The article is organized as follows: the first section contains a brief, non-exhaustive review of the main measures of the agreement at a qualitative scale of reaction. The approach is then described with the Model of Dirichlet Multinom and a generalization based on the mixture. Two applications are submitted below. The first refers to the diagnosis of schizophrenia, while the second considers a sensory analysis for food. The last part sets out the findings. Finally, two online annexes contain the definition of the main measures of the agreement and code for the application of the method proposed in the software. Health diagnoses are expected to be very accurate. Taylor and Abrams` method seems fair and reliable. The low overall agreement, as well as the results of the paired convention, shows that the other three methods can be used to diagnose certain types of schizophrenia, but not separately or as a gold standard. We opted for multivariate probability functions when developing Bayes models. There are many examples of Bayaise hierarchical multivariate analysis of repeated measurement results that have already been used successfully in the medical literature [19-21]. We believe that, for a number of reasons, the use of multivariate models is preferable to the Bland and Altman successive pair comparison approach.
These include the fact that more than two methods are often considered, that all information is analyzed at the same time (which gives greater power and greater statistical robustness and efficiency), marginal distributions of parameters of interest can be easily deduced without an asymptomatic approach and the probabilistic approach is closer to how researchers think. The modelling and implementation of bayesen`s multivarien hierarchical model with vague preliminary information was also simple. With Free WinBUGS  software and inclusive computer programs, the calculations usually ended in minutes. If, in fact, one of the main reasons for so few repeated comparison studies with measurement methods is due to the lack of easily available statistical mechanisms for analyzing these data, we hope that our careful presentation of statistical analysis and computer code for two examples will help to circumvent this barrier for future researchers.