Evaluating score- and feature-based likelihood ratio models for multivariate continuous data: applied to forensic MDMA comparison

Likelihood ratio models are increasingly used in forensic science to evaluate evidence. When identical raw data are used, feature-based and score-based models produce fundamentally different results. Score-based models yield much lower absolute likelihood ratios and are more stable than feature-based models, because they reduce multivariate information to a univariate distance or similarity score. Feature-based models retain the full multivariate structure of original feature values. The paper explains how data pre-treatment and dimension reduction affect both methods, using chemical profiles of MDMA as an example.

Study at a glance

Abstract

Likelihood ratio (LR) models are moving into the forefront of forensic evidence evaluation as these methods are adopted by a diverse range of application areas in forensic science. We examine the fundamentally different results that can be achieved when feature- and score-based methodologies are employed to calculate likelihood ratio as a measure for the strength of evidence in forensic comparison, especially when comparable hypotheses and identical raw data are used. The focus is on LR based on multivariate continuous data. As an example of this, chemical profiles used in MDMA (illicit drugs) comparisons, will be investigated. The two model types, feature based and score based, are shown to perform differently when identical raw data are used. Score-based models provide much lower absolute LR values than feature-based models and demonstrate greater stability than feature-based models. This is the result of using different information of the raw data as evidence. Score-based models reduce multivariate information to a univariate distance or similarity score between items, whereas feature-based models use the multivariate structure of all the original feature values (and their combinations) of individual items as evidence. We discuss the different results and provide an explanation of the effects of data pre-treatment and dimension reduction on both methods.