Bayesian methods for analysing pesticide contamination with uncertain covariates

Abstract

Two chemical properties of pesticides are thought to control their environmental fate. These are the adsorption coefficient k(_oc) and soil half-life t(^soil_1/2). This study aims to demonstrate the use of Bayesian methods in exploring whether or not it is possible to discriminate between pesticides that leach from those that do not leach on the basis of their chemical properties, when the monitored values of these properties are uncertain, in the sense that there are a range of values reported for both k(_oc) and t(^soil_1/2) - The study was limited to 43 pesticides extracted from the UK Environment Agency (EA) where complete information was available regarding these pesticides. In addition, analysis of data from a separate study, known as "Gustafson's data”, with a single value reported for k(_oc) and t(^soil_1/2) was used as prior information for the EA data. Bayesian methods to analyse the EA data are proposed in this thesis. These methods use logistic regression with random covariates and prior information derives from (i) available United States Department of Agriculture (USDA) data base values of k(_oc) and t(^soil_1/2) for the covariates and (ii) Gustafson's data for the regression parameters. They are analysed by means of Markov Chain Monte Carlo (MCMC) simulation techniques via the freely available WinBUGS software and R package. These methods have succeeded in providing a complete or a good separation between leaching and non-leaching pesticides