Calibration performance of DNAStatistX in a laboratory setting and considerations for likelihood ratio thresholds

A previous study examining calibration and discrimination performance highlighted the need for caution when interpreting "low" likelihood ratios (LRs) derived from maximum likelihood estimate-based models DNAStatistX and EuroForMix [1]. The study reported that calibration performance was dependent on the dataset, dataset size and the subpopulation correction factor (Fst). In the worst case scenario (smallest dataset and Fst 0.01) miscalibration of LRs occurred up to LR ∼1000. In the best case scenario (largest dataset and Fst 0.03) there were signs of miscalibration up to LR ∼100 but not above. In the current study, the discrimination power and calibration performance were examined for DNAStatistX using a dataset that more closely reflects our casework practice. This involved analysing PowerPlex® Fusion 6C data using two different analytical threshold sets and up to three PCR replicate profiles in the LR calculation. The results showed calibration performance that was comparable or better than previous findings for maximum likelihood based (MLE) models. The use of two different sets of analytical thresholds yielded similar results. Calibration performance decreased when replicate profiles were combined in the LR calculation. Additionally, this study demonstrates that using per-dye LRs to assess calibration performance can be beneficial, especially when the dataset size is limited. Overall, the findings support previous research, suggesting that setting a lower threshold for reporting is useful when using MLE-based models. Ideally, the threshold is as low as possible as that may avoid overlooking valuable evidence. An LR value of 1000 seems supported by the data.