120 Cross-validation of breeding values and future phenotypes for heifer pregnancy in Red Angus cattle using the LR method

Abstract Heifer pregnancy (HP) in beef cattle is a heritable and economically relevant trait but is challenging to predict using best linear unbiased prediction (BLUP) methodology due to the binary nature of the phenotypes. The observed categorical responses (e.g., 1 = pregnant; 0 = non-pregnant) are due to an animal either exceeding a particular threshold for pregnancy or not on an underlying genetic distribution which is why this binary trait is commonly evaluated using a threshold model (TM). However, TM are susceptible to scenarios where only 1 category of observation is predominant. In this study we compare genetic evaluations including “partial” and “whole” data to cross-validate methodologies for the prediction of HP. The objectives of this study were then to evaluate estimators of prediction accuracy, bias, and dispersion. Heifer fertility data were obtained from the Red Angus Association of America. Evaluations for HP were performed using ASREML3.0 and a univariate, BLUP, traditional TM; and a linear continuous BLUP animal model (LM) on both reference (whole) and validation (partial) datasets. Reference data included all animals with HP phenotypes, and validation data had censored the most recent 5 yr of HP phenotypes (~16% of HP phenotypes). Reference and validation estimated breeding values (EBV) were then compared using linear regression (LR) methods. Heritability estimates were 0.05 (± 0.007) and 0.15 (± 0.009) for LM and TM respectively. The average BIF accuracy of censored validation animals (n = 10,175) using a LM on whole and partial datasets were 0.07 and 0.03 respectively, and using a TM on whole and partial datasets were 0.09 and 0.04 respectively. Comparing the mean EBV from the partial and whole datasets can be a good estimator of bias in the model as the difference is expected to be 0. The difference between the mean EBV from partial and mean EBV from whole datasets were –0.003 and –0.016 for LM and TM respectively, suggesting that the TM is a slightly more biased evaluation. The regression of EBV obtained with whole data on EBV estimated with partial data has an expected value of 1 if there is no over/under dispersion which can be a useful tool in determining model correctness and gauge bias in an evaluation. The closer a regression coefficient is to 1, the less biased the evaluation is. The slope of the regression of EBV estimated from whole data on partial data for validation animals were 0.98 (P < 0.0001) and 0.91 (P < 0.0001) for LM and TM respectively. Although both regression coefficients were above 0.90 suggesting both evaluations are performing well, these results suggest that TM models may be a more biased predictor of HP than LM despite a small benefit in prediction accuracy for the trait.