Fuente: JUSTICE & SCIENCE VOL 50 MARZO 2010
C.P. Saunders, A.B. Hepler, L.J. Davis, J. Buscaglia
In forensic science, two competing hypotheses are often of interest: the suspect left the trace or the suspect is a random man. With respect to forensic handwriting comparisons, these hypotheses reduce to ‘the suspect wrote the questioned document’ versus ‘the suspect did not write the questioned document.’ The likelihood ratio is formed by taking the ratio of the relevant probabilities (or densities) of the evidence under each hypothesis. We propose that, when unknown parameters are present, the commonly reported likelihood ratio is only a point estimate of the likelihood ratio of interest. For example, in handwriting it is necessary to estimate a suspect's writing profile based upon a sample; thus, the resultant likelihood ratio has additional uncertainty associated with it. We illustrate various interval estimates for the likelihood ratio that naturally account for this uncertainty. The first, a parametric approach, assumes that a handwriting profile can be modelled by a high dimensional multinomial distribution and is an extension of Lindley's likelihood ratio . The second, a nonparametric method, makes use of a biometric score and modifies the technique summarized by Meuwly . This modification requires an additional exchangeability assumption on characters, which leads to a U-Statistic representation. Each technique will be illustrated and compared using a research database of about 500 writers, with approximately five writing samples per writer.