ROC Regression Using Percentile Values for Event Time Outcomes

• Activity Number: 189
• Type: Invited
• Date/Time: Monday, August 2, 2010 : 10:30 AM to 12:20 PM
• Sponsor: Section on Statistics in Epidemiology
• Author(s): Yuying Jin*+ and Margaret Sullivan Pepe and Yingye Zheng
• Companies: University of Washington and Fred Hutchinson Cancer Research Center and Fred Hutchinson Cancer Research Center
• Keywords: ROC curve ; event time outcomes ; percentile value ; regression ; censoring

Abstract

Biomarkers have been widely used to aid medical screening, diagnosis and prognosis of disease. Receiver operating characteristic (ROC) curve is a popular summary for evaluating the accuracy of such biomarkers. Pepe and Cai (2004) and Cai (2004) considered fitting a regression model for the distribution of one minus percentile values to make inference about ROC curves. Here we extend the percentile value based ROC regression model to censored event time outcome data. The additional dimension of time is incorporated when estimating the covariate-specific percentile values using semiparametric and nonparametric models. A weighted GLM form of ROC regression is proposed to accommodate censoring. Our new approaches are fairly easy to implement using standard software. The performance of proposed estimators is investigated by simulations and is illustrated using Seattle Heart Failure Study.