Poul Svante Eriksen

Estimating the probability of allelic drop-out of STR alleles in forensic genetics

In crime cases with available DNA evidence, the amount of DNA is often sparse due to the setting of the crime. In such cases, allelic drop-out of one or more true alleles in STR typing is possible. We present a statistical model for estimating the per locus and overall probability of allelic drop-out using the results of all STR loci in the case sample as reference. The methodology of logistic regression is appropriate for this analysis, and we demonstrate how to incorporate this in a forensic genetic framework.

Statistical model for degraded DNA samples and adjusted probabilities for allelic drop-out

DNA samples found at a scene of crime or obtained from the debris of a mass disaster accident are often subject to degradation. When using the STR DNA technology, the DNA profile is observed via a so-called electropherogram (EPG), where the alleles are identified as signal peaks above a certain level or above a signal to noise threshold. Degradation implies that these peak intensities decrease in strength for longer short tandem repeat (STR) sequences. Consequently, long STR loci may fail to produce peak heights above the limit of detection resulting in allelic or locus drop-outs. In this paper, we present a method for measuring the degree of degradation of a sample and demonstrate how to incorporate this in estimating the probability of allelic drop-out. This is done by extending an existing method derived for non-degraded samples. The performance of the methodology is evaluated using data from degraded DNA, where cases with varying amounts of DNA and levels of degradation are investigated.

Allelic drop-out probabilities estimated by logistic regression—Further considerations and practical implementation

We discuss the model for estimating drop-out probabilities presented by Tvedebrink et al. [7] and the concerns, that have been raised. The criticism of the model has demonstrated that the model is not perfect. However, the model is very useful for advanced forensic genetic work, where allelic drop-out is occurring. With this discussion, we hope to improve the drop-out model, so that it can be used for practical forensic genetics and stimulate further discussions. We discuss how to estimate drop-out probabilities when using a varying number of PCR cycles and other experimental conditions.

Decomposition and Invariance of Measures and Statistical Transformation Models

1. Introduction.- 2. Topological groups and actions.- 3. Matrix Lie groups.- 4. Invariant, relatively invariant, and quasi-invariant measures.- 5. Decomposition and factorization of measures.- 6. Construction of invariant measures.- 7. Exterior calculus.- 8. Statistical transformation models.- Further results and exercises.- References, with author index.- Notation index.

Identifying contributors of DNA mixtures by means of quantitative information of STR typing

Estimating the weight of evidence in forensic genetics is often done in terms of a likelihood ratio, LR. The LR evaluates the probability of the observed evidence under competing hypotheses. Most often, probabilities used in the LR only consider the evidence from the genomic variation identified using polymorphic genetic markers. However, modern typing techniques supply additional quantitative data, which contain very important information about the observed evidence. This is particularly true for cases of DNA mixtures, where more than one individual has contributed to the observed biological stain. This article presents a method for including the quantitative information of short tandem repeat (STR) DNA mixtures in the LR. Also, an efficient algorithmic method for finding the best matching combination of DNA mixture profiles is derived and implemented in an on-line tool for two- and three-person DNA mixtures. Finally, we demonstrate for two-person mixtures how this best matching pair of profiles can be used in estimating the likelihood ratio using importance sampling. The reason for using importance sampling for estimating the likelihood ratio is the often vast number of combinations of profiles needed for the evaluation of the weight of evidence. Online tool is available at http://people.math.aau.dk/~tvede/dna/.

3D visual data mining: goals and experiences

The visual exploration of large databases raises a number of unresolved inference problems and calls for new interaction patterns between multiple disciplines--both at the conceptual and technical level. We present an approach that is based on the interaction of four disciplines: database systems, statistical analyses, perceptual and cognitive psychology, and scientific visualization. At the conceptual level we offer perceptual and cognitive insights to guide the information visualization process. We then choose cluster surfaces to exemplify the data mining process, to discuss the tasks involved, and to work out the interaction patterns.

