ABSTRACT
In the previous chapter we considered some inferential aspects associated with single ROC curves, but there we focussed exclusively on estimation. It was assumed that empirically gathered data sets constitute samples from two populations (P and N), that for a given classifier there exists a “true” ROC curve representing the scores of individuals from these populations, and that the sample data will be used to estimate either this ROC curve itself or some summary characterization of it such as AUC or PAUC. A variety of assumptions about the population scores can be made, and hence a variety of different estimators and estimation methods were reviewed. As well as finding estimates of the relevant parameters it is important to quantify the variability of these estimates from sample to sample, so this led to consideration of confidence intervals for the various parameters.
