Comparison of statistical multivariate analysis methods for chemical forensics profiling of a carbamate chemical warfare agent precursor

In chemical forensics, statistical multivariate methods are used to assist in the source determination of samples with an unknown origin. Several suitable statistical methods exist and have been used in chemical forensics studies related to chemical warfare agents. When using these methods, the performance of the classification can often be improved by determining which variables, e.g. impurities, are most important for separating the classes. A variable selection method is therefore used to select the most important variables to be included in further classification analyses. If multiple statistical and variable selection methods are used, they should yield similar results to ensure the comparability of analyses. In this study, the performance and comparability of statistical classification methods and variable selection methods were evaluated, using data relevant for chemical forensics. The statistical classification methods compared were principal component analysis (PCA), hierarchical cluster analysis (HCA), partial least squares discriminant analysis (PLS-DA), orthogonal partial least squares discriminant analysis (OPLS-DA), k-nearest neighbours ( k −NN), and linear discriminant analysis (LDA). The variable selection methods included were Fisher-ratio/Degree-of-class-separation (F-ratio/DCS), model weight values (w*), and variable importance in projection (VIP). The data sets obtained by using the three variable selection methods were used for the classification analyses, resulting in over 90 classification analyses. The classification results of PLS-DA, OPLS-DA, k-NN, and LDA were evaluated using test set samples class prediction, and these classification methods obtained highly similar results. However, the use of different variable selection methods led to higher variability in results. • Chemometric methods evaluated for chemical forensics profiling of chemical weapons. • Variable selection methods F-ratio/DCS, model weight values, and VIP compared. • Comparison of six classification methods: PCA, HCA, PLS-DA, OPLS-DA, k-NN, and LDA. • A total of 92 classification analyses using key variable selection methods • Classification method results were comparable, variable selection results varied.