Cluster analysis of European Y-chromosomal STR haplotypes using the discrete Laplace method

The European Y-chromosomal short tandem repeat (STR) haplotype distribution has previously been analysed in various ways. Here, we introduce a new way of analysing population substructure using a new method based on clustering within the discrete Laplace exponential family that models the probability distribution of the Y-STR haplotypes. Creating a consistent statistical model of the haplotypes enables us to perform a wide range of analyses. Previously, haplotype frequency estimation using the discrete Laplace method has been validated. In this paper we investigate how the discrete Laplace method can be used for cluster analysis to further validate the discrete Laplace method. A very important practical fact is that the calculations can be performed on a normal computer. We identified two sub-clusters of the Eastern and Western European Y-STR haplotypes similar to results of previous studies. We also compared pairwise distances (between geographically separated samples) with those obtained using the AMOVA method and found good agreement. Further analyses that are impossible with AMOVA were made using the discrete Laplace method: analysis of the homogeneity in two different ways and calculating marginal STR distributions. We found that the Y-STR haplotypes from e.g. Finland were relatively homogeneous as opposed to the relatively heterogeneous Y-STR haplotypes from e.g. Lublin, Eastern Poland and Berlin, Germany. We demonstrated that the observed distributions of alleles at each locus were similar to the expected ones. We also compared pairwise distances between geographically separated samples from Africa with those obtained using the AMOVA method and found good agreement.

Weight of the evidence of genetic investigations of ancestry informative markers

Ancestry-informative markers (AIMs) are markers that give information about the ancestry of individuals. They are used in forensic genetics for predicting the geographic origin of the investigated individual in crime and identification cases. In the exploration of the genogeographic origin of an AIMs profile, the likelihoods of the AIMs profile in various populations may be calculated. However, there may not be an appropriate reference population in the database. The fact that the likelihood ratio (LR) of one population compared to that of another population is large does not imply that any of the populations is relevant. To handle this phenomena, we derived a likelihood ratio test (LRT) that is a measure of absolute concordance between an AIMs profile and a population rather than a relative measure of the AIMs profiles likelihood in two populations. The LRT is similar to a Fishers exact test. By aggregating over markers, the central limit theorem suggests that the resulting quantity is approximately normally distributed. If only a few markers are genotyped or if the majority of the markers are fixed in a given population, the approximation may fail. We overcome this using importance sampling and show how exponential tilting results in an efficient proposal distribution. By simulations and published AIMs profiles, we demonstrate the applicability of the derived methodology. For the genotyped AIMs, the LRT approach achieves the nominal levels of rejection when tested on data from five major continental regions.

Asserting Real-Time Emotions through Cued-Recall: Is it Valid?

Asserting emotions through free-recall is commonly used to evaluate user experience (UX) of interactive systems. From psychology we know that free-recall of emotions leads to a significant memory bias where participants rely on a few of the most intense episodes when asserting an overall experience. It is argued that cued-recall can reduce the memory bias in UX evaluations. Yet, this has not been studied empirically. We present a systematic empirical study based on 38 participants. We measured emotions in terms of objective galvanic skin responses (GSR) and subjective Self-Assessment Manikin (SAM) ratings. We found significant correlations between emotions experienced in real-time and those experienced during cued-recall. This validates the use of cued-recall for UX evaluations. An implication is that HCI researchers and practitioners now have cued-recall as an alternative that significantly reduces the memory bias and enables highly detailed measurements of emotions while not disturbing participants during system interaction.

The multivariate Dirichlet-multinomial distribution and its application in forensic genetics to adjust for subpopulation effects using the θ-correction.

In this paper, we discuss the construction of a multivariate generalisation of the Dirichlet-multinomial distribution. An example from forensic genetics in the statistical analysis of DNA mixtures motivates the study of this multivariate extension. In forensic genetics, adjustment of the match probabilities due to remote ancestry in the population is often done using the so-called θ-correction. This correction increases the probability of observing multiple copies of rare alleles in a subpopulation and thereby reduces the weight of the evidence for rare genotypes. A recent publication by Cowell et al. (2015) showed elegantly how to use Bayesian networks for efficient computations of likelihood ratios in a forensic genetic context. However, their underlying population genetic model assumed independence of alleles, which is not realistic in real populations. We demonstrate how the so-called θ-correction can be incorporated in Bayesian networks to make efficient computations by modifying the Markov structure of Cowell et al. (2015). By numerical examples, we show how the θ-correction incorporated in the multivariate Dirichlet-multinomial distribution affects the weight of evidence.