An overview of ROC analysis

Abstract

The Receiver Operating Characteristic (ROC) curve is considered to be one of the best developed statistical tools to analyze the performance of binary classifier systems, such as medical diagnostic tests, and to classify elements of a population into two groups, namely “diseased” and “healthy”. In this dissertation, the theoretical construct of the ROC curve is studied, along with summary measures derived from the ROC curve, particularly the Area under the Curve (AUC). Different techniques from three estimation frameworks, namely the non-parametric, parametric and semi-parametric frameworks, are discussed, and their performance is compared using simulation studies. The best estimation techniques, according to the simulation studies, are then applied to two real medical datasets, obtained from Mater Dei hospital. Methodology on how an optimal threshold value is derived from the ROC curve is discussed, and is then used to determine the best threshold for both applications.