Elevating theoretical insight and predictive accuracy in business research: Combining PLS-SEM and selected machine learning algorithms

We propose a routine for combining partial least squares-structural equation modeling (PLS-SEM) with selected machine learning (ML) algorithms to exploit the two method’s causal-predictive and causal-exploratory capabilities. Triangulating these two methods can improve the predictive accuracy of research models, enhance the understanding of relationships, assist in identifying new relationships and therewith contribute to theorizing. We demonstrate the advantages and challenges of triangulating the two methods on an illustrative example along a four-step-routine: (1) Develop a PLS-SEM on a baseline conceptual model and use its standards to assess measurement model quality and generate latent variables scores. (2) Apply specific ML algorithms on the extracted data to validate relationships and identify new (linear) relationships that may go beyond the initial hypotheses; similarly, assess model advancements in the form of nonlinearities and interaction effects. (3) Evaluate the theoretical plausibility of alternative models. (4) Integrate alternative models in PLS-SEM and compare these with the baseline model using a recently proposed prediction-oriented test procedure in PLS-SEM.