Modelling the prevalence of wildlife diseases using simulated diagnostic test data

Abstract

Bayesian Latent Class Models (BLCMs) are algorithms that are used to infer disease prevalence when true disease statuses and gold-standard diagnostic tests are not available. However, limited attention has been given to the specification and validation of BLCMs, which are necessary if credible estimates of diagnostic test performance and disease prevalence are to result.

Across six technical chapters, this thesis investigates the fundamental principles of specification and validation via a series of experiments that apply BLCMs to ante-mortem diagnostic test data. To achieve this, simulated arrays of diagnostic test data are generated to reflect the reality of the imperfect trapping and testing efforts that take place in nature. Moreover, the classic Hui-Walter algorithm is generalised within a Bayesian framework to unlock the capability of BLCMs to handle both varying prior information and varying hypotheses simultaneously.

Methods to validate BLCMs are developed and then scaled up across a wide range of possible diagnostic testing scenarios via the creation of procedures to explore high-dimensional parameter spaces. For the first time, it is demonstrated that the credibility of BLCM inferences is in fact predictable.

Among the key findings discovered are dependence structures that are critical to the identifiability of BLCMs; these structures are uncovered at the limits of parameter spaces, and between the means and variances of the inferred statistics. Accordingly, methods are explored to mitigate for these structures as a further prerequisite to obtaining credible estimates.

Attention then turns to testing the core assumptions used to specify the generalised Hui-Walter algorithm. The assumptions about where the true values of diagnostic test performance and disease prevalence exist are removed, and the resulting sensitivity analyses provide confirmation that the findings reported throughout the thesis are indeed generalisable, even to unusual testing scenarios.

With a rigorous validation protocol in place, a novel class of time-dependent BLCMs is specified, and then provided with data from one of the world’s longest running wildlife studies. New and rigorously validated inferences of disease prevalence are revealed, and anecdotal trends are corroborated, highlighting the real-world applications of this thesis.