18 Evaluating Confidence Interval Coverage in Correlated AUC Comparisons: DeLong vs. Bootstrap Methods

Oregon Convention Center 

Intro: Technological advances have bolstered biomarker discovery in personalized medicine. Evaluating biomarkers' ability to distinguish disease states is critical. DeLong's method, commonly used for comparing correlated AUCs (Area Under the Receiver Operating Curve) in diagnostic tests on the same subjects, has been reported to underestimate confidence interval coverage in small samples with high correlations between tests, as indicated by recent studies[1]. Methods: We compared DeLong's method with a Bootstrap normal approximation via simulations using logistic regression models under various conditions. Variations included sample sizes (20-200), case-control ratios (1:1 to 1:5), AUC levels (0.5-0.8), and test correlations (0-0.75). Results: Though results suggest poor coverage probability for both DeLong and Bootstrap normal approaches at small sample sizes, the Bootstrap approach consistently outperformed DeLong's method, especially at higher correlations. This pronounced improvement at higher correlations advocates the Bootstrap method as a superior alternative for AUC comparison in small samples with correlated biomarkers.

DeLong's test

bootstrap-resampling

Area Under the Curve

coverage probability

correlated AUC

small sample 

Section on Medical Devices and Diagnostics