Studies on effects of parental selection on estimation of genetic parameters and breeding values of metric traits
This study reports on effects of parental selection on estimation of variance components (VC) and breeding values (BV). Bayesian estimation via Gibbs sampling, restricted maximum likelihood (REML), and Method using an individual animal model were employed for estimating VC. Bayesian and empirical BLUE/BLUP estimation were considered for predicting BV. Methods were evaluated according to their empirical sampling properties via stochastic simulation assuming a univariate infinitesimal additive genetic model. Combinations of two simulation models (MO) (with or without effect of contemporary groups-CG), random and non-random selection (phenotypic and based on BLUP), two levels of heritability (0.20 and 0.50) and two levels of pedigree information (PI) (0% and 15% randomly missing) were considered. Non-random phenotypic and BLUP selection caused similar effects on the VC estimates from all methods. Bayesian and REML estimation showed the same pattern with respect to effects of selection, which was not the case of Method . When all information was available (no missing ), REML and Bayesian estimates were not biased by non-random selection for both MO. Method estimates were strongly biased by non-random selection when the model included CG. The bias was empirically shown to be a consequence of not fully account for gametic phase disequilibrium. The joint effect of non-random selection and missing PI yielded highly biased estimates of VC from all methods. Missing PI did, in general, not cause bias in randomly selected populations. Mean square error (MSE) of Bayesian, REML and Method estimates decreased with non-random selection when data did not include CG effects, but increased when CG were present. Method estimates usually had greater MSE than did REML and Bayesian estimates. Bayesian and empirical BLUE/BLUP estimation of CG effects and BV did not differ in their rank correlations with the true values for all simulated populations. The average deviations and average squared deviations of the estimated CG effects and predicted BV from their true values were generally different for Bayesian and empirical BLUE/BLUP analyses, but differences were usually trivial and likely due to the prior information on the VC in the Bayesian analyses.