Partial least-squares methods for spectral analyses. 2. Application to simulated and glass spectral data

Abstract

Partial least-squares (PLS) methods for quantitative spectral analyses are compared with classical least-squares (CLS) and principal component regression (PCR) methods by using simulated data and infrared spectra from bulk seven-component, silicate-based glasses. Analyses of the simulated data sets show the effect of data pretreatment, base-line variations, calibration design, and constrained mixtures on PLS and PCR prediction errors and model complexity. Analyses of the simulated data sets also illustrate some qualitative differences between PSL and PCR. PLS and PCR predicted concentration errors from the simulated data sets and a set of the Fourier transform infrared spectra of silicate-based glasses (S-glass) show that prediction errors are not statistically different between these two methods for these individual data sets with limited numbers of samples. However, PLS and PCR are both superior to CLS methods in the case of the analysis of S-glass where only one analyte is known in the calibration samples and the components of unknown concentration overlap all the spectral features of the analyte components. CLS analysis precision significantly improves when the three known analyte concentrations (B/sub 2/O/sub 3/, P/sub 2/O/sub 5/, and OH) are used in the calibration. In this latter case, PLS and PCR concentration predictions are unchanged, andmore » although they each still yield a lower standard error of prediction than the CLS method, there is no longer strong statistical evidence that these differences between PLS or PCR and CLS are outside experimental error for the B/sub 2/O/sub 3/ component. The ability of CLS and PLS methods to provide chemically useful estimates of the pure-component spectra is also demonstrated.« less