National Institute of Standards and Technology
Assessing High Dimensional Evidence: Scores, Probability, and Scientific Validity
Many types of forensic evidence are high dimensional (e.g. patterns like fingerprints or tool marks, mass or Raman spectroscopy). Analyses for high-dimensional evidence often involve comparing a crime scene sample (Q) and a control sample (C) collected from a person (or object) of interest and summarizing the observed degree of correspondence in a low (generally one) dimensional statistic, or score (S). The appropriate probabilistic analysis of such scores is a subject of ongoing debate among the statistical forensics community. The most commonly proposed method is to assess a score-based likelihood ratio (SLR), which seeks to represent the probability of observing score S when Q and C are known to come from the same source divided by the probability of observing score S when Q and C are known to come from different sources. This approach has received strong criticism for several reasons. In this talk, I will propose a different role for SLRs to improve both discrimination power and the degree of scientific validity by considering control samples from a large pool of sources without knowledge of which control sample corresponds to the person (or object) of interest.
Refreshments at 3:45pm in Snedecor 2101.