Combining qualitative and quantitative diagnostic tests in the absence of a gold standard – what to do with missing data? GB virus C (GBV-C) viremia as an example

Objectives: Using multiple methods to test for a virus when no standardized test exists introduces several potential sources of variation. We investigated how discrepancies from multiple tests and/or missing data can be evaluated and reconciled statistically.

Methods: Bayesian Latent Class Analysis was used to model the responses from the different tests. Each qualitative response had three categories (positive, negative, missing) and the quantitative response was divided into four categories (none or low, moderate, high, missing). The model was parameterized by the prevalence of virus, sensitivity and specificity of each test, and probability of each test being missing. A WinBUGS program was developed and used to estimate parameters in the model and to classify each sample.

Results: RT-PCR amplification of four GBV-C genome regions (E2, 5'-NTR, NS3 and NS5A) was performed on 381 serum samples from 139 HIV infected subjects: not all tests were run on all samples. Additionally, all samples were tested with a quantitative real-time RT-PCR. GBV-C RNA prevalence based on each single test was 49%, 77%, 78%, and 79% respectively. Under the Bayesian Latent Class Model, the estimated prevalence of GBV-C is 45% in this population; the estimated sensitivities of the four qualitative tests were 92%, 83%, 93%, 99%, and the specificities were 98%, 31%, 32%, 35% respectively. 175 out of 381 samples were classified as positive.

Conclusions: Bayesian Latent Class Analysis can be used to evaluate and reconcile different tests in the absence of a gold standard. E2 is the best single test in the GBV-C testing.