12/4/2015

Post date: Dec 17, 2015 10:59:45 PM

Title: A placement value based approach to correlated and concave ROC curves with order constraints

Speaker: Dr. Zhen Chen, Biostatistics Branch, NICHD, NIH

[Abstract] In many diagnostic accuracy studies with multiple correlated tests, it is often the case that the tests are ordered in their performance a priori. In these situations, it is important to incorporate such a priori constraint in estimation of ROC curves, as this can improve statistical efficiency. Further, it is often desirable to consider ROC curves that are concave, since concavity is consistent with optimal decision principles. Motivated by these considerations, we propose a new approach to estimating multiple correlated ROC curves using placement values. The concavity constraint is constructed by recasting the placement value as a product of a uniform and another arbitrary random variable, while the ordering constraint is achieved through product measures in the context of mixture distributions. To allow flexible distributions for the test scores and for the derived placement values, we use Dirichlet process mixture priors. We demonstrate the good performance of the proposed approach through extensive simulation studies and illustrate its application by analyzing data from the Physician Reliability Study that investigated diagnosis of endometriosis using different combinations of clinical information.