Bayesian spatial panel models: a flexible Kronecker error component approach

We introduce a class of spatial panel data models with correlated error components that can simultaneously handle cross-sectional and temporal correlation. These models are based on Gaussian Markov Random Fields with a Kronecker product of separable error covariance matrices, which allows capturing correlations both in time and space while reducing the number of parameters being estimated. We then propose a unified approach for estimating these models using a novel Bayesian approach, known as integrated nested Laplace approximations. An empirical illustration using U.S. cigarette consumption data is given, and we find that the most general model outperforms its competitors in both in-sample fit and forecast performance.