Statistical modelling methodology for investigating risk factors of antimicrobial use and resistance

: applied to UK dairy farms

Abstract

Antimicrobial resistance (AMR) in food-producing animals is a key concern for global health and food security. To tackle this problem, we must first understand the risk factors for antimicrobial use and AMR development. This thesis presents new Bayesian models for examining risk factors for both antimicrobial use and resistance, and then applies these models to data from dairy farms in Southwest England, UK.

Veterinary sales and milk recording data were used to investigate trends in antimicrobial sales to 124 dairy farms between 2010-18 from routinely collected data. A natural language processing algorithm was harnessed for semi-automated linking of sales data to antimicrobial product specifications. Significant reductions in sales of AMs were observed over the study period (41%). Predictive projective feature selection was employed to identify potential risk factors for antimicrobial sales and strong evidence was found for associations with five predictors which included antimicrobial purchase frequency and average parity.

Secondly, a non-linear Bayesian model coupled to a microbiology technique for analysing proportional resistance was developed to enable the statistical modelling of environmental AMR load for the first time. This model was compared against existing methods and was shown to better incorporate uncertainty and reduce the biasing of risk factor estimates produced by the varying bacterial abundance in the samples. Using data from a previous study into antimicrobial resistance E. coli in the environment of dairy farms, the model was then used to investigate the effect of climate on antimicrobial resistance, finding evidence that both temperature and relative humidity, as well as the interaction between them, were associated with resistance to four of the antimicrobials tested.

These findings improve our understanding of both antimicrobial use and resistance in UK dairy farming, and the flexible Bayesian modelling methods presented have the potential to underpin future research into AMR and its surveillance.