Skip to main content

Cox Proportional-hazards Regression

Covariate effects on event rates without specifying the baseline hazard

Statistics
Advanced Modeling
Survival Analysis
Regression
The Cox model, partial likelihood, hazard-ratio interpretation, uncertainty, assumptions, diagnostics, and design-aware bioinformatics use.
Author

Jiahong Sun

Published

November 6, 2026

Modified

November 6, 2026

1 A Regression Model for Event Rates

For subject \(i\), let \(T_i\) be an event time, \(C_i\) a censoring time, \(Y_i=\min(T_i,C_i)\) the observed time, and \(\Delta_i=I(T_i\le C_i)\) the event indicator. Given a covariate vector \(x_i\), the Cox proportional-hazards model is

\[ \boxed{h(t\mid x_i)=h_0(t)\exp(x_i^\mathsf{T}\beta).} \]

The unspecified baseline hazard \(h_0(t)\) may change freely over time. The regression coefficients \(\beta\) multiply that common time pattern on the hazard scale. This makes the model semiparametric: the covariate effects are finite dimensional, but the baseline hazard is not assigned a parametric shape.

The observation process matters as much as the event process. Standard independent-censoring arguments require, at least conditionally on modeled covariates, that being censored at time \(t\) does not reveal the subject’s future event risk. Administrative study closure can plausibly satisfy this condition; dropout caused by worsening health often cannot.

2 What the Coefficients Mean

For two subjects whose covariates differ by \(d=x_a-x_b\),

\[ \frac{h(t\mid x_a)}{h(t\mid x_b)} =\exp(d^\mathsf{T}\beta). \]

The baseline hazard cancels, and the ratio does not depend on \(t\). For a one-unit difference in predictor \(x_j\), holding the other modeled predictors fixed, the hazard ratio is \(\exp(\beta_j)\). A hazard ratio of 2 means twice the instantaneous event rate among subjects still event-free at that time. It does not mean twice the event probability, half the median survival time, or a constant difference between survival curves.

Because

\[ H(t\mid x)=H_0(t)\exp(x^\mathsf{T}\beta), \qquad S(t\mid x)=S_0(t)^{\exp(x^\mathsf{T}\beta)}, \]

the same hazard ratio can correspond to very different absolute risks as \(H_0(t)\) changes. Report adjusted survival probabilities or risk differences at scientifically meaningful times when absolute prognosis matters.

Interpretation also depends on coding:

  • a coefficient for a continuous predictor is linear on the log-hazard scale unless splines or transformations are included;
  • a factor coefficient compares its level with the chosen reference level;
  • an interaction makes a main-effect hazard ratio conditional on the other interacting variable; and
  • a baseline-adjusted association is not automatically a causal effect.

3 Partial Likelihood

Suppose distinct event times are \(t_{(1)}<\cdots<t_{(D)}\), subject \(i(d)\) has the event at \(t_{(d)}\), and \(R(t_{(d)})\) is the risk set immediately before that time. Conditional on one member of the risk set failing then, the model assigns probability

\[ \Pr\{i(d)\text{ fails}\mid R(t_{(d)}),\text{ one failure}\} = \frac{\exp(x_{i(d)}^\mathsf{T}\beta)} {\sum_{j\in R(t_{(d)})}\exp(x_j^\mathsf{T}\beta)}. \]

Multiplying these conditional contributions gives Cox’s partial likelihood,

\[ \boxed{ L_p(\beta)=\prod_{d=1}^{D} \frac{\exp(x_{i(d)}^\mathsf{T}\beta)} {\sum_{j\in R(t_{(d)})}\exp(x_j^\mathsf{T}\beta)}.} \]

The nuisance baseline hazard cancels from each comparison. Maximizing \(L_p(\beta)\) yields the maximum partial-likelihood estimator \(\hat\beta\). Although the formula resembles a likelihood, its information comes from covariate contrasts within risk sets, not from treating observed event times as independent responses.

3.1 Score and information

Write

\[ \bar x(\beta,t)= \frac{\sum_{j\in R(t)}x_j\exp(x_j^\mathsf{T}\beta)} {\sum_{j\in R(t)}\exp(x_j^\mathsf{T}\beta)}. \]

The score is

\[ U(\beta)=\sum_{d=1}^{D} \{x_{i(d)}-\bar x(\beta,t_{(d)})\}. \]

Thus estimation balances the covariates of subjects who experienced events against risk-score-weighted covariates of everyone who could have experienced those events. With little overlap inside risk sets, coefficients can be unstable even when the original sample is large.

4 Ties, Baseline Survival, and Prediction

When several events share the same recorded time, the simple partial likelihood does not specify their internal ordering. Common implementations use the Efron or Breslow approximation; an exact partial likelihood is available for genuinely discrete event times. The choice can matter when ties are frequent, so the method and time resolution should be reported. R’s survival::coxph() uses Efron’s method by default.

After estimating \(\beta\), a Breslow-type estimator of cumulative baseline hazard is

\[ \widehat H_0(t)= \sum_{t_{(d)}\le t} \frac{d_d}{\sum_{j\in R(t_{(d)})}\exp(x_j^\mathsf{T}\hat\beta)}, \]

where \(d_d\) is the number of events at \(t_{(d)}\) under the corresponding tie convention. Predicted survival follows from

\[ \widehat S(t\mid x)= \exp\{-\widehat H_0(t)\exp(x^\mathsf{T}\hat\beta)\}. \]

Prediction therefore requires both the relative-risk model and an estimated baseline hazard. Calibration must be checked at relevant times and in the target population; a good ranking statistic alone is insufficient.

