Techniques for maximizing discrimination between groups in norm referenced measurement to reflect sensitivity to group differences inherent in the evaluation of multilevel educational systems are discussed. Data from the Beginning Teacher Evaluation Study (BTES) were used to examine relationships between reading and mathematics achievement and instructional variables. Intraclass correlation, correlations of class means on items and on total scale, and correlations of class means on items and on alloted time in instruction were used as item selection techniques to form subscales of the fraction test of the BTES. By constructing the subscales on the basis of the between-class relationship of the items to instruction, the sensitivity of the scale to between-class differences in instruction was found to increase and the sensitivity to the same two variables within class decreased. These results indicate the usefulness of selecting items on the basis of their relationship to the variable of interest. (Author/CM)