Multidimensional Item Response Theory Parameter Estimation With Nonsimple Structure Items

Abstract

Estimation of multidimensional item response theory (MIRT) model parameters can be carried out using the normal ogive with unweighted least squares estimation with the normal-ogive harmonic analysis robust method (NOHARM) software. Previous simulation research has demonstrated that this approach does yield accurate and efficient estimates of item discrimination and difficulty across a variety of conditions. However, these studies have been limited primarily to the case of simple structure, where each item is associated only with one of the latent traits underlying the data. The current simulation study seeks to extend this earlier work by comparing NOHARM and unidimensional IRT-based estimates of difficulty and discrimination values for items that do not conform to simple structure, that is, are associated with more than one latent trait. The outcomes of interest were relative bias and standard error values for parameter estimates under a variety of conditions, including the degree of nonsimple structure present. Results demonstrate that both bias and standard error tended to be larger for items that do not conform to simple structure than for those that do, but that the degree of such differences was influenced by factors such as correlation between the latent traits, sample size and distribution of the latent trait, among others.