On the Feasibility of Metaheuristic Algorithms to Estimate Growth Mixture Models: A Methodological Investigation

• In Event: Methods and Advancements in Mixture Modeling

Abstract

Direct applications of growth mixture models are prevalent for modeling unknown population heterogeneity of growth characteristics through the extension to latent classes. Maximum likelihood via the expectation-maximization (EM) algorithm remains a predominant estimation approach in this context chiefly because mixture models can naturally be configured as missing data problems where class membership is unobserved. This research focuses on the utility of metaheuristic optimization algorithms for model estimation. Several metaheuristic algorithms, including a hybrid algorithm, are outlined in detail and their extensions to estimating growth mixture models are presented. A Monte Carlo simulation is conducted to compare these approaches with maximum likelihood, vis-a-vis the EM algorithm, under real-world conditions found in applied research.

Authors

• Jeff Harring, University of Maryland

• Charles Fisk, University of Maryland