Cox Proportional Hazards Model

The model defines the hazard for an individual as the product of two distinct components:

• Baseline Hazard Function (h₀(t)): An unspecified, non-negative function of time. It represents the hazard when all covariates are zero. Because it is left unspecified, the Cox model is semiparametric—it makes no assumptions about the shape of the underlying risk over time.
• Partial Hazard (exponential component): A parametric function that models the multiplicative effect of covariates on the baseline hazard. It is expressed as exp(β₁X₁ + β₂X₂ + ... + βₚXₚ), where β coefficients represent the log-hazard ratios. This separation allows estimation of covariate effects without needing to model the often complex baseline risk.

The model's validity hinges on the assumption that the hazard ratio (HR) between any two individuals is constant over time. This means the effect of a covariate is multiplicative and does not change as time progresses.

• Interpretation: A hazard ratio of 2.0 for a treatment group means the risk of the event is double that of the control group at any given time point.
• Verification: This assumption is formally tested using Schoenfeld Residuals. A significant correlation between scaled Schoenfeld residuals and time indicates a violation.
• Remedy: If violated, strategies include using an extended Cox model with time-varying covariates or switching to an Accelerated Failure Time (AFT) model.

The regression coefficients (β) are estimated using maximum partial likelihood, a brilliant statistical innovation by Sir David Cox. This method eliminates the unknown baseline hazard h₀(t) from the estimation process.

• Mechanism: The likelihood is constructed by considering the risk set at each distinct event time—the set of individuals still at risk just before the event.
• Conditional Probability: It calculates the probability that the specific individual who experienced the event was the one to do so, given the risk set and their covariates.
• Advantage: This allows for valid inference on the β coefficients without ever specifying or estimating the shape of the baseline hazard function.

When multiple subjects experience the event at the exact same recorded time, the standard partial likelihood is undefined. Several approximation methods exist to handle these ties:

• Breslow Method: The default in many packages. It approximates the exact marginal likelihood and works well when ties are few.
• Efron Method: A more accurate approximation than Breslow, especially with moderate to heavy ties. It adjusts the risk set contributions for tied individuals.
• Exact Method: Computes the exact marginal likelihood by considering all possible orderings of the tied events. It is computationally intensive for large datasets with many ties but provides the most precise estimates.

The standard Cox model assumes a linear relationship between continuous covariates and the log-hazard. Martingale residuals are the primary diagnostic tool to assess this assumption.

• Purpose: They represent the difference between the observed event indicator and the model's predicted cumulative hazard.
• Application: A plot of martingale residuals against a continuous covariate reveals the correct functional form. A random scatter around zero supports linearity.
• Non-Linearity: A systematic curved pattern suggests the need for a transformation (e.g., log, square root) or the use of smoothing splines or fractional polynomials to correctly model the covariate's effect.

When a categorical predictor violates the proportional hazards assumption but is not the primary variable of interest, a stratified Cox model provides a robust alternative.

• Mechanism: The data is partitioned into distinct strata based on the non-proportional variable. A separate, unspecified baseline hazard function h₀ₖ(t) is estimated for each stratum.
• Key Constraint: The regression coefficients (β) for all other covariates are assumed to be identical across strata. The model estimates a common set of β coefficients, but allows the underlying risk to vary freely across the stratification variable.
• Use Case: Commonly used to adjust for site effects in multicenter clinical trials where baseline risk differs by hospital but the treatment effect is assumed constant.

The defining constraint of the Cox model: the hazard ratio between any two individuals is constant over time. Covariate effects do not change as the study progresses.

• Schoenfeld Residuals: The primary diagnostic tool to test this assumption. A non-zero slope over time indicates a violation.
• Grambsch-Therneau Test: A formal statistical test derived from scaled Schoenfeld residuals.
• Remedy: If violated, use an Extended Cox Model with time-varying coefficients or stratify by the offending variable.

The exponentiated coefficient exp(β) from the Cox model, representing the instantaneous relative risk of the event.

• Interpretation: A hazard ratio of 2.0 means the treatment group has twice the risk of the event at any moment compared to the control group.
• Binary Contrast: It is a relative measure; it does not provide the absolute probability of survival.
• Confidence Intervals: Always report the 95% CI to assess precision. A CI crossing 1.0 indicates a non-significant effect.

Standard Cox regression fails when predictor values change during observation (e.g., a patient's blood pressure measured monthly).

• Counting Process Format: Restructure data so each subject has multiple rows representing intervals (start, stop, event].
• Immortal Time Bias: A critical error where future treatment status is incorrectly coded as a baseline variable. Time-varying analysis prevents this.
• Landmark Analysis: A simpler alternative that resets the baseline at a specific time point for dynamic prediction.

When a patient can experience a different event that precludes the primary outcome (e.g., death from a car accident before cancer recurrence), the standard Cox model overestimates risk.

• Cause-Specific Hazard: Models the rate of the primary event in subjects still at risk.
• Fine-Gray Subdistribution Hazard: Models the Cumulative Incidence Function (CIF) directly, keeping competing event subjects in the risk set.
• CIF: The absolute probability of the primary event occurring by time t in the presence of competing risks.

Assessing how well a Cox model predicts outcomes requires specialized metrics for censored data.

• Concordance Index (C-Index): Measures discrimination. A value of 0.5 is random guessing; 1.0 is perfect ordering of risk. It answers: 'Are high-risk patients dying sooner?'
• Brier Score: Measures calibration and discrimination. It is the mean squared error between predicted survival probability and actual status at time t. Lower is better.
• Time-Dependent ROC: Evaluates sensitivity and specificity for event occurrence by a specific horizon.

When the linearity assumption of the Cox model is too restrictive, non-linear machine learning methods offer greater flexibility.

• DeepSurv: A deep feed-forward neural network that outputs a risk score optimized via the Cox partial likelihood. It captures complex interactions automatically.
• Random Survival Forests (RSF): An ensemble of survival trees that handles non-linear effects and high-order interactions without parametric assumptions.
• Lasso Cox Regression: Applies an L1 penalty to shrink coefficients to zero, performing automatic feature selection in high-dimensional genomic data.