Adjusting for Nonignorable and Partially Ignorable Verification Bias in Estimating ROC Curve and Its Area

• Activity Number: 528
• Type: Contributed
• Date/Time: Wednesday, August 5, 2009 : 2:00 PM to 3:50 PM
• Sponsor: Biopharmaceutical Section
• Author(s): Danping Liu*+ and Xiao-Hua (Andrew) Zhou
• Companies: University of Washington and University of Washington
• Keywords: Alzheimer's disease ; nonignorable missing data ; partially ignorable missing data ; ROC curve ; verification bias

Abstract

In estimating the ROC curve, it is often the case that not all the patients receive disease verification. When the missingness of the true disease status is not ignorable, the observed likelihood involves the missing mechanism given by a selection model. Two situations are discussed here: (1) nonignorable missing: all the missingness depends on the missing disease status; (2) partially ignorable missing: only a portion of the missingness depends on the missing disease status. We specify a parametric selection model and a disease model in order to obtain the maximum likelihood estimates. With the estimated verification and disease probabilities, we construct the empirical estimates of the ROC curve and its area. The proposed method is illustrated using a data set from research in the Alzheimer's disease.