Improving IRT item bias detection with iterative linking and ability scale purification

The effectiveness of several iterative methods of item response theory (IRT) item bias detection was examined in a simulation study. The situations employed were based on biased items created using a two-dimensional IRT model. Previous research demonstrated that the non-iterative application of some IRT parameter linking procedures produced unsatisfactory results in a simulation study involving unidirectional item bias. A modified form of Drasgow’s iterative item parameter linking method and an adaptation of Lord’s test purification procedure were examined in conditions that simulated unidirectional and mixed-directional forms of item bias. The results illustrate that iterative linking holds promise for differentiating biased from unbiased items under several item bias conditions. In addition, a combination of methods, involving cycles of iterative linking followed by ability scale purification, was found to be even more effective than iterative linking alone. This combination of procedures totally eliminated false positive misidentifications for the most pervasive item bias condition, and false negative misidentifications were also reduced. Combining iterative linking with ability scale purification appears to be a viable method for analyzing multidimensional IRT data with unidimensional IRT item-bias methods. Index terms: ability scale purification, item bias, item response theory, iterative linking, iterative methods, metric linking, multidimensional IRT model