Estimation of Variance Components with Large-Scale Dominance Models

The cost of computations for dominance models using the algorithm of Hoeschele and VanRaden (8) can be substantially decreased. The number of equations can be reduced by applying the recurrence equation to a limited number of ancestors. Computing disk space and central processing unit time can be lowered when the mixed model equations are solved by iteration on the data using the method of modified second-order Jacobi. Estimation of the dominance variance for very large animal models is feasible with method R. Tests involved up to 3 million first lactation records for stature of Holsteins. The model included effects of management, regression on inbreeding percentage, age, animal additive, and dominance. Estimates of the dominance variance ranged from 2.9 to 13.5% of the phenotypic variance for a data file with 0.40 million animals and 7.5 to 8% for a data file with 3.3 million animals. Genetic evaluations with dominance effect in the model required less than twice the memory and central processing unit time of evaluations with the additive model. Computations with dominance models are now feasible for large data files.