I have used Robert Centors ROC analyzer for
calculating the non-parametric ROC area of even binary
diagnostic values.
The ROC area is useful when comparing the
discriminating power of diagnostic variables
independent of the incidence of the disease, even for
binary variables.
I think this reference can be of interest:
The Area under an ROC Curve with Limited Information
Wilbert B. van den Hout
Another reference, which explains why the ROC area is a good measure of
predictive power for general continuous and discrete predictor variables,
is my own Stata Journal article (Newson 2002). A pre-publication draft of
this can be downloaded from my website (see my signature below). The
article contains an example of calculating confidence limits in Stata for
the difference between 2 ROC areas for 2 different "continuous" predictors
and the same binary disease outcome. The method used there will work
equally well for binary and other "non-continuous" predictors.
