Identifying Latent Interaction Effects in Situations With Many Predictor Variables: Challenges and Solutions

• In Event: 44.078 - Advanced Issues in Structural Equation Modeling

Abstract

Methodological research on latent interaction models within the structural equation modeling framework has mainly focused on small models, that is models with only one latent dependent variable and two latent predictor variables. However, in applied research, for example in motivation theories typically many (correlated) variables are used to predict an outcome. In this presentation, we introduce a Bayesian lasso estimator to identify latent linear and nonlinear effects in situations with many predictor variables and high multicollinearity. We show in two simulation studies that the lasso has an increased power in situations with highly correlated predictor variables compared to standard approaches, and that the analysis of separate submodels leads to a highly inflated type I error rate.

Authors

• Holger Brandt, University of Kansas

• Jenna Cambria, University of Arkansas

• Augustin Kelava, Eberhard Karls Universität Tübingen