Estimação de medidas de desempenho de testes diagnósticos via Inferência Bayesiana

Abstract

The use of diagnostic tests to identify the presence or absence of a particular disease in an individual is essential in the area of health. The performance of diagnostic tests is measured by their sensitivity and specificity. These measures are obtained by comparing the results of the diagnostic test with a reference test, called Gold Standard, which correctly classifies both diseased and non-diseased individuals. In the absence of a Gold Standard test, the sensitivity and specificity need to be estimated by alternative methods since the number of diseased individuals in the population is unknown. The objective of the study was to introduce the Bayesian methodology proposed by JOSEPH, GYORKOS and COUPAL (1995) to estimate sensitivity (S), specificity (E) and disease prevalence (P), in the absence of a gold standard test. In addition, we aimed to compare the influence of an informative prior distribution and the sample size on the results. In order to meet the objective, six cases or situations were considered. In three of the cases, prior uniform distributions were used for the parameters with different sample sizes, and in the other three cases, prior beta distributions with different sample sizes. To illustrate the methodology in a real situation, in two of the six cases, mentioned previously, we used real data from an experiment in which the Comparative Cervical Test was performed to detect bovine tuberculosis. From the results obtained, the size of the sample did not have a significant influence on the results, but the estimates were more precise in cases which prior beta distributions were used for the parameters S, E and P. It was concluded that the Bayesian methodology is adequate to estimate the parameters, but there is a need to use prior distributions with some information so that the identifiability problem is solved.