Cox Proportional Hazards & Diagnostic Accuracy (ROC-AUC)
Multivariable survival adjustment, Schoenfeld residual checks, and evaluating diagnostic biomarkers with Sensitivity, Specificity, and Youden's Index.
In this lesson you will learn
- 1Verify the Proportional Hazards (PH) assumption in Cox regression visually (non-crossing curves) and formally (Schoenfeld residuals p > .05).
- 2Compute and interpret Sensitivity, Specificity, PPV, NPV, and Youden's Index (J = Sens + Spec − 1) for a diagnostic test.
- 3Interpret the Area Under the ROC Curve (AUC) from 0.50 (chance coin flip) to 1.00 (perfect discrimination).
On this page▼
Cox Proportional Hazards & ROC Diagnostic Thresholds
This section covers checking the Proportional Hazards assumption via Schoenfeld residuals and selecting optimal biomarker cutoffs via ROC analysis and Youden's J.
Reporting a single Cox Hazard Ratio when survival curves cross over each other hides the fact that a treatment has early hazard (e.g. surgical mortality) followed by late benefit.
1. Cox Proportional Hazards (1972): Models the hazard rate h(t) = h₀(t) × exp(β₁X₁ + β₂X₂ + …). Its key assumption is that the Hazard Ratio between any two covariate profiles stays constant across time. Check this via:
- Visual inspection: Kaplan-Meier curves should not cross, and
log(−log S(t))curves should remain parallel. - Formal test: Schoenfeld Residuals test (
p > .05confirms proportional hazards). When violated, use stratified Cox regression, time-varying covariates, or Restricted Mean Survival Time (RMST).
2. Diagnostic Accuracy & ROC Curves: When evaluating a continuous biomarker against a gold-standard diagnosis, shifting the cutoff trades off Sensitivity (TP / [TP + FN]) against Specificity (TN / [TN + FP]). Plotting Sensitivity against 1 − Specificity across all cutoffs produces the ROC curve, and Youden's Index (J = Sensitivity + Specificity − 1) identifies the threshold that maximises balanced classification.
Diagnostic Accuracy & ROC Threshold Lab (Sensitivity vs Specificity)
Explore how shifting the biomarker cutoff trades off Sensitivity against Specificity (100 diseased, 200 healthy)
Clinical Example
A cardiology diagnostic study evaluates high-sensitivity Troponin I against coronary angiography in 300 chest-pain patients (100 confirmed MI, 200 non-MI). At the Youden-optimal cutoff, TP = 86, FN = 14, TN = 176, FP = 24.
Interpretation: Sensitivity (86.0%) and Specificity (88.0%) are intrinsic properties of the assay at that cutoff, whereas PPV (78.2%) and NPV (92.6%) depend on disease prevalence in the tested population.
Common Pitfall
Interpreting an ROC-AUC of 0.52 (95% CI [0.45, 0.59]) as '52% accurate', without realising that AUC = 0.50 is the diagonal line of a random coin toss.
Included with Any Analysez AI Study Package
This deep interactive chapter (Cox Proportional Hazards & Diagnostic Accuracy (ROC-AUC)) and its clinical simulation labs are unlocked for everyone who has bought any Analysez AI study package.