Estimation of the sensitivity and specificity of assays for screening antibodies to HIV: a comparison between the frequentist and Bayesian approaches.

Bayesian and frequentist methods have been applied rarely to the same sets of data for evaluating assays for screening antibodies to HIV, especially for assays with relatively high sensitivities and/or specificities of 100% compared with reference assays. In this study, 95% confidence intervals and 95% Bayesian credible intervals were calculated for sensitivity and specificity for the evaluation of the accuracy of HIV antibody assays using data from China, WHO UNAIDS, USA, Australia, Tanzania, and India. When the sensitivity and/or specificity were 100%, a Bayesian approach obtained reasonable interval estimates of assays for screening antibodies to HIV, whereas frequentist methods express objectively the accuracy of each individual assay. It is suggested that the two types of estimates be reported simultaneously to evaluate more comprehensively a set of highly accurate antibodies for HIV assays.

Bayesian evaluation of the human immunodeficiency virus antibody screening strategy of duplicate enzyme-linked immunosorbent assay in Xuzhou Blood Center, China.

BACKGROUND: Accurate estimation of the risk of human immunodeficiency virus (HIV) infection through transfusion is essential for monitoring blood safety. The risk, however, is so low that it can only be estimated by mathematical modeling. With the Bayesian dependence model, this study evaluates the HIV antibody screening strategy of duplicate enzyme-linked immunosorbent assay (ELISA) in Xuzhou Blood Center and therefore estimates part of the total risks of transfusion-transmitted HIV infection. STUDY DESIGN AND METHODS: Data from Xuzhou Blood Center between 2004 and 2008 were used. Information was obtained on donor profiles and screening and confirmatory test results. The portion of the risks of HIV infection through transfusion concerned was estimated by evaluating the screening algorithm in terms of its accuracy and predictive power with the Bayesian dependence model. RESULTS: A total of 234,602 donations from voluntary blood donors in Xuzhou Blood Center were screened for HIV antibody. For the study screening algorithm, its sensitivity, specificity, false-positive predictive value (FPPV), and false-negative predictive value (FNPV) were 0.9951 (95% Bayesian credible interval [BCI], 0.9763-0.9997), 0.9991 (95% BCI, 0.9990-0.9992), 0.9647 (95% BCI, 0.9018-0.9923), and 1.52 × 10(-7) (95% BCI, 7.31 × 10(-9) -1.15 × 10(-6) ), respectively. For the positive detection rate (9.60 × 10(-4) ) and FPPV (0.9647), the differences between their own Bayesian median estimates and real values were 2.70 × 10(-5) and -0.0033, respectively. CONCLUSIONS: The HIV antibody screening algorithm of duplicate ELISA is well evaluated in its accuracy and predictive power with the Bayesian dependence model. The FNPV measures the part of the risks of transfusion-associated HIV transmission concerned.

Evaluation of 30 commercial assays for the detection of antibodies to HIV in China using classical and Bayesian statistics.

The purpose of this study was to evaluate the 30 commercial HIV-antibody (HIV-Ab) assays in the nationwide assessment program of China using classical and Bayesian statistical methods. The classical estimates of sensitivity and specificity varied from 95.9% to 100% and from 94.6% to 100%, respectively. The proportions of assays with 100% sensitivity and with 100% specificity reached 63.3% (19/30) and 3.3% (1/30), respectively. Using the Bayesian logit hierarchical model, the overall estimates of sensitivity and specificity were 99.8% (95% Bayesian credible interval [BCI]: 99.4-100%) and 98.1% (95% BCI: 97.4-98.7%), respectively, for the 17 ELISAs under evaluation. For the 13 rapid assays, the corresponding overall estimates were reported to be 99.2% (95% BCI: 98.5-99.8%) and 98.4% (95% BCI: 97.8-98.9%), respectively. In addition, given the prevalences of HIV infection among the general population of China and the intravenous drug user group in China, the positive predictive values were estimated for each individual assay in the framework of the two schools of statistical thought. Furthermore, by comparing the two types of estimates, it is concluded that the two types of statistical methods were complementary for the evaluation of very accurate HIV-Ab assays.