Correction of non-linearities in spectroscopic multivariate calibration by using transformed original variables. Part II. Application to principal component regression Original Research Article

The addition of non-linear transformations of the original variables to the original data set is proposed to model strong non-linearities with PCR. This yields always better results than the ones obtained by using only the original variables. In some cases predictive ability increases, in other cases the same predictive ability is achieved with however smaller complexity, giving more parsimonious and robust models. A correct selection of the PCs to be included in the model improves the results obtained by using the top–down selection, giving equivalent models to those obtained by using polynomial PCR, but generally with a lower needed complexity, or by using linear PLS regression with transformed variables. Three real data sets, with different sources and degrees of non-linearity have been tested.