Abstract
Chapter 5 presents thresholding methods in the context of dynamic multiscale spatio-temporal models (DMSTMs) for Gaussian univariate data. DMSTMs naturally accommodate data observed on irregular grids and heteroscedasticity. In addition, DMSTMs lead to computations that are parallelizable, scalable and fast. These models assume a spatial hierarchical structure. Based on such structure, DMSTMs decompose spatial areal data with a multiscale decomposition. This decomposition yields empirical multiscale coefficients that are assumed to be noisy versions of latent multiscale coefficients. Typically, the DMSTM specification is completed by assuming the latent multiscale coefficients evolve through time following state-space equations. Conversely, in this chapter we also consider thresholding methods that consider possible cases when some latent multiscale coefficients may be fixed through time. In that case, some of the latent multiscale coefficients may be thresholded to zero vectors, allowing clustering of subregions. Another possible case estimates some of the latent multiscale coefficients to be different than zero but fixed through time. Finally, we illustrate the application of our DMSTM thresholding methods with an analysis of agriculture production data.