Significance and bias - noosing the best model for variance composition

BLUP methodology has accelerated the dissemination of genetic progress, enabling the expansion of poultry production. The progress achieved relies on the accuracy of breeding values (BVs), which in turn depends on the validity of models used to estimate variance components. Aside from significance testing, the fit of models can be assessed by bias and prediction error variance (PEV) of EBVs. Analyses involved estimation of variance components of hen housed production (HHP), body weight (BWT) and egg weight (EWT) and identification of the best model using log-likelihood ratio test. The population consisted of 1.3M chicks over 24 generations. For HHP, the only significant random term was the direct genetic effect (σ2A) of the chick (M1). A typical model used in commercial setting (M2) includes σ2A and permanent environment (σ2p.e) effects. For BWT, the best model included σ2A, maternal genetic effects and σ2p.e (M3), for EWT it also included covariance between σ2A, and maternal genetic effects (M4). Prediction errors through bias from the models were estimated by regressing phenotypes on BVs predicted from ancestral information only. Slopes for BWT and EWT were significantly different from 1. There was no statistical difference between slopes for HHP due to large SE, caused by small numbers (n=1,800) and low h2. The slopes for BWT predictions from the best model M3 (β=0.91, SE 0.02) and M2 (β=0.94, SE 0.02) were not statistically different (n=70,614). For M1 the slope was 0.77 (SE 0.02). The slopes for EWT were similar for the best model M4 (β=0.9, SE 0.09) and M2 (β=0.91, SE 0.08). Slopes from models M1 and M3 were estimated at 0.89 and 0.88 (SE 0.08) respectively. Differences in PEV were small, with lowest values found for M3 in BWT, M2 in HHP and M1 in EWT. Although best models for BWT and EWT were not statistically different from M2, they hold the potential benefit of estimating maternal BVs and thus expanding breeding objectives.