Interpreting Cox Model Coefficients

From a Cox Coefficient to a Hazard Ratio

A Cox regression coefficient is measured on the log-hazard scale, so its numerical value is not itself a direct description of how much the hazard changes. Interpretation requires exponentiating the coefficient and specifying the covariate change being compared. That change might be one unit of a continuous predictor, a clinically meaningful contrast such as ten years of age, or membership in one category rather than a reference category.

The Cox proportional hazards model represents the hazard for subject \(i\), conditional on covariates \(x_i\), as

Here \(\lambda_0(t)\) is the baseline hazard, \(x_i\) is a vector of covariate values, and \(\beta\) is the vector of regression coefficients. The hazard has units of inverse time; a hazard ratio is unitless. The model assumes that covariates multiply the baseline hazard by an exponential factor, so a given covariate contrast has a hazard ratio that does not change with time.

For one continuous predictor \(x\), a change of \(\Delta x\) units has hazard ratio \(\exp(\Delta x\beta)\). In particular, the hazard ratio for a one-unit increase is \(\exp(\beta)\). For a binary predictor coded 1 for one group and 0 for the reference group, \(\exp(\beta)\) compares the coded group with the reference group, holding the other modeled covariates fixed.

Assumptions Behind the Interpretation

The coefficient interpretation is conditional on the Cox model being an adequate description of the data. Important assumptions include proportional hazards for the covariate effects being interpreted, an appropriate functional form for continuous predictors on the log-hazard scale, correct coding of categorical predictors, and a suitable treatment of censoring. In particular, censoring is assumed to be independent of the event time conditional on the modeled information, in the sense required for the analysis.

A proportional hazards coefficient describes a constant hazard ratio over time. If the actual association changes with time, one coefficient averages or summarizes a pattern that may not be constant; interpreting it as the hazard ratio at every time would then be misleading. For a continuous predictor, the usual linear term also assumes that a one-unit increase has the same effect on log hazard regardless of the predictor's starting value. These assumptions concern the model, not the arithmetic of exponentiation.

An interpretation must also name the comparison and units. If age is measured in years, \(\exp(\beta)\) is the hazard ratio per additional year. If the coefficient is instead for age in decades, it is the ratio per additional ten years. The numerical coefficient changes under rescaling, while the underlying comparison can remain the same.

Confidence Intervals on the Hazard-Ratio Scale

Suppose \(\hat\beta\) estimates one Cox coefficient and \(SE(\hat\beta)\) is its estimated standard error. For a large enough sample, a conventional Wald interval on the coefficient scale is \(\hat\beta\pm z_{1-\alpha/2}SE(\hat\beta)\). Since exponentiation is increasing, exponentiating both endpoints gives the corresponding interval for the hazard ratio:

For a 95% interval, \(z_{1-\alpha/2}\) is approximately 1.96. This interval describes uncertainty in the estimated hazard ratio under the model and the large-sample approximation. It is not a range containing 95% of individual hazard ratios.

For a contrast involving multiple coefficients, the uncertainty depends on their covariance. If the contrast is \(c^\top\beta\), its estimated standard error is \(\sqrt{c^\top\widehat{\mathrm{Var}}(\hat\beta)c}\). Using only the separate standard errors and ignoring covariance can produce an incorrect interval.

Contrasts, Scaling, and Covariance

A coefficient is easiest to interpret when the change represented by one unit is useful in context. Rescaling a continuous predictor changes the coefficient's numerical value but not fitted hazard ratios for the same underlying contrast. For example, expressing age in decades rather than years makes the coefficient correspond to a ten-year increase. This can make results easier to communicate and avoid very small coefficients.

For multiple covariates, a comparison can change more than one value at once. The log hazard ratio is then the sum of the relevant coefficient changes. Its standard error must account for covariance among coefficient estimates. This is especially important when interpreting a contrast not represented by a single coefficient.

Why a Hazard Ratio Is Not an Event Probability

A hazard ratio is a relative comparison of instantaneous event rates among subjects still event-free just before time \(t\). It does not by itself specify the probability that an event has occurred by a particular time. Absolute event probabilities depend on the baseline hazard, follow-up time, and covariate values. Therefore, a Cox coefficient cannot be read directly as a risk difference or a risk ratio.

Under proportional hazards, if \(S_0(t)\) is the survival probability at time \(t\) for the reference covariate values and \(\eta=x^\top\beta\) is a subject's linear predictor relative to those values, then

This relation shows why a hazard ratio requires a baseline survival level and a time point to translate into an absolute event probability. Estimating the baseline survival is a distinct step from interpreting coefficients.

Diagnostics, Alternatives, and Common Pitfalls

Before reporting an interpretation, check that the coefficient's direction matches the coding and that its unit matches the way the predictor was entered. Confirm which group is the reference, whether the contrast changes one or several predictors, and whether an interaction means that the effect depends on another covariate. Wide intervals may reflect limited information, sparse events, or correlated predictors; exponentiation does not remove that uncertainty.

A single constant hazard ratio may be inadequate if proportional hazards fails. Schoenfeld residuals are a standard tool for assessing that assumption. A continuous predictor may also need a nonlinear form on the log-hazard scale rather than a single linear coefficient. These issues affect whether a single coefficient has the simple interpretation given here; they cannot be corrected merely by expressing the result as a percentage.

The Cox model is often used for adjusted relative comparisons, whereas Kaplan–Meier estimates describe survival in observed groups without covariate adjustment. Neither summary automatically answers every absolute-risk question. Absolute survival or event probabilities require an estimated baseline survival and a specified covariate pattern. Odds ratios from logistic regression and risk ratios are different measures; they should not be substituted for a Cox hazard ratio.

• Wrong: “A hazard ratio of 1.5 means 50% more people will have the event.” Correction: It means an estimated 50% higher instantaneous hazard at each time under proportional hazards, not a 50% increase in event probability.
• Wrong: “A coefficient of \(-0.30\) means the hazard is 30% lower.” Correction: Exponentiate: \(\exp(-0.30)=0.741\), an estimated 25.9% lower hazard.
• Wrong: “The treatment coefficient compares treated and control subjects at any age, even if the contrast also changes age.” Correction: State the exact covariate contrast; include every changed predictor in the linear contrast.
• Wrong: “Exponentiating the endpoints is unnecessary because the log-scale interval is already a hazard-ratio interval.” Correction: The coefficient interval is on the log-hazard scale; exponentiate both endpoints to report a hazard-ratio interval.