Abstract:
|
In the traditional latent class (LC) model, multiple categorical responses are assumed to be independent within categories of a latent classification variable. This model has recently been extended to incorporate categorical and continuous covariates as predictors of class membership through multinomial logistic regression. Routines for maximum-likelihood (ML) estimation are currently available in Mplus (Muthen & Muthen, 1998) and Latent GOLD (Vermunt & Magidson, 2000). In many examples, however, the likelihood function exhibits unusual features, causing ML estimates and their associated standard errors to behave erratically. In this talk, we explore a variety of theoretical and practical issues surrounding the use of the LC model with covariates, including Bayesian alternatives to ML estimation. We illustrate these issues with an example from adolescent substance use: tracking changes in marijuana use and attitudes among American high school seniors from 1977 to the present, using data from Monitoring the Future.
|