Quantifying the weight of fingerprint evidence using an ROC-based Approximate Bayesian Computation algorithm
share
The Bayes factor has been advocated to quantify the weight of forensic evidence; however, in many situations, the likelihood functions required to characterise complex pattern data do not exist and Bayes factors cannot be evaluated directly. Approximate Bayesian Computation allows assigning Bayes factors in these settings. We propose a novel algorithm that relies on the Receiver Operating Characteristic curve to address issues associated with using ABC for model selection. Our algorithm produces more stable results than traditional ones and makes it easier to monitor convergence as the number of simulations increase. We use our method to revisit a previously published fingerprint model.
READ FULL TEXT
