allFit: Refit a fitted model with all available optimizers

Description

Attempt to re-fit a [g]lmer model with a range of optimizers. The default is to use all known optimizers for R that satisfy the requirements (i.e. they do not require functions and allow box constraints: see ‘optimizer’ in lmerControl). These optimizers fall in four categories; (i) built-in (minqa::bobyqa, lme4::Nelder_Mead, nlminbwrap), (ii) wrapped via optimx (most of optimx's optimizers that allow box constraints require an explicit gradient function to be specified; the two provided here are the base R functions that can be accessed via optimx), (iii) wrapped via nloptr (see examples for the list of options), (iv) ‘dfoptim::nmkb’ (via the (unexported) nmkbw wrapper: this appears as ‘nmkbw’ in meth.tab)

Usage

allFit(object, meth.tab = NULL, data=NULL,
       verbose = TRUE,
       show.meth.tab = FALSE,
       maxfun = 1e5,
       parallel = c("no", "multicore", "snow"),
       ncpus = getOption("allFit.ncpus", 1L), cl = NULL,
       catch.errs = TRUE)

Value

an object of type allFit, which is a list of fitted merMod objects (unless show.meth.tab is specified, in which case a data frame of methods is returned). The

summary method for this class extracts tables with a variety of useful information about the different fits (see examples).

Arguments

object

a fitted model

meth.tab

a matrix (or data.frame) with columns

method: the name of a specific optimization method to pass to the optimizer (leave blank for built-in optimizers)

optimizer

the optimizer function to use

data

data to be included with result (for later debugging etc.)

verbose

logical: report progress in detail?

show.meth.tab

logical: return table of methods?

maxfun

passed as part of optCtrl to set the maximum number of function evaluations: this is automatically converted to the correct specification (e.g. maxfun, maxfeval, maxit, etc.) for each optimizer

parallel

The type of parallel operation to be used (if any). If missing, the default is taken from the option "boot.parallel" (and if that is not set, "no").

ncpus

integer: number of processes to be used in parallel operation: typically one would choose this to be the number of available CPUs. Use options(allFit.ncpus=X) to set the default value to X for the duration of an R session.

An optional parallel or snow cluster for use if parallel = "snow". If not supplied, a cluster on the local machine is created for the duration of the boot call.

catch.errs

(logical) Wrap model fits in tryCatch clause to skip over errors? (catch.errs=FALSE is probably only useful for debugging)

Details

Needs packages optimx, and dfoptim to use all optimizers
If you are using parallel="snow" (e.g. when running in parallel on Windows), you will need to set up a cluster yourself and run clusterEvalQ(cl,library("lme4")) before calling allFit to make sure that the lme4 package is loaded on all of the workers
Control arguments in control$optCtrl that are unused by a particular optimizer will be silently ignored (in particular, the maxfun specification is only respected by bobyqa, Nelder_Mead, and nmkbw)
Because allFit works by calling update, it may be fragile if the original model call contains references to variables, especially if they were originally defined in other environments or no longer exist when allFit is called.

Examples

Run this code

if (interactive()) {
library(lme4)
  gm1 <- glmer(cbind(incidence, size - incidence) ~ period + (1 | herd),
             data = cbpp, family = binomial)
  ## show available methods
  allFit(show.meth.tab=TRUE) 
  gm_all <- allFit(gm1)
  ss <- summary(gm_all)
  ss$which.OK            ## logical vector: which optimizers worked?
  ## the other components only contain values for the optimizers that worked
  ss$llik                ## vector of log-likelihoods
  ss$fixef               ## table of fixed effects
  ss$sdcor               ## table of random effect SDs and correlations
  ss$theta               ## table of random effects parameters, Cholesky scale
} 
if (FALSE) {
  ## Parallel examples for Windows
  nc <- detectCores()-1
  optCls <- makeCluster(nc, type = "SOCK")
  clusterEvalQ(optCls,library("lme4"))
  ### not necessary here because using a built-in
  ## data set, but in general you should clusterExport() your data
  clusterExport(optCls, "cbpp")
  system.time(af1 <- allFit(m0, parallel = 'snow', 
                          ncpus = nc, cl=optCls))
  stopCluster(optCls)
}

Run the code above in your browser using DataLab

Last chance! 50% off unlimited learning