ABSTRACT
Metz and Kronman introduced the bivariate binormal model around 1980 to analyze paired ROC datasets. Its software implementation (CORROC2) has been used in over 100 publications, but it is not well-documented in the archival literature. While not necessary to analyze MRMC datasets, it is the only known method, until a recent advance, to measure correlations at the latent decision variable level between paired interpretations. The correlations are needed to design a calibrated simulator – an example is given in Chapter 23. A calibrated simulator is essential to proper validation of any proposed analysis method. The bivariate sampling model and visualizing the bivariate probability density function are illustrated with R examples. Parameter estimation of the bivariate binormal model is described. Practical details of CORROC2 software are provided as well as how this Windows software can be run in an OS X environment. The details of the code are in online appendices. A recent advance, which replaces the binormal model with the contaminated binormal model, is described. It too yields the desired correlations but is robust with respect to degenerate datasets that cannot be fitted by CORROC2.
