PLS-Based Model Selection: The Role of Alternative Explanations in Information Systems Research

Exploring theoretically plausible alternative models for explaining the phenomenon under study is a crucial step in advancing scientific knowledge. This paper advocates model selection in information systems (IS) studies that use partial least squares path modeling (PLS) and suggests the use of model selection criteria derived from information theory for this purpose. These criteria allow researchers to compare alternative models and select a parsimonious yet well-fitting model. However, as our review of prior IS research practice shows, their use—while common in the econometrics field and in factor-based SEM—has not found its way into studies using PLS. Using a Monte Carlo study, we compare the performance of several model selection criteria in selecting the best model from a set of competing models under different model set-ups and various conditions of sample size, effect size, and loading patterns. Our results suggest that appropriate model selection cannot be achieved by relying on the PLS criteria (i.e., R2, Adjusted R2, GoF, and Q2), as is the current practice in academic research. Instead, model selection criteria—in particular, the Bayesian information criterion (BIC) and the Geweke-Meese criterion (GM)—should be used due to their high model selection accuracy and ease of use. To support researchers in the adoption of these criteria, we introduce a five-step procedure that delineates the roles of model selection and statistical inference and discuss misconceptions that may arise in their use.