Abstract
The area under the receiver operating characteristic (ROC) curve is often used to summarize and compare the discriminatory accuracy of a diagnostic test or modality, and to evaluate the predictive power of statistical models for binary outcomes. Parametric maximum likelihood methods for fitting of the ROC curve provide direct estimates of the area under the ROC curve and its variance. Nonparametric methods, on the other hand, provide estimates of the area under the ROC curve, but do not directly estimate its variance. Three algorithms for computing the variance for the area under the nonparametric ROC curve are commonly used, although ambiguity exists about their behavior under diverse study conditions. Using simulated data, we found similar asymptotic performance between these algorithms when the diagnostic test produces results on a continuous scale, but found notable differences in small samples, and when the diagnostic test yields results on a discrete diagnostic scale.