One Table, Two Tables, More Tables...Principal Component Analysis, Partial Least Square and Multi‐ Table Approach for brain imaging and genomics?
July 13, 2012
Hervé Abdi , Full Professor, School of Behavioral and Brain Sciences, University of Texas at Dallas, adjunct Professor of Radiology, University of Texas Southwestern Medical Center at Dallas
Brain imaging generates very large data sets and these data are often collected to understand behavioral data. Recently the complexity of these data has been increased by including new types of data such as genomics data. This talk will present an overview of the multivariate approaches used to analyze such data. When the data are collected in one table in which observations are described by a large number of variables, they can be analyzed with principal component analysis (PCA), a technique
that combines all the variables to get new variables called factor scores that optimally describe the observations. Often the observations are plotted on factor maps that can be used to visualize the data. When dealing with two data tables (e.g., brain imaging and behavior), the problem is to find the information common to these two tables and PCA is extended to partial least square analysis (PLS). When dealing with more than two tables (e.g., genomics, brain imaging, and behavior), several methods, regrouped under the label of multi‐table analysis (e.g., multiple factor analysis, STATIS, sum‐ PCA), can be used. All these methods will be described and illustrated with examples from the brain imaging literature.