Various procedures have been developed to select efficient storage structures; all of these procedures, however, require certain statistics to describe the use of attributes by problem programs. A very practical problem faced by anyone attempting to use any of these selection techniques is the determination of values for the attribute use statistics. This paper reports on one method which has been empirically tested for determination of attribute access probabilities. Central to this method is recognition of five kinds of statistic collection biases which are systematically considered by the proposed method but which do confound collection by previously developed methods. Finally, experimental results are reported which indicate specific comparative advantages of the proposed method over ones used by the DP staff of a medium-sized manufacturer.