WHICH PRINCIPAL COMPONENTS TO UTILIZE FOR PRINCIPAL COMPONENT REGRESSION

Principal components (PCs) for principal component regression (PCR) have historically been selected from the top down for a reliable predictive model. That is, the PCs are arranged in a list starting with the most informative (PC associated with the largest singular value) and proceeding to the least informative (PC associated with the smallest singular value). PCs are then chosen starting at the top of this list. This paper discusses an alternative procedure of treating PC selection as an optimization problem. Specifically, without any regard to the ordering, the optimal subset of PCs for an acceptable predictive model is desired. Five data sets are analyzed using the conventional and alternative approaches. Two data sets are spectroscopic in nature, two data sets deal with quantitative structure-activity relationships (QSARs) and one data set is concerned with modeling. All five data sets confirm that selection of a subset without consideration to order secures the best results with PCR. One data set is also compared using partial least squares 1.