CMStatistics 2016: Start Registration
View Submission - CMStatistics
B1421
Title: Nonparametric predictive inference for diagnostic test thresholds Authors:  Manal Alabdulhadi - Durham Uinverstiy (United Kingdom) [presenting]
Frank Coolen - Durham University (UK)
Tahani Coolen-Maturi - Durham University (United Kingdom)
Abstract: The accuracy of diagnostic test relates to the ability of the test to distinguish between diseased and healthy individuals. Providing good methods for defining the accuracy of diagnostic tests assist physicians to detect the probability of disease for their patients. In 2-Group and 3-Group ROC analysis, setting thresholds for classification is often the most important decision. The standard uses the maximisation of the Youden index, a global measurement of diagnostic accuracy. We consider an alternative to the maximisation of the Youden index, by explicitly considering the use of the classification procedure for a specific number of future patients. We consider nonparametric predictive inference (NPI), which is a powerful statistical framework that yields direct probabilities for one or $m$ future observations, based on $n$ observations for related random quantities. We introduce 2-Group and 3-Group predictive method to select optimal diagnostic thresholds in order to have the best classification of one or more future persons on the basis of their test results. We find that the optimal thresholds sometimes lead to other value of thresholds with different number of $m$ future patients. We generalize the Youden index by applying our method to the Youden index and maximising the sum of the probabilities of correct classification for the different groups. Comparison between our method and generalization of Youden index is discussed.