Statistics and Risk Management Seminar - 9/6/2014 - 3 p.m.

Monday, 9 June 2014, 3:00 p.m.

Lecturer: Miguel de Carvalho, Assistant Professor, Department of Statistics, Pontificia Universidad Católica de Chile e membro do CMA

Title: "Bayesian Nonparametric Youden Index Modeling"

Local: Sala de Seminários - Edifício VII
Faculdade de Ciências e Tecnologia, Quinta da Torre, Caparica

Abstract: Accurate diagnosis of disease is of crucial importance in health care and medical research. The major goal of a diagnostic test is to distinguish diseased patients from nondiseased patients and, before a test is routinely used in practice, its ability to distinguish between these two states must be rigorously evaluated. Without loss of generality, we assume that a patient is diagnosed as diseased (positive) if the test’s value is greater than a given threshold value and as nondiseased (negative) in the opposite case. The accuracy of the test at any given threshold can be measured by the probability of a true positive (sensitivity) and a true negative (specificity). The receiver operating characteristic (ROC) curve, a popular graphical tool for evaluating the discriminatory ability of a continuous scale diagnostic test, plots the sensitivity, Se(c), against 1-specificity, 1 − Sp(c), as the threshold c varies through the range of possible test results. To evaluate the discriminatory ability of a test it is common to summarize the information of the ROC curve into a single global value or index. The two most popular summary indices of diagnostic accuracy are the area under the ROC and the Youden index. In this work we focus on the Youden index (YI), which can be defined as YI = maxc{Se(c) + Sp(c) − 1} and has the attractive feature of providing a criterion for choosing the optimal threshold value c∗ to screen subjects in practice. With the aim of having a flexible model that can handle skewness, multimodality and other nonstandard features of the data, without the need of knowing in advance their existence, we propose to estimate the Youden index and its associated optimal threshold c∗, using Bayesian nonparametric techniques, namely, Dirichlet process mixtures. The performance of the estimator is evaluated through a simulation study and a real data application is provided. This is joint work with V. Inácio de Carvalho (PUC Chile) and A. Branscum (Oregon State University).