Grey Coupland

The development of inspection protocols and mitigation strategies are critical components of plant and animal biosecurity measures. The inclusion of available detection data can greatly enhance the evidence base for these types of decisions. However, a key step in analysing these data is the choice of an appropriate statistical model. This paper focuses on determining an appropriate model for biosecurity border and post-border detection of non-indigenous species for Barrow Island, Australia under stringent biosecurity controls. This is a flagship biosecurity project that is under close national and international scrutiny. A range of standard models were compared, including standard and zero-inflated Poisson, negative binomial and log-normal alternatives. These models failed to adequately describe the key characteristics of the data, namely an excess of zero and single organism detections, a range of detections between two and a hundred organisms, and a few extreme values, ranging between 250 and 1000 organisms. Alternative models were explored, including: (i) modelling the censored data ignoring the zero and extremely large detections, (ii) a component where detections were modelled ignoring the zero detections and including outliers and finally, (iii) zero-inflated model with the complete data set inclusive of zero detections. The negative binomial based models consistently gave the best outcomes under different degrees of inflation or over-dispersion, but are limited in that they cannot provide information about the mechanisms underlying zero-inflation. A three component log normal mixture model was found to be the best fit as it addressed these issues. This study demonstrates the importance of model choice in analysing biosecurity data, and suggests that mixture models may be more appropriate than more standard distributions. The data set gathered at Barrow Island is, however, unique for biosecurity border and post-border detection biosurveillance. Given this, general inferences about the underlying phenomena made based on the model should be made with caution. Comparable datasets should be obtained and tested to validate the choice of model for this type of data.