CMStatistics 2021: Start Registration
View Submission - CMStatistics
B0305
Title: Disentangling zero-inflation from overdispersion in statistical biodosimetry Authors:  Jochen Einbeck - Durham University (United Kingdom) [presenting]
Adam Errington - Durham University (United Kingdom)
Jonathan Cumming - Durham University (United Kingdom)
Paul Wilson - University Of Wolverhampton (United Kingdom)
Abstract: It is well known that one of the ways that overdispersion can be triggered in count data (regression) models is by the presence of `excess' zeros, i.e. more zeros than could be expected under the employed count data distribution. This phenomenon is exploited in the detection of partial body exposures in radiation biomarkers. For instance, counts of dicentric chromosomes per cell usually adhere well to the Poisson law, unless the exposure of the body to the ionizing radiation was only partial. In that case, the non-irradiated cells will generally contribute just zero counts of chromosomal aberrations, resulting overall in an overdispersed distribution of counts. So, any overdispersion of the biomarker can be taken as evidence of partial exposure. The situation is more difficult when the radiation biomarker is overdispersed per se (that is, even under full body exposure), as, for instance, for gamma-H2AX protein foci. In this case, overdispersion is not indicative of partial exposure. Hence, what is required are methods that can separate the overdispersion resulting from the excess zeros from the `stem' overdispersion contributed by the exposed part. We will present several approaches to this problem, among these a negative binomial version of the `contaminated Poisson method', and a recently proposed diagnostic tool known as the `quantile band plot'.