Differential item functioning (DIF) is usually regarded as a test fairness issue in high-stakes tests. In low-stakes tests, it is more of an accuracy problem. However, in low-stakes tests, the same method, deleting items that demonstrate significant DIF, is still employed to treat DIF items. When political concerns are not important, such as in low-stakes tests and instruments that are not used to make decisions about people, deleting items might not be optimal. Computerized adaptive testing (CAT) is more and more frequently used in low-stakes tests. The DIF-CAT method evaluated in this research is designed to cope with DIF in a CAT environment. Using this method, item parameters are separately estimated for the focal group and the reference group in a DIF study, then CATs are administered based on different sets of item parameters for the focal and reference groups.

To evaluate the performance of the DIF-CAT procedure, it was compared in a simulation study to (1) deleting all the DIF items in a CAT bank and (2) ignoring DIF. A 300-item flat item bank and a 300-item peaked item bank were simulated using the three-parameter logistic IRT model with D = 1,7. 40% of the items in each bank showed DIF. The DIF size was b and/or a = 0.5 while original b ranged from -3 to 3 and a ranged from 0.3 to 2.1. Three types of DIF were considered: (1) uniform DIF caused by differences in b, non-uniform DIF caused by differences in a, and non-uniform DIF caused by differences in both a and b. 500 normally distributed simulees in each of reference and focal groups were used in item parameter re-calibration. In the Delete DIF method, only DIF-free items were calibrated. In the Ignore DIF method, all the items were calibrated using all simulees without differentiating the groups. In the DIF-CAT method, the DIF-free items were used as anchor items to estimate the item parameters for the focal and reference groups and the item parameters from recalibration were used. All simulees used the same item parameters in the Delete method and the Ignore method. CATs for simulees within the two groups used group-specific item parameters in the DIF-CAT method. In the CAT stage, 100 simulees were generated for each of the reference and focal groups, at each of six discrete q levels ranging from -2.5 to 2.5. CAT test length was fixed at 40 items. Bias, average absolute difference, RMSE, standard error of θ estimates, and person fit, were used to compare the performance of the DIF methods. DIF item usage was also recorded for the Ignore method and the DIF-CAT method.

Generally, the DIF-CAT method outperformed both the Delete method and the Ignore method in dealing with DIF items in CAT. The Delete method, which is the most frequently used method for handling DIF, performed the worst of the three methods in a CAT environment, as reflected in multiple indices of measurement precision. Even the Ignore method, which simply left DIF items in the item bank, provided θ estimates of higher precision than the Delete method. This poor performance of the Delete method was probably due to reduction in size of the item bank available for each CAT.

Session Video