Application of an empirical approach for predicting accuracy for genomic evaluations

Including genomics in genetic evaluations can effectively increase selection response, especially for hard to measure, sex limited, and late in life traits. Modelling the increase in accuracy is useful when designing reference data projects and when breeders choose animals to genotype. Theoretical equations exist to predict the EBV accuracy of un-phenotyped animals. However, there are anecdotal reports that the accuracy obtained in practice was often lower than theoretical predictions. This paper validated an empirical approach to predicting accuracy in Australian Brahman data for nine traits. The empirical approach required the accuracy of reference and target animals from a standard pedigree BLUP genetic evaluation and the accuracy of reference animals from a GBLUP genetic evaluation. Using this information, a series of equations were applied to obtain the predicted GBLUP accuracy for target animals. Forward cross-validation showed that the empirical predicted GBLUP was comparable to the actual GBLUP accuracy observed for target animals (accuracy differed between 0.9% and 3.6%). In contrast, theoretical predictions differed from the observed GBLUP accuracy between 5.2% and 21.8%. For smaller (<4,000) reference populations, the theoretical accuracy was closer to the observed GBLUP accuracy, with differences ranging from 5.2% to 11.6%. The theoretical accuracy was overestimated by between 20.7% and 21.8% for larger reference populations. Empirical estimates of the effective number of chromosome segments (M_e) were between 2.0 and 3.9 times that of theoretical M_e, with the greatest difference being for the traits with larger reference sizes. This suggests that the theoretical M_e is the reason for overestimated theoretical accuracy predictions.