NCJ Number
226843
Journal
Problems of Forensic Sciences Volume: 69 Dated: 2007 Pages: 62-70
Date Published
2007
Length
9 pages
Annotation
This paper proposes an evaluation methodology for any likelihood-ratio (LR) that can be used in a LR-based forensic discipline.
Abstract
The LR framework for the interpretation of evidence is a mathematical and logical tool that provides many advantages in a forensic context. The objective is to compute the LR as the degree of support for one hypothesis compared with a hypothesis with a different conclusion. The proposed LR evaluation method is based on information theory, as it shows the performance of the system in the form of cross-entropy as a measure of uncertainty and deviation between probability distributions. This assessment methodology takes into account not only the discrimination of the technique that was used but also the calibration of LR values. It measures the loss of information due to the deviation of the forensic scientist’s LR values from the LR values computed with knowledge of the true answers. The proposed evaluation method is illustrated with an example that uses different LR computation techniques with a speaker recognition system. The proposed evaluation techniques are compared to discrimination measures as DET (detection error tradeoff) plots, using a clear and standard protocol such as those developed by the American National Institute of Standards and Technology (NIST) in its 2006 Speaker Recognition Evaluation. This is an example of a methodology for evaluating the performance of a set of LR values in a controlled and transparent way. 2 figures and 15 references