Evaluation of Bayesian Hui-Walter and logistic regression latent class models to estimate diagnostic test characteristics with simulated data.

Estimation of the accuracy of diagnostic tests in the absence of a gold standard is an important research subject in epidemiology (Dohoo et al., 2009). One of the most used methods the last few decades is the Bayesian Hui-Walter (HW) latent class model (Hui and Walter, 1980). However, the classic HW models aggregate the observed individual test results to the population level, and as a result, potentially valuable information from the lower level(s) is not fully incorporated. An alternative approach is the Bayesian logistic regression (LR) latent class model that allows inclusion of individual level covariates (McInturff et al., 2004). In this study, we explored both classic HW and individual level LR latent class models using Bayesian methodology within a simulation context where true disease status and true test properties were predefined. Population prevalences and test characteristics that were realistic for paratuberculosis in cattle (Toft et al., 2005) were used for the simulation. Individual animals were generated to be clustered within herds in two regions. Two tests with binary outcomes were simulated with constant test characteristics across the two regions. On top of the prevalence properties and test characteristics, one animal level binary risk factor was added to the data. The main objective was to compare the performance of Bayesian HW and LR approaches in estimating test sensitivity and specificity in simulated datasets with different population characteristics. Results from various settings showed that LR models provided posterior estimates that were closer to the true values. The LR models that incorporated herd level clustering effects provided the most accurate estimates, in terms of being closest to the true values and having smaller estimation intervals. This work illustrates that individual level LR models are in many situations preferable over classic HW models for estimation of test characteristics in the absence of a gold standard.