Accuracy of Item Parameter Linking Accounting for Uncertainty in Parameter Estimates
In item response theory, the estimated item parameters of different administrations should theoretically be the same if random samples are drawn from a population. However, in practical testing, different administrations of the same inventory will have slightly different item parameters. Linking methods put item parameters on the same scale when administered to different random groups. This helps to prevent unfair advantages or disadvantages for different groups. This study compares the accuracy of the recently proposed Barrett and Van der Linden (BVL) linking method, the Haebara (HB) linking method, and concurrent calibration (CC) to link data that fit the three-parameter data-generating model as well as in misspecified data from the reduced heteroscedasticity (RH) model. In this study, The HB method performed best in the data simulated from the three-parameter model, and CC also performed better than the BVL method. In the data produced from the RH data-generating model, both HB and CC outperformed BVL. This information can inform further research regarding optimal methods for linking data in different data-generating conditions. This is a preliminary step toward creating guidelines for using linking methods in the analysis of real data, even when the real data is unlikely to be similar to a three-parameter distribution.
Quantitative psychology|Educational tests & measurements|Statistics
Plourde, Jessica Marie, "Accuracy of Item Parameter Linking Accounting for Uncertainty in Parameter Estimates" (2021). ETD Collection for Fordham University. AAI28652632.