THE USE OF THE BINORMAL MODEL FOR PARAMETRIC ROC ANALYSIS OF QUANTITATIVE DIAGNOSTIC-TESTS

The binormal model is widely used for parametric receiver operating ch aracteristic (ROC) analyses of data concerning the accuracy of medical diagnostic tests. Empirical evaluation of the performance of this mod el in the face of departures from binormality has been limited to inte rpretations of radiology-type examinations recorded on a rating scale. This paper extends the investigation to the performance of the model with biochemical and other tests recorded on an interval scale. In ord er to describe non-binormal pairs of distributions, a useful standardi zed graphical display is developed; this display also illustrates seve ral features of ROC curves. We consider non-binormal pairs of distribu tions with or without a monotone likelihood ratio and show that by tra nsformation of the underlying scale, one can make many such pairs rese mble closely the binormal model. These findings justify Metz's use of the binormal model in the 'LABROC' software for ROC analyses of labora tory type data even when the raw data may 'look' decidedly non-Gaussia n.