Joint analysis of semicontinuous data with latent variables

Abstract

A two-part latent variable model is proposed to analyze semicontinuous data in the presence of latent variables. The proposed model comprises two major components. The first component is a structural equation model (SEM), which characterizes latent variables using corresponding multiple attributes and examines the interrelationships among them. The second component is a two-part model to assess a semicontinuous response of interest. The semicontinuous variable is characterized by a mixture of zero values and continuously distributed positive values. The two-part model manages this semicontinuous variable by splitting it into two random variables; one is a binary indicator to determine whether the response is zero, another is a continuous variable to determine the actual level of the positive response. A full Bayesian approach coupled with spike-and-slab lasso prior is developed for simultaneous variable selection and parameter estimation. The proposed methodology is demonstrated by a simulation study and applied to the analysis of the Chinese General Social Survey dataset. New insights into the interrelationships among non-cognitive ability, education level, and annual income are obtained.