Many applications of mixed linear statistical models for genetic evaluation of dairy cattle assume that genetic and residual components of variance are each constant across environments. However, this assumption is violated for production and conformation traits, which can reduce accuracy of selection and cause biases in the proportions of breeding animals chosen from each environment. Best linear unbiased prediction can accommodate heterogeneous variances if the appropriate variance components are known. Variance components may need to be estimated within individual herds using Bayesian or empirical Bayes methods, but such approaches may not yet be computationally feasible on a national basis. For this study, a structural log-linear model for sire and residual variances was used to identify various management factors associated with differences in within-herd variance components. Increases of herd size and within-herd mean were associated with significant increases of within-herd residual variance for milk and fat yields, but residual variance of milk yield decreased slightly as the proportion of registered animals in the herd increased. Type of milking system, silage storage system, DHI testing program, use or nonuse of a TMR, and use or nonuse of automatic milking machine removal devices also significantly affected residual variances. However, differences in sire variances across levels of management factors were not significant.