Comparative assessment of three common algorithms for estimating the variance of the area under the nonparametric receiver operating characteristic curve
Mario A. Cleves, Ph.D.
Arkansas Center for Birth Defects Research and Prevention
Department of Pediatrics, University of Arkansas for Medical Sciences
Little Rock, Arkansas
|
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.
View all articles by this author:
Mario A. Cleves, Ph.D.
View all articles with these keywords:
receiver operating characteristic (ROC) curve, trapezoidal rule, sensitivity, specificity, discriminatory accuracy, predictive power
Download citation: BibTeX RIS
Download citation and abstract: BibTeX RIS
|