Use of the reversible jump Markov chain Monte Carlo algorithm to select multiplicative terms in the AMMI-Bayesian model.

The model selection stage has become a central theme in applying the additive main effects and multiplicative interaction (AMMI) model to determine the optimal number of bilinear components to be retained to describe the genotype-by-environment interaction (GEI). In the Bayesian context, this problem has been addressed by using information criteria and the Bayes factor. However, these procedures are computationally intensive, making their application unfeasible when the model's parametric space is large. A Bayesian analysis of the AMMI model was conducted using the Reversible Jump algorithm (RJMCMC) to determine the number of multiplicative terms needed to explain the GEI pattern. Three a priori distributions were assigned for the singular value scale parameter under different justifications, namely: i) the insufficient reason principle (uniform); ii) the invariance principle (Jeffreys' prior) and iii) the maximum entropy principle. Simulated and real data were used to exemplify the method. An evaluation of the predictive ability of models for simulated data was conducted and indicated that the AMMI analysis, in general, was robust, and models adjusted by the Reversible Jump method were superior to those in which sampling was performed only by the Gibbs sampler. In addition, the RJMCMC showed greater feasibility since the selection and estimation of parameters are carried out concurrently in the same sampling algorithm, being more attractive in terms of computational time. The use of the maximum entropy principle makes the analysis more flexible, avoiding the use of procedures for correcting prior degrees of freedom and obtaining improper posterior marginal distributions.

Shrinkage in the Bayesian analysis of the GGE model: A case study with simulation.

The genotype main effects plus the genotype × environment interaction effects model has been widely used to analyze multi-environmental trials data, especially using a graphical biplot considering the first two principal components of the singular value decomposition of the interaction matrix. Many authors have noted the advantages of applying Bayesian inference in these classes of models to replace the frequentist approach. This results in parsimonious models, and eliminates parameters that would be present in a traditional analysis of bilinear components (frequentist form). This work aims to extend shrinkage methods to estimators of those parameters that composes the multiplicative part of the model, using the maximum entropy principle for prior justification. A Bayesian version (non-shrinkage prior, using conjugacy and large variance) was also used for comparison. The simulated data set had 20 genotypes evaluated across seven environments, in a complete randomized block design with three replications. Cross-validation procedures were conducted to assess the predictive ability of the model and information criteria were used for model selection. A better predictive capacity was found for the model with a shrinkage effect, especially for unorthogonal scenarios in which more genotypes were removed at random. In these cases, however, the best fitted models, as measured by information criteria, were the conjugate flat prior. In addition, the flexibility of the Bayesian method was found, in general, to attribute inference to the parameters of the models which related to the biplot representation. Maximum entropy prior was the more parsimonious, and estimates singular values with a greater contribution to the sum of squares of the genotype + genotype × environmental interaction. Hence, this method enabled the best discrimination of parameters responsible for the existing patterns and the best discarding of the noise than the model assuming non-informative priors for multiplicative parameters.