Comparison of international dairy sire evaluations from meta-analysis of national estimated breeding values and direct analysis of individual animal performance records

Our objective was to assess the predictive ability of different methodologies for international genetic evaluation of milk yield and to determine the magnitude of differences in the resulting sire estimated breeding values (EBV). Data included first lactation records of 16,057,335 Holstein-sired cows from 237,049 herds in 14 countries. Meta-analysis of national sire EBV using the multiple-trait across country evaluation (MACE) procedure, single-trait analysis of individual animal performance records, multiple-trait analysis of individual animal performance records, and borderless herd cluster model were compared by assessing predictive ability. Comparisons were based on root mean square error of sire EBV from a subset of records from cows calving between 1990 and 1995 and corresponding pedigree indices for sires that received their first genetic evaluations in 1996 or 1997. The number of bulls first evaluated in 1996 or 1997 that were in common between the top 25, 100, and 250 for pedigree index and the top 25, 100, and 250 for EBV were also determined for each method. Average root mean square error of prediction was 10.3 kg 2 for the borderless single-trait model, 6.6 kg2 for the borderless herd cluster model, and 6.7 kg2 for both the borderless multiple-trait model and meta-analysis of national sire EBV using MACE. The mean numbers of common bulls among the top 25, 100, and 250, respectively, when selected on pedigree index and subsequent EBV were 11, 48, and 154 for the borderless single-trait model; 16, 66, and 176 for the borderless multiple-trait model; 16, 66, and 178 for the borderless herd cluster model; and 15, 66, and 178 for meta-analysis of national sire EBV using MACE. Rank correlations between sire EBV from different models ranged from 0.77 for the single-trait borderless model and the meta-analysis using MACE to 0.92 for the borderless multiple-trait and the borderless herd cluster models.