Estimation and Inference in Latent Variable Models

Pennsylvania State Univ University Park, University Park PA

Investigators

Abstract

Under certain patterns of constraints used to identify latent variable models, the quality of inference can be affected by peculiar properties of the likelihood. The project will examine issues in maximum likelihood (ML) estimation of latent variable models, including identification, convergence, and standard errors of the parameters. The work will begin by studying simple factor models and extend to more complicated models such as growth curves, behavioral genetics models, and multi-method, multi-trait matrices. In a Bayesian setting, some of the difficulties with ML estimation are avoided, but other problems emerge that require new considerations. Specifically, researchers must ensure that posterior means and variances are not biased by mode-switching. The project will explore the effectiveness of various approaches to applying constraints and prior information to ensure proper Bayesian inferences. Latent variable models are commonly used to model complex phenomena in psychology, sociology, economics, education, and human development, and software for maximum likelihood based analysis is widely available. The project will lead to recommendations for best practices for estimation and inference in these models.