Computational Aspects of Forensic Evidence Interpretation

• Activity Number: 252
• Type: Contributed
• Date/Time: Monday, August 4, 2014 : 2:00 PM to 3:50 PM
• Sponsor: Section on Statistics in Defense and National Security
• Author(s): Christopher Saunders and Danica Ommen*+ and Cedric Neumann
• Companies: MITRE/South Dakota State University and South Dakota State University and South Dakota State University
• Keywords: Forensic Science ; Bayesian ; Computational Statistics ; Asymptotic ; Markov-Chain Monte-Carlo

Abstract

Forensic evidence interpretation requires the statement of two models, corresponding to the defense and prosecution, about how the evidence in a situation has arisen. Often, the information that a forensic scientist has available to evaluate between the two models includes: samples from a trace of unknown origin, samples from a specific source, and samples from an alternative source population. The forensic scientist is required to interpret and present the value of evidence, typically a Bayes Factor. There are many computational difficulties in calculating the Bayes Factor, a ratio of two marginal likelihoods, in real-world situations. In the case that the marginal likelihoods cannot be evaluated directly, they are approximated using Monte Carlo integration methods. However, research has illustrated the inefficiency of these methods in approximating the Bayes Factor. This inefficiency has a significant negative impact on the implementation of these methods in forensics. Also, a Gibbs sampler is often needed to sample values from the approximate distribution for parameters of interest. We will look at an example relating to glass fragments to demonstrate the computational methods a