Robustez na capacidade preditiva dos modelos AMMI e Fatoriais Analíticos no estudo de dados multi-ambientais desbalanceados

Abstract

The present work aimed to verify the robutness of the AMMI predictive ability through using several Bayesian and Frequentist approaches, and Analytical Factor (FA) in the study of unbalanced multi-environmental data (MET), using simulated data. To verify the eficiency of these methods, random unbalanced was performed using 10%, 33% and 50% of loss. To evaluate the predictive ability of the missing data in proposed models, the PRESS statistics (prediction error sum square) and the correlations between the observed and predicted values were used, through cross-validation methods. The results showed that inpredictive terms, at the level of 10% of unbalance the Bayesian AMMI models with variance heterogeneity (AMMIB-D) and AMMI models through EM algorithm for random effects of genotype and fixed environment (EM-AMMIM) were superior followed by Bayesian AMMI models with homogeneity of variances (AMMIB-I) and FA2. At 30% of dataloss, the AMMIB-I was superior, followed by EM-AMMIB-D models, AMMI models through EM algorithm for fixed effects of environment and genotype (EM-AMMIF) and FA2. At 50% dataloss, the AMMIB-I and AMMIB-D models were superior, followed by the FA2model. It can be concluded that the AMMI models are frequentist or Bayesian and Factorial Analytical were robust in the study of MET data with high levels of loss of genotypes in the environments.