Adjusting for Covariate Effects on Classification Accuracy Using the Covariate-Adjusted ROC Curve

• Activity Number: 504
• Type: Contributed
• Date/Time: Thursday, August 2, 2007 : 8:30 AM to 10:20 AM
• Sponsor: Biometrics Section
• Author(s): Holly Janes*+ and Margaret Pepe
• Companies: Johns Hopkins University and Fred Hutchinson Cancer Research Center
• Keywords: classification accuracy ; sensitivity ; specificity ; ROC curve ; covariate adjustment

Abstract

In evaluating markers for disease screening or diagnosis, it is often necessary to account for factors associated with the marker, such as subject characteristics or attributes of the test operator or test procedure. Analogous to approaches taken in therapeutic and etiologic studies, we adjust for covariates by characterizing the performance of the marker among subjects with fixed covariate value. We demonstrate that covariate adjustment is fundamentally different from other roles for covariates in marker evaluation. We propose the adjusted ROC curve (AROC), which is the common covariate-specific ROC curve, when covariates do not affect discrimination, and a weighted average of covariate-specific ROC curves more generally. The AROC is used to describe the age-adjusted discriminatory accuracy of prostate-specific antigen as a marker for prostate cancer in the Physicians' Health Study.