Epidemiological studies assessing disease prevalence are critically important to both identification of new or re-emerging pathogens and control of endemic pathogens in humans and animals (including zoonosis and food borne outbreaks) and are therefore ubiquitous worldwide.
Interpretation of the resulting data and comparison between studies is often challenging due to differences in (i) the sampling scheme and (ii) diagnostic methods that are used, which often are of varying accuracy. An additional key challenge lies in the definition of ‘disease’ that is targeted in each case.
Countries typically collect data in a way that is best suited for their specific needs, and non-standardised sampling strategies and diagnostic methods produce prevalence estimates that cannot be directly compared.
Hence, the need for harmonization in terms of monitoring and reporting disease occurrence will eventually improve comparability between the analytical findings derived from different studies.
Diagnostic tests should be evaluated in each population and for each targeted condition. Proper evaluation of tests using simple gold-standard methods requires data that are extremely expensive to obtain and may not even be feasible for infections with a long latent period. When a reference standard does not exist a viable alternative for robust evaluation of diagnostic test accuracy involves the use of latent class models (LCMs) that do not require knowledge of the true disease status of individuals or the use of a gold-standard reference test.
These latent class methods were pioneered by Hui and Walter (1980) and thorough discussion of their applicability in diagnostic accuracy studies was first given by Walter and Irwig (1988). Within the last decade extension of these methods for prevalence estimation and certification of disease freedom has evolved considerably.
Moreover, Bayesian LCMs (BLCMs) for diagnostic accuracy and prevalence estimation studies have been successfully implemented in a Bayesian framework for over 20 years. A Bayesian approach is necessary when using non-identifiable LCMs (i.e., statistical models that, broadly speaking, contain test accuracy and/or prevalence parameters that cannot be uniquely estimated by the data alone). BLCMs are widely used because of their flexibility, the ease of interpretation of their results, and the availability of user-friendly software.
Key elements that must be addressed in BLCMs are:
- the absence of a reference test,
- the need to identify the condition that the tests under evaluation are targeting,
- an explicit description of the specified BLCM structure and (iv) a clear and justified specification of priors.
Importantly, standards for the reporting of diagnostic accuracy studies that use Bayesian Latent Class Models– termed STARD-BLCM – have been recently published (Kostoulas et al., 2017) with the aim to facilitate improved quality of reporting on the design, conduct and results of diagnostic accuracy studies that use BLCMs.