control.inla: Control variables in control.inla

Description Usage Arguments Value See Also

Description

Control variables in control.inla for use in inla

Usage

1
2
inla.set.control.inla.default(...)
control.inla(adapt.hessian.max.trials, adapt.hessian.mode, adapt.hessian.scale, adaptive.max, adjust.weights, cmin, correct, correct.factor, correct.strategy, correct.verbose, cpo.diff, cutoff, diagonal, diff.logdens, dz, fast, force.diagonal, global.node.degree, global.node.factor, h, huge, int.design, int.strategy, interpolator, lincomb.derived.correlation.matrix, lincomb.derived.only, linear.correction, mode.known, npoints, numint.abserr, numint.maxfeval, numint.relerr, optimiser, print.joint.hyper, reordering, restart, skip.configurations, stencil, step.factor, step.len, strategy, stupid.search, stupid.search.factor, stupid.search.max.iter, tolerance, tolerance.f, tolerance.g, tolerance.step, tolerance.x, verbose)

Arguments

...

Possible arguments

strategy

Character The strategy to use for the approximations; one of 'gaussian', 'simplified.laplace' (default), 'laplace' or 'adaptive'

int.strategy

Character The integration strategy to use; one of 'auto' (default), 'ccd', 'grid', 'eb' (empirical bayes), 'user' or 'user.std'

int.design

Matrix Matrix of user-defined integration points and weights. Each row consists theta values and the integration weight. (EXPERIMENTAL!)

interpolator

Character The interpolator used to compute the marginals for the hyperparameters. One of 'auto', 'nearest', 'quadratic', 'weighted.distance', 'ccd', 'ccdintegrate', 'gridsum', 'gaussian'. Default is 'auto'.

fast

Logical If TRUE, then replace conditional modes in the Laplace approximation with conditional expectation (default TRUE)

linear.correction

Logical Default TRUE for the 'strategy = laplace' option.

h

Numerical The step-length for the gradient calculations for the hyperparameters. Default 0.01.

dz

Numerical The step-length in the standarised scale for the integration of the hyperparameters. Default 0.75.

diff.logdens

Numerical The difference of the log.density for the hyperpameters to stop numerical integration using int.strategy='grid'. Default 6.

print.joint.hyper

Logical If TRUE, the store also the joint distribution of the hyperparameters (without any costs). Default TRUE.

force.diagonal

Logical If TRUE, then force the Hessian to be diagonal. (Default FALSE)

skip.configurations

Logical Skip configurations if the values at the main axis are to small. (Default TRUE)

mode.known

Logical If TRUE then no optimisation is done. (Default FALSE.)

adjust.weights

Logical If TRUE then just more accurate integration weights. (Default TRUE.)

tolerance

Numerical The tolerance for the optimisation of the hyperparameters. If set, this is the default value for for 'tolerance.f^(2/3)', 'tolerance.g' and 'tolerance.x'; see below.

tolerance.f

Numerical The tolerance for the absolute change in the log posterior in the optimisation of the hyperparameters.

tolerance.g

Numerical The tolerance for the absolute change in the gradient of the log posterior in the optimisation of the hyperparameters.

tolerance.x

Numerical The tolerance for the change in the hyperparameters (root-mean-square) in the optimisation of the hyperparameters.

tolerance.step

Numerical The tolerance for the change in root-mean_squre in the inner Newton-like optimisation of the latent field.

restart

Numerical To improve the optimisation, the optimiser is restarted at the found optimum 'restart' number of times.

optimiser

Character The optimiser to use; one of 'gsl' or 'default'.

verbose

Logical Run in verbose mode? (Default FALSE)

reordering

Character Type of reordering to use. (EXPERT OPTION; one of "AUTO", "DEFAULT", "IDENTITY", "REVERSEIDENTITY", "BAND", "METIS", "GENMMD", "AMD", "MD", "MMD", "AMDBAR", "AMDC", "AMDBARC", or the output from inla.qreordering. Default is 'auto'.)

