Abstract
Chapter 9 proposes ensembles of dynamic multiscale spatio-temporal models (DMSTMs) for univariate Gaussian data. Even though DMSTMs allow fast, scalable, and parallelizable computations, the resulting estimates of the latent processes may be spatially blocky. However, in practice many spatio-temporal processes are spatially smooth. To obtain smoother estimates of the latent processes, our ensemble approach considers mixtures of DMSTMs defined on spatially shifted versions of the multiscale partition. However, this is a case that suffers from the M-open multiple optima paradox presented in Chapter 8. Consequently, model posterior probabilities tend to incorrectly concentrate on a few DMSTMs of the ensemble and are not reliable weights for Bayesian model averaging. To deal with this issue, we set the model prior probabilities to be the same for all the DMSTMs in the ensemble, and do not update them. In an analysis of data simulated from a spatio-temporal model with a smooth spatial covariance function, when compared to any of the individual DMSTMs in terms of estimating the underlying latent process, the ensemble estimate has a much smaller mean squared error. Finally, an ensemble analysis of temperature data demonstrates the usefulness and flexibility of our ensemble of DMSTMs approach.