DIF DETECTION WITH THE LIKELIHOOD RATIO TEST WITHIN THE FRAMEWORK OF ITEM RESPONSE THEORY
 
Wang, W.C. and Yeh, Y.L., National Chung Cheng University, Taiwan
 
This study investigates the Type I error rates and statistical power in DIF detection using the likelihood ratio test within the framework of item response theory. Several factors are manipulated: (a) scale constraint (between factor, all-other-item-anchored, 1-item-anchored, 4-item-anchored, and 10-item-anchored), (b) model (between factor, the two-, three- parameter logistics models and the graded response model), (c) percentage of DIF items in a test (within factor, 0%, 12%, 20%, and 32%), and (d) DIF favor (within factor, one-sided and both-sided). One hundred replications are made in each condition. The all-other-item-anchored method yields good Type I error rates and power if the mean item difficulty between reference and focal groups approaches zero, regardless of the percentage of DIF items in a test is as high as 32%. The 1-, 4-, and 10-item-anchored method yields satisfactory Type I error rates and good power across all the conditions. For practical DIF analysis, the all-other-item-anchored method is not recommended because it is applicable only under very stringent conditions. The item-anchored method is recommended. As long as one anchored item is selected correctly, the 1-item-acnhored method yields reasonable type I error and power. If four anchored items could be selected correctly, the 4-item-anchored method would yield even better DIF results. Although the 10-item-anchored method yields somewhat better DIF result than the 4-item-anchored method, the former is not highly recommended because it may not be easy for DIF analysts to locate 10 anchored items correctly.