Abstract
Although previous studies have offered procedures to perform predictive analytics on PLS path models, there are no guidelines for practitioners on what good or bad predictive characteristics are, what kinds of PLS models might benefit from predictive analysis, and how predictive characteristics of PLS models should be displayed and reported in publications. We seek to rectify this gap in academic practice by systematically analyzing the predictive performance of simulated PLS models with varying measurement and structural properties, and by visually examining their predictive performance. We used the SimSem package to simulate data for these models. This package was created for the R Statistical Environment in order to perform MonteCarlo simulations of Structural Equation Models. We generated three simulated models: (a) a model with near perfect measurement and structural properties; (b) a model with deteriorated measurement more typical to real world studies; and (c) a model using data following a discrete Likert-type scale typically used in psychometric and management studies. Our findings show that under excellent conditions, PLS models can indeed generate highly precise case-level predictions. However, we also illustrate there are lower limits of measurement quality below which studies simply cannot offer practical case-level claims. We also found that Likert scales can be accommodated by predictive PLS, though we caution that there is still room for debate on what prediction means in these situations. Finally, we show that all the above situations can be found in empirical studies, and offer ways to interpret these results.