Study of Regularized Generalized Structured Component Analysis to Overcome Multicollinearity in Component-Based SEM

Abstract

Multicollinearity is one of the issues that may arise in the analysis of Structural Equation Modeling (SEM). An indication of multicollinearity is the high correlation between latent variables and the correlation between indicators forming the latent construct. Multicollinearity causes the interpretation of SEM analysis to be inappropriate. In this study, Regularized generalized structured component analysis (RGSCA) is used as a solution to overcome multicollinearity in component-based SEM. The research aims to apply RGSCA to East Java poverty data, which contains multicollinearity. The first step is analyze data using GSCA, however the weights of and indicators are not significant, and the three estimated path coefficients are also not significant at the 95% confidence interval. The high correlation value between the indicators further indicates the presence of multicollinearity. Futhermore, the data are analyzed using RGSCA with ridge parameters namely which provides minimum prediction error (CV). The results of the analysis reveal that all estimation of loading factors, weights and path coefficients are significant at 95% confidence intervals. The interpretation of the path coefficient results suggests that education, health, and economy significantly influence poverty, while health and economy also have a significant effect on education, and health additionally exhibits a significant effect on economy. The overall model evaluation results obtained a FIT value of 0.662, indicating that the model can explain about 66.2% of the data variation.