables, without requiring specification of the underlying hazard function. This semiparametric property, making no assumptions about the shape of the baseline hazard, gives the Cox model its flexibility and widespread appeal.

Survival analysis is central to clinical research, epidemiology, and engineering, and the Cox model is a standard method for analyzing time to event data with multiple covariates. It allows researchers to assess whether factors such as treatment assignment, age, disease stage, or biomarker levels affect the rate at which an event occurs while accounting for censoring (participants who are lost to follow up or whose event has not occurred by the study's end). The hazard ratio produced by Cox regression is one of the most frequently reported effect measures in clinical trials and observational studies, and understanding its estimation and interpretation is essential for both producers and consumers of health research.

The Cox model expresses the hazard (instantaneous event rate) at time t for an individual as the product of a baseline hazard function (common to all individuals) and an exponential function of the linear predictor (the weighted sum of covariates). Because the baseline hazard cancels out in the partial likelihood function used for estimation, it does not need to be specified; only the regression coefficients (and their exponentiated forms, the hazard ratios) are estimated. A hazard ratio greater than 1 indicates that the covariate is associated with an increased event rate; a hazard ratio less than 1 indicates a decreased event rate; and a hazard ratio of 1 indicates no association. The proportional hazards assumption, that the hazard ratio is constant over time, should be checked using Schoenfeld residuals, log log survival plots, or interaction terms with time. When the assumption is violated, stratified models, time varying coefficients, or alternative models (such as the accelerated failure time model) may be used.

Use Cox regression when the outcome is time to an event and you need to adjust for multiple covariates. It is the standard method for estimating the treatment hazard ratio in randomized controlled trials with survival endpoints, often alongside a log rank test, for prognostic factor studies in oncology and cardiovascular research, and for analyses of registry and claims data where event timing is important. Cox regression handles right censoring naturally and can accommodate time varying covariates, multiple events, and competing risks with appropriate extensions.

A critical mistake is ignoring the proportional hazards assumption. If hazard ratios change over time, a single summary hazard ratio is misleading. Another common error is treating censored observations as events or excluding them from the analysis, both of which introduce bias. Including too many covariates relative to the number of events leads to overfitting. A common rule of thumb is at least 10 events per candidate predictor, but it is only a rule of thumb: Riley and colleagues (2020) show that the sample size needed to develop a prediction model depends on the number of candidate predictor parameters, how common the outcome is and the expected model fit. Researchers sometimes confuse hazard ratios with relative risks or odds ratios; while related, they measure different things and apply to different time frames. Collinearity among covariates can destabilize coefficient estimates, and researchers should check variance inflation factors before interpreting results. Finally, Cox regression models the rate of events, not the cumulative probability; survival probabilities require additional estimation steps (such as the Breslow estimator of the baseline cumulative hazard, from which adjusted survival curves are derived).

Oncologists use Cox regression to analyze data from a phase III trial comparing a new immunotherapy with standard chemotherapy in 600 patients with advanced non small cell lung cancer. The primary endpoint is overall survival. The model adjusts for age, performance status, tumor stage, and PD L1 expression level. The hazard ratio for immunotherapy versus chemotherapy is 0.72 (95% confidence interval: 0.58 to 0.89), indicating a 28% reduction in the rate of death at any given time. Schoenfeld residual tests show no evidence that the proportional hazards assumption is violated for any covariate. A forest plot of subgroup hazard ratios shows consistent benefit across age groups and tumor stages, strengthening the case for the new therapy.