Spring 2002 Seminar Series - February 7, 2002
University of Minnesota
School of Statistics
College of Liberal Arts

Bi-Distribution Association Marginal Models for Evaluating Diagnostics

Joseph B. Lang
Dept. of Statistics and Actuarial Science
University of Iowa

Thursday, February 7, 2002
4:00 PM, B10 Ford Hall
Minneapolis, East Bank Campus
Social at 3:30 PM, 300 Ford Hall

Abstract

Diagnostic (or signal-detection) systems can be used to decide which of two classes, e.g. "normal" or "abnormal," a subject falls in. These systems are typically imperfect classifiers. The classification rules are often based on system-generated ordinal ratings, e.g. (1=definitely normal,...,5=definitely abnormal). In this presentation, we consider empirical studies of diagnostic systems whereby each subject in a sample is rated twice, once using system 1 and once using system 2. The resulting ratings, which are correlated by study design, are used to address the following objectives. Objective(1): Compare the diagnostic capacities of the two systems using receiver operating characteristic (ROC) analysis; and Objective(2): Model and describe the rating pair association or agreement.
The presentation can be divided into two main parts. Part I gives the basics of ROC analysis, highlighting the difference between manifest and latent ROC curves. It also describes existing methods for addressing Objective (1). Part II introduces the new class of Bi-distribution Association Marginal (BAM) models, which can be used to simultaneously address Objectives (1) and (2). A maximum-likelihood fitting algorithm, BAMROC, is described, and data from a recent imaging diagnostic study are analyzed. Current and future work on BAM models is described.