Allow the user to set restrictions in the buildScoreCache
for both the Bayesian and the MLE approach.
Control function similar to fit.control
.
Usage
build.control(
method = "bayes",
max.mode.error = 10,
mean = 0,
prec = 0.001,
loggam.shape = 1,
loggam.inv.scale = 5e-05,
max.iters = 100,
epsabs = 1e-07,
error.verbose = FALSE,
trace = 0L,
epsabs.inner = 1e-06,
max.iters.inner = 100,
finite.step.size = 1e-07,
hessian.params = c(1e-04, 0.01),
max.iters.hessian = 10,
max.hessian.error = 0.5,
factor.brent = 100,
maxiters.hessian.brent = 100,
num.intervals.brent = 100,
n.grid = 250,
ncores = 1,
cluster.type = "FORK",
max.irls = 100,
tol = 1e-08,
tolPwrss = 1e-07,
check.rankX = "message+drop.cols",
check.scaleX = "message+rescale",
check.conv.grad = "message",
check.conv.singular = "message",
check.conv.hess = "message",
xtol_abs = 1e-06,
ftol_abs = 1e-06,
trace.mblogit = FALSE,
catcov.mblogit = "free",
epsilon = 1e-06,
seed = 9062019L
)
Arguments
- method
a character that takes one of two values: "bayes" or "mle". Overrides
method
argument frombuildScoreCache
.- max.mode.error
if the estimated modes from INLA differ by a factor of
max.mode.error
or more from those computed internally, then results from INLA are replaced by those computed internally. To force INLA always to be used, thenmax.mode.error=100
, to force INLA never to be usedmax.mod.error=0
. See alsofitAbn
.- mean
the prior mean for all the Gaussian additive terms for each node. INLA argument
control.fixed=list(mean.intercept=...)
andcontrol.fixed=list(mean=...)
.- prec
the prior precision (\(\tau = \frac{1}{\sigma^2}\)) for all the Gaussian additive term for each node. INLA argument
control.fixed=list(prec.intercept=...)
andcontrol.fixed=list(prec=...)
.- loggam.shape
the shape parameter in the Gamma distribution prior for the precision in a Gaussian node. INLA argument
control.family=list(hyper = list(prec = list(prior="loggamma",param=c(loggam.shape, loggam.inv.scale))))
.- loggam.inv.scale
the inverse scale parameter in the Gamma distribution prior for the precision in a Gaussian node. INLA argument
control.family=list(hyper = list(prec = list(prior="loggamma",param=c(loggam.shape, loggam.inv.scale))))
.- max.iters
total number of iterations allowed when estimating the modes in Laplace approximation. passed to
.Call("fit_single_node", ...)
.- epsabs
absolute error when estimating the modes in Laplace approximation for models with no random effects. Passed to
.Call("fit_single_node", ...)
.- error.verbose
logical, additional output in the case of errors occurring in the optimization. Passed to
.Call("fit_single_node", ...)
.- trace
Non-negative integer. If positive, tracing information on the progress of the "L-BFGS-B" optimization is produced. Higher values may produce more tracing information. (There are six levels of tracing. To understand exactly what these do see the source code.). Passed to
.Call("fit_single_node", ...)
.- epsabs.inner
absolute error in the maximization step in the (nested) Laplace approximation for each random effect term. Passed to
.Call("fit_single_node", ...)
.- max.iters.inner
total number of iterations in the maximization step in the nested Laplace approximation. Passed to
.Call("fit_single_node", ...)
.- finite.step.size
suggested step length used in finite difference estimation of the derivatives for the (outer) Laplace approximation when estimating modes. Passed to
.Call("fit_single_node", ...)
.- hessian.params
a numeric vector giving parameters for the adaptive algorithm, which determines the optimal stepsize in the finite-difference estimation of the hessian. First entry is the initial guess, second entry absolute error. Passed to
.Call("fit_single_node", ...)
.- max.iters.hessian
integer, maximum number of iterations to use when determining an optimal finite difference approximation (Nelder-Mead). Passed to
.Call("fit_single_node", ...)
.- max.hessian.error
if the estimated log marginal likelihood when using an adaptive 5pt finite-difference rule for the Hessian differs by more than
max.hessian.error
from when using an adaptive 3pt rule then continue to minimize the local error by switching to the Brent-Dekker root bracketing method. Passed to.Call("fit_single_node", ...)
.- factor.brent
if using Brent-Dekker root bracketing method then define the outer most interval end points as the best estimate of \(h\) (stepsize) from the Nelder-Mead as \(h/factor.brent,h*factor.brent)\). Passed to
.Call("fit_single_node", ...)
.- maxiters.hessian.brent
maximum number of iterations allowed in the Brent-Dekker method. Passed to
.Call("fit_single_node", ...)
.- num.intervals.brent
the number of initial different bracket segments to try in the Brent-Dekker method. Passed to
.Call("fit_single_node", ...)
