Beef cattle feed intake and growth: empirical Bayes derivation of the Kalman filter applied to a nonlinear dynamic model.

ABSTRACT

Feed intake and growth rate of a single group of growing-finishing feedlot beef cattle are difficult to predict. Subsequent performance can be projected more precisely from past performance of a group of cattle. Using an adaptation of the statistical procedure called the empirical Bayes (EB) derivation of the Kalman filter, estimates from any dynamic model (M) can be adjusted based on past performance. The model may be either linear or nonlinear. With this procedure, predictions of intake and body weight gain are periodically updated by multiplying the estimates from M by statistically weighted factors. These factors are derived from the ratio of performance in each period to the performance predicted by M. For comparison to the EB adjustment, weighting of factors by least-squares (LS) adjustment also was tested to predict subsequent feed intake and gain. The test data base consisted of periodic feed intake and gain observations (usually 28 d) for 200 pens of feedlot steers. Bias of prediction was lower for EB than for M or LS for feed intake and (usually) gain. Intake and gain prediction errors averaged for the whole feeding period were .42 kg/d for intake and .14 kg/d for gain by EB, being .84 and .18 kg/d more precise than M and .12 and .33 kg/d more precise than LS predictions. More than two observations were needed before LS produced accurate prediction but after about 80 d, LS and EB estimates converged. Accuracy of both estimates continued to improve as days on feed increased.(ABSTRACT TRUNCATED AT 250 WORDS)