Premium
Bayesian semi‐parametric ROC analysis
Author(s) -
Erkanli Alaattin,
Sung Minje,
Jane Costello E.,
Angold Adrian
Publication year - 2006
Publication title -
statistics in medicine
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 1.996
H-Index - 183
eISSN - 1097-0258
pISSN - 0277-6715
DOI - 10.1002/sim.2496
Subject(s) - frequentist inference , bayesian probability , computer science , prior probability , gibbs sampling , dirichlet process , receiver operating characteristic , parametric statistics , kernel density estimation , statistics , gold standard (test) , kernel (algebra) , posterior probability , mathematics , artificial intelligence , bayesian inference , machine learning , combinatorics , estimator
This paper describes a semi‐parametric Bayesian approach for estimating receiver operating characteristic (ROC) curves based on mixtures of Dirichlet process priors (MDP). We address difficulties in modelling the underlying distribution of screening scores due to non‐normality that may lead to incorrect choices of diagnostic cut‐offs and unreliable estimates of prevalence of the disease. MDP is a robust tool for modelling non‐standard diagnostic distributions associated with imperfect classification of an underlying diseased population, for example, when a diagnostic test is not a gold standard. For posterior computations, we propose an efficient Gibbs sampling framework based on a finite‐dimensional approximation to MDP. We show, using both simulated and real data sets, that MDP modelling for ROC curve estimation closely parallels the frequentist kernel density estimation (KDE) approach. Copyright © 2006 John Wiley & Sons, Ltd.