Mixtures of g-priors for Bayesian Model Averaging with Economic Application

Ley, Eduardo ; Steel, Mark F. J.

关键词: ALGORITHMS;    ARCHIVE;    AREA;    ASPECT;    BAYES FACTOR;   

DOI  :  10.1596/1813-9450-5732
RP-ID  :  WPS5732

学科分类：社会科学、人文和艺术（综合）

来源: World Bank Open Knowledge Repository

【 摘 要 】

This paper examines the issue ofvariable selection in linear regression modeling, wherethere is a potentially large amount of possible covariatesand economic theory offers insufficient guidance on how toselect the appropriate subset. In this context, BayesianModel Averaging presents a formal Bayesian solution todealing with model uncertainty. The main interest here isthe effect of the prior on the results, such as posteriorinclusion probabilities of regressors and predictiveperformance. The authors combine a Binomial-Beta prior onmodel size with a g-prior on the coefficients of each model.In addition, they assign a hyperprior to g, as the choice ofg has been found to have a large impact on the results. Forthe prior on g, they examine the Zellner-Siow prior and aclass of Beta shrinkage priors, which covers most choices inthe recent literature. The authors propose a benchmark Betaprior, inspired by earlier findings with fixed g, and showit leads to consistent model selection. Inference isconducted through a Markov chain Monte Carlo sampler overmodel space and g. The authors examine the performance ofthe various priors in the context of simulated and realdata. For the latter, they consider two importantapplications in economics, namely cross-country growthregression and returns to schooling. Recommendations forapplied users are provided.