This submission is the final year of the work in the original proposal (award number 15PNIJ-23-GG-04232-MUMU). The proposal timelines and deliverables have been adjusted to account for a two-year progress report and a final report at the end of the supplemental funding time period.
Weight of evidence (WoE) approaches have been suggested as alternatives to the identification paradigm to forensic evidence interpretation. Three statistical definitions of probability have spawned the three most used methods to quantify the WoE: the Two-Stage approach, Bayesian methods, and machine-learning or artificial intelligence (AI) methods. However, it is often difficult to define the probability functions necessary to conduct these analyses. Therefore, score-based methods have been proposed for numerous evidence types such as fingerprints, handwriting, and firearms, among others. Traditionally, score-based methods focus on univariate distributions for assessing the evidential value instead of the high-dimensional distributions of the original features. Scores are easily incorporated into all three methods of quantifying WoE, but rarely in a statistically rigorous manner. It has been shown that score-based methods for assigning the WoE do not share the same basic properties as the feature-based methods.
This research project proposes to extend basic research components of the score-based methods developed under previous NIJ Awards in a statistically rigorous manner for the interpretation of evidence for the classical and Bayesian paradigms. This project will develop a formal set of guiding principles for applying score-based methods, provide a rigorous foundation for the use of multivariate scores that map the full evidence- including the background population- to a set of scores, design statistical methods to construct formal Bayes Factors and Two-Stage approaches when the score is an unknown monotonic transform of the likelihood ratio, and develop strategies that allow for the adaptive estimation of the score from the observed evidence.
This final research effort will allow for forensic science researchers to leverage new machine learning methods from the AI community in a statistically rigorous manner to the interpretation of complex evidence forms. By developing rigorous score-based methods that have WoE-like properties, we will be able to fuse sub-WoEs for different forensic modalities into an omnibus WoE. NCA/NCF
