Latent variable modeling for mixed-type outcomes

Abstract

Latent variable models provide an important tool for the analysis of multivariate data due to reduction of dimensionality and diverse applications of latent quantities. Most previous works discuss the latent variable models with all manifest variables of the same type--- that is, all continuous or all discrete. However, owing to the nature of the problems and the design of questionnaires, mixed discrete and continuous data are very common in behavioral, medical and social research. In this thesis, we propose using a Bayesian framework to handle such mixed data. On the basis of appropriate prior distributions, to avoid heavy computation in evaluating the multiple integrals, the Markov Chain Monte Carlo (MCMC) method is implemented to obtain the posterior distributions.