Title and affiliation: Assistant Professor, Department of Psychology, The Ohio State University
Website: https://psychology.osu.edu/people/cho.1240
Title: Constructs may or may not be latent: Studies on two domains of structural equation modeling
Abstract
Structural equation modeling (SEM) enables empirical testing of hypothetical relationships among observed variables and underlying constructs. Traditionally, SEM has assumed all constructs to be latent, existing independently of their indicators, and represents them with (common) factors. This domain is known as factor-based SEM. However, some constructs, such as socioeconomic status and genes, are not inherently latent but instead correspond to a summary or cluster of their indicators. To handle such constructs, component-based SEM has emerged, representing constructs as composite indexes of indicators, termed components.
In this talk, I begin with a systematic comparison of the two SEM domains and briefly introduce two novel SEM methods—structured factor analysis (SFA) and deep learning generalized structured component analysis (DL-GSCA). SFA addresses two long-standing issues in factor-based SEM—improper solutions and factor score indeterminacy—by using a single cost function to estimate model parameters and the probability distribution of candidate factor scores. In contrast, DL-GSCA extends component-based SEM by employing artificial neural networks to identify nonlinear components that maximize predictive power for target outcomes. I will demonstrate their technical foundations and practical utility through empirical applications.