Agreement and kappa-type indices

Kappa-type indices use the concept of agreement to express the reproducibility of nominal measurements. This article grounds kappa-type indices in statistical modeling, making explicit the underlying premises and assumptions. We critically review whether the interpretation of the kappa index as a chance-corrected probability of agreement can be substantiated. Further, we show that the so-called paradoxical behavior of the kappa index is explained from the fact that it is a measure of predictive association, rather than a pure measure of reproducibility. We discuss a number of alternative forms, critically examining whether they can be translated in tangible real-life interpretations.