Low accuracy of Bayesian latent class analysis for estimation of herd-level true prevalence under certain disease characteristics—An analysis using simulated data

Estimation of the true prevalence of infected individuals involves the application of a diagnostic test to a population and adjusting according to test performance, sensitivity and specificity. Bayesian latent class analysis for the estimation of herd and animal-level true prevalence, has become increasingly used in veterinary epidemiology and is particularly useful in incorporating uncertainty and variability into analyses in a flexible framework.

However, the approach has not yet been evaluated using simulated data where the true prevalence is known. Furthermore, using this approach, the within-herd true prevalence is often assumed to follow a beta distribution, the parameters of which may be modelled using hyperpriors to incorporate both uncertainty and variability associated with this parameter.

Recently however, the authors of the current study highlighted a potential issue with this approach, in particular, with fitting the distributions and a tendency for the resulting distribution to invert and become clustered at zero. Therefore, the objective of the present study was to evaluate commonly specified models using simulated datasets where the herd-level true prevalence was known.

The specific purpose was to compare findings from models using hyperpriors to those using a simple beta distribution to model within-herd prevalence. A second objective was to investigate sources of error by varying characteristics of the simulated dataset. Mycobacterium avium subspecies paratuberculosis infection was used as an example for the baseline dataset.

Data were simulated for 1000 herds across a range of herd-level true prevalence scenarios, and models were fitted using priors from recently published studies. The results demonstrated poor performance of these latent class models for diseases characterised by poor diagnostic test sensitivity and low within-herd true prevalence.

All variations of the model appeared to be sensitive to the prior and tended to overestimate herd-level true prevalence. Estimates were substantially improved in different infection scenarios by increasing test sensitivity and within-herd true prevalence. The results of this study raise questions about the accuracy of published estimates for the herd-level true prevalence of paratuberculosis based on serological testing, using latent class analysis.

This study highlights the importance of conducting more rigorous sensitivity analyses than have been carried out in previous analyses published to date.