A Comparison between the Structural Equation Model and IRT Graded Response Model on Item Selection of Personality Scale

Objective To compare the application of the structural equation model and the IRT graded response model in item selection of personality scale. Methods Lisrel 8.70 and Multilog 7.03 were applied respectively to estimate parameters and indexes of goodness-of-fit in the structural equation model and the IRT graded response model in terms of the second factor of "forthright" in the Chinese College Student Personality Scale, on the basis of 7229 measured cases. Results The results of the two models indicated that neither was good enough in terms of the goodness-of-fit on item 5, 6, 7, and 8. On the one hand, the typical feature was found to be a too low factor load and an unsatisfactory overall fit index in term of the structural equation model. In term of the IRT graded response model, the major faults were found to have a rather unsatisfactory differentiation parameter and a bad location parameter with too poor a shape of the characteristic curve and the information curve for the related items. Items 6 and 8 tend to be much worse in term of the structural equation model while Items 5 and 7 are found to be worse in term of the IRT graded response model. Conclusion The inference results of both models were consistent in general, although there were slight differences on a few specific items. These two models could be applied in combination since each one has its advantages.