Diagnostic Thresholds for Two and Three Ordinal Groups

• Activity Number: 580
• Type: Contributed
• Date/Time: Wednesday, August 1, 2012 : 2:00 PM to 3:50 PM
• Sponsor: Section on Statistics in Epidemiology
• Author(s): Kristopher Attwood*+ and Lili Tian
• Companies: SUNY at Buffalo and SUNY at Buffalo
• Keywords: Diagnostic Thresholds ; Youden index ; ROC ; Cut-points

Abstract

The ROC curve and AUC are standard tools for assessing the overall ability of a biomarker to discriminate between healthy (H) and disease (D). In practice, however, researchers are still tasked with selecting a threshold for classifying patients. The Youden index (J) is often used for this purpose, whose threshold maximizes the sum of the correct classification rates (CCRs). Unfortunately, J's threshold estimates can have high variability and produce unbalanced CCRs. We propose a new criterion, maximum area (MA), whose threshold maximizes the product of the CCRs. The proposed method produces more precise threshold estimates with minimal loss in relative overall CCR (< 3%). The procedure can be generalized to the more practical non-binary disease cases, where an intermediate stage (IS) exists between H and D. There is limited discussion on thresholds in this setting, with a generalized J recently proposed. The threshold estimates of the generalized J tend to under-utilize the available sample and produce limited coverage of the IS. The proposed extensions of the MA and closest to perfection methods to this setting, produce less variable threshold estimates and more balanced CCRs.