Effects of dependent errors in the assessment of diagnostic test performance

Latent class models can be used to assess diagnostic test performance when there is no perfectly accurate gold standard test available for comparison. These models usually assume independent errors between the tests, conditional on the true disease state of the subject. Maximum likelihood estimates of the prevalence of the disease and the error rates of diagnostic tests are then obtained. This paper examines the impact of error dependencies between binary diagnostic tests on the parameter estimates obtained from the latent class models. The independence model often gives parameter estimates having relatively small bias, but in some situations (for example, when disease prevalence is low and the tests have low specificity, such as in population screening) the bias may be more serious.