Diagnostic research in the absence of a gold standard

Abstract

A key problem faced in many diagnostic studies is the absence of a single reference standard that can accurately distinguish between patients with and without the target condition. This problem is commonly referred to as: absence of a gold standard. In the absence of a gold standard, the classification of the target condition is prone to error. When these classification errors are ignored, the evaluation of the accuracy of the test(s) under evaluation or the estimation of the prevalence of the target condition can be severely biased. This thesis examines potential solutions that are aimed at alleviating the problems associated with the absence of a gold standard. In particular, we focus on latent class modeling and composite reference standards: two commonly used methods in diagnostic test evaluation literature. There is currently no consensus on how to best address the frequently encountered absence of a gold standard problem. In this thesis we show that composite reference standards can cause large biases in inferences about diagnostic test accuracy and target condition disease prevalence. It is also shown that latent class analysis should be used with caution. We suggest various approaches to improve latent class analysis by augmenting the diagnostic test data. Future research should be directed at further elucidating the merits and pitfalls of statistical modeling to account for the absence of a gold standard.