The Stability of a Partial Correlation Index for Identifying Items that Perform Differentially in Subgroups

Abstract

The stability was evaluated of a partial correlation index and two other methods-comparisons of item characteristic curves and comparisons of item difficulties-in assessing race (white vs. black) and sex differences in the performance of verbal items on the GRE Aptitude Test. In general, the partial correlation index, like the other indexes, exhibited substantial consistency in identifying the same items in different samples from the same populations and displayed appreciable congruence in its values in these samples. The major exception concerned the general instability of the indexes in the analyses of sex differences for blacks, probably stemming from the small samples of blacks involved. It was tentatively recommended that all these indexes be restricted to samples of a least 1,500, with no fewer than 300 in the smallest subgroup.