Add 4 vignettes introducing fixed/group sequential design simulations

[LittleBeannie]

This issue adds 4 vignettes introducing fixed/group sequential design simulations.

• Simulate Fixed Designs with Ease via sim_fixed_n
• Simulate Group Sequential Designs with Ease via sim_gs_n
• Custom Fixed Design Simulations: A Tutorial on Writing Code from the Ground Up
• Custom Group Sequential Design Simulations: Crafting from Scratch

[LittleBeannie]

Hi @jdblischak, I have a question regarding parallel computing. My current implementation is as follows, and I’m wondering if both the library(doFuture) and library(doRNG) are both necessary.

tic()
registerDoFuture()
registerDoRNG()
plan("multisession", workers = 4)
set.seed(2025)

ans <- foreach(
  sim_id = seq_len(n_sim),
  .combine = "rbind",
  .errorhandling = "stop"
  ) %dorng% {
    ans_new <- one_sim(...) # for one single simulation
    ans_new
  }

plan("sequential")
toc()

[LittleBeannie]

Reference 1: https://github.com/Merck/simtrial/blob/main/R/sim_fixed_n.R#L253

Also reduce the number of workers to 2, per CRAN requirements.

[jdblischak]

I have a question regarding parallel computing. My current implementation is as follows, and I’m wondering if both the library(doFuture) and library(doