Predictive characteristics of lactation models for pasture-based Holstein-Friesian dairy cows

Mathematical functions to describe a series of milk test day records have the advantage of minimizing random variation, while simultaneously summarizing the lactation profile. Five empirical functions and two mechanistic models were used to model herd and individual milk yield profiles of multiparous Holstein-Friesian cows on 113, 290 milk yield records (8438 lactations) collected from 1994-2005. The models tested were the incomplete gamma (IG), a modified gamma (MG), an exponential (EXP), a polynomial regression (PR), a mixed log (ML), the bi-compartmental (BC), and Dijkstra (DJ) functions, the latter two being mechanistic models. Each model was fitted using the non-linear (NLIN) function of SAS. Model accuracy was evaluated based on residual mean square (RMS), the magnitude and distribution of residuals, and the correlation between the observed and predicted values. All the models, except MG, did equally well in portraying the lactation profile. Parameter estimates were significant (P<0.05), with large serial correlations indicating biased predictions, especially during mid-lactation. Correlations of residuals and observed herd average lactations ranged between -0.13 (MG) to 0.19 (IG), while that between observed and predicted was between 0.76 and 0.99 for the same models. Lactation curves of individual cow milk yields were more varied, exhibited the tendency for a second peak which were not accurately modeled. Mechanistic models performed best with herd data, the PR model fitted overall best, while the MG model fitted the profile least accurately in this study.