A Quantitative Comparison of the Performance of Likelihood Ratio Systems in Trace-Reference Problems

Master thesis (2024)

Authors

W.G. Versteegh Electrical Engineering, Mathematics and Computer Science

Contributors

R.J. Fokkink Applied Probability - EEMCS (mentor)
N. Parolya Statistics - EEMCS (graduation committee member)
R.J.F. Ypma Nederlands Forensisch Instituut (NFI) (graduation committee member)
Faculty
Electrical Engineering, Mathematics and Computer Science, Electrical Engineering, Mathematics and Computer Science
More Info
expand_more

Published Date

08-05-2024

Language

English

Reuse Rights

Other than for strictly personal use, it is not permitted to download, forward or distribute the text or part of it, without the consent of the author(s) and/or copyright holder(s), unless the work is under an open content license such as Creative Commons.

Faculty

Electrical Engineering, Mathematics and Computer Science

Abstract

In forensic science, the strength of evidence is calculated mainly by statistical models called likelihood ratio systems. In court cases, the specific-source likelihood ratio system is used by forensic scientists to determine if a trace originates from a known reference, called the trace-reference problem. However, collecting sufficient data to create a specific source model may be time-consuming and costly. If the number of court cases becomes too high this could be problematic. Therefore there is a need for other models that can perform as well as a specific-source model if it is infeasible.

A common-source model could be a solution, as this model can be re-used over cases. To this end, we introduce two common-source systems: a common-source feature-based system and a common-source score-based system. We compare their performance to a specific-source score-based system in a trace-reference setting. The simulations show that the common source feature-based method is the best-performing likelihood ratio system if the dimensionality is not too high, and the sources are equally variable. The analysis shows that the common-source score-based method can work as effectively as a specific-source score-based model in certain scenarios.

Additionally, we researched a preprocessor, known as percentile rank, which aims to consider typicality for score-based methods. For the common-source score-based system, using a percentile-rank preprocessor can improve the performance for large sample sizes, while considering the rarity of the measurements.