Title: Biomarker accuracy determination via an affinity-based measure
Authors: Miguel de Carvalho - CEAUL (Centro de Estatistica e Aplicacoes), Universidade de Lisboa (Portugal)
Bradley Barney - University of Utah (United States) [presenting]
Garritt Page - Brigham Young University (United States)
Abstract: The area under the receiver operating characteristic curve (AUC) and Youden index are popular metrics for quantifying the ability of a biomarker to predict the presence/absence of a disease. These metrics can perform poorly when the biomarker distribution with/without the condition does not stochastically dominate the other. We propose using a measure based on Hellinger's affinity which overcomes this potential pitfall, providing a global summary of the similarity between biomarker densities from individuals with versus without a disease. We also illustrate the application of nonparametric Bayesian methods to flexibly estimate a covariate-dependent version of the affinity-based measure. We investigate the performance of our suggested metric via simulation. We also apply the methodology to a prostate cancer study to compare the potential diagnostic ability of two prostate-specific antigen (PSA)-derived biomarkers.