Identifying Hidden Structures in Marketing's Structural Models Through Universal Structure Modeling An Explorative Bayesian Neural Network Complement to LISREL and PLS

Abstract

Universal Structure Modeling (USM), introduced herein, offers an exploratory complement to confirmatory structural equation modeling methods such as covariance-based structural equation modeling (e.g., LISREL) and partial least squares. Because USM combines the iterative methodology of partial least squares with a Bayesian neural network approach involving a multilayer perceptron architecture, it enables researchers to identify “hidden” structures within their models and highlights theoretically unproposed model paths, nonlinear relations among model variables, and interactive effects. Using synthetic data, the authors demonstrate USM’s ability to identify linear and nonlinear relationships and provide evidence that the method does not overfit the original data. They also find hidden paths, nonlinearity, and interactions in two structural models published in the Journal of Marketing, which illustrates the practical relevance of USM. They provide recommendations for researchers regarding in which conditions and how USM should be used.