Schoenfeld Residuals

A Schoenfeld residual is defined for each covariate at each observed event time. It is the difference between the observed covariate value for the individual who experienced the event and the expected value of that covariate, given the risk set at that time. Formally, for subject i with covariate vector X_i who fails at time t_i, the residual is r_i = X_i - E[X | R(t_i)], where the expectation is a weighted average over all individuals still at risk. These residuals are not defined for censored observations, only for actual event times.

The core purpose of Schoenfeld residuals is to test whether a covariate's effect is constant over time. If the proportional hazards assumption holds, the residuals should exhibit no systematic trend when plotted against time. A non-zero slope indicates a time-varying coefficient, violating the assumption. The formal Grambsch-Therneau test computes a correlation coefficient between the scaled residuals and a function of time (typically the Kaplan-Meier survival estimate). A significant p-value (< 0.05) provides strong evidence against the null hypothesis of proportional hazards.

Raw Schoenfeld residuals are difficult to interpret directly because their variance depends on the covariate distribution. Scaled Schoenfeld residuals multiply the raw residual vector by an estimate of the inverse of the Cox model's information matrix. This scaling produces residuals that are:

• Standardized: Easier to visualize and compare across covariates.
• Additive: The scaled residual for a subject approximates the change in the estimated coefficient if that subject were removed.
• Directly Interpretable: A smoothed plot of scaled residuals plus the original coefficient estimate reveals the functional form of the time-varying effect.

Visual inspection of Schoenfeld residuals is standard practice. A plot of scaled residuals vs. time with a loess smoother overlaid reveals the nature of any violation:

• Horizontal line: Supports proportional hazards.
• Upward/downward trend: Indicates a covariate effect that increases or decreases over time.
• Non-linear pattern: Suggests a complex time interaction. The cox.zph() function in R's survival package generates these plots automatically, displaying the smoother and a confidence band. Outliers in the plot may also identify influential observations that disproportionately affect the model fit.

When Schoenfeld residuals reveal a violation, several modeling strategies can address the non-proportionality:

• Stratification: Fit separate baseline hazards for levels of the offending categorical covariate, allowing the baseline hazard to differ while assuming proportional hazards for other covariates.
• Time-by-Covariate Interaction: Include an explicit interaction term between the covariate and a function of time, such as log(t) or a step function, in the Cox model.
• Accelerated Failure Time (AFT) Model: Switch to a parametric model that does not assume proportional hazards, such as Weibull or log-normal regression.
• Landmark Analysis: Analyze survival conditional on surviving to a specific time point, effectively allowing effects to vary across time windows.

Schoenfeld residuals are mathematically connected to martingale residuals, another key diagnostic in survival analysis. While martingale residuals assess the overall adequacy of the model's functional form for continuous covariates, Schoenfeld residuals specifically target the time-constancy of effects. A key distinction:

• Martingale residuals: Defined for every subject (censored and uncensored), range from -∞ to 1, and detect non-linearity in covariate effects.
• Schoenfeld residuals: Defined only at event times, sum to zero asymptotically, and detect time-varying coefficients. Together, they form a complementary diagnostic suite for Cox model validation.

The formal statistical hypothesis test that quantifies the visual patterns seen in Schoenfeld residual plots. It computes a correlation coefficient between scaled residuals and a function of time (often the Kaplan-Meier transform or log-time). A significant p-value (< 0.05) provides strong evidence that the proportional hazards assumption has been violated for that specific covariate. The test can be applied globally to the entire model or individually to each predictor, making it the definitive diagnostic companion to Schoenfeld residual visualization.

While Schoenfeld residuals test proportionality, Martingale residuals diagnose the correct functional form of continuous covariates. They are calculated as the difference between the observed event indicator and the model's predicted cumulative hazard. Plotting Martingale residuals against a continuous predictor reveals non-linearity:

• A random scatter around zero suggests a correct linear fit
• A curved pattern indicates the need for a transformation (e.g., log, quadratic) or spline
• Extreme negative values identify potential outliers with unexpectedly long survival times

Symmetrically transformed Martingale residuals designed to identify outliers and influential observations more effectively. Deviance residuals are centered around zero and approximately normally distributed when the model fits well. Observations with absolute deviance residuals exceeding 2 or 3 warrant investigation—they may represent patients whose survival experience is poorly predicted by the model. Unlike Schoenfeld residuals which focus on covariate effects over time, deviance residuals highlight individual subjects that the model systematically fails to characterize.

The foundation for assessing overall model goodness-of-fit. Cox-Snell residuals are defined as the estimated cumulative hazard for each subject at their event or censoring time. If the model is correctly specified, these residuals should follow a unit exponential distribution with a constant hazard of 1. A plot of the Nelson-Aalen estimator of the residuals against the residuals themselves should approximate a 45-degree line through the origin. Systematic departures indicate global model misspecification that Schoenfeld residual analysis alone may miss.