5 Estimand, Estimate, and Uncertainty

The scientific estimand should be stated before the model is fit. It may be a conditional hazard ratio, an adjusted survival probability at 24 months, or a contrast standardized over a target covariate distribution. These are not the same target.

Under correct model specification, independent censoring, identifiable coefficients, and regularity conditions, \(\hat\beta\) is approximately normal. Model-based standard errors come from the inverse observed partial-likelihood information. Wald, score, and partial-likelihood-ratio procedures then give large-sample tests or intervals. Their effective information is driven more by the number and distribution of events than by the number enrolled.

A sandwich covariance can account for clustered score contributions or certain forms of model misspecification. It does not repair a wrong mean structure, informative censoring, unmeasured confounding, or nonproportionality; it changes uncertainty for the coefficient target to which the working fit converges.

6 Model Structure and Assumptions

Assumption or choice Why it matters Assessment or sensitivity analysis
Proportional hazards A coefficient represents one time-constant log hazard ratio Scaled Schoenfeld residuals; time interactions; interval-specific effects
Correct covariate form Linearity on the log-hazard scale can distort effects Martingale residual plots; splines; prespecified transformations
Independent censoring conditional on covariates Biased risk sets bias coefficient and survival estimates Compare dropout patterns; inverse-probability-of-censoring sensitivity analyses
Independent units or modeled clustering Repeated events and related subjects invalidate naive uncertainty Identify subjects/families; cluster-robust variance or multilevel event model
Correct time origin and risk sets Delayed entry and immortal time change who is at risk Encode start–stop intervals; audit eligibility at each time
Adequate overlap and events Sparse risk-set contrasts cause instability Inspect events by covariate pattern; profile estimates; penalized sensitivity analysis
Competing-event handling matches the estimand Censoring a competing event changes interpretation State cause-specific versus cumulative-incidence target

See Proportional-hazards diagnostics for the residual definitions and a structured diagnostic workflow.

7 Extensions Change the Question

  • Stratified Cox model: each stratum has its own baseline hazard but shares selected coefficients. It handles a categorical variable whose baseline hazards need not be proportional, but it does not estimate that variable’s hazard ratio.
  • Time-varying coefficient: \(\beta_j(t)\) directly models a changing effect, such as \(x_jg(t)\). The resulting effect must be reported over time rather than collapsed into one number.
  • Time-dependent covariate: \(x_i(t)\) changes during follow-up. It must use information available just before time \(t\); coding a future response into the past creates immortal-time or reverse-causation bias.
  • Clustered or recurrent events: a counting-process representation and cluster-aware variance can describe repeated at-risk intervals, but the scientific estimand and dependence assumptions must remain explicit.
  • Penalized Cox model: ridge or lasso penalties can stabilize high-dimensional prediction, but tuning, selection, validation, and post-selection inference become part of the analysis.

8 Bioinformatics Example: Prognostic Gene Expression

Suppose pretreatment expression of a prespecified gene signature is measured in 180 patients, and the endpoint is progression-free survival. A defensible model might adjust for clinical stage and study center, use a spline for age, and estimate the signature hazard ratio per prespecified standard-deviation increase.

  • Experimental unit: the patient, not a sequencing read, gene, aliquot, or repeated scan.
  • Biological replication: independent patients sampled from the target clinical population.
  • Technical replication: lanes or aliquots improve measurement precision but do not increase the number of independent survival outcomes.
  • Design: define time zero, event, censoring, delayed entry, covariates, and analysis plan before inspecting outcome associations.
  • Statistical interpretation: the fitted coefficient is an adjusted conditional hazard ratio under the model.
  • Biological interpretation: association does not show that the signature is causal, druggable, or transportable to another assay or population.

If thousands of genes were screened to select the signature in these same 180 patients, the final Cox standard error ignores that search. Feature selection and all preprocessing must be repeated inside resampling, followed by external validation when the aim is prediction.

The following syntax is illustrative and deliberately non-executable in this site, whose publication workflow installs Quarto but not R packages:

library(survival)
library(splines)

fit <- coxph(
  Surv(entry_month, exit_month, progressed) ~
    signature_z + ns(age, df = 3) + stage + strata(center),
  data = patients,
  ties = "efron",
  x = TRUE
)

summary(fit)
cox.zph(fit)

9 Common Misinterpretations

  1. “Semiparametric” does not mean assumption-free. The baseline shape is unrestricted, but proportional hazards, covariate form, censoring, and design assumptions remain.
  2. A hazard ratio is conditional on survival to time \(t\). The composition of the risk set changes over time, so causal interpretations require additional assumptions and careful estimand definition.
  3. Adjustment is not automatic control of confounding. It depends on measured covariates, their timing, their functional form, and whether adjustment opens biasing pathways.
  4. A nonsignificant coefficient is not evidence of no relevant effect. Check its interval, time pattern, event count, and the scale on which scientific importance is judged.
  5. One fitted hazard ratio does not validate a prediction model. Absolute-risk calibration, discrimination, optimism, and transportability need separate evaluation.

10 Connections and Take-home Messages

The Cox model estimates relative event rates by comparing subjects within risk sets while leaving the baseline hazard unspecified. Its usefulness depends on a clear estimand, valid time origin and risk sets, conditionally independent censoring, adequate covariate form, proportional effects, and correct biological replication. Hazard ratios should be accompanied by uncertainty and, where scientifically relevant, absolute-risk summaries.

11 References

Back to top