| The goal of all chemometricians is to devise methods which extract all of the information from a data set. Analytical chemists have found that chemometrics has the potential to enhance analysis and provide more information; however, the need for expertise to use chemometric methods is frequently prohibitive. It is the premise of this work that use of additional available information allows chemometric methods to be devised such that more information is extracted than previous methods. In addition, the proposed methods reduce or eliminate the need for user interaction and expertise. The primary form of additional information employed are noise estimates of the spectral and reference values.; In this research, the noise addition method, a criterion for selecting factors which makes use of noise estimates, was devised. The method provides significance levels and minimizes user interaction. In order to determine its effectiveness, the method was applied to several areas of chemometrics. In the first application, the noise addition method was used to select factors in principal components analysis (PCA). Most current PCA selection methods assume homoscedastic noise, an assumption that often fails in analytical data sets. The noise addition method is based on the actual noise characteristics of the data so that the method assumptions are fulfilled. Matching data characteristics widens the applicability of the noise addition method and enhances selection of principal components.; In a second application, the noise addition method was used to select factors for calibration. The noise addition method performed well for data sets in which spectroscopic noise was the dominant source of noise, but was less successful when the reference values had significant error.; In order to remove the need for expert-supervised factor selection in performing calibration, parsimonious regression (PR) was investigated as a potential alternative. Parsimonious regression calculates calibration models using an optimization algorithm such that prediction ability is optimized. The method also has the additional benefit of not requiring supervision to select factors. Several method of estimating predictive ability were used as optimization criteria and the resultant PR models compared favorably to best case PLS calibration. |
