Comparison of single-trait to multi-trait national evaluations for yield, health, and fertility

Abstract

Flexible software was designed to replace the current animal model programs used for national genetic evaluations. Model improvements included (1) multi-trait processing, (2) multiple fixed class and regression variables, (3) differing models for different traits, (4) random regressions, and (5) foreign data included using pseudo-records. Computational improvements included (6) parallel processing, (7) renumbering class variables to equation numbers within the program so that estimated effects are output with original identification numbers, and (8) reliability computed within the same program. When applied to 3 fertility traits of 27,971,895 cows and heifers, the new model used daughter pregnancy rate as a correlated trait to improve heifer and cow conception rate evaluations for older animals and in herd-years where records are missing, and also added information from crossbreds. When applied to 7 traits and 76,846,327 lactation records of 30,064,300 cows, gains in accuracy were small for yield and somatic cell score, moderate for daughter pregnancy rate, and larger for productive life for recent bulls compared with single-trait evaluations. For very old bulls, multi-trait gains were also large for protein because lactation records were available only for milk and fat. Multi-trait productive life was computed with exact rather than approximate methods; however, correlated information from conformation was excluded, reducing advantages of the new model over the previous software. Estimates of breed differences, inbreeding depression, and heterosis were similar to previous estimates; new estimates were obtained for conception rates. Predictions were compared by truncating 4 yr of data, and genetic trend validation was applied to all breed-trait combinations. The estimates of trend account for increases in inbreeding across time. Incorporation of foreign data gave correlations above 0.98 for new with previous evaluations of foreign Holstein bulls, but lower for other breeds. The 7-trait model required 35 GB of memory and 3 d to converge using 7 processors. The new software was implemented for fertility traits in 2013 and is scheduled for implementation with yield, somatic cell score, and productive life in 2014. Further revision of the models and software may be needed in the near future to account for genomic preselection.

Keywords: animal model; best linear unbiased prediction; correlated trait; genetic evaluation.