An evaluation of estimation and inference methodologies in roc analysis

Abstract

One major problem that is addressed in this work is the construction of an analytical framework for comparing and evaluating some statistical properties of classification processes. The term classification process is a very general term which refers to a large class of processes differing in certain aspects from each other, depending on the area in which they are applied The general objective of a classification process is to discriminate between the possible states of the entities under study and to decide in which state such elements actually subsist. In view of the importance of the decisions taken with the aid of such processes and the large costs, whether financial, physical or medical, which are usually incurred for applying and developing such processes, evaluating their performance has become crucial. Amongst the most popular methods for measuring the performance of a classification processes are those provided by the receiver operating characteristic (ROC) curve and the summary indices derived from it; which are collectively referred to as ROC analysis. We shall start by providing a detailed mathematical description of the ROC Curve and the summary measures derived from it. We shall then proceed to evaluate a number of estimation techniques, available in literature, for ROC curves and their summary measures. Results from simulation studies conducted to analyze and compare the performance of the estimation methods are presented. Finally an application of ROC analysis for evaluating various screening tests for p- Thalassaemia using data supplied by the molecular genetics lab of the University of Malta is considered.