Regression test of various versions of STRmix

TL;DR

This paper compares different versions of STRmix over seven years, showing that despite incremental improvements, the core model's stability results in consistent likelihood ratios for true contributors.

Contribution

It provides a regression analysis demonstrating the stability of STRmix's core models across multiple versions from 2.3 to 2.9 over seven years.

Findings

High similarity in likelihood ratios for true contributors across versions

Core models of STRmix have remained stable over time

Incremental improvements did not significantly alter core results

Abstract

STRmix has been in operational use since 2012 for the interpretation of forensic DNA profiles. During that time incremental improvements have been made to the modelling of variance for single and composite peaks, and drop-in. The central element of the algorithm has remained as the Markov chain Monte Carlo (MCMC) based on the Metropolis-Hastings algorithm. Regression experiments are reported comparing versions V2.3 to V2.9. There is a high degree of similarity in LRs for true contributors across multiple versions of STRmix that span 7 years of development (from V2.3 to V2.9). This is due to the fact that the underlying core models of STRmix that are used to describe DNA profile behaviour have remained stable across that time.