On partial least squares dimension reduction for microarray-based classification: a simulation study

In microarray tumor tissue classification studies, the expressions of thousands of genes (variables) are simultaneously measured across a few tissue samples. Standard statistical methodologies in classification do not work well when the dimension, p, is greater than the sample size, N. One approach to classification problems, when pmuch greater thanN, is to first apply a dimension reduction method and then perform the classification in the reduced space. In this paper, we study dimension reduction for classification in high dimension based on partial least squares (PLS) and principal components analysis (PCA). In addition, we propose and explore two hybrid-PLS methods for dimension reduction. PLS components are linear combinations of the original predictors, but the weights are nonlinear functions of both the predictors and response variable. This makes it difficult to study the PLS classification methodologies analytically, so, in this paper, we turn to a numerical study using simulation. (C) 2003 Elsevier B.V. All rights reserved.