Abstract
In this thesis, we examined and applied several Bayesian methods in veterinary clinical epidemiological studies, both for their modelling flexibility in handling complex empirical problems and for the possibility to include external evidence through prior distributions.
We started our research in the context of
... read more
diagnostic test characteristics estimation and investigated two latent class methods: the classical Bayesian Hui-Walter (HW) and the multilevel logistic regression (LR) latent class models. In chapter 2 we examined both approaches within simulations. Results from various scenarios showed that LR models were in many situations preferable over HW models to estimate test characteristics in the absence of a gold standard. In addition, LR models with the animal level risk factor provided robust estimate for the odds ratio of the risk factor. This conclusion was verified in the real-world example in chapter 3 where test characteristics of the single intradermal comparative cervical tuberculin (SICCT) test for the diagnosis of bovine tuberculosis (bTB) in Northern Ireland were evaluated. Both methods showed comparable posterior medians for the SICCT test sensitivity and specificity, but the LR model provided narrower 95% posterior credible intervals.
We further explored the use of Bayesian methods in the context of multilevel diagnostic prediction models. In chapter 4 we developed a new method to include prior knowledge through the random effects distribution for prediction in future clusters. Results showed that Bayesian prediction models using expert opinion as informative priors had better predictive performance than the frequentist model that removed the random effects. Results also indicated that incorporation of more precise expert opinion led to better predictions at the individual as well as at the cluster level. This prediction approach was subsequently applied in a clinical study in chapter 5 to diagnose subclinical ketosis (SCK) in early lactation dairy cows. The elicited expert opinion from a bovine specialist on the SCK risk for each farm was incorporated as informative priors for the random effects. Results showed however no clinically important improvement in prediction at the individual cow as well as at the herd level.
The possibility to incorporate external evidence as informative priors in clinical analyses was further investigated in minor species. In chapter 6, we adopted the Bayesian power prior method to re-analyze a Dutch equine clinical trial and evaluate the effect of oral Glucosamine (GS) and Chondroitin Sulfate (CS) on stiff joints in elderly horses. Nine cross-species studies were weighted by an equine expert for their clinical relevance on the equine research question and used to specify the power prior. Results showed no effect of the nutraceuticals which supported the original conclusion from the equine trial. However, sensitivity analyses showed mixed results, which reflected the current heterogeneity and uncertainty on the clinical effect size of GS and CS for stiff joints in elderly individuals.
This thesis shows that in addition to the well-accepted Bayesian analysis with non-informative priors, informative priors could be a useful tool as well in the clinical practice of veterinary epidemiology.
show less