Macrolevel Discriminant Analysis: An Extension of Linear Discriminant Analysis for Nested Data

• Activity Number: 367
• Type: Contributed
• Date/Time: Tuesday, August 6, 2013 : 10:30 AM to 12:20 PM
• Sponsor: Section on Statistical Learning and Data Mining
• Author(s): Jose-Miguel Yamal*+ and E. Neely Atkinson and Getie Zewdie and Dennis Cox
• Companies: UT School of Public Health and UT MD Anderson Cancer Center and UT School of Public Health and Rice University
• Keywords: linear discriminant analysis ; classification ; hierarchical data classification ; quantitative cytology ; cervical cancer ; cervical intraepithelial neoplasia

Abstract

Here we consider the problem of using data with a nested structure to classify subjects. We assume that multiple measurements are available for each subject and that the goal of the analysis is to classify each subject into one of K classes based on the measurements for that subject. Whereas there has been a considerable amount of research dedicated to classification algorithms and models, they do not account for correlations between measurements within a subject. One of the best-known classification methods for the usual classification problem is Fisher's linear discriminant analysis (LDA). We extend this popular method to the classification problem of nested data by including a random effect into the LDA model. Simulation results are presented as well as an application to the classification of tissue samples from quantitative measurements on cell nuclei.