Spatio-Temporal Expectation-Maximization Models
Estimation of hierarchical spatio-temporal models through the
Expectation-Maximization (EM) algorithm within a likelihood-based
framework, following Fasso and Cameletti (2010)
and the heterogeneous-network model of
Fasso, Cameletti and Nicolis (2007) , with the EM
machinery of McLachlan and
Krishnan (2008, ISBN:9780470191613). Parameters are estimated by maximum
likelihood via Kalman filtering and smoothing; standard errors and spatial
prediction uncertainty are obtained by a spatio-temporal parametric
bootstrap. The package also provides the spatially-clustered STEM (SC-STEM)
models, an extension in which regression coefficients, variance components
and latent dynamics vary across spatial regimes estimated from the data by
maximizing a Potts-penalized likelihood, in the spirit of the
spatially-clustered regression of Sugasawa and Murakami (2021)
and of the spatially-clustered spatial
autoregressive models of Cerqueti, Maranzano and Mattera (2025)
. SC-STEM comes with likelihood-based
information criteria, a two-step rule based on the Bayesian information
criterion for selecting the number of regimes and the spatial penalty,
and a refit-with-clustering parametric bootstrap that propagates the
uncertainty of the estimated partition into intervals and tests.