Abstract
It is known that there is a very powerful and detailed methodology for a reliability analysis of tests that are used for the measurement in psychological diagnostics. However, there are also various diagnostic classification procedures in which observed variable is measured only on a nominal scale.
For such classifications, the information about their reliability is available very rarely. Of course, one reason for this is the fact that there is no sufficiently developed methodology for the nominal classification, unlike for the test measurements.
The analysis of reliability in these cases is usually based on the Cohen-Fleiss kappa conception, i.e. the concordance analysis of two or more replications of classification. The aim of this paper is to present new, innovative methods of the analysis of reliability of classification procedures, based on a probabilistic model of error in classification.
This model is based on analogous principles as the "true-error" model of classical test theory.