A Comparison of IRT (Item Response Theory) Common-Scale Development Methods for Mixed-Format Test With Nonnormal Latent Trait Distribution

• In Event: Approaches to Scaling, Equating, and Linking

Abstract

This paper compares the performance of three IRT linking methods with application of unidimensional IRT models in the presence of non-normal latent trait distribution and multidimensionality associated with format effects through computer simulation and real data. The IRT methods investigated in this study are: separate calibration with linking, fixed parameter calibration, and concurrent calibration. The preliminary results show that when the latent trait distributions are skewed, considerable amount of error can be introduced to the estimation and linking of item parameters. The findings also confirm the literature that non-normal latent distribution has small effects on estimation and linking accuracy when data were not extremely skewed and when the sample sizes are sufficiently large and tests are long.

Authors

• Zhen Li, Government of Newfoundland and Labarador

• Tianli Li, ACT, Inc.