These change notes reflect changes in functionality relative to the
implementation of Laplace/AGHQ approximation in the nimble
package, version 1.3.0.
Move Laplace/AGHQ approximation from nimble to
nimbleQuad package.
Add a nested approximation method that provides approximate
posterior inference using methodology similar to the well-known INLA
approach (implemented in the R-INLA package) and to the related methods
for extended Gaussian latent models (EGLMs) of Stringer et al. (2023),
implemented in the aghq R package.
Greatly improve efficiency of Laplace/AGHQ approximation through major revision of derivative calculations, including removal of “computeMethod” options (PR 52).
Greatly improve efficiency of Laplace/AGHQ approximation when
using the multivariate normal distribution, via a new
dmnormAD version of the distribution customized for
efficient computation with NIMBLE’s AD system.
Greatly improve efficiency of building Laplace/AGHQ when
determining parameter dependencies in setup_OneAGHQuad (PR
48).
Greatly improve efficiency of Laplace/AGHQ summary computation of Hessian.
Automatically make use of analytic gradient and Hessian with respect to the random variables for univariate and multivariate normal distributions rather than AD in Laplace/AGHQ approximation. This also causes the “outer” Laplace/AGHQ optimization to not use the AD-based gradient.
Greatly improve the efficiency of building overall Laplace approximation when using many individual Laplace approximations (PR #59).
Inform user when overall Laplace approximation is decomposed into multiple Laplace approximations.
Improve messaging when Laplace/AGHQ approximation does not converge.
Provide a variety of quadrature grids, extending from standard AGHQ, including grid pruning, sparse grids, and allowing user-provided grids.
Automatically remove random effect nodes with no dependent data from Laplace (Issue #56).
Add ability to find posterior mode (as well as prior-penalized MLE, as used in fisheries) via Laplace approximation.
Fix handling of fnscale when
updateSettings is used.
Fix buildAGHQ for case with no random
effects.
Fix length of latent nodes in Laplace/AGHQ under non-1:1 transformation (PR #41).
Fix seg fault when “model” option passed as
innerOptimStart (Issue #62).
mvQuad.