A comparative study of qualitative and quantitative models used to interpret complex STR DNA profiles.

The investigation of the performance of models to interpret complex DNA profiles is best undertaken using real DNA profiles. Here we used a data set to reflect the variety typically encountered in real casework. The "crime-stains" were constructed from known individuals and comprised a total of 59 diverse samples: pristine DNA/DNA extracted from blood, 2-3 person mixtures, degradation/no-degradation, differences in allele sharing, dropout/no dropout, etc. Two siblings were also included in the test-set in order to challenge the systems. Two kinds of analyses were performed, namely tests on whether a person of interest is a contributor based on weight-of-evidence (likelihood ratio) calculations, and deconvolution test to estimate the profile of unknown constituent parts. The weight-of-evidence analyses compared LRmix Studio with EuroForMix including exploration of the effect of applying an ad hoc stutter-filter. For the deconvolution analysis we compared EuroForMix with LoCIM-tool. When we classified persons of interests into being true contributors or non-contributors, we found that EuroForMix, overall, returned a higher true positive rate for the same false positive levels compared to LRmix. In particular, in cases with an unknown major component, EuroForMix was more discriminating for mixtures where the person of interest was a minor contributor. Comparing deconvolution of major contributors we found that EuroForMix overall performed better than LoCIM-tool.