Additive genetic, permanent environmental, and residual components of variance were estimated at three herd production levels by restricted maximum likelihood for all lactation milk and fat yields and natural logarithm of yields. Data consisted of 121,136 mature equivalent, 2x, 305 day first and later lactation yields for 91,206 artificially sired Holstein cows calving between 1979 and 1984 throughout the United States. A total of 485 of 526 sires represented had first crop daughters in the data. Three production levels were defined by mean mature equivalent milk yield of all cows freshening in the same herd-year. The univariate model of analysis included fixed herd-year-season and sire genetic group and random sire nested within group, cow nested within sire, and residual effects;The analyses of untransformed records resulted in components of variance that in general increased as the level of production increased. For both milk and fat yield, genetic variance increased at the greatest relative rate, resulting in the largest estimates of heritability at the high level of production. Correlations of sire solutions across production levels were close to expected values, indicating the absence of genotype environment interaction. Variance components remained heterogeneous after log transformation of yields. Estimates of residual components on the log scale decreased as production level increased. Estimates of heritability were not changed by log transformation of yields;After variance component estimation, cows and sires were evaluated for milk yield using three mixed models. In two models, variances estimated at the medium production level were used for all records and the analysis used either untransformed or log transformed yields. In the third model, untransformed yields were used and heterogeneous variances at the three production levels were considered. Rank correlations of sire and cow evaluations from the three models were close to unity. Ranks of top sires were similar, but differences in ranks of top cows were large across the three models. A model accounting for heterogeneous variances at several production levels is feasible if variance estimates are available. If heterogeneity of variances are ignored, untransformed yields should be used instead of log transformed yields.