Location: Animal Genomics and Improvement LaboratoryTitle: Improved genomic validation including extra regressions
Submitted to: Interbull Annual Meeting Proceedings
Publication Type: Proceedings
Publication Acceptance Date: 9/9/2021
Publication Date: 9/9/2021
Citation: Van Raden, P.M. 2021. Improved genomic validation including extra regressions.Interbull Bulletin. 56:65–69.
Technical Abstract: Genomic predictions (GEBV) are often validated by predicting later deregressed conventional evaluations or daughter yield deviations (dEBV or DYD) from earlier GEBV. Predicting later GEBV from earlier GEBV could be easier for the public to understand and to verify than standard validation and could be applied to single-step models where the GEBV account for genomic preselection but the later dEBV do not. Genomic validations could also predict deregressed GEBV (dGEBV) that include only the new information from the gain in reliability. Changes in genetic trend or rank can also be tested as in validation of conventional EBV by including extra regressions such as on birth year, parent average (PA), or expected future inbreeding (EFI) from the earlier evaluation. The new validation can compute model squared correlations (R2) that ideally should be high, indicating stable evaluations, and predict GEBV difference (final GEBV – earlier GEBV) to give residual R2 that ideally should be low, indicating that changes in evaluations are not a function of other known factors or the earlier GEBV. The new validation methods were applied to U.S. GEBV and for 7 main traits. For most, the regressions on birth year indicated that genetic trend decreased as daughters were added, the regressions on PA were negative, indicating too much blending of PA with direct genomic value, the regressions on EFI were not significant, and regressions on earlier GEBV were > 1.0 when the extra regressions were included. The model R2 ranged from 48 to 79% and the residual R2 ranged from 3 to 18%. These new, more flexible methods give a more complete picture of GEBV properties and how models may be improved to reduce bias and improve prediction accuracy.