American Statistical Association New York City Metropolitan Area Chapter Memorial Sloan-Kettering Cancer Center Biostatistics Seminar

ROC curves are routinely used to express the discriminatory power of a marker in terms of sensitivity and specificity. Corresponding methods for positive and negative predictive values are less commonly used, possibly because they are functions of the prevalence. In this talk, I will discuss a method to decompose the predictive values as a product of two functions: one involving prevalence and the other involving the cumulative distribution of the marker conditional on the gold standard. This decomposition lends itself to a graphical summary and it can also be used to model the predictive values as a function of covariates. I will use examples from the literature as well as using MSK data to illustrate various uses of this decomposition.