Utilising Allelic Dropout Probabilities Estimated by Logistic Regression in Casework

For the two previously proposed models (Tvedebrink et al., 2009; Tvedebrink et al., 2012), one (T1) uses a proxy for template that is constant across loci, and the other model (T2) uses an exponential curve. In the current paper, Tvedebrink et al. introduce a third model (T'2) that is a variant of the exponential curve model. In almost all tests performed on the Identifiler sets, T1 and T2 models produced the highest mean log(likelihood) of allelic dropout probabilities in the training sets; however, T'2 produced the highest mean log(likelihood) in the test sets. The mean in the training set was always higher than in the test set, regardless of the model used. The authors interpret this as meaning that a locus effect exists, but this changes either batch to batch of mutlimix, or profile to profile, or over time by ageing of the camera or other laboratory changes, such as cleaning agents. For the PowerPlex 21 data, T1 regularly gave the highest mean log(likelihood) in both the training and test sets. The authors interpret this as meaning that pristine source data are too good to show the expected degradation effect, and are not suitable to train these logistic models. Of the three models, T'2 trained on casework data is narrowly the best model for immediate use in casework due to its portability, since it produced the highest log(likelihood) in test sets more often; however, the authors advise that further development is required in the application of locus specific effects, which are likely to vary from profile to profile or time to time. 2 tables and 7 references