Abstract

We consider the problem of learning parameters of latent variable models frommixed (continuous and ordinal) data with missing values. We propose a novelBayesian Gaussian copula factor (BGCF) approach that is consistent undercertain conditions and that is quite robust to the violations of theseconditions. In simulations, BGCF substantially outperforms two state-of-the-artalternative approaches. An illustration on the `Holzinger & Swineford 1939'dataset indicates that BGCF is favorable over the so-called robust maximumlikelihood (MLR) even if the data match the assumptions of MLR.

 

Quick Read (beta)