Sampling variability of nonparametric estimates of the areas under receiver operating characteristic curves: an update.

RATIONALE AND OBJECTIVES

Several methods have been proposed for calculating the variances and covariances of nonparametric estimates of the area under receiver operating characteristic curves (AUC). The authors provide an explanation of the relationships between them and illustrate the factors that determine sampling variability.

METHODS

The authors investigated the algebraic links between two methods, that of "placements" and that of "pseudovalues" based on jackknifing. They also performed a numerical investigation of the comparative performance of the two methods.

RESULTS

The "placement" method has a simple structure that illustrates the determinants of the sampling variability and does not require specialized software. The authors show that the pseudovalues used in the jackknife method are directly linked to the placement values.

CONCLUSION

Because of the close link, borne out in a numeric investigation of the sampling variation, and because of the ease of computation, the choice between the two methods can be based on users' preferences. For indexes other than the AUC, however, the use of pseudovalues holds greater promise.