Steele, CD;
(2016)
Statistical issues surrounding the analysis of forensic low-template DNA samples.
Doctoral thesis , UCL (University College London).
Preview |
Text
Steele__thesis.pdf.REDACTED.pdf Download (3MB) | Preview |
Abstract
Increased sensitivity of forensic DNA profiling over the last decade has led to increased stochasticity in the resulting profiles, causing difficulties for interpretation that were acknowledged by the Caddy Report [Caddy et al., 2008]. These difficulties were largely overcome with the adoption of statistical models allowing for dropout and dropin, but interpretation issues remain, several of which are tackled in this thesis. One such issue concerns the choice of allele frequency databases when the ethnic background of the true source of the crime scene DNA is unknown. I propose a heuristic for choosing a single database and adjusting the likelihood ratio calculations to allow for the possibility that a different database may be more appropriate. Another issue in general, and specifically for the database choice heuristic, is the choice of an appropriate value for the population genetics parameter FST to account for distant relatedness between the alleged contributor and an alternative source of the DNA. I present empirical estimates of FST in worldwide populations, relative to the continental-scale reference databases that are used for UK forensic DNA profiles. In the last few years many software packages for the evaluation of low-template DNA samples have emerged, including likeLTD originally developed by my supervisor Prof Balding but greatly improved and reprogrammed by myself as part of my PhD work. There remains little consensus on how to validate these software packages. I present a method of validation based on the use of multiple-replicate crime stain profiles. It relies on the intuition that sufficient replicates of even very noisy DNA profiling runs eventually generate the same information as a single high-quality replicate. I show that likeLTD performs well when assessed by this approach. Finally, I present a new statistical model that extends likeLTD to incorporate the peak height information in a crime scene profile. I show results based on simulation and laboratory trials verifying the good performance of the new model in improved discrimination between true and false hypotheses.
Archive Staff Only
 |
View Item |
