Location: Animal Genomics and Improvement LaboratoryTitle: Methods to compute reliabilities for genomic predictions of feed intake
Submitted to: American Dairy Science Association Proceedings
Publication Type: Abstract Only
Publication Acceptance Date: 3/11/2018
Publication Date: 6/24/2018
Citation: Van Raden, P.M., Hutchison, J.L. 2018. Methods to compute reliabilities for genomic predictions of feed intake [abstract]. Journal of Dairy Science. 101(Suppl. 2):370-371(abstr. 392).
Technical Abstract: For new traits without historical reference data, cross-validation is often the preferred method to validate reliability (REL). Time truncation is less useful because few animals gain substantial REL after the truncation point. Accurate cross-validation requires separating genomic gain from pedigree contributions and assuming that other animals with pedigrees less connected to the reference data will have less REL than validation cows. Fiveway cross-validation of residual feed intake (RFI) used data from 80% of the 3,965 US research cows to predict the other 20% and repeated the process 5 times to test predictions for all cows. However, RFI records were excluded from validation data if the cow had progeny in the reference data to ensure correct prediction direction. Pedigree REL for the validation cows was 13%, and their genomic REL was only 18% compared with 21% expected. Pedigree REL for elite young calves was 3%. After adjusting the discount factor to match expected with observed REL, their genomic REL was only 9%, which was less than the 12% previously estimated. Research cows often have paternal sibs, maternal sibs, or dams with RFI records, whereas most calves in other herds are >2 generations removed from any relatives with RFI records and thus have lower REL. Cross-validation of SCS records for these same 3,965 research cows gave REL estimates similar to those for RFI. Reliability of SCS was also estimated as genomic REL of national SCS predictions (72%) multiplied by the correlation of research herd predictions with national predictions (0.39) squared, which resulted in 11% REL. Some evaluations report feed saved, which includes RFI plus the economic value for body weight composite (BWC). Inclusion of BWC added only about 7% to the REL of feed saved because RFI contributed much more genetic variance than BWC. Correlated yield traits contribute much variance to feed intake but not to RFI, which is independent of yield. Increases in genomic REL with further RFI data can be forecast. For young calves, such REL could be 12% with 5,000, 19% with 10,000, 31% with 20,000, and 52% with 50,000 cows in the reference population.