On April 9, 2013, the American Hereford Association (AHA) released new genomic-enhanced expected progeny differences (GE-EPDs) based on new correlations updated because of the addition of more than 2,000 animals with 50K genotypes.
These new correlations made several non-parent animals change substantially in various traits. These changes caused some conversation within the industry because these differences were even outside potential changes from the inclusion of phenotypes into National Cattle Evaluation (NCE).
So, why did this happen and what is a reasonable solution? We first must think about how data is used in NCE. Traditionally, the data submitted by breeders through Whole Herd Reporting is used to build EPDs and accuracy on Hereford animals.
We currently set some limitations on the data that is used to help protect the integrity of the data. That would include eliminating animals that perform at certain levels above and below the average of the contemporaries. In addition, breeders separate groups through management codes. Finally, there are heritabilities associated with each trait and with maternal milk (MM), the heritability is at .10 or below for Hereford.
This means that the differences from high to low are not as extreme and it takes lots of phenotypic records to make changes in the current NCE. The genomic component that was added lifted these limitations and identified markers associated with genes that had a pretty major affect on traits. This is why we saw such a huge change is some non-parent animals.
So, we addressed this issue during the weekend of April 13 and Dorian Garrick and Mahdi Saatchi, Iowa State University animal scientists, developed a pipeline that set some limitations on how much affect the genomic component could have on an animal’s EPD.
On April 15, the AHA Board of Directors voted to endorse the changes and on April 16 these updated EPDs were released. The only animals affected are those that have GE-EPDs.
There is no re-ranking of the animals, just a slight adjustment of the GE-EPD. You can find these updated GE-EPDs on Hereford.org.
Below is a more detailed explanation of the changes from Dorian Garrick. If you have any questions, please contact Jack Ward at firstname.lastname@example.org or 816-842-3757.
Modification to Genomic Prediction for AHA
by Dorian Garrick
Expected progeny differences (EPDs) are designed to assist in selection decisions in two ways. First, the ranking of candidates on EPD is the most accurate ranking based on the available pedigree and performance information. Second, the EPDs themselves are presented in units of measurement and should therefore be able to be interpreted in terms of the effect on phenotype.
This means that if two bulls differ in EPD by 10 lb., the offspring difference should also average 10 lb. If the actual offspring difference was more than 10 lb. (say 12 lb.), it indicates that there is not enough variation in the EPDs. If the actual offspring difference was less than 10 lb. (say 8 lb.) it indicates that there is too much variation in the EPDs.
We can check that the evaluation is behaving as we would expect by regressing the performance in a validation population on the EPDs. The regression slope should be 1, meaning that for every 1 lb. increase in EPD, on average there is a 1 lb. increase in observed performance. When we do this for real EPDs, either from conventional pedigree-based evaluation or genomic evaluation, we often see these regressions are a little different from 1. In genomic evaluations, they are often (but not always) a little less than 1. In the past, we have never made any adjustments for this, either for traditional EPDs or genomic EPDs.
The most recent evaluation produced one prediction that was as extreme as has ever been seen using conventional prediction. Accordingly, we have inspected the data, the procedures and these regression coefficients (of genomic EPD on actual performance) and confirmed many are slight under the expected value of 1. This makes animals at either extreme have EPDs that are a little more extreme than they should be.
Accordingly, we have decided to adjust the scale of the genomic predictions, to force the regression to be 1. In the case of birth weight for example, the adjustment is a coefficient of 0.8. This means a genomic prediction of 10 lb. would be scaled back to 8, as would a genomic prediction of -10 lb. be scaled back to -8 lb. This adjustment has no effect on the genomic rankings for any particular trait, and has no effect on the correlation. It is only the scales that are being slightly modified. The coefficient for milk was 0.47 whereas the other coefficients are less extreme (ie. closer to 1).
These changes will only impact blended EPDs on animals that had low accuracy on their conventional EPD. Among those animals, it will make those that were extreme at either end of the distribution a bit less extreme, the amount of reduction being directly dependent upon the estimates from the regression analysis.