| dc.description.abstract | Under common-item nonequivalent groups design linking, the functionality of common items as bridge between two parallel forms entails strict content and statistical restrictions on these items. Despite representativeness to the whole test, and similar positions of item placement on each of the parallel forms, common items could present differential item functioning (DIF) effects between test takers of different forms, especially when the groups of test takers are not equivalent in ability profiles. DIF common items under such scenario should impair the adequacy of linking if they were used as linking items instead of being examined and then taken care of. However, the menace of DIF common items on linking has not been substantiated by research yet.
In the current study, I have reviewed the related literature in item response theory, equating, and differential item functioning with emphases on linking methods, forms of DIF, and DIF detection methods. Responding to the scarcity of research on DIF common-item effects on linking, a series of Monte Carlo simulation studies were conducted under common-item nonequivalent groups design linking, testing potential influential factors in empirical research, i.e., sample size, ratio of common items, ratio of DIF items, magnitude of DIF, form of DIF, and direction of DIF. Recovery of equating slope A and equating intercept B, and item discrimination a and item location b was evaluated using signed bias and root mean square error (RMSE).
My results show that generally as sample size went up, the bias and RMSE went down, an effect tended to level off at 1000 participants in each group. The number of DIF common items and the magnitude of uniform DIF items were testified as more influential factors than number of common items and the direction of DIF. As the number of DIF common items increased, and/or the magnitude of uniform DIF increased, the bias and RMSE increased quickly. Bias and RMSE of equating intercept B was mostly related to the uniform DIF common items against or in favor of group 2 test takers, while bias and RMSE of equating slope A was mostly related to the nonuniform DIF common items against or in favor of group 2 test takers. Only B was seriously biased when having uniform DIF. Both B and A were seriously biased when having uniform and nonuniform DIF at the same time. Overall, the mean bias and mean RMSE of item discrimination a, and the mean bias and mean RMSE of item location b were small on most simulation conditions. Within common items, the mean bias and mean RMSE of item discrimination a, and the mean bias and mean RMSE of item location b were sensitive to simulation condition changes. Results were canvassed and limitations were pointed out at the end of this dissertation with recommendations for future research. | en |
