Integrative linear discriminant analysis with guaranteed error rate improvement. (22nd October 2018)

• Record Type: Journal Article
• Authors: Li, Quefeng; Li, Lexin
• Is Part Of: Biometrika. Volume 105:Number 4(2018:Dec.)
• Page Start: 917
• Page End: 930
• Publication Date: 2018-10-22
• Subjects: Bayes error -- High-dimensional classification -- Integrative analysis -- Linear discriminant analysis -- Multi-type data -- Regularization; Biometry -- Periodicals; 570.1519505
• DOI: 10.1093/biomet/asy047
• Languages: English
• ISSNs: 0006-3444

Abstract

Summary: Multiple types of data measured on a common set of subjects arise in many areas. Numerous empirical studies have found that integrative analysis of such data can result in better statistical performance in terms of prediction and feature selection. However, the advantages of integrative analysis have mostly been demonstrated empirically. In the context of two-class classification, we propose an integrative linear discriminant analysis method and establish a theoretical guarantee that it achieves a smaller classification error than running linear discriminant analysis on each data type individually. We address the issues of outliers and missing values, frequently encountered in integrative analysis, and illustrate our method through simulations and a neuroimaging study of Alzheimer's disease.