VEMIRT: A Family of Variational Methods for Multi-dimensional IRT Applications (Chun Wang)

In psychological and behavioral science, the increasing availability of rich survey data and the emerging needs of assessing multifaceted constructs pose great challenges to existing techniques used to handle and analyze heterogeneous assessment data. The multiple-group multidimensional IRT (MIRT) model that relaxes strict measurement invariance assumption is a viable psychometric tool to establish commensurate measures for the constructs of interest. To calibrate this model more efficiently, we developed a family of innovative Gaussian variational expectation-maximization (GVEM) methods. In this talk, I will briefly introduce GVEM followed by two latest developments, including (1) using a multiple-group GVEM with regularization to detect differential item functioning (DIF); and (2) combining GVEM with a novel penalty to detect intersectional DIF. I'll share applications of both methods.