Comparison of feed evaluation models on predictions of milk protein yield on Québec commercial dairy farms.

Feed evaluation models (FEM) are a core part in dairy cow feeding. As these models are developed using different biological and mathematical approaches mainly tested in a research context, their abilities to predict production in commercial farms need to be validated, even more so when they are used outside the context of their development. Four FEM-National Research Council, 2001 (NRC_2001); Cornell Net Carbohydrate and Protein System, 2015 (CNCPS); NorFor, 2011; and INRA, 2018 (INRA_2018)-were evaluated on their abilities to predict daily milk protein yield (MPY) of 541 cows from 23 dairy herds in the province of Québec, Canada. The effects of cow and diet characteristics were tested on the residuals of MPY. Sensitivity and uncertainty analyses were then performed to evaluate the influence of the uncertainty of the main characteristics of cows and feed ingredients measured on the farm and used in the 4 FEM on the predictions of metabolizable protein (MP) supply and MPY. The 4 models had acceptable predictions of MPY, with concordance correlation coefficients (CCC) ranging from 0.75 to 0.82 and total bias ranging from 12.8% to 19.3% of the observed mean. The Scandinavian model NorFor had the best predictions with a CCC of 0.82, whereas the 3 other models had similar CCC at 0.75 to 0.76. The INRA_2018 and NRC_2001 models presented strong central tendency biases. Removing herd effect put the 4 FEM at the same level of performance, with 11.9 to 12.4% error. Analyzing model behavior within a herd seems to partly negate the effect of using predicted dry matter intake (DMI) in the comparison of models. Diet energy density, days in milk, and MPY estimated breeding value were related to the residual in the 4 models, and Lys and Met (as percent of MP) only in NRC_2001 and NorFor. This suggests that inclusion of these factors in these models would improve MPY predictions. From the sensitivity analysis, for the 4 FEM, DMI and factors affecting its prediction had the greatest influence on the predictions of MP supply and MPY. Of the feed ingredients, forage composition had the greatest effect on these predictions, including a strong effect of legume proportion with NorFor. Diet acid detergent fiber concentration had a very strong effect on MP supply and MPY predictions only in INRA_2018, because of its effect on organic matter digestibility estimation. The range of predictions of MP supply and MPY when combining all these potential uncertainties varied depending on the models. The INRA_2018 model presented the lowest standard deviation (SD) and NorFor the highest SD for the predictions of both MP supply and MPY. Overall, despite the fact that FEM were developed in a research context, their use in a commercial context yields acceptable predictions, with NorFor yielding the best predictions overall, although within-herd responses varied similarly for the 4 tested models.