Prediction error improvements using variable selection on small calibration sets—a comparison of some recent methods

ISSN: 0967-0335 © IM Publications LLP 2012 doi: 10.1255/jnirs.966 All rights reserved Regression is probably the most widely studied and applied statistical analysis method in the chemometric literature. The aim is to develop models which can be used to predict properties of interest based on measurements of the chemical system, such as spectroscopic data. Multivariate calibration techniques such as multiple linear regression (MLR), principal component regression (PCR) and partial least squares regression (PLS)1 can then be used to compute a mathematical model. It correlates the multivariate measurement (spectrum) to the concentration of the analyte of interest and such a model can be used to predict the concentrations of new samples. When the number of measured predictor variables is large and it is not known beforehand which specific predictors are Prediction error improvements using variable selection on small calibration sets—a comparison of some recent methods