SELECTION OF Coffea arabica L. HYBRIDS USING MIXED MODELS WITH DIFFERENT STRUCTURES OF VARIANCE-COVARIANCE MATRICES

Abstract

This study aimed to evaluate different structures of variance-covariance matrices in modeling of productive performance of coffee genotypes over the years, and select hybrids of <em>Coffea arabica</em> using mixed models. A mixed linear model was used to estimate variance components, heritability coefficients, and prediction of genetic values of hybrids and cultivars. Three commercial cultivars and eight hybrids of <em>C. arabica</em> L. were evaluated. The field production after acclimatization of seedlings was conducted in March 2006. The yield averages from 2009, 2010, 2011, 2013, and 2014 agricultural years were evaluated. The selection criteria of models were used to test 10 structures of variance-covariance matrices, and later a model was chosen to estimate the components of variance, heritability coefficients, and prediction of genetic values. According to Bayesian information criterion (BIC), the best structure was ARMA (Autoregressive Moving Average); however, considering the Akaike Information Criterion (AIC) and corrected Akaike Information Criterion (AICC), the CSH (Heterogeneous Composite Symmetric) was indicated. The Spearman correlation between the genotypic values obtained in the models with ARMA and CSH type R matrix was 0.84. The high and positive correlation indicates that the best model could involve the R matrix with ARMA or CSH structure. The heritability of individual genotypes differed from heritability in broad sense, which considers the independence among agricultural years. Hybrids with higher performance were identified by ordering the genotypic effects, among them, H 2.2, H 4.2, and H 6.1 hybrids were highlighted.