Efficient estimation of item response theory models with multiple groups in large-scale educational assessments

In large-scale educational assessment programs, students in different countries or regions are assessed in subject domains such as mathematics, reading and science. In these programs, the underlying model is an item response theory model which defines the relationship between a hypothesized latent variable vector, the background variables and the observed item responses. Due to the large amount of data involved in these assessment programs, several simplifying assumptions are usually made when estimating the parameters of the underlying model. These assumptions include a unidimensional latent variable and measurement invariance across regions. This talk presents a new estimation method using a second-order Laplace approximation of the likelihood for multidimensional multiple group item response theory models which enables the use of more realistic models in large-scale educational assessment programs. We illustrate how the proposed method can be used to improve the estimation of population parameters using large-scale assessment data and suggest ways in which the operational procedures can be modified to better assess the performance of students in individual countries or regions.