Using pseudogold standards and latent-class analysis in combination to evaluate the accuracy of three diagnostic tests

We previously reported our use of latent-class models to estimate the sensitivity (Se) and specificity (Sp) for each of three tests used to monitor farmed salmon for infectious salmon anaemia virus (ISAv). Those tests were reverse-transcriptase polymerase chain reaction (RT-PCR), virus isolation (VI), and an indirect immunofluorescent-antibody test (IFAT). We used tissues from 403 salmon from four populations presumed to have different prevalence of ISAv. However, no formal evaluation of the assumptions of conditional independence and constant accuracy had been carried out. In our present study, we "adjusted" that and used two pseudogold standards (a composite reference standard and a study pseudogold), as indicative of the true health status of each fish. The assumption of constant accuracy across populations was evaluated using separate random-effects logistic-regression models for fish classified as D+ or D− (disease positive or negative, according to the pseudogold standards) with study population included in the model to determine if it affected the probability of a positive test result. Where there was evidence of variation in test accuracy across populations, the issue was further investigated using separate latent-class models with informative priors for each study population. Our results suggested that only one PCR test had an accuracy that varied across populations. The assumption of conditional independence among tests was first evaluated using log-linear models of D+ and D− fish with significant interaction between test results indicative of conditional dependence. Latent-class models which incorporated up to two pairs of between-test dependencies were also fit using Bayesian methods. The two approaches showed considerable evidence of dependence between IFAT and VI and some evidence of dependence between one PCR and IFAT. Results obtained from both maximum-likelihood models and from Bayesian analyses of models allowing for conditional dependence between two pairs of tests were consistent with those obtained with the pseudogold standards. The results suggest that pseudogold standards can help in choosing a correct dependence structure and should be used in combination with latent-class models.