SMABS 2004 Jena University
SMABS 2004 Home Organization About Jena Sponsors Links Imprint / Contact SMABS Home

European Association of Methodology

Department of methodology and evaluation research

Jena University

Contributions: Abstract

Latent class, finite mixture, latent trait, and random-effects modeling with LATENT GOLD

Jeroen K. Vermunt
Tilburg University
The Netherlands

LATENT GOLD was developed as a tool for the estimation of the most important types of latent class and finite mixture models. New in version 4.0 is the possibility to specify models containing continuous latent variables, thus yielding factor analytic, latent trait, and random-effects models, as well as mixture variants of these. The program has a user-friendly SPSS-like graphical interface, reads SPSS and text data files, and provides a convenient method for organizing and managing the various tabular and graphic output sections. The program contains four modules labeled Cluster, Factor, Regression, and Choice.

The Cluster module implements a latent class model with a single nominal latent variable: latent classes can be viewed as unobserved clusters that are homogenous with respect to their responses to the observed indicators. The Factor module can be used for latent class analysis with several ordinal latent variables, and is strongly related to latent trait analysis. In both modules, observed indicators can be nominal, ordinal, continuous, Poisson counts, binomial counts, or any combination of these. Other important features are the easy detection and inclusion in the model of local dependencies, the full information maximum likelihood treatment of missing data, the parametric bootstrap procedure, the possibility to use of covariates to describe/predict class membership, and the interactive graphical displays to facilitate interpretation of the results. New in version 4.0 is the possibility to include continuous latent variables in the Cluster and Factor models, yielding mixture variants of IRT and factor analysis. Other new features are the multilevel latent class model, sample design corrected SEs, and the possibility to include truncated and censored variables.

The Regression and Choice modules implement what is known as latent class or finite mixture regression analysis: rather than assuming that regression coefficients are homogenous in the population, one allows them to differ across classes. In Regression, the dependent variable can be specified to be nominal, ordinal, continuous, a Poisson count, or a binomial count; and in Choice, it can be a discrete choice, a ranking, or a rating. The data set that is analyzed may contain replications, repeated measurements, or other types of dependent observations. This means that the latent class regression and choice models can be used to analyze two-level data structures, such as those obtained in event history or conjoint studies, using a flexible non-parametric random-coefficients approach. Also in the Regression and Choice modules, one can include continuous latent variables, yielding standard and mixture multilevel GLMs. Other new features are the three-level extension of the mixture regression model, sampling design corrected SEs and Wald statistics, as well as an option to specify the dependent variable to be truncated, censored, or zero inflated.