On the Choice of Prior in Bayesian Model Averaging

Bayesian model averaging attempts to combine parameter estimation and model uncertainty in one coherent framework. The choice of prior is then critical. Within an explicit framework of ignorance we define a ‘suitable’ prior as one which leads to a continuous and suitable analog to the pretest estimator. The normal prior, used in standard Bayesian model averaging, is shown to be unsuitable. The Laplace (or lasso) prior is almost suitable. A suitable prior (the Subbotin prior) is proposed and its properties are investigated