Theory to Prediction: Hybrid Models for Binary Endogenous Variables in PLS-SEM Using Supervised Machine Learning

Abstract

Theoretical explanation and out-of-sample prediction are both important in applied structural modeling. Partial least squares structural equation modeling (PLS-SEM) estimates relationships among latent-variable scores, but its ordinary least squares structural regressions are not directly suited to a binary endogenous outcome. We evaluate a hybrid workflow in which PLS-SEM latent scores are supplied to logistic regression (LR), linear and radial-kernel support vector machines (SVMs), and random forests (RFs). Eight open-source datasets were analyzed after the focal continuous endogenous score was dichotomized at the mean or median. Models were compared using 5-, 7-, and 10-fold cross-validation and 90:10, 80:20, and 70:30 train-test splits. Performance was assessed with sensitivity, specificity, accuracy, area under the receiver operating characteristic curve, Cohen's kappa, and the Matthews correlation coefficient. RF most frequently produced the highest cross-validation values, whereas held-out test performance varied across datasets and split ratios. Radial-kernel SVM was competitive in several datasets. The results show that supervised classifiers can extend a PLS-SEM workflow for constructed binary outcomes, but no classifier was uniformly superior. Model selection should therefore rely on prespecified tuning and rigorous out-of-sample validation.