A comparison of a pseudo-likelihood estimation and full information maximum likelihood estimation for fitting multivariate longitudinal ordinal data

We consider a latent variable model for multivariate ordinal responses accounting for dependencies among items. Time-dependent latent variables and random effects account for the inter-dependencies of the multivariate ordinal items. Model estimation is usually obtained using the full maximum likelihood via the EM algorithm. However, computationally problems can arise due to the calculation of multiple integrals involved in the likelihood. The paper proposes a pseudolikelihood approach which involves only bivariate marginal probabilities. The proposed estimation method is evaluated by means of a little simulation study. A real data example illustrates the performance of both the full and the limited information estimation method