|Van Doormaal, Brian -|
Submitted to: Interbull Annual Meeting Proceedings
Publication Type: Proceedings
Publication Acceptance Date: March 10, 2011
Publication Date: March 10, 2011
Citation: Van Raden, P.M., Van Doormaal, B. 2011. Genomic validation of national systems. Interbull Annual Meeting Proceedings. Guelph, ON, Canada, February 27-28. 4 pp. Technical Abstract: National evaluation centers began testing estimated breeding values (EBVs) several decades ago to determine optimum statistical methods and convince breeders that the chosen methods work properly. In 1998, validation of traditional EBVs became more formal when 3 tests of genetic trend were required by Interbull. However, those tests measure only trend bias and not EBV accuracy. Genomic EBVs (GEBVs) for young animals are now a major focus of genomic selection. Bias removal is important because traditional parent averages (PAs) for elite young stock often were inflated because of overevaluated dams. Statistical methods and data sets used to compute GEBVs are evolving rapidly, and breeders have little experience with calculations before they are revised or new data introduced. Tests for genomic evaluations help ensure that optimal methods are chosen and that breeders can be confident that the predictions are accurate. Validation methods should be convenient, convincing, and helpful in choosing an accurate model. The current Interbull validation method detects problems with scaling but not necessarily accuracy or bias. One goal of validation is to ensure that consistent data are input into genomic multiple-trait across-country evaluations (GMACE). Secondly, validation is used as a trade barrier by the European Union (EU). Embryos, cows, heifers, and live young bulls can be marketed internationally using PAs or EBVs, but if a country’s GEBV validation for protein is not within tolerance, semen from its genomically tested young bulls is banned from the EU. Interbull now has an official role in setting EU standards. Most other countries allow open importation of young bull semen. The following questions were raised for discussion: 1) what experiences have evaluation centers had with the current validation test, 2) are expected regressions and adjustments for selection well understood, 3) how can recently introduced traits be validated or inclusion of cows in the reference population tested if most genotypes are from young animals, 4) should the validation model include only a regression and no intercept so that a single parameter captures both bias and slope problems, 5) for countries that share reference bulls, should validation bulls include only domestic or also foreign bulls to improve power of test, and 6) what other tests or demonstrations could help breeders to understand GEBV properties. Group discussion disclosed that 1) use of a 4-year data cutoff in the current test is not optimal for new traits or small populations; 2) expected regressions account for selection only within test bulls but not for preselection from the training population; 3) independent predictions could be computed (e.g., cows only versus bulls only) to check if different data sources are consistent; 4) inclusion of both a regression and intercept is recommended, but a problem in just the regression may also cause the intercept to differ from 0; 5) usefulness of foreign validation bulls to improve power may be limited by suboptimal PAs on domestic scales from 4 years ago, highly selected 2nd-country bulls, and potential evaluation bias; and 6) extension and communication to the industry is a major opportunity. Group recommendations were that 1) validation tests should be useful to other researchers in documenting evaluation properties; 2) validations, documentation, and educational materials should also be useful to breeders and breeding companies in purchasing decisions, and national evaluation centers need to build more confidence by improving and refining the whole genomic evaluation system; and 3) ongoing monitoring is needed for 2-3 years after implementing official genomic evaluations and after changing models or input data to verify that calculations work as intended. After each evaluation or at least each year, predictions from initial GEBVs should be compared with PAs to demonstrate how effective each is in predicting new data of new bulls. Such comparisons can use simple means of selected bulls instead of the parametric tests used in model selection and validation.