|
Activity Number:
|
283
|
|
Type:
|
Contributed
|
|
Date/Time:
|
Tuesday, July 31, 2007 : 10:30 AM to 12:20 PM
|
|
Sponsor:
|
Biometrics Section
|
| Abstract - #309373 |
|
Title:
|
ROC Curve of a Generalized Odds-Rate Model
|
|
Author(s):
|
Huining Kang*+ and Edward J. Bedrick and Susan R. Atlas and Cheryl L. Willman
|
|
Companies:
|
University of New Mexico and University of New Mexico and University of New Mexico and University of New Mexico
|
|
Address:
|
HSC DOIM MSC 10 5550, Albuquerque, NM, 87131,
|
|
Keywords:
|
ROC curve ; Proportional odds-rate model ; Pseudo maximum likelihood estimator
|
|
Abstract:
|
Many approaches to estimating a receiver operating characteristic (ROC) curve assume a binormal form for it. However, one caveat is that the binormal ROC curve is not concave when the slope parameter is not equal to one, which is unlikely for most practical problems. We propose a new functional form for the ROC curve derived from the generalized proportional odds-rate (GPO) model. While sharing some properties with the binormal ROC curve, the GPO ROC curve has an advantage of being concave everywhere. One of the parameters can be interpreted as the odds ratio of diseased versus non-diseased subjects, which makes the GPO ROC curve useful in analyzing the effect of covariates on diagnostic tests. We have developed a pseudo maximum likelihood estimator for estimating the GPO ROC curve and illustrate the method with a data set from a gene expression microarray study on acute leukemia.
|
