A New Approach for Statistically Matching Hydrocarbon Profiles using Partial Least Squares (PLS) and Principal Component Analysis (PCA)

Multivariate statistics is a powerful tool for the analysis of forensic evidence such as oil samples. Part of this process involves the use of a combination of chemical and statistical techniques which allow samples (evidence) to be grouped or matched to pre-existing groups. Chemical profiles can be created using chromatography, this produces massive datasets with thousands of points. Currently the most common data reduction method for these chemical profiles is the use of peaks that occur within the data. These peaks are usually integrated to find peak area and other information. The software commonly used to pick the reference points for the start and end of a peak, does not always generate reproducible points. A new method for data reduction of hydrocarbon profiles is proposed that utilises the entire dataset by averaging the instrument response into appropriate sized bin-widths. The data produced from these chemical profiles is both high dimensional and correlated. Usually there are more variables than observations so traditional techniques like Linear Discriminant Analysis (LDA) cannot be directly applied. In this paper, data reduction using Partial Least Squares (PLS) and Principal Components Analysis (PCA) followed by classification using LDA will be compared. Simulated data was used to compare statistical methods and different bin-widths for the averaging method. It was found when group difference does not dominate inter-observational differences PLS is superior to PCA. Results from the case study and simulation data show the averaging method is a viable alternative to the traditional peak area method.