galamm_control: Control values for galamm fit

Description

This function can be called for controling the optimization procedure used when fitting GALAMMs using galamm.

Usage

galamm_control(
  optim_control = list(),
  method = c("L-BFGS-B", "Nelder-Mead"),
  maxit_conditional_modes = 10,
  pirls_tol_abs = 0.01,
  reduced_hessian = FALSE
)

Value

Object of class galamm_control, which typically will be provided as an argument to galamm.

Arguments

optim_control: List containing optimization parameters. If method = "L-BFGS-B" it is passed on to stats::optim's control argument and if method = "Nelder-Mead", it is passed on to lme4::Nelder_Mead's control argument. If not otherwise specified, and method = "L-BFGS-B", the following arguments are set to non-default values: fnscale = -1 and lmm = 20.
method: Character string defining the algorithm to be used for maximizing the marginal log-likelihood. The default is "L-BFGS-B", which uses the limited memory Broyden-Fletcher-Goldfarb-Shanno algorithm with box constrained as implemented in stats::optim. The other options is "Nelder-Mead", which calls the Nelder-Mead algorithm with box constraints implemented in lme4::Nelder_Mead. The argument is case sensitive.
maxit_conditional_modes: Maximum number of iterations in penalized iteratively reweighted least squares algorithm. Ignored if family = "gaussian" for all observations, since then a single step gives the exact answer.
pirls_tol_abs: Absolute convergence criterion for penalized iteratively reweighted least squares algorithm. Defaults to 0.01, which means that when the reduction in marginal likelihood between two iterations is below 0.01, the iterations stop.
reduced_hessian: Logical value. Defaults to TRUE, which means that the full Hessian matrix at the maximum marginal likelihood solution is computed. If FALSE, a reduced Hessian matrix with second order partial derivatives with respect to fixed regression coefficients and factor loadings. The latter can help is the full Hessian is not positive definite.

References

batesFittingLinearMixedEffects2015galamm

broydenConvergenceClassDoublerank1970galamm

byrdLimitedMemoryAlgorithm1995galamm

fletcherNewApproachVariable1970galamm

goldfarbFamilyVariablemetricMethods1970galamm

nelderSimplexMethodFunction1965galamm

shannoConditioningQuasiNewtonMethods1970galamm

Examples

Run this code

# Define control object with quite a high degree of verbosity (trace = 6)
# and using the last 20 BFGS updates to estimate the Hessian in L-BFGS-B.
control <- galamm_control(optim_control = list(trace = 6, lmm = 20))