Adjusting for Covariate Effects on Classification Accuracy Using the Covariate-Adjusted ROC Curve
Recent scientific and technological innovations have produced an abundance of potential markers which are being investigated for their use in disease screen- ing and diagnosis. In evaluating these markers, it is often necessary to account for covariates which are associated with the marker of interest. These covariates may include subject characteristics, expertise of the test operator, test proce- dures, or aspects of specimen handling. In this paper, we propose the AROC, a covariate-adjusted measure of the classification accuracy. The AROC is the common covariate-specific ROC curve, when the covariate does not affect dis- crimination, and a weighted average of covariate-specific ROC curves, when the covariate does affect discrimination. We propose non-parametric and semi- parametric estimators for the AROC, provide asymptotic distribution theory for these estimators, and investigate their finite sample performance. We illus- trate our methods using data from the Physicians’ Health Study. The AROC is used to characterize the age-adjusted discriminatory accuracy of prostate- specific antigen as a biomarker for prostate cancer.
Janes, Holly and Pepe, Margaret S., "Adjusting for Covariate Effects on Classification Accuracy Using the Covariate-Adjusted ROC Curve" (March 2006). UW Biostatistics Working Paper Series. Working Paper 283.