Comparison between different statistical models for the prediction of direct genetic component on embryo establishment and survival in Italian Brown Swiss dairy cattle

The aims of this study were to infer variance components and heritability for the direct component on embryo establishment and survival related traits and to compare different statistical models in terms of goodness-of-fit and predictive ability. Embryo establishment and survival (EES) was defined as the outcome of an AI event, its direct effect was represented as the effect of the service sire from which semen was taken. Indicators of EES were calving per service (CS) and non-return at 56. d after service (NR56). Insemination records from the Italian Brown Swiss population reared in the Alps were used. Data included 124,206 inseminations performed by 86 technicians on 28,873 cows in 1400 herds. Services were recorded from 1999 to 2008. Linear-sire, linear-animal, threshold-sire, and threshold-animal models were used to estimate (co)variance components for CS and NR56. Four levels of complexity within each model were tested, so that 16 different models were compared for each of the two fertility traits. Comparison was assessed on the basis of the goodness-of-fit and predictive ability. Paternal half-sibs groups were created as average outcome of the inseminations from a given service sire. Goodness-of-fit was evaluated by regressing the service sire estimated breeding value from each model to paternal half-sibs average CS or NR56. Predictive ability was assessed through sums of chi-squared and percentage of wrong predictions. Predictors were the respective service sire's estimated breeding values constructed on a reduced (independent) training dataset, including years from 1999 to 2005, and predictands were the paternal half-sibs means for every bull in the remaining years (2006-2008). Prediction of EES was considered differently according to whether service sires had observations in the training dataset (prediction of proven bulls) or they had not (prediction of young bulls). Estimates of heritability ranged from 0.011 to 0.119 for CS, and from 0.005 to 0.054 for NR56. In general, threshold models explained a larger proportion of additive genetic variance than linear models, and animal models yielded higher heritabilities than sire models. Calving per service was much more predictable than NR56, but no significant differences were found among models. Although heritabilities were low, the prediction of future EES of a paternal half-sib group is feasible