Description

Apply Integrated Nested Laplace Approximation (INLA, Rue et al. 2009) to models estimable by gam or bam, using the INLA variant described in Wood (2019). Produces marginal posterior densities for each coefficient, selected coefficients or linear transformations of the coefficient vector.

Usage

Arguments

Details

Let b, h and y denote the model coefficients, hyperparameters/smoothing parameters and response data, respectively. In principle, INLA employs Laplace approximations for p(b_i|h,y) and p(h|y) and then obtains the marginal posterior distribution p(b_i|y) by intergrating the approximations to p(b_i|h,y)p(h|y) w.r.t h (marginals for the hyperparameters can also be produced). In practice the Laplace approximation for p(b_i|h,y) is too expensive to compute for each b_i and must itself be approximated. To this end, there are two quantities that have to be computed: the posterior mode b*|b_i and the determinant of the Hessian of the joint log density log p(b,h,y) w.r.t. b at the mode. Rue et al. (2009) originally approximated the posterior conditional mode by the conditional mode implied by a simple Gaussian approximation to the posterior p(b|y). They then approximated the log determinant of the Hessian as a function of b_i using a first order Taylor expansion, which is cheap to compute for the sparse model representaiton that they use, but not when using the dense low rank basis expansions used by gam. They also offer a more expensive alternative approximation based on computing the log determiannt with respect only to those elements of b with sufficiently high correlation with b_i according to the simple Gaussian posterior approximation: efficiency again seems to rest on sparsity. Wood (2018) suggests computing the required posterior modes exactly, and basing the log determinant approximation on a BFGS update of the Hessian at the unconditional model. The latter is efficient with or without sparsity, whereas the former is a ‘for free’ improvement. Both steps are efficient because it is cheap to obtain the Cholesky factor of H[-i,-i] from that of H - see choldrop. This is the approach taken by this routine.

The approx argument allows two further approximations to speed up computations. For approx==1 the exact posterior conditional modes are not used, but instead the conditional modes implied by the simple Gaussian posterior approximation. For approx==2 the same approximation is used for the modes and the Hessian is assumed constant. The latter is quite fast as no log joint density gradient evaluations are required.

Note that for many models the INLA estimates are very close to the usual Gaussian approximation to the posterior, the interactive argument is useful for investigating this issue.

bam models are only supported with the disrete=TRUE option. The discrete=FALSE approach would be too inefficient. AR1 models are not supported (related arguments are simply ignored).

Value

A list with elements beta and density, both of which are matrices. Each row relates to one coefficient (or linear coefficient combination) of interest. Both matrices have nb columns. If int!=0 then a further element reml gives the integration weights used in the CCD integration, with the central point weight given first.

WARNINGS

This routine is still somewhat experimental, so details are liable to change. Also currently not all steps are optimally efficient.

The routine is written for relatively expert users.

ginla is not designed to deal with rank deficient models.

Author(s)

Simon N. Wood simon.wood@r-project.org

References

Rue, H, Martino, S. & Chopin, N. (2009) Approximate Bayesian inference for latent Gaussian models by using integrated nested Laplace approximations (with discussion). Journal of the Royal Statistical Society, Series B. 71: 319-392.

Wood (2019) Simplified Integrated Laplace Approximation. In press Biometrika.

Examples