AMMI Bayesiano para dados ordinais

Abstract

In this thesis we study the implementation of Bayesian AMMI for ordinal data. Initially we revisited theoretical aspects of Bayesian analysis, Multi Environment Trials (MET) and threshold models. In the last two sections are presented papers for scientific journals. The first is a review on Bayesian-AMMI literature folowed by a case study of the state of the art implementation. The model has shown flexibility to fit unbalanced, non-orthogonal and heteroscedastic data, but depends on continuous response in which Gaussian assumption is reasonable after scaling. The second deals with Bayesian AMMI to ordinal data. An ordinal data set on MET was artificially generated from continuous responses. This allows for a gold standard on ordinal data analysis, that is not available in actual ordinal data. A latent underlying continuous variable modeled with cumulative probit link allows for a suitable implementation of the analysis. This has shown to be efficient in telling stable from unstable genotypes. Using ordinal models interpretation is less powerful but more rigorous and consistent with continuous data analysis.