The stability of selected indices for detecting differential item performance (item bias), from one randomly equivalent sample to another, is addressed. Some recent research has criticized these indices as too unreliable for utility in measuring bias in achievement test items. Using data from a national testing of the ACT Assessment, however, this study suggests that the reliability of the indices is situation-specific. Bias detection indices may be viewed as most reliable in testing situations that involve large sample sizes and some item heterogeneity. A preference is also stated for assessing reliability based on signed rather than unsigned indices. (Author)