cpo.diff

Numerical Threshold to define when the cpo-calculations are inaccurate. (EXPERT OPTION.)

npoints

Numerical Number of points to use in the 'stratey=laplace' approximation (default 9)

cutoff

Numerical The cutoff used in the 'stratey=laplace' approximation. (Smaller value is more accurate and more slow.) (default 1e-4)

adapt.hessian.mode

Logical Should optimisation be continued if the Hessian estimate is void? (Default TRUE)

adapt.hessian.max.trials

Numerical Number of steps in the adaptive Hessian optimisation

adapt.hessian.scale

Numerical The scaling of the 'h' after each trial.

adaptive.max

Selecting strategy="adaptive" will chose the default strategy for all fixed effects and model components with length less or equal to adaptive.max, for others, the gaussian strategy will be applied.

huge

Logical If TRUE then try to do some of the internal parallisations differently. Hopefully this will be of benefite for 'HUGE' models. (Default FALSE.) [THIS OPTION IS OBSOLETE AND NOT USED!]

step.len

Numerical The step-length used to compute numerical derivaties of the log-likelihood

stencil

Numerical Number of points in the stencil used to compute the numerical derivaties of the log-likelihood (3, 5, 7 or 9). (default 5)

lincomb.derived.only

Logical If TRUE the only compute the marginals for the derived linear combinations and if FALSE, the and also the linear combinations to the graph (Default TRUE)

lincomb.derived.correlation.matrix

Logical If TRUE compute also the correlations for the derived linear combinations, if FALSE do not (Default FALSE)

diagonal

Numerical Expert use only! Add a this value on the diagonal of the joint precision matrix. (default 0.0)

numint.maxfeval

Numerical Maximum number of function evaluations in the the numerical integration for the hyperparameters. (Default 100000.)

numint.relerr

Numerical Relative error requirement in the the numerical integration for the hyperparameters. (Default 1e-5)

numint.abserr

Numerical Absolute error requirement in the the numerical integration for the hyperparameters. (Default 1e-6)

cmin

Numerical The minimum value for the negative Hessian from the likelihood. Increasing this value will stabalise the optimisation but can introduce bias in some estimates unless -Inf is used. (Default -Inf)

step.factor

Numerical The step factor in the Newton-Raphson algorithm saying how large step to take (Default 1.0)

global.node.factor

Numerical The factor which defines the degree required (how many neighbors), as a fraction of n-1, that is required to be classified as a global node and numbered last (whatever the reordering routine says). Here, n, is the size of the graph. (Disabled if larger than 1.) (default 2.0)

global.node.degree

Numerical The degree required (number of neighbors) to be classified as a global node and numbered last (whatever the reordering routine says). (default .Machine$integer.max)

stupid.search

Logical Enable or disable the stupid-search-algorithm, if the Hessian calculations reveals that the mode is not found. (Default TRUE.)

stupid.search.max.iter

Numerical Maximum number of iterations allowed for the stupid-search-algorithm. (default 1000)

stupid.search.factor

Numerical Factor (>=1) to increase the step-length with after each new interation. (default 1.05)

correct

Logical Add correction for the Laplace approximation. (default FALSE)

correct.factor

Numerical Factor used in adjusting the correction factor (default=10) if correct=TRUE

correct.strategy

Character The strategy used to compute the correction; one of 'simplified.laplace' (default) or 'laplace'

correct.verbose

Logical Be verbose when computing the correction? (default FALSE)

Value

The function control.inla is used to TAB-complete arguments and returns a list of given arguments. The function inla.set.control.inla.default returns a list with all the default values of all parameters within this control statement.

See Also

control.update, control.lincomb, control.group, control.mix, control.link, control.expert, control.compute, control.gev2.default, control.family, control.fixed, control.inla, control.predictor, control.results, control.mode, control.hazard, inla


inbo/INLA documentation built on Dec. 6, 2019, 9:51 a.m.