Bayesian latent class analysis when the reference test is imperfect

Abstract

Latent class analysis (LCA) has allowed epidemiologists to overcome the practical constraints faced by traditional diagnostic test evaluation methods, which require both a gold standard diagnostic test and ample numbers of appropriate reference samples. Over the past four decades, LCA methods have expanded to allow epidemiologists to evaluate diagnostic tests and estimate true prevalence using imperfect tests over a variety of complex data structures and scenarios, including during the emergence of novel infectious diseases. The objective of this review is to provide an overview of recent developments in LCA methods, as well as a practical guide to applying Bayesian LCA (BLCA) to the evaluation of diagnostic tests. Before conducting a BLCA, the suitability of BLCA for the pathogen of interest, the availability of appropriate samples, the number of diagnostic tests, and the structure of the data should be carefully considered. While formulating the model, the model's structure and specification of informative priors will affect the likelihood that useful inferences can be drawn. With the growing need for advanced analytical methods to evaluate diagnostic tests for newly emerging diseases, LCA is a promising field of research for both the veterinary and medical disciplines.

Validation of an Indirect Immunofluorescence Assay and Commercial Q Fever Enzyme-Linked Immunosorbent Assay for Use in Macropods

“…There are a number of challenges related to diagnostic test validation for wildlife species, including access to reference positive and negative control samples representative of the target population(s) for which the test is being validated, the availability of sufficient numbers of samples to allow parameters to be estimated with certainty, the availability of a species-appropriate gold standard reference test, and detailed knowledge of species-specific pathophysiology and immunology ( 53 , 54 ). In this study, we used a Bayesian latent class analysis to overcome the lack of known positive control samples, the absence of a gold standard reference test, and uncertainty regarding the true disease status of many of the individual animals sampled ( 38 ). The IFA developed for this study clearly outperformed the commercially available IDVet ELISA in terms of diagnostic sensitivity, with highly comparable diagnostic specificity.…”

“…A Bayesian latent class model was developed to estimate the diagnostic sensitivity and specificity of each test in the absence of a gold standard, assuming conditional dependence between the tests ( 38 ). Prior information about the diagnostic specificity and sensitivity of each assay was modeled using unimodal beta distributions based on published data.…”

Evaluation of diagnostic test procedures for SARS‐CoV‐2 using latent class models

“…These models work by formulating a mathematical expression (a likelihood), for the observed test results with the diagnostic test parameters to be estimated. LCM based on the Hui–Walter paradigm was proposed over 40 years ago 3 and has since become widely used for evaluating diagnostic tests within the veterinary literature 4–7 . More recently, LCM has also begun to be used within the human medical field 8,9 .…”

“…LCM based on the Hui–Walter paradigm was proposed over 40 years ago 3 and has since become widely used for evaluating diagnostic tests within the veterinary literature. 4 , 5 , 6 , 7 More recently, LCM has also begun to be used within the human medical field. 8 , 9 The LCM is fit to a table of frequencies for each test type, result, and population group.…”

Serological Hendra Virus Diagnostics Using an Indirect ELISA-Based DIVA Approach with Recombinant Hendra G and N Proteins

“…For estimating the diagnostic specificity (DSp) with 95% confidence and absolute precision of ±2%, assuming sensitivity (Se) of the DIVA assays (HeV-G and HeV-N) could be >99% and specificity (Sp) to be 98%, the minimum required sample size was estimated a priori to be 95 samples for estimating DSe and 188 samples for estimating DSp, using the R package 'epiR' [34]. A Bayesian latent class analysis is the OIE-recommended approach [35], where insufficient reference samples are available for such an analysis, making no assumption of the true disease status of the animals from which the samples are derived.…”