Variable Selection for Partial Least-Squares Calibration of Near-Infrared Data from Orthogonally Designed Experiments

Abstract

A method of variable selection for use with orthogonally designed calibration data sets, such as factorial or partial factorial designs, is described. The procedure works by assessing the degree of correlation between each X variable (e.g., wavelength) and each Y variable (concentration, composition, etc.). The X variables are then ranked according to the degree of correlation. A forward selection method is then used to determine the optimum number of high-ranked variables to be used for calibration purposes. The algorithm was tested on a data set obtained by near infrared (NIR) spectrometry. Significant improvements in the prediction accuracy of partial least-squares (PLS) models were observed for two of the components in the chemical mixture by using the selected wavelengths in the NIR spectra rather than all the variables in the original spectra or the variables obtained using a spectral-variance-based variable selection method.