CMStatistics 2021: Start Registration
View Submission - CMStatistics
Title: Calibration of likelihood ratios systems in forensic science Authors:  Jan Hannig - University of North Carolina at Chapel Hill (United States) [presenting]
Hari Iyer - National Institute of Standars and Technology (United States)
Abstract: Many computer programs and software systems used in the interpretation of forensic evidence have as their output Bayes factors, also commonly referred to as likelihood ratios. For example, it is not unusual to see it reported that the DNA recovered at the crime scene is a million times more likely under the assumption that the defendant is a contributor to the crime stain than under the assumption that the defendant is not a contributor. We summarize existing approaches for examining the validity of likelihood ratio systems. We will see that what is used in the current practice has significant drawbacks related to uncertainty quantification. We then discuss a new statistical methodology based on generalized fiducial inference for empirically examining the validity of such likelihood ratio assessments. Using data from a number of sources, such as glass, paint and DNA evidence, we illustrate our approach by examining LR values calculated using standard approaches in the forensic literature. We also use the new tool to show limitations of a common method of post-hoc re-calibrating of outputs