A designed experiment for the examination of techniques used in the analysis of near-infrared spectra. Part 2. Derivation and testing of regression models

Abstract

Using mixtures of four pure chemicals, the predictive abilities of wavelength regression models produced by three different selection strategies are compared with models based on principal components. While both types of model predicted composition equally well, component models were less likely, based on calibration data, to produce an overoptimistic impression of predictive performance. No single wavelength selection strategy consistently produced better models. While wavelength models tended to be difficult to interpret in terms of absorbance bands, consistent patterns analagous to measurements of derivatives could be identified.