COMPSTAT 2016: Start Registration
View Submission - COMPSTAT
A0257
Title: Nonparametric and parametric confidence intervals for the Youden index and its associated cutoff point Authors:  Benjamin Reiser - University of Haifa (Israel) [presenting]
Leonidas Bantis - The University of Texas MD AndersonCancer Center (United States)
Christos T Nakas - University of Bern (Switzerland)
Abstract: The receiver operating characteristic (ROC) curve is commonly used to evaluate a continuous biomarker that attempts to discriminate between a healthy and a diseased population. The ROC curve is a plot of the sensitivity $\{sens(c)\}$ versus 1-specificity $\{1-spec(c)\}$ over all possible threshold values $c$ of the marker. To evaluate the discriminatory ability of a marker it is common to summarize the information of the ROC curve into a single global value or index. Although the area under the ROC curve is the most frequently used global index of diagnostic accuracy the maximum of the Youden Index, being defined by maximizing over all possible threshold values as $J= \max \{sens(c) + spec(c) - 1\}$, is also often used. $J$ is equivalent to the Kolmogorov-Smirnov distance between the two populations. In practice, clinicians are often interested in determining a cutoff for classification purposes. Frequently the optimal cutoff value $c^{*}$ is chosen as the value of $c$ for which $J$ is maximized. In the applied literature confidence intervals for $J$ and $c^{*}$ are typically ignored. We provide new nonparametric kernel density-based and parametric delta method-based methods for constructing confidence intervals for both $J$ and $c^{*}$. We compare our methods to currently available techniques through simulations and discuss some real examples.