SAMPLING VARIABILITY OF AREA UNDER ROC-CURVE AND OF D'E

By computer simulation, random samples were drawn from underlying Gaussian, uniform, and negative exponential distributions which differed in terms of their mean and variance. The conditional probability of an observation of a given magnitude, or higher, was computed for each of 2 distributions. Resulting conditional probabilities for the 2 distributions trace out a receiver operating characteristic (ROC) curve. The area under the ROC-curve, which provides a distribution-free measure of performance, was calculated over a wide range of possible performances. It is shown that the sampling variability of the area measure is dependent upon its mean value and is relatively independent of the complex of conditions which lead to the given mean value. A similar conclusion is reached for the sampling variability of the intersection of the ROC-curve with the negative diagonal. This measure leads to an estimate of the sampling variability of the measure d'e, an index of O performance, often employed in the signal detectability approach to performance measurement. (PsycINFO Database Record (c) 2006 APA, all rights reserved). © 1969 American Psychological Association.