A1255
Title: Group shrinkage for spatial autoregressive models with convex combinations of spatial weights matrices
Authors: Xiaoyi Han - Xiamen University (China) [presenting]
Abstract: Spatial autoregressive models with convex combinations of spatial weights matrices have been employed to capture the relative importance of spillovers from different connectivity matrices in a variety of economic applications. We investigate the scenario where different spatial weights matrices exhibit group structure, with each group representing a spatial spillover channel consisting of multiple connectivity matrices. We propose a new group shrinkage prior, the Group Inverse-Gamma Gamma with Bayesian Group Lasso prior, to identify the relevant versus irrelevant groups, pinpoint the influential connectivity matrices and mitigate multicollinearity within each group. We develop a computationally tractable Markov Chain Monte Carlo (MCMC) algorithm that enables accurate estimation and inference. We further propose an approximated exchange algorithm to enhance the computational efficiency. Simulation results demonstrate that our shrinkage prior outperforms existing priors in both estimation and variable selection, while the exchange algorithm proves computationally efficient. Empirical analysis examining channels of sovereign debt risk spillover also demonstrates the effectiveness of the proposed method.
