On Desiderata for Score-Based Likelihood Ratios for Forensic Evidence

• Activity Number: 337
• Type: Topic Contributed
• Date/Time: Tuesday, August 6, 2013 : 10:30 AM to 12:20 PM
• Sponsor: Committee of Representatives to AAAS
• Author(s): Christopher Saunders*+ and John J. Miller
• Companies: South Dakota State University and George Mason University
• Keywords: Forensic Evidence ; Likelihood ; Bayes Factor ; Identification ; Approximation ; Biometrics

Abstract

This presentation offers some opinions on the desirable features of score-based likelihood ratios (SLRs) for interpreting and presenting forensic evidence. Let E denote all of the available evidence, with decomposition E={ E_s, E_u} where E_s denotes the evidence sample(s) obtained from a suspect, E_u denotes the evidence sample of unknown source obtained. We consider an arbitrary, but fixed score function, s, serving to reduce the evidence to the following form E'={s(E_s, E_u)}. Several score-based interpretations of the likelihood ratio have appeared in the literature providing a method for evaluating the weight of E' in light of two competing hypotheses, H_p and H_d. Recent studies in writer identification have shown that when E' is held constant, subtle changes in conditioning arguments regarding the defense proposition often lead to radically different values of the SLR. Each proposed SLR has advantages and disadvantages and, in our estimation, it is best to resist the idea of a "universally correct" SLR. We instead have concentrated our efforts on enumerating desirable theoretical properties for SLRs in general and the evaluation of proposed SLRs against each property.