The objective of this study was to compare Gaussian and Robust linear mixed models for the estimation of variance components by REML and Gibbs Sampling, using data from fifty simulated populations consisting of 1,000 animals distributed in 5 generations. Two levels of fixed effect and three hypothetical phenotypic values for a trait, with different levels of contamination were used in the simulations. Additive and residual variance estimates were similar for both REML and Bayesian inference using the Gaussian and Robust model. The best estimates of residual variance in the presence of contaminants were obtained by the Robust model. Estimates of heritability were similar for all models, but regression analyses indicated that predicted genetic values obtained by the robust model were more similar to real breeding values. These results suggest that the contaminated normal linear model is a flexible alternative for robust estimation in animal breeding.
Bayesian inference; breeding value; robust estimation; variance components