High Dimensional Linear Regression via the R2-D2 Shrinkage Prior

We propose a new class of priors for linear regression, the R-square induced Dirichlet Decomposition (R2-D2) prior. The prior is induced by a Beta prior on the coefficient of determination, and then the total prior variance of the regression coefficients is decomposed through a Dirichlet prior. We demonstrate both theoretically and empirically the advantages of the R2-D2 prior over a number of common shrink- age priors, including the Horseshoe, Horseshoe+, and Dirichlet-Laplace priors. The R2-D2 prior possesses the fastest concentration rate around zero and heaviest tails among these common shrinkage priors, which is established based on its marginal density, a Meijer G-function. We show that its Bayes estimator converges to the truth at a Kullback-Leibler super-efficient rate, attaining a sharper information theoretic bound than existing common shrinkage priors. We also demonstrate that the R2-D2 prior yields a consistent posterior. The R2-D2 prior permits straightforward Gibbs sampling and thus enjoys computational tractability. The proposed prior is further investigated in a mouse gene expression application.