.- n.grid
recompute density on an equally spaced grid with
n.grid
points.- ncores
The number of cores to parallelize to, see ‘Details’. If >0, the number of CPU cores to be used. -1 for all available -1 core. Only for
method="mle"
.- cluster.type
The type of cluster to be used, see
?parallel::makeCluster
.abn
then defaults to"PSOCK"
on Windows and"FORK"
on Unix-like systems. With "FORK" the child process are started withrscript_args = "--no-environ"
to avoid loading the whole workspace into each child.- max.irls
total number of iterations for estimating network scores using an Iterative Reweighed Least Square algorithm. Is this DEPRECATED?
- tol
real number giving the minimal tolerance expected to terminate the Iterative Reweighed Least Square algorithm to estimate network score. Passed to
irls_binomial_cpp_fast_br
andirls_poisson_cpp_fast
.- tolPwrss
numeric scalar passed to
glmerControl
- the tolerance for declaring convergence in the penalized iteratively weighted residual sum-of-squares step. Similar totol
.- check.rankX
character passed to
lmerControl
andglmerControl
- specifying ifrankMatrix(X)
should be compared withncol(X)
and if columns from the design matrix should possibly be dropped to ensure that it has full rank. Defaults tomessage+drop.cols
.- check.scaleX
character passed to
lmerControl
andglmerControl
- check for problematic scaling of columns of fixed-effect model matrix, e.g. parameters measured on very different scales. Defaults tomessage+rescale
.- check.conv.grad
character passed to
lmerControl
andglmerControl
- checking the gradient of the deviance function for convergence. Defaults tomessage
but can be one of "ignore" - skip the test; "warning" - warn if test fails; "message" - print a message if test fails; "stop" - throw an error if test fails.- check.conv.singular
character passed to
lmerControl
andglmerControl
- checking for a singular fit, i.e. one where some parameters are on the boundary of the feasible space (for example, random effects variances equal to 0 or correlations between random effects equal to +/- 1.0). Defaults tomessage
but can be one of "ignore" - skip the test; "warning" - warn if test fails; "message" - print a message if test fails; "stop" - throw an error if test fails.- check.conv.hess
character passed to
lmerControl
andglmerControl
- checking the Hessian of the deviance function for convergence. Defaults tomessage
but can be one of "ignore" - skip the test; "warning" - warn if test fails; "message" - print a message if test fails; "stop" - throw an error if test fails.- xtol_abs
Defaults to 1e-6 stop on small change of parameter value. Only for
method='mle', group.var=...
. Default convergence tolerance for fitted(g)lmer
models is reduced to the value provided here if default values did not fit. This value here is passed to theoptCtrl
argument of(g)lmer
(see help oflme4::convergence()
).- ftol_abs
Defaults to 1e-6 stop on small change in deviance. Similar to
xtol_abs
.- trace.mblogit
logical indicating if output should be produced for each iteration. Directly passed to
trace
argument inmclogit.control
. Is independent ofverbose
.- catcov.mblogit
Defaults to "free" meaning that there are no restrictions on the covariances of random effects between the logit equations. Set to "diagonal" if random effects pertinent to different categories are uncorrelated or "single" if random effect variances pertinent to all categories are identical.
- epsilon
Defaults to 1e-8. Positive convergence tolerance \(\epsilon\) that is directly passed to the
control
argument ofmclogit::mblogit
asmclogit.control
. Only formethod='mle', group.var=...
.- seed
a non-negative integer which sets the seed in
set.seed(seed)
.
Details
Parallelization over all children is possible via the function foreach
of the package doParallel. ncores=0
or ncores=1
use single threaded foreach
. ncores=-1
uses all available cores but one.
See also
Other buildScoreCache:
buildScoreCache()
Examples
ctrlmle <- abn::build.control(method = "mle",
ncores = 0,
cluster.type = "PSOCK",
max.irls = 100,
tol = 10^-11,
tolPwrss = 1e-7,
check.rankX = "message+drop.cols",
check.scaleX = "message+rescale",
check.conv.grad = "message",
check.conv.singular = "message",
check.conv.hess = "message",
xtol_abs = 1e-6,
ftol_abs = 1e-6,
trace.mblogit = FALSE,
catcov.mblogit = "free",
epsilon = 1e-6,
seed = 9062019L)
ctrlbayes <- abn::build.control(method = "bayes",
max.mode.error = 10,
mean = 0, prec = 0.001,
loggam.shape = 1,
loggam.inv.scale = 5e-05,
max.iters = 100,
epsabs = 1e-07,
error.verbose = FALSE,
epsabs.inner = 1e-06,
max.iters.inner = 100,
finite.step.size = 1e-07,
hessian.params = c(1e-04, 0.01),
max.iters.hessian = 10,
max.hessian.error = 0.5,
factor.brent = 100,
maxiters.hessian.brent = 100,
num.intervals.brent = 100,
tol = 10^-8,
seed = 9062019L)