An Approach to Evaluate Nontrivial Misfit in Item Response Theory

• In Event: Innovative Approaches to Item Response Theory

Abstract

In IRT, testing the exact model-data fit is less useful because no model fits the data perfectly. Therefore, this study proposed the use of the range-H_0 testing approach to assessing non-negligible model-data fit at the item level in a large-scale testing context. To test the range-H_0, we introduced a new fit statistic as well as applied Orlando and Thissen’s (2001) S-X^2 fit statistic. The results of a simulation showed that the range- H_0 testing approach with the new fit statistic demonstrated controlled Type I error rates and reasonable power whereas S-X^2 was under-power even with a very large sample size.

Authors

• Hwanggyu Lim, The Graduate Management Admission Council

• Craig S. Wells, University of Massachusetts - Amherst

• Hongyu Diao, Educational Testing Service