Feature-based models for forensic likelihood ratio calculation

Abstract

The likelihood ratio is a generally accepted measure for the strength of evidence in forensic comparison problems. These problems concern comparisons where it is investigated whether at least two items come from the same source or not, e.g. whether the DNA on the crime scene comes from the suspect or not. The use of likelihood ratios by forensic experts in practical forensic casework demands for a unified system to compute likelihood ratios. Therefore, the EU funded the ''ENFSI-LR’’ project that aims to construct software which helps forensic experts to calculate likelihood ratios based on validated scripts and harmonized models. In this thesis some problems concerning the ENFSI-LR project are addressed. Solutions to these problems are useful for unification, validation and (future) development of the software. Throughout this thesis, the emphasis is on continuous two-level feature-based models. In the literature, these underlying models have led to two likelihood ratio formulas. In this thesis it is proved that these two formulas are exactly the same. This thesis also explores several parameter estimation methods for the two-level model. Standard estimation methods are compared with estimation methods which have not been used in forensic statistics until now: a generalized weighted mean or maximum likelihood estimation. As an extension of existing feature-based models, a model is introduced that combines discrete- and continuous evidence into one likelihood ratio.