Bayesian Statistics Seminar
North Carolina State University
presents
Dr. Jiezhun (Sherry) Gu
Duke Clinical Research Institute
"Bayesian ROC curve estimation with partial gold standard"
ABSTRACT
Receiver Operating Characteristic (ROC) curve has been widely used in medical science for its ability to measure the accuracy of diagnostic or prognostic tests under the gold standard. However, in complicated medical practice, the gold standard test could be invasive, expensive, or even unavailable. Complete lack of availability of gold standard will result in non-identifiability in the classical sense, so any inference will be prior driven. In most publications, gold standard is observed for some subjects, but not for all. Under such a partial gold standard scenario, we develop a method based on a binormal model which takes the cases without gold standard into account in ROC curve estimation. The binormality assumption allows us to identify the parameters of the distribution of diagnostic measurements for non-diseased and diseased populations up to some monotone transformation H. We propose an indirect approach to treat H as latent function and
construct estimates using a rank-based likelihood. This method is implemented using MCMC techniques.
This is a joint work with Subhashis Ghosal, NC State University
Thursday, April, 29, 2010
4:00 - 5:00 pm
1216 Sas Hall