Statistical Inference of Two Classifiers by Partial Area Under the ROC Curve with Empirical Likelihood

• Activity Number: 125 - New Nonparametric Statistical Methods for Multivariate and Clustered Data
• Type: Contributed
• Date/Time: Monday, July 30, 2018 : 8:30 AM to 10:20 AM
• Sponsor: Section on Nonparametric Statistics
• Author(s): Xue Ding* and Mai Zhou
• Companies: University of Kentucky and University of Kentucky
• Keywords: binary classification; pAUC; two-sample test; nonparametric approach

Abstract

The Receiver Operating Characteristic (ROC) curve has been extensively used in the assessment of a binary classifier. As a quantitative summary index, the area under ROC curve (AUC) measures the overall accuracy of classifying the positive subjects from negative subjects. Under the practical concern, people set the restrictions on the false positive rate to focus on the partial area under the curve (pAUC). Using two-sample empirical likelihood methodology by Owen (2001), we investigate the difference between two ROC curves by the pAUC with paired data. Unlike previously proposed methods, the empirical likelihood ratio in our study is asymptotically chi square distributed without any adjustment. Moreover, compared to the existing nonparametric approaches, our test procedure avoids having to estimate the variance, resulting in a more accurate test statistic. In addition, the corresponding confidence interval is invariant to transformation. We illustrate our approach in a real data example and evaluate its performance in the simulation studies.