Benchmark priors for Bayesian model averaging

Abstract

In contrast to a posterior analysis given a particular sampling model, posterior model probabilities in the context of model uncertainty are typically rather sensitive to the specification of the prior. In particular, "diffuse" priors on model-specific parameters can lead to quite unexpected consequences. Here we focus on the practically relevant situation where we need to entertain a (large) number of sampling models and we have (or wish to use) little or no subjective prior information. We aim at providing an "automatic" or "benchmark" prior structure that can be used in such cases.We focus on the Normal linear regression model with uncertainty in the choice of regressors. We propose a partly noninformative prior structure related to a Natural Conjugate g-prior specification, where the amount of subjective information requested from the user is limited to the choice of a single scalar hyperparameter g 0j . The consequences of different choices for g 0j are examined. We investigate theoretical properties, such as consistency of the implied Bayesian procedure. Links with classical information criteria are provided. In addition, we examine the finite sample implications of several choices of g 0j in a simulation study. The use of the MC 3 algorithm of Madigan and York (1995), combined with efficient coding in Fortran, makes it feasible to conduct large simulations. In addition to posterior criteria, we shall also compare the predictive performance of different priors. A classic example concerning the economics of crime will also be provided and contrasted with results in the literature. The main findings of the paper will lead us to propose a "benchmark" prior specification in a linear regression context with model uncertainty.

Spatial Growth Regressions: Model Specification, Estimation and Interpretation

“…The importance of the initial level of income and human capital is consistent with other studies of economic growth in non-spatial settings where these variables also appeared as the most important (e.g. Fernández et al, 2001b).…”

Robust linear static panel data models usingε-contamination

“…Likewise, rather than specifying a Wishart distribution for the variance-covariance matrices as is customary, Zellner's g-prior (Λ β = (τ gX X) −1 for β or Λ b = (τ hW W ) −1 for b) has been widely adopted because of its computational efficiency in evaluating marginal likelihoods and because of its simple interpretation as arising from the design matrix of observables in the sample. Since the calculation of marginal likelihoods using a mixture of g-priors involves only a one dimensional integral, this approach provides an attractive computational solution that made the original g-priors popular while insuring robustness to misspecification of g (see Zellner (1986) and Fernandez, Ley and Steel (2001) to mention a few). To guard against mispecifying the distributions of the priors, many suggest considering classes of priors (see Berger (1985)).…”

On describing multivariate skewed distributions: A directional approach

“…The interpretation of LPS can be facilitated by considering that in the case of i.i.d. sampling, LPS approximates the sum of the Kullback-Leibler divergence between the actual sampling density and the predictive density and the entropy of the sampling distribution (Fernández, Ley and Steel 2001). So LPS captures uncertainty due to a lack of fit plus the inherent sampling uncertainty.…”