Bayesian Lasso Regression Modeling via Model Averaging

Abstract

Lasso estimates regression models by imposing the sum of absolute values (L₁ norms) of regression coefficients as a constraint on the sum of squared errors. The Bayesian lasso modeling is a Bayesian approach for the lasso, which specifying the Laplace (i.e., double-exponential) distribution as a prior for regression parameters. The resulting model depends on values of hyper-parameters included in the prior distribution and is often unreliable in model selection, since model estimation is by numerical computation. We propose a Bayesian lasso regression modeling by combining resampling methods and model averaging from a predictive perspective. The efficiency of our modeling is investigated through Monte Carlo simulations and real data analysis.