Comparing Out-of-Sample Predictive Ability of PLS, Covariance, and Regression Models

摘要

Partial Least Squares Path Modelling (PLSPM) is a popular technique for estimating structural equation models in the social sciences, and is frequently presented as an alternative to covariance-based analysis as being especially suited for predictive modeling. While existing research on PLSPM has focused on its use in causal-explanatory modeling, this paper follows two recent papers at ICIS 2012 and 2013 in examining how PLSPM performs when used for predictive purposes. Additionally, as a predictive technique, we compare PLSPM to traditional regression methods that are widely used for predictive modelling in other disciplines. Specifically, we employ out-of-sample k-fold cross-validation to compare PLSPM to covariance-SEM and a range of a-theoretical regression techniques in a simulation study. Our results show that PLSPM offers advantages over covariance-SEM and other prediction methods.