Abstract
Bayesian finite mixture models, frequently referred to as Bayesian latent class models have become increasingly common for diagnostic test data in the absence of a gold standard test. Most Bayesian analyses in the veterinary literature have dealt with a dichotomised diagnostic outcome. The use of Bayesian finite mixture models for continuous test outcomes, such as sample to positive (S/P) ratios produced by an ELISA, is much less common, despite continuous models taking advantage of all of the information captured in the test outcome. This paper revisits the idea of the Bayesian finite mixture model and provides a practical guide for researchers who would like to use this approach for modelling continuous diagnostic outcomes as it preserves all information from the observed data. Synthetic datasets and a dataset from literature were analysed to illustrate that a mixture model with continuous diagnostic outcomes can be used to estimate true prevalence and to evaluate test sensitivity and specificity. In addition, directly modelling the continuous test outcomes rather than dichotomising them, means that optimal cut-offs can be defined based on the test purpose rather than being determined before testing. Moreover, as animals with higher scores are more likely to be infected, using continuous data allows test interpretation to be made at the individual animal level. In contrast, dichotomization treats all animals above a cut-off as having the same infection risk. This study demonstrates that dichotomisation is not a 'must' when using Bayesian latent class analysis for diagnostic test data, and suggests that latent class analysis using continuous test outcomes should be favoured when evaluating veterinary diagnostic tests producing continuous outcomes.