Statistical approaches to the analysis of receiver operating characteristic (ROC) curves

In this article we review published and some unpublished work in statistical analyses of ROC curves. We describe both single and joint indices and indicate the approaches that have been taken to consider between-reader variations and correlations, within-reader variations, and variations and correlations between cases. We then discuss in detail a single index, the TP ratio at a fixed FP ratio (designated TPFP), or the FP ratio at a fixed TP ratio (designated FPTP). We show how to calculate confidence limits around any point on the curve; we further show, using the conventional Dorfman and Alf program and the jackknifing technique, how to calculate these confidence limits for multiple curves derived from the same sample of patients.