Validation of genomic and genetic evaluations in 305d production traits of Nordic Holstein cattle

Proceedings of the 2017 Interbull Meeting, Tallinn, Estonia, August 25 - 28 2017

As genomic selection has been used already for several years, it has become evident that the validation of genomic evaluations relying on traditional animal models is becoming unsuitable. The GEBV validation test recommended by Interbull is cross-validation based on the forward prediction. It was designed at the time when the multi-step genomic evaluation was the standard method. The aim of this study was to take a closer look on accuracy and stability of (G)EBVs. Validations for GEBVs were done using yield deviations (YD) or daughter yield deviations (DYD) calculated with single-step GBLUP instead of EBV model. Moreover, we studied the stability of (G)EBV estimations in consecutive evaluations. We used Nordic Holstein 305 days production data containing ca. 7.3 million cows with 15.6 million observations. Genotypes were available for 30056 animals which had either records or offspring in the full 305d data. The test setup consisted of four data sets: the full data, called data0, included calvings up to March 2016. Three reduced data sets were data-1, data-2, and data-3, from which one year of calvings was deleted at a time. This allowed studying the accuracy of predictions by production years, and also the stability of (G)EBV estimates across lactations. The bull validation was a regression of DYDEBV on PAdata-3 or, for GEBVdata-3, regression of DYDGEBV on GEBVdata-3. The results suggested that after use of genomic selection the DYD from EBV model become biased and that GEBVs validated using DYDs from the BLUP model might receive too low reliability. The validation reliability for protein GEBV (r2) was 0.34 using DYD from EBV model and 0.36 using DYD from ssGBLUP. Similarly, when making cow validations, it might be advisable to use YDs calculated from ssGBLUP for the validation. The r2 in GEBV validations using YD from ssGBLUP were on average 5 % units higher compared to validations using YDs from the EBV model.

2017