Sample sizes for evaluation of diagnostic tests for bovine paratuberculosis in the absence of a gold standard

摘要

Test evaluation studies for bovine paratuberculosis are challenging because of the difficulties of correctly defining the true infection status of cows in infected herds. An alternative approach involves methods that don't rely on having gold standard information but instead use test results from 2 conditionally independent tests in 2 or more populations. These populations could be herds or subpopulations within the same herds, such as cows of different parity/lactation number. We have developed frequentist and Bayesian approaches to sample size calculations for studies to estimate sensitivity and specificity with desired precision. For the frequentist approach, we constructed an Excel spreadsheet template (available at http://www.epi.ucdavis.edu/diagnostictests/) to perform the calculations following the Hui and Walter (1980) model that assumes asymptotic normality of ML estimates of parameters. In the Bayesian approach, we determine a sample size that yields high predictive probability with respect to the future study data, of precise estimates of sensitivity and specificity. The method is implemented using the Splus/R library emBedBUGS together with WinBUGS. Comparison of both methods for estimation of the sensitivity and specificity of ELISA and fecal culture tests for bovine paratuberculosis is presented in 2 populations with assumed prevalences of 1% and 15% and where the estimates of the desired interval width for the sensitivity of ELISA and fecal culture are +- 10%, and for the specificity of ELISA and fecal culture are +- 2% and +- 1%, respectively. Findings from a range of other plausible scenarios indicate that large sample sizes (> 2000 animals / population) are needed to obtain reasonably precise estimates and these sample size requirements increase as prevalences in the 2 populations become closer to one another. Such large sample sizes might be impractical in many circumstances. The Bayesian approach is more flexible because it avoids limitations in the Hui and Walter model, when one or more estimates is close to 1.