PSXII-8 - Gaussian Kernel Based on Geographic Information to Model Farm Effects in Genetic Evaluation of Pasture-raised Beef Cattle

Abstract: The notion of contemporary groups (CG) is routinely used in beef cattle breeding programs for adjusting phenotypic observations for environmental effects. Although CG include the effect of farms, they do not allow information borrowing across farms as the CG are generally modeled either as fixed effects or as statistically independent random effects. Nonetheless, farms that are located closer to each other may share similarities in terms of climate as well as soil and forage quality. We proposed a Gaussian kernel approach to model farm effects based on their geographical position. The data consisted of 487,905 yearling weight observations of Nellore cattle from 394 farms located throughout Brazil and a pedigree file with 868,003 animals from the Embrapa-Geneplus Breeding Program. The benchmark model (M0) included the fixed effects of sex, age (linear covariate), and season at measurement, as well as direct genetic and residual effects as random, independently distributed with variance Aσ_a² and Iσ_e², respectively, where A is the numerator relationship matrix. From the M0, three models were analyzed using REML/BLUP method. M1, M2, and M3 included, respectively, the farm effect as fixed, as independent-random, and correlated-random using a Gaussian kernel with different bandwidths (h; ranging from 0.05 to 10). The AIC was used to compare the models’ goodness-of-fit. The incorporation of farm as a random effect had a negative impact on the overall fit and, consequently, the M2 and M3 presented the largest AIC (5,159,839, and 5,159,838, respectively). The M3 presented the best fit using h equal to 0.05. For this model, when the h value was increased (values close to one off-diagonal), the model fit was proportionally penalized. No gain was achieved with the correlated farm effect because a more informative kernel may have resulted in multi-collinearity issues. In the best model (M1, AIC= 2,555,877) the estimated genetic and residual variance were equal to 391.5, and 1,586 kg², respectively.