Comparison of Estimation Algorithms for Latent Class Models

Abstract

Latent class models are used to identify unobserved subgroups in a population, but estimation ischallenged by multimodal likelihood surfaces that can produce local solutions. This study employed Monte Carlo simulations of four-class models with varying sample sizes, class prevalences, and measurement error to investigate the prevalence, proximity, and interpretability of local optima, as well to compare the behavior of two estimation algorithms: Expectation- Maximization and Newton-Raphson. Local solutions often emerged in difficult conditions and yielded qualitatively different class interpretations, highlighting potential instability in parameter recovery. In addition, the two algorithms exhibited different estimation behavior despite being initialized with the same sets of starting values. These findings support the use of exploring local solutions and employing multiple estimation strategies to ensure robust and reliable inference in latent class analysis.