To investigate the effect of violating the assumption of equal item difficulty on Kuder-Richardson (KR) Formula 21 reliability coefficient, 670 eighth-and ninth- grade students were administered 26 short, homogeneous "tests" of mathematics concepts and skills. Both KR Formula 20 and KR Formula 21 were used to estimate reliability on each test. The 26 tests were sorted into a high item difficulty variability group and a low item difficulty variability group, and the magnitude of differences in KR20 and KR21 reliability coefficients were compared for the two groups. The difference in KR20 and KR21 reliability coefficients was significantly greater when the range of item difficulty values was .30 or more. Nevertheless, KR21 was a good estimate of KR20 when the range of item difficulty was relatively narrow. Implications for test selection are suggested. When KR21 has been used to estimate a test's reliability, the user should note that the test has a lower bound of internal consistency reliability, particularly when the item difficulty range is great. (Author/GDC)