Abstract

Co-author: Mary R. Williams (National Center for Forensic Science, University of Central Florida)

This presentation will examine the calculation of a likelihood ratio to assess the evidentiary value of fire debris analysis results. Models based on support vector machine (SVM), linear and quadratic discriminant analysis (LDA and QDA ) and k-nearest neighbors (kNN) methods were examined for binary classification of fire debris samples as positive or negative for ignitable liquid residue (ILR). Computational mixing of data from ignitable liquid and substrate pyrolysis databases was used to generate training and cross validation samples. A second validation was performed on fire debris data from large-scale research burns, for which the ground truth (positive or negative for ILR ) was assigned by an analyst with access to the gas chromatography-mass spectrometry data for the ignitable liquid used in the burn. The probabilities of class membership were calculated using an uninformative prior and a likelihood ratio was calculated from the resulting class membership probabilities . The SVM method demonstrated a high discrimination, low error rate and good calibration for the cross-validation data; however, the performance decreased significantly for the fire debris validation data, as indicated by a significant decrease in the area under the receiver operating characteristic (ROC) curve. The QDA and kNN methods showed performance trends similar to those of SVM . The LDA method gave poorer discrimination, higher error rates and slightly poorer calibration for the cross validation data; however the performance did not deteriorate for the fire debris validation data.