Accounting for Model Uncertainty in the Prediction of University Graduation Rates

Empirical analysis requires researchers to choose which variables to use as controls in their models. Theory should dictate this choice, yet often in social science there are several theories that may suggest the inclusion or exclusion of certain variables as controls. The result of this is that researchers may use different variables in their models and come to disparate conclusions with respect to predicted effects and their statistical significance. In such cases one is uncertain of which particular set of regressors forms the model that represents the data. The approach used below accounts for uncertainty in variable selection by using Bayesian model averaging (BMA). Accounting for uncertainty, we demonstrate that BMA provides better out-of-sample prediction for university graduation rates than results based on alternative variable selection methods.