Utilising allelic dropout probabilities estimated by logistic regression in casework

Abstract

Some advanced methods for DNA profile interpretation require a probability for the event of dropout. Methods have been suggested based on logistic regression. Two of these respectively use a proxy for template that is constant across loci and one that is modelled using an exponential curve. Both of these methods allow different modelling constants from each locus. A variant of the model using an exponential curve is discussed. This variant constrains the constants to be the same for every locus. We test these two methods and the variant by developing the constants (training) on one set of data and testing them on another. This mimics the likely use in casework. We find that the new variant appears to be the most useful in that it performs better than the other two options when trained on one data set and used on another. The hypothesised reason for this is that locus to locus variation in amplification efficiency varies with time, multimix batch, or from sample to sample.