ROC analysis: comparison between the binormal and the Neyman-Pearson model

ROC analysis, a technique widely used to evaluate performance of human observers and medical diagnostic equipment is based on a binormal model, which leads to unrealistic results under many conditions. This paper develops a binormal Neyman-Pearson (BNP) formulation for ROC curves. Comparison shows significant superiority of the BNP model: the results are more accurate, more reliable and applicable to various relationships among the populations' parameters. The formulation is well suited to the statistical estimation of ROC parameters from categorical data from computer and human observer studies.