BAYESIAN MODEL AVERAGING AND THE OPTIMIZATION OF WORKFORCE PREDICTIONS

Abstract

This thesis explores the viability of Bayesian model averaging (BMA) as an alternative to traditional general linear models (GLMs) and ensemble-based machine-learning methods commonly used to predict workforce outcomes. In contrast to both the practices that focus on selecting a single “best” GLM and set of predictors, and the ensemble-based machine-learning methods that combine many simpler models, the BMA approach explores the space of all models to be considered and assigns probabilistic weights to each. These posterior model probabilities (PMPs) can then be used to generate optimal predictions regarding future data observations via a weighted-average of the model-specific predictions. By averaging over models, BMA routines are well-suited to addressing the model uncertainty that arises when a researcher has numerous potential predictor variables. Rather than condition inferences upon a single model and set of predictors, or upon a collection of poorer models and simpler subsets, the BMA routine can average predictions across all possible combinations of predictor variables. Focusing upon this form of model uncertainty, this thesis demonstrates how BMA might be employed to optimally forecast workforce outcomes in both classification and regression contexts by way of two illustrative case studies related to the prediction of employee turnover intentions.