Date of Award

Spring 2021

Document Type

Open Access Dissertation

Department

Chemistry and Biochemistry

First Advisor

Stephen L. Morgan

Abstract

Numerous disciplines in forensic science utilize various types of spectral data when analyzing evidence. Spectral techniques are particularly critical for the analysis of trace evidence, as these methods are normally non-destructive. Preserving evidence, especially trace evidence, is of high priority in a forensic laboratory setting. Once a piece of evidence has been fully consumed, no more analyses can be performed. Typically, visual comparison, or spectral overlay, is performed to compare questioned samples (evidence) to standards or knowns. However, such an approach may not be optimal in distinguishing the subtle, yet highly important, discriminating characteristics present in the spectra. As statistical analysis becomes increasingly influential in the forensic science community, multivariate chemometric approaches may aid in overcoming the major downfall of spectral overlay to classify and identify samples. More traditional approaches allow for dimension reduction and classification of samples. However, multivariate data sets can pose a problem with having far fewer samples than variables to build the classification model. Sparse statistical approaches overcome this limitation by reducing the number of variables retained in the final model. Only a few, significant parameters remain. This reduces model complexity and increases prediction accuracy of the model. Here, logistic regression with Lasso regularization is the sparse approach that was compared to traditional classification techniques to group fiber and lipstick samples based on their spectral data.

Included in

Chemistry Commons

Share

COinS