The Mantel-Haenszel procedure (N. Mantel and W. Haenszel, 1959) and its extension to constructed response items, the Generalized Mantel Haenszel (A. Agresti, 1990), compare performance of subgroups across different score groups to determine differential item functioning (DIF). At each level of comparison, or score group, the subgroups are considered comparable, and the expectation is that the two groups should perform equally well on the item. If there is a significant difference between the subgroups in terms of performance, the item is potentially biased in favor of one subgroup or the other. Traditionally, the levels of comparison have been determined by total raw score. In this paper, an alternative method is proposed for determining the levels of comparison based on an examinee's ability as estimated by the Graded Response Model of F. Samejima (1969). A comparison between the methods using simulated data demonstrates: (1) the high percent of agreement between them; (2) the similar patterns and magnitudes of the overall hit rates and the hit rates for adjusted items; (3) similar patterns and magnitudes for the Type I and Type II error rates; and (4) agreement between the types of items detected by item type and amount of adjustment. (Contains 6 figures, 10 tables, and 7 references.) (Author/SLD)