Effect of modelling unknown parent groups and metafounders on the historical genetic trend of fertility traits

Abstract

Unknown parent groups (UPG) allow modelling unobserved selection in unknown parents. Because UPG are defined at least partly by year of birth, biased estimates can also bias estimates of environmental trends like management. Different modelling of UPG can reduce biases and standard errors. The 4 fertility traits daughter pregnancy rate (DPR), cow conception rate (CCR), heifer conception rate (HCR) and early first calving (EFC) in US dairy cattle make a good study case, because those have been affected by selection on correlated traits such as milk yield and have greatly differing recording patterns. Traits DPR and CCR are strongly correlated but DPR was recorded since ~1960 and CCR was recorded since ~2000. For missing traits, current traditional evaluation compress UPG definitions for missing years to avoid solving for UPGs with no direct information, and treat UPG as correlated across traits but uncorrelated across years (RandomUPGs). New models included: Fixed UPGs; Metafounders fitting average coancestry across UPGs based on year of birth (MFDeltaF); or including expected magnitude of change due to selection on a correlated trait (MFDeltaG). The data set consisted in 94 million records with potentially large numbers of missing values depending on trait and year, a pedigree including 94 million animals. Genetic evaluations were by BLUP and results are presented for Holstein.  In all cases UPGs are treated as “a priori” correlated across traits. Genetic trends resulted in all cases in a fast decrease of DPR from 1960 until 2000. For DPR, this descent was most pronounced with RandomUPGs, closely followed by MFDeltaG and MFDeltaG, which yielded slightly less change because the inclusion of average coancestry results in smaller a priori changes. Similar trends but with larger differences across methods were observed for the correlated trait CCR, where the trend is inferred from correlations because of absence of records. Trends from 2000 to 2020 for both CCR and DPR were positive, with MFDeltaF showing slightly faster increases. Solutions of UPGs/MFs were most noisy with FixedUPGs, followed by RandomUPGs, followed by MFDeltaF which was the smoothest. Overall, for traits with years of missing records and with selection due to correlated traits not included in the data, modelling UPGs as random, and possibly correlated across years, is useful for correct genetic trends.