Testing methods for quantifying Monte Carlo variation for categorical variables in Probabilistic Genotyping

Abstract

Two methods for applying a lower bound to the variation induced by the Monte Carlo effect are trialled. One of these is implemented in the widely used probabilistic genotyping system, STRmix Neither approach is giving the desired 99% coverage. In some cases the coverage is much lower than the desired 99%. The discrepancy (i.e. the distance between the LR corresponding to the desired coverage and the LR observed coverage at 99%) is not large. For example, the discrepancy of 0.23 for approach 1 suggests the lower bounds should be moved downwards by a factor of 1.7 to achieve the desired 99% coverage. Although less effective than desired these methods provide a layer of conservatism that is additional to the other layers. These other layers are from factors such as the conservatism within the sub-population model, the choice of conservative measures of co-ancestry, the consideration of relatives within the population and the resampling method used for allele probabilities, all of which tend to understate the strength of the findings.

Uncertainty in probabilistic genotyping of low template DNA: A case study comparing STRMix™ and TrueAllele™

“…To compute a combined LR across all loci, TrueAllele simply multiplied the LRs for each locus and reported the product as the overall LR for the case. As shown in of the MCMC analysis [27,28]. The HPD procedure produces a probability interval (similar to a confidence interval).…”

“…The second adjustment made by STRMix involved computation of the highest posterior density (HPD) of the LR. This procedure is designed to account for uncertainty about the true value of the LR that arises from the possibility of sampling error in estimating allele frequencies and the possibility of run‐to‐run variation in the results of the MCMC analysis [27, 28]. The HPD procedure produces a probability interval (similar to a confidence interval).…”