|
Abstract:
|
The paper introduces a general class of global-local shrinkage priors for variable selection. Included in this general class is the now famous ``horseshoe prior'' and many of its generalizations. These priors are classified into two classes: those with exponential tails and those with polynomial tails. For orthogonal designs, all these priors select asymptotically the correct non-zero regression variables. However, while estimators of the non-zero regressors based on polynomial tailed priors attain asymptotic normality at the correct rate, the same is not true for exponential tailed priors.
|
