Predicting farm-level animal populations using environmental and socioeconomic variables

Abstract

Accurate information on the geographic distribution of domestic animal populations helps biosecurity authorities to efficiently prepare for and rapidly eradicate exotic diseases, such as Foot and Mouth Disease (FMD). Developing and maintaining sufficiently high-quality data resources is expensive and time consuming. Statistical modelling of population density and distribution has only begun to be applied to farm animal populations, although it is commonly used in wildlife ecology. We developed zero-inflated Poisson regression models in a Bayesian framework using environmental and socioeconomic variables to predict the counts of livestock units (LSUs) and of cattle on spatially referenced farm polygons in a commercially available New Zealand farm database, Agribase. Farm-level counts of cattle and of LSUs varied considerably by region, because of the heterogeneous farming landscape in New Zealand. The amount of high quality pasture per farm was significantly associated with the presence of both cattle and LSUs. Internal model validation (predictive performance) showed that the models were able to predict the count of the animal population on groups of farms that were located in randomly selected 3km zones with a high level of accuracy. Predicting cattle or LSU counts on individual farms was less accurate. Predicted counts were statistically significantly more variable for farms that were contract grazing dry stock, such as replacement dairy heifers and dairy cattle not currently producing milk, compared with other farm types. This analysis presents a way to predict numbers of LSUs and cattle for farms using environmental and socio-economic data. The technique has the potential to be extrapolated to predicting other pastoral based livestock species.

Keywords: Biosecurity; Markov Chain Monte Carlo simulation; Spatial epidemiology; Species distribution modelling; Zero-Inflated Poisson Regression.