A Skew-Gaussian‎ ‎Spatio-Temporal Process with Non-Stationary Correlation Structure

This paper develops a new class of spatio-temporal process models that can simultaneously capture skewness and non-stationarity. The proposed approach which is based on using the closed skew-normal distribution in the low-rank representation of stochastic processes, has several favorable properties. In particular, it greatly reduces the dimension of the spatio-temporal latent variables and induces flexible correlation structures. Bayesian analysis of the model is implemented through a Gibbs MCMC algorithm which incorporates a version of the Kalman filtering algorithm. All fully conditional posterior distributions have closed forms which show another advantageous property of the proposed model. We demonstrate the efficiency of our model through an extensive simulation study and an application to a real data set comprised of precipitation measurements.