Hadgu raises concerns about how diagnostic tests for detecting Chlamydia trachomatis are evaluated and states that 'discrepant analysis strongly biases evaluations in favour of the new test'. He questions the sensitivity and specificity figures that are generated by discrepant analysis. It may be useful to put Hadgu's concerns in perspective. He certainly is correct in stating that evaluations of diagnostic tests overstate sensitivity because they are always based on positive test results as compared with an imperfect gold standard. Only a comparison using a perfect test would provide true sensitivity. It has long been recognized by researchers that new and more sensitive diagnostic tests will identify positive specimens in the D cell of a typical two-by-two box (double negatives by previous technology). This is apparent in the evaluation of antigen detection methods as compared to amplified DNA tests, which do identify many more positives, but, let us review our paper, which Hadgu used as an example, and see what precisely is the demonstrated scope of this problem.
We found 139 specimens positive by cell culture and LCR, with another 13 that were positive by cell culture and negative by LCRprobably due to some form of inhibition. There were 84 specimens positive by LCR and negative by cell culture, and 1896 specimens negative by both LCR and cell culture (see Tables 1 and2). Using simple ratios -139/13 and 84/Xone can estimate that there are approximately 8 positive specimens in the D cell that could test negative due to inhibition, assuming that inhibitors occur at the same rate as they did in the isolation positive specimens. That would change the total number of true-positive specimens from the 234 that were initially identified by tissue culture and confirmed LCR results to 242. If these 8 are included into the calculations, the sensitivity of LCR would be 91)3 per cent (221/242) instead of the 94)4 per cent (221/234) presented in the paper. However, since these 8 specimens were negative by all tests, the corresponding values of other diagnostic tests would also be reduced. (For example, cell culture would not be 64)9 per cent (152/234) but 62)8 per cent (152/242)). Thus, the order of sensitivities of diagnostic test being evaluated or compared would not change but the absolute numbers would.
Discrepant analysis focuses on correct categorization of specimens that are identified as positive in the initial run. No new positive results are identified through discrepant analysis, but results may be reclassified from false-positive to true-positive status by confirmation. As microbiologists, we were faced with a challenge of what to do when a newly developed diagnostic test is more sensitive than the previous gold standards. We and others have used discrepant analysis with the DNA amplification tests because for C. trachomatis this approach provides the most accurate and practical means for understanding performance of new and more sensitive assays. The Hadgu alternatives are to stick with the old imperfect gold standard, which classifies the increment in sensitivity offered by the new technology as false-positive, or to use the tie-breakers (those we use for discrepant analysis) for all specimens. There are practical problems to the latter option that make it untenable. It would require the manufacturer of a new and much more sensitive test to develop two new tests to be evaluated in parallel (thus greatly increasing expense and laboratory workloads) even though only one of the tests will ultimately be used in the marketplace.