Two-level analysis of covariance structures for unbalanced designs with small level-one samples.