Likelihood ratio model for classification of forensic evidence

One of the problems of analysis of forensic evidence such as glass fragments, is the determination of their use-type category, e.g. does a glass fragment originate from an unknown window or container? Very small glass fragments arise during various accidents and criminal offences, and could be carried on the clothes, shoes and hair of participants. It is therefore necessary to obtain information on their physicochemical composition in order to solve the classification problem. Scanning Electron Microscopy coupled with an Energy Dispersive X-ray Spectrometer and the Glass Refractive Index Measurement method are routinely used in many forensic institutes for the investigation of glass. A natural form of glass evidence evaluation for forensic purposes is the likelihood ratio--LR=p(E|H(1))/p(E|H(2)). The main aim of this paper was to study the performance of LR models for glass object classification which considered one or two sources of data variability, i.e. between-glass-object variability and(or) within-glass-object variability. Within the proposed model a multivariate kernel density approach was adopted for modelling the between-object distribution and a multivariate normal distribution was adopted for modelling within-object distributions. Moreover, a graphical method of estimating the dependence structure was employed to reduce the highly multivariate problem to several lower-dimensional problems. The performed analysis showed that the best likelihood model was the one which allows to include information about between and within-object variability, and with variables derived from elemental compositions measured by SEM-EDX, and refractive values determined before (RI(b)) and after (RI(a)) the annealing process, in the form of dRI=log(10)|RI(a)-RI(b)|. This model gave better results than the model with only between-object variability considered. In addition, when dRI and variables derived from elemental compositions were used, this model outperformed two other classification methods in classifying test set observations into car or building windows.