Computer simulation was used to analyze the effect of considering contemporary groups as fixed (Model [1]) or random (Model [2]) on genetic value predictions. Eighty-eight types of populations were generated from four different ratios of error to contemporary group variances, two types of association (random and nonrandom) among sires’ breeding value and contemporary groups, and 11 contemporary group sizes. Non-random association among sires and contemporary groups resulted in a positive correlation (.6) between them. Both sires’ additive value and contemporary groups were considered as random variables in the simulation process. Prediction error variance, mean squared error, correlation between predicted and real breeding values, and rank correlation were measured. When sires were randomly distributed across contemporary groups, Model [2] produced equal or smaller values of prediction error variance and mean squared error and correlations larger than or equal to Model [1] in all cases. Advantages of Model [2] were enhanced when contemporary group size was small. When sires were nonrandomly distributed across contemporary groups, Model [2] yielded smaller prediction error variance and larger correlations than Model [1] in all populations. Mean squared error was larger under Model [2] than under Model [1] when contemporary group size was small (2 and 3) and error to contemporary group variance ratio was low (1.5 and 2.75). For the worst case, when contemporary group size was equal to 2 and error to contemporary group variance ratio was equal to 1.5, mean squared error was 41.6% larger under Model [2] than under Model [1]. © 1992, American Dairy Science Association. All rights reserved.