Monitoring Daily Milk Yields with a Recursive Test Day Repeatability Model (Kalman Filter)

Abstract

Mixed model methodology and recursive estimation techniques (Kalman filter) were combined to detect significant changes in the performance level of both individual cows and an entire herd. Yields were predicted for the next time of recording using all data available up to that time. Predicted yields were compared with actual measurements. If the error of prediction, or innovation, exceeded ±2 times its standard deviation, the observation was considered to be significantly different from the former yield level. The data comprised 30,199 records for 135 cows and 366 d. Effects fitted in the model were test day, breed, and lactation class as fixed effects and cow within lactation number as a random effect. A lactation curve was fitted within lactation class. Of the observed milk yields, 9.2% deviated significantly from the expected value in a negative direction. None of the innovation of the fixed-day effect exceeded the threshold of two standard deviations. Compared with the results of rolling average, which were calculated as the average of a 10-d period, over 20% of the observations of the daily milk yield were classified differently by the two methods.The mixed-model method for recursive estimation takes better account of environmental and lactational effects influencing daily milk yields as the rolling average. The mixed-model recursive estimation method was applicable for the detection of suspicious (i.e., outside a specified prediction interval) observations of individual cows at the time of the actual recording.