Statistical discrimination of liquid gasoline samples from casework.

The intention of this study was to differentiate liquid gasoline samples from casework by utilizing multivariate pattern recognition procedures on data from gas chromatography-mass spectrometry. A supervised learning approach was undertaken to achieve this goal employing the methods of principal component analysis (PCA), canonical variate analysis (CVA), orthogonal canonical variate analysis (OCVA), and linear discriminant analysis. The study revealed that the variability in the sample population was sufficient enough to distinguish all the samples from one another knowing their groups a priori. CVA was able to differentiate all samples in the population using only three dimensions, while OCVA required four dimensions. PCA required 10 dimensions of data in order to predict the correct groupings. These results were all cross-validated using the "jackknife" method to confirm the classification functions and compute estimates of error rates. The results of this initial study have helped to develop procedures for the application of multivariate analysis to fire debris casework.