Parallelize 'lme4' functions

+ =

The futurize package allows you to easily turn sequential code into parallel code by piping the sequential code to the futurize() function. Easy!

TL;DR

library(futurize)
plan(multisession)
library(lme4)

gm <- glmer(cbind(incidence, size - incidence) ~ period + (1 | herd),
             data = cbpp, family = binomial)
gm_all <- allFit(gm) |> futurize()

Introduction

This vignette demonstrates how to use this approach to parallelize lme4 functions such as allFit() and bootMer().

The lme4 package fits linear and generalized linear mixed-effects models. Its allFit() function fits models using all available optimizers to check for convergence issues, and bootMer() performs parametric bootstrap inference. Both are excellent candidates for parallelization.

Example: Fitting with multiple optimizers

The allFit() function fits a model with each available optimizer, which can be done in parallel:

library(lme4)

## Fit a generalized linear mixed model
gm <- glmer(cbind(incidence, size - incidence) ~ period + (1 | herd),
            data = cbpp, family = binomial)

## Try all available optimizers
gm_all <- allFit(gm)

Here allFit() evaluates sequentially, but we can easily make it evaluate in parallel by piping to futurize():

library(futurize)
library(lme4)

gm <- glmer(cbind(incidence, size - incidence) ~ period + (1 | herd),
            data = cbpp, family = binomial)
gm_all <- allFit(gm) |> futurize()

This will distribute the optimizer fits across the available parallel workers, given that we have set up parallel workers, e.g.

plan(multisession)

Unlike other parallel backends in R, futurize() relays standard output, messages, and warnings produced by the parallel workers back to your main R session. For instance, when running the above, you will see the progress output from each optimizer as it completes:

> gm_all <- allFit(gm) |> futurize()
bobyqa : [OK]
Nelder_Mead : [OK]
nlminbwrap : [OK]
nmkbw : [OK]
optimx.L-BFGS-B : [OK]
nloptwrap.NLOPT_LN_NELDERMEAD : [OK]
nloptwrap.NLOPT_LN_BOBYQA : [OK]

This output originates from the parallel workers and is relayed to your R session, so you get the same informative feedback as when running sequentially.

The built-in multisession backend parallelizes on your local computer and works on all operating systems. There are other parallel backends to choose from, including alternatives to parallelize locally as well as distributed across remote machines, e.g.

plan(future.mirai::mirai_multisession)

and

plan(future.batchtools::batchtools_slurm)

Example: Parametric bootstrap

The bootMer() function performs parametric bootstrap inference on fitted models:

library(futurize)
plan(multisession)
library(lme4)

## Fit a linear mixed model
fm <- lmer(Reaction ~ Days + (Days | Subject), data = sleepstudy)

## Bootstrap the fixed-effect coefficients
boot_coef <- function(model) fixef(model)
b <- bootMer(fm, boot_coef, nsim = 100) |> futurize()

Supported Functions

The following lme4 functions are supported by futurize():

allFit()
bootMer()
influence() for ‘merMod’
profile() for ‘merMod’

Without futurize: Manual PSOCK cluster setup

For comparison, here is what it takes to parallelize bootMer() using the parallel package directly, without futurize:

library(lme4)
library(parallel)

## Fit a linear mixed model
fm <- lmer(Reaction ~ Days + (Days | Subject), data = sleepstudy)

## Set up a PSOCK cluster
ncpus <- 4L
cl <- makeCluster(ncpus)

## Bootstrap the fixed-effect coefficients
boot_coef <- function(model) fixef(model)
b <- bootMer(fm, boot_coef, nsim = 100,
             parallel = "snow", ncpus = ncpus, cl = cl)

## Tear down the cluster
stopCluster(cl)