[Estimation by the Hui and Walter method of the sensitivity and specificity of a diagnostic test in the absence of a reference test: results of a simulation study]

Abstract

Hui and Walter proposed an approach to directly estimate the parameters of diagnostic tests (sensitivity and specificity for dichotomous tests) and the prevalence of the disease in the absence of a reference error-free test, and without imposing any constraint. We performed a simulation study to investigate the finite-sample properties of the proposed estimators for the situation of three tests. We obtained the following results: i) in screening situations (prevalence < or = 0.10), severe bias and poor precision in the parameter estimates occur and this approach should not be used except for studies with a very large sample size (n > or = 2500); ii) in diagnostic situations (prevalence > or = 0.30), the approach is much more reliable and very little or no bias occurs even for small sample sizes (n = 50, 100), except when both the prevalence is very high (p > or = 0.90) and the performances of the tests mediocre (sensitivity and specificity lower than 0.80).