Bias remains a challenge for genomic evaluations
Bias remains a challenge for genomic evaluations
by Chad Dechow
The author is an associate professor of dairy cattle genetics at Penn State University.
Genomics was a large focus at the Advancing Dairy Cattle Genetics workshop this past February in Phoenix, Ariz. The issue of bias in genomic evaluations was not on the official agenda, but it was discussed among participants and briefly during the presentation of USDA scientist Paul VanRaden. I’d like to explore the issue of bias more fully because many breeders have concerns about the stability of genomic evaluations. I think it is important to acknowledge their concerns so that breeders have confidence in the effectiveness of our genetic evaluation system.
Young versus older sires
The chart demonstrates how proofs for genomic young sires and for daughter-proven bulls have changed from April 2011 to April 2014. I’ve included the top 150 genomic young sires for $NM (Lifetime Net Merit) in April of 2011 that had at least 50 daughters by April 2014. Likewise, I included the top 150 bulls that had at least 50 daughters in April 2011 and that added at least 500 daughters to their proof since that time. We could produce a similar type of chart for many traits, and the general observations would be much the same.
The genomic young sires are represented by the blue dots and, as you can see, they had much higher $NM in 2011 than our daughter-proven bulls, as the blue dots are further to the right in our chart than the red dots. Our estimate in 2011 was that the genomic young sires should have been approximately $150 superior to the daughter-proven bulls. However, the difference between our two groups of bulls was not as large in 2014 after the genomic young sires added daughters to their proof.
The green line in the chart represents the expected $NM in April 2014 based on the 2011 evaluations. Dots above the line represent bulls whose $NM gained in value from 2011 to 2014, whereas those below the line represent bulls whose $NM declined. If our evaluations are unbiased, we expect an equal number of bulls above and below our line.
It is readily apparent that the genomic young sires are almost all below the line, which means there was an upward bias in the proof before they had milking daughters. In contrast, we see that the daughter-proven bulls had minimal bias. I’ve limited the chart to the top 150 bulls. The amount of bias is less when all bulls are considered but still evident.
I will point out that even though the difference between our two groups of bulls has shrunk to an average of $60, there was still an advantage for the genomic young sires. The changes for $CM (Lifetime Cheese Merit) and $FM (Lifetime Fluid Merit) were very similar to those presented for $NM.
It is also important to note that the “risk” of using genomic young sires was not extreme for commercial producers — many bulls were not as good as advertised but few were terrible. For breeders who invest in embryo transfer and in-vitro fertilization, on the other hand, even a small drop can cause a substantial loss in the value of an investment, and that is where frustration occurs.
Changes have been made to help reduce bias to some extent. The largest change involved rescaling the traditional PTA (predicted transmitting ability) of cows in 2010 to better match the range of PTA typically observed for sires. Before that change, some cows had extremely high PTAs for many traits and that biased their sons’ genomic predictions upward.
Unfortunately, there are other problems that may cause bias to persist in our genomic evaluations.
No. 1. One issue is genomic preselection. We genomically test many bulls, but only the best ever have daughters that enter an A.I. program.
In a perfect world, we would generate daughters from both the best bulls and the worst so that we could more accurately estimate genomic effects. I’m not aware of too many producers interested in breeding their cows to lousy bulls for the sake of more accurate genomic evaluations! This type of bias could actually result in an underestimate of the genetic merit for genomic young sires over time.
No. 2. A second issue is one of non-random mating when highly ranked young sires first have semen available. We increasingly see semen from such bulls be reserved for special matings to elite females to try and generate the next extreme high-ranking bull or heifer. It makes perfect sense to do so, but it means that a bull’s first set of daughters is not random.
There have been some high-profile bulls whose genetic merit estimates dropped after their initial daughter proof. Nonrandom mating gets the blame; keep in mind that a bull’s initial daughter proof is still heavily influenced by his genomic test because his daughters are young. For some bulls who dropped, the issue appears to have been inflation from the genomic test more than the effect of nonrandom mating.
No. 3. I will be interested to evaluate the amount of bias observed for proofs from young sires that are sired by young sires. Our genomic predictions are reasonably accurate one generation out, but we’re not entirely sure how accurate they will be two generations in advance.
What should be done
It is clear that bias does occur in genomic evaluations for a variety of reasons that are tough to predict. The challenge becomes how breeders should address this bias.
Some believe we should calculate the average drop from genomic young sire proofs to daughter proofs and subtract that amount from the new genomic young sires. I’m concerned that type of historical approach may not be accurate because of the improvements that are made to the evaluation system over time and because it is difficult to predict the direction and magnitude of future bias. A second suggestion has been to simply publish statistics on how proofs have changed over time and allow breeders to use the information as they see fit.
From my perspective, I try not to compare young sires and daughter-proven sires directly against each other. If producers are selecting from the top of either list, they are going to be using a high-quality bull in most instances. Some bulls will be overvalued, but the herd will still make genetic progress.