[Comparative analysis of three approaches for rater agreement]

This work is a comparative analysis of three general approaches for rater agreement measurement which produce three different statistical coefficients: descriptive (sigma, pi y kappa coefficients), based on log-linear models and on mixture models. Although loglinear and mixture approaches assume a different concept of chance correction, their flexibility allow using its parameters to analyze agreement and disagreement patterns. We show that all descriptive measures can also be obtained by introducing constraints to a basic loglinear (or mixture) quasi-independence model. Due to the different measurement scale used, both loglinear and mixture approaches have a similar interpretation when diagonal parameters are positive but a different interpretation otherwise.