Bayesian non-parametric models for zoonotic disease applications

Miller, Poppy and Diggle, Peter and Jewell, Christopher (2020) Bayesian non-parametric models for zoonotic disease applications. PhD thesis, Lancaster University.

Text (2020MillerPhD)
2020MillerPhD.pdf - Published Version
Available under License Creative Commons Attribution-NonCommercial-NoDerivs.
Download (17MB)

Abstract

Advanced statistical models are a key tool in developing interventions to reduce disease incidence, particularly in low resource settings. Infectious diseases often have complex infection processes and pathways, particularly zoonotic diseases which often have direct and indirect routes of infection. This makes epidemiological studies aimed at identifying and/or quantifying risk factors challenging, as they typically include complexities such as multi-level dependency structures, correlated covariates, missing data, and high noise. Often, data are only partially observed due to censoring and structurally missing information, and are often observational rather than the result of direct treatments. This thesis explores novel methods and models designed to tease out pathways and factors that contribute to risk of disease in humans for zoonotic pathogens. Chapter 2 develops a Bayesian non-parametric model to estimate the proportion of cases attributable to known sources of disease, and identify sub-types of pathogens which are unusually dangerous. This model was applied to a campylobacteriosis data set from New Zealand with results showing chicken from a single supplier was likely the source of approximately 70% of cases in the data set, and identified 9 particularly dangerous subtypes. Chapter 3 widens the scope to consider the relative contributions of many potential risk factors for disease (causal or not). Our model considers many environmental and social risk factors for leptospirosis in complex urban environments, including rat exposure. We estimate both rat exposure and leptospirosis risk using a Bayesian non-parametric cut model which correctly accounts for uncertainty in the rat exposure predictions. The results identify groups of high risk individuals, based on socio-economic data and environmental risk factors, that could be targeted using interventions. This chapter highlighted a significant limitation in many epidemiological studies which use inaccurate diagnostic techniques. Chapter 4 develops methodology to address this limitation by modelling within-host immune responses to pathogenic challenge. This was done by integrating a mechanistic ordinary differential equation model with a Bayesian censored noise model. Our model estimates expected changes in antibody levels after challenge with different pathogens, and indicates possible differences in immune response that may be responsible. The model is also able to estimate time of challenge at an individual level. The model is applied to a leptospirosis challenge data set in sheep, and shows significant differences in immune response to serovars Pomona and Hardjobovis. Integration of this methodology with epidemiological studies (like those in Chapters 2 and 3) will allow for more accurate estimation of relative risks and enable more effective intervention strategies to be developed.