Prediction: Coveted, Yet Forsaken? Introducing a Cross-Validated Predictive Ability Test in Partial Least Squares Path Modeling

Abstract

Management researchers often develop theories and policies that are forward-looking. The prospective outlook of predictive modeling, where a model predicts unseen or new data, can complement the retrospective nature of causal-explanatory modeling that dominates the field. Partial least squares (PLS) path modeling is an excellent tool for building theories that offer both explanation and prediction. A limitation of PLS, however, is the lack of a statistical test to assess whether a proposed or alternative theoretical model offers significantly better out-of-sample predictive power than a benchmark or an established model. Such an assessment of predictive power is essential for theory development and validation, and for selecting a model on which to base managerial and policy decisions. We introduce the cross-validated predictive ability test (CVPAT) to conduct a pairwise comparison of predictive power of competing models, and substantiate its performance via multiple Monte Carlo studies. We propose a stepwise predictive model comparison procedure to guide researchers, and demonstrate CVPAT's practical utility using the well-known American Customer Satisfaction Index (ACSI) model.