Avaliação de métodos de estimação de componentes de variância com aplicação em bovinos Nelore

Abstract

Four classical methods to estimate variance components were evaluated for the one-way classification random model, two-way nested random model and two-way crossed mixed model with interaction. A simulation study was carried out in a variety of situations involving five values of variance ratio A(a)2/A(c)2,being A(b)2 the random effect variance (except error) and A(c)2 the error variance, three ascending levels of unbalanced ness and three sample sizes. Theoretical aspects were shown about Bayesian method, followed by an illustrative example. Variance components estimates for each model were obtained by analysis of variance, ANOVA; maximum likelihood, ML; restricted maximum likelihood, REML; and minimum variance quadratic unbiased estimation, MIVQUE(0).The estimation methods were also illustrated with an application to Nelore cattle. The estimators were compared through their bias and means quared error, MSE. ML estimator was indicated for the one-way classification random model with variance components ratio equal to 0.05, for it showns maller bias and was more efficient than other estimators. It was also recommended for the other two models since it performed well inalivariance ratios, degrees of unbalanced ness and sample sizes, although it has been revealed more biased, with some convergence problems and there quired time computing was high and expensive. REML and ANOVA estimators may be other choices, since they shown good properties such as small bias and adequate mean squared errors. MJVQUE(0) was the worst estimator. Bayesian method, available in SAS software, seems to be a viable, flexible and promising alternative for not familiarized Bayesian methodology users.