Challenges and Opportunities in the Development of the Latent Gold Program

• Activity Number: 444
• Type: Topic Contributed
• Date/Time: Wednesday, August 6, 2014 : 8:30 AM to 10:20 AM
• Sponsor: Section on Statistical Computing
• Author(s): Jay Magidson*+ and Jeroen K. Vermunt
• Companies: Statistical Innovations and Tilburg University
• Keywords: latent class ; mixture models ; Latent GOLD ; latent Markov ; multilevel ; random effects

Abstract

Latent GOLD (LG) began as a GUI program for latent class (LC) analysis and mixture regression, and was extended to include known classes, continuous latent variables (CFactors), multilevel variants with higher level discrete and/or continuous latent variables (GClasses and GCFactors), complex sampling options, and extensive monte carlo capabilities such as power calculations for any LG model. Our primary challenges were to maintain a unified structure with a simple beginner's interface while letting advanced users customize highly complex models. For the latter, we added a syntax language with extensive equation writing capabilities and numerous permitted parameter restrictions, and dynamic latent variables to allow mixture latent Markov model estimation.

In this presentation we discuss the various technical and design challenges we faced along the way towards today's version 5.0 which allows latent variables as clusters, underlying factors, or random effects, with response variables of many different scale types, and extensive capabilities to relate the resulting latent variables to external independent and distal outcome variables in a Step 3 analysis, and to score new cases.