MODELING LACTATION CURVES IN DAIRY COWS WITH LINEAR, NON-LINEAR AND MIXED MODELS

PICCARDI, M.; BALZARINI, M.; MACCHIAVELLI, R.; CAPITAINE FUNES, A.; BÓ, G.

Conferencia; International Biometric Conference; 2012

The modeling of daily milk yield at different days from calving is crucial to improve management decisions in dairy herds. A total of 95,079 monthly test-day milk records collected from 9,390 primiparous cows from 35 Argentinean herds were used to statistically model the milk production along lactation. Models of lactation curves have been traditionally fitted in the context of nonlinear and linearized fixed effect models. More complex approaches, such as random coefficient regression models represent a most innovative initiative to adjust individual milk yield data. Mixed models are proved to be efficient to handle heteroscedasticity, but in dairy science the problem of heterogeneous variance through time (days in milk after calving) is rarely considered. In this study we fitted two nonlinear models (Wood and MilkBot model). Each model was fitted with and without a random subject specific (cow) effect to describe the 305-day milk yield lactation curve and to predict the peak yield, the peak time and to estimate the total milk yield. Additionally we fit generalized linear mixed models. We used the nonlinear mixed (NLMIXED) and the linear mixed (MIXED) procedures in SAS. Comparisons of the models were made based on the square root of the mean. Results suggest that modeling standard nonlinear functions for lactation curves incorporating a random specific (cow) effect improved the fittings and predictions. An alternative good adjust, either in summer and winter lactations, was obtained with the linearized version of Wood´s model that take into account heteroscedasticity by different lactation stages (e.g. less than 50 days in milk, between 50 and 100 days in milk, and more than 100 days in milk).