Six item bias detection techniques (distractor analysis, the Mantel-Haenszel procedure, Angoff's delta plot, Angoff's modified delta plot, Stricker's partial correlation index, and direct comparison of item difficulties) were compared for their ability to identify biased items at three stages of test development. The techniques were applied to items administered during the spring 1989 item pilot (n=about 1,000 for each of 36 forms), the fall 1989 form pilot (n=about 460 for each of three tests), and the spring 1990 final test version (n=about 46,000 each for the reading, mathematics, and language tests) of the Grade 9 Alabama Basic Competency Tests. Four subgroup pairs were analyzed for the reading, mathematics, and language tests; each administration; and each technique. The subgroup pairs included black versus white females, black versus white males, black females versus black males, and white females versus white males. The method with the greatest stability across administrations was the direct difficulty difference, followed closely by distractor analysis. The Mantel-Haenszel procedure and the partial correlation index both showed weak to moderate stability. The poorest results concerning stability were found for both Angoff techniques. The modified Angoff version is not a desirable technique, since it can identify inappropriate items. Results suggest the possible use of more than one index to examine a set of items. A 60-item list of references is included. (SLD)