ABSTRACT
A development in the use of approximations to Bayesian models has been proposed in a sequence of papers by Rue and coworkers. The basic idea is that a wide range of models which have a latent Gaussian structure can be approximated via integrated nested Laplace approximation (INLA). As in the case of CARBayes, INLA is based on definition of a formula and model fit statements. Uncorrelated Heterogeneity models can be fitted straightforwardly on INLA. All that is needed is the index vector assigned to each spatial unit. Spatial structure can be modeled in INLA straightforwardly. The package allows the specification of Improper Conditional autoregressive (ICAR) and convolution models. INLA requires the use of a special graph file that includes the adjacencies of the study regions. The ICAR model can be fitted using the “besag” model specification. The convolution or BYM model can be fitted in two different ways in INLA.
