Analyses of variance were conducted on first lactation milk, fat and protein production records in England and Wales of daughters of British Friesian sires. Herds were split on milk yield into high and low levels of mean production and, in subsequent analyses, into high and low levels of within herd variance and coefficient of variation using all first lactation records. Data were then extracted on daughters of 798 young sires undergoing progeny test and on 118 widely used proven sires to generate connections. Least squares analyses were conducted within levels and genetic correlations estimated from the covariance of sire effects.W ith data split on mean yield, the heritability of milk yield was 0-24 at the low level and 0-30 at the high level, that of log transformed yield being 0-25 and 0-35 respectively. With data split on variance the corresponding figures were 0-24, 0-30, 0-27 and 0-36 respectively, and when split on coefficient of variation, 0-22,0-26,0-26 and 0-32. There were similar increases for fat and protein yield, proportionately smaller increases for fat and protein content.Genetic correlations were close to 1-0 between high and low levels for all traits on all criteria of data splitting. As a consequence progeny testing of bulls is rather more accurate at high mean or variance of production levels and data can be combined optimally without scaling. Cows of the highest predicted value using an index will be found in high variance herds.