IRT item bias detection procedures: Issues of model misspecification, robustness, and parameter linking

Abstract

This article examines the consequences of employing IRT item bias detection procedures with multidimensional IRT item data. Parameter linking methods used in previous studies of item bias were investigated in a simulation that minimized the need for such linking. The results illustrate shortcomings of two linking methods that have been employed in IRT item bias detection studies. The effectiveness of these methods depended on several factors, including the number of biased items in a fixed-length test, whether bias was against only one group or more than one group, and the correlation between the two latent abilities. The findings indicated that some current IRT-based statistical procedures for detecting item bias were not generally effective at differentiating biased from unbiased items. Index terms: item bias, item response theory, multidimensional IRT data, parameter linking, reverse bias, statistical artifacts.