Performance of Sandwich Estimator in Item Response Theory When Normality Assumption of Latent Trait Is Violated

• In Event: 58.028 - Applications of Item Response Theory

Abstract

Sandwich estimator has been implemented in many item response theory (IRT) software as it theoretically produces consistent standard estimate even if model misspecification is present. This study compared the performance of Sandwich estimator against several other standard error procedures commonly used in IRT model parameter estimation, under the conditions where the normality assumption of latent trait is violated. Preliminary results showed that sandwich-type covariance matrix failed to provide more consistent standard error than cross-product moment approximation and supplemented EM approach, given the current simulation conditions.

Authors

• Zhongtian Lin, Florida State University

• Insu Paek, Florida State University