Test Reproducibility in Relation to Identification Sneath, P. H. A.,, 24, 508-523 (1974), doi = https://doi.org/10.1099/00207713-24-4-508, publicationName = Microbiology Society, issn = 1466-5026, abstract= Test errors are much greater than usually thought and their investigation requires good statistical design. Analysis of variance of blind randomized trials is especially valuable. For representative selections of tests currently used in bacterial taxonomy and identification, the discrepancies within one laboratory are usually less than 4%; with care they can be reduced to 2% or less. Between-laboratory variation is much greater; discrepancies of about 8% are usual under routine conditions, and even 15% is common. The effect of errors on identification is due to (a) error in the reference descriptions of taxa, (b) error in the description of the unknown strain to be identified. Simultaneous polythetic methods (e.g., using matching coefficients or analogues of these) are robust to both types of error, and misidentifications are usually minor in degree, i.e., the unknown is allocated to a taxon close to the correct one. The rate of failure is expected to be greatest when the reference descriptions are based on the tightest clusters. Sequential identification methods are especially sensitive to errors because the unknown may then be allocated to a taxon far from the correct one. Monothetic sequential keys are sensitive to errors of type (b), whereas type (a) error is usually screened out during construction of the key., language=, type=