Latent GOLD 4.5 is a powerful latent class and finite mixture program. Latent GOLD contains separate modules for estimating three different model structures: Latent Class Cluster models; Discrete Factor (DFactor) models; Latent Class Regression models. The Advanced option includes additional advanced features for continuous latent variables (CFactors), multilevel modeling, and survey options for complex sample data.

Features
- Full windows implementation - point and click
- Interactive graphics provide new insights into data and powerful model diagnostic capabilities
- Flexible model structures can handle variables of different metrics
- Automatic generation of sets of random starting values
- Fast, efficient maximum likelihood and posterior mode estimation based on EM and Newton Raphson algorithms
- Use of Bayes constants to eliminate boundary solutions
- Bivariate residual diagnostic for local dependencies

Capabilities

Known Class Indicator
This feature allows more control over the segment definitions by pre-assigning selected cases (not) to be in a particular class or classes.

Conditional Bootstrap p-value
Model difference bootstrap can be used to formally assess the significance in improvement associated with adding additional classes, additional DFactors and/or an additional DFactor levels to the model, or to relax any other model restriction.

Overdispersed (Count and Binomial Count in Regression)
Overdispersion is a common phenomenon in count data. It means that, as a result of unobserved heterogeneity, the variance of the count variable is larger than estimated by the Poisson (binomial) model. The overdispersed option makes it possible to account for unobserved heterogeneity by assuming that the rates (success probabilities) follow a gamma (beta) distribution. This yields a negative-binomial model for overdispersed Poisson counts and a negative-binomial model for overdispersed binomial counts. Note that this option is conceptually similar to including a normally distributed random intercept in a regression model for a count variable.

The overdispersion option is useful if one wishes to analyze count data using mixture or zero-inflated variants of (truncated) negative-binomial or beta-binomial models (Agresti, 2000; Long, 1997; Simonoff, 2003). The negative-binomial model is a Poisson model with an extra error term coming from a gamma distribution. The beta-binomial model is a variant of the binomial count model that assumes that the success probabilities come from a beta distribution. These models are common in fields such as criminology, political sciences, medicine, biology, and marketing.

Home Page - http://www.statisticalinnovations.com/products/latentgold.html