Loglinear Models with Missing Data: A Latent Class Approach

Abstract

This article discusses latent class models as an approach to categorical data analysis when some variables have missing data. In contrast to standard latent class models in which each variable is either latent or observed for all sample observations our models include variables that are latent (missing) for some observations and manifest (not missing) for others. Particular attention is devoted to models in which the probability that an observation is missing on a variable depends on the level of that variable itself; in other words it focuses on models for noningnorable nonresponse. It discusses the estimation of loglinear models using a latent class approach. It shows how this framework applies to various missing data models and indicates how Habermans DNEWTON program can be used to estimate missing data models.