Advances in Methods, Theories, and Philosophy
Loading...
Paper Number
2126
Paper Type
short
Description
The use of interaction terms in partial least squares structural equation modeling (PLS-SEM) risks overfitting models to small samples and producing poor out-of-sample generalizability. But the added complexity of interactions in PLS-SEM is not captured by in-sample fit metrics, and we propose that interaction terms in PLS-SEM should be assessed by out-of-sample methods and metrics. However, out-of-sample predictive methods like PLSpredict do not yet account for interaction terms. We start by providing a formal procedure for generating out-of-sample predictions from such models. We then empirically demonstrate that interactions produce far higher Type I error than that expected by researchers, and that out-of-sample predictive metrics indeed offer more accurate assessment of the validity of interaction terms for PLS-SEM. We also show that two-stage estimation of interactions is superior to other popular methods of operationalizing interactions in PLS-SEM, when the generalizability of interactions is of concern.
Recommended Citation
Danks, Nicholas and Ray, Soumya, "Predictive Validation of Interaction Terms in PLS-SEM" (2023). ICIS 2023 Proceedings. 4.
https://aisel.aisnet.org/icis2023/adv_theory/adv_theory/4
Predictive Validation of Interaction Terms in PLS-SEM
The use of interaction terms in partial least squares structural equation modeling (PLS-SEM) risks overfitting models to small samples and producing poor out-of-sample generalizability. But the added complexity of interactions in PLS-SEM is not captured by in-sample fit metrics, and we propose that interaction terms in PLS-SEM should be assessed by out-of-sample methods and metrics. However, out-of-sample predictive methods like PLSpredict do not yet account for interaction terms. We start by providing a formal procedure for generating out-of-sample predictions from such models. We then empirically demonstrate that interactions produce far higher Type I error than that expected by researchers, and that out-of-sample predictive metrics indeed offer more accurate assessment of the validity of interaction terms for PLS-SEM. We also show that two-stage estimation of interactions is superior to other popular methods of operationalizing interactions in PLS-SEM, when the generalizability of interactions is of concern.
When commenting on articles, please be friendly, welcoming, respectful and abide by the AIS eLibrary Discussion Thread Code of Conduct posted here.
Comments
20-Theory