Comparing Item Parameter Recovery of the Compensatory MIRT Model by MH-RM and MML across Model Structures and Numbers of Latent Traits

• In Event: Applications of Statistical Methodology, Psychometrics, and Research-in-Progress

Abstract

This study compares the item parameter recovery by MH-RM and MML under the 2PL multidimensional item response model across model structures (simple and complex) and number of latent dimensions (2 and 4 dimensions). Test length was fixed at 15 items per dimension in the simple structure, and 5 items were specified to measure an additional trait in the complex condition. Results indicated that within-item dimensional contamination brought about a negative impact on item discrimination recovery regardless of methods, as evidenced by the increase in total estimation errors by both MML and MH-RM. As the MIRT model has more latent dimensions, MML recovery of item slopes worsened with more negatively biased estimates and higher total errors, whereas MH-RM estimations continued to be stable and little difference was observed between the two- and four-latent trait model estimations. Estimations of the threshold parameters were generally unbiased, and both MML and MH-RM offered similar accuracy. Overall, MH-RM outperformed MML when the MIRT model involves a large number of latent traits, as expected, but MML performance was relatively strong even with up to four latent dimensions to estimate, at which point MML estimations of item slopes more erroneous but still acceptably accurate.

Authors

• Hoan Do, Clinical Outcomes Solutions

• Gordon Brooks, Ohio University