Spatio-Temporal Modeling with INLA
Andrew B. Lawson · 2021
While much space has been devoted to Spatio-Temporal (ST) modeling with Markov chain Monte Carlo, it is certainly possible to avoid the use of scripted program code, and to employ alternative approximations at least for the simpler ST models. This chapter demonstrates a range of models that can be fitted straightforwardly using integrated nested Laplace approximation (INLA). It is instructive to consider differences in computation time between the different packages, as this could influence users’ choice of computational platform. The SCRCST data example was used throughout. This has 276 units with 46 regions and 6 time periods. For two models with Improper Conditional autoregressive convolution and random walk temporal effect, one with IID interaction, and one without, system.time was employed to assess overall time for execution. The INLA models assumed weakly informative precision prior distributions for the spatial effects and a weakly informative prior distribution for the precision of the temporal effect.