written by Matt Woolfolk, ASA Director of Performance Programs
As seedstock producers, we talk plenty about EPDs, but less often to we get into discussion about the accompanying piece to each EPD: its accuracy value. By book definition, accuracy of an EPD is the relationship between estimated breeding value of the animal (the published EPD) and the “true” breeding value of the animal. The accuracy of an EPD is calculated simultaneously with the EPDs during the weekly IGS runs. Accuracy values range from 0.00 to 1.00. Each accuracy value also has an associated possible change value for each trait. Think of this possible change value like a standard deviation. The possible change values demonstrate how wide our variation can be for an animal at that accuracy level. You may not realize it, but you have seen this number before in Digital Beef. It is labeled as the “+/- Chg” directly below the EPD. Selection indexes do not have accuracy values, as they are a combination of multiple EPDs in a formula. The methodology used for calculating accuracy on a single EPD does not fit for selection indexes.
The best visualization of this concept is by looking at a bell curve for various accuracy levels. “The curve” may have helped some of you pass college, and now it will help you visualize what the accuracy of an EPD is telling us. On these curves, the center number of the curve is the example animal’s weaning weight EPD (equal to 50 in all our scenarios). With that as our estimated breeding value for the animal, each curve will demonstrate various levels of accuracy (0.20, 0.40, 0.60, and 0.80). When you look at a bell curve with standard deviations, approximately 69% of the area of the curve is within one standard deviation of the mean (middle), and 95% is within 2 standard deviations of the mean. Our WW EPD is the mean, and our possible change values are our standard deviations.
For an animal with a WW EPD accuracy of 0.20, you’re likely looking at a young calf with maybe a weaning weight turned in on himself and no genomics. The possible change value at this 0.20 accuracy level is +/- 13. That means that we’re confident his true genetic value for weaning weight is between 37 and 63, and we’re almost certain that it is between 24 and 76.
Let’s say we get genomically enhanced EPDs on the calf and accuracy improves to a 0.40. Then, our ranges shrink to one standard deviation being 40 to 60, and that we’re 95% confident that our breeding value is between 30 and 70. Yes, that’s still a good bit of variation, but we are drawing the window in closer. This is where genomic enhanced EPDs on young cattle show their value.
Now we take this same bull and he has a couple calf crops with weaning data turned in on them to the point that his accuracy rises to 0.60. With our change value dropping to 7, we now are largely confident his breeding value is between 43 and 57. At this point, we’ve shrunk the window to the point we should feel like we have a good indication of how this bull breeds.
Finally, the bull becomes popular and is used via AI in several herds for a few years. His WW EPD accuracy climbs to 0.80. With a change value of 3 at this accuracy level, we’re now confident his true breeding value is between 47 and 53, while being almost certain his true value is between 44 and 56.
From this exercise, you can see that the key to increasing accuracy in EPDs is data: individual, genomic, and progeny data. With our move to IGS and the BOLT genetic evaluation system four years ago, the calculation of accuracies has become more precise due to increased computing power that wasn’t available in the early days of EPD calculations. Today’s accuracy values are lower than what we saw 10 years ago, but the experts in the field assure us that they have more faith in current accuracy arithmetic compared to previous models.
Studying accuracy of EPDs is almost as important as studying the EPD values themselves, as this exercise shows that not all EPD of equal value are truly created equal. Even with the challenges associated with lower accuracy values, EPDs are still our best available statistical tool for genetic selection. It’s important to keep in mind that like any tool for any job, they have their limitations and only work when properly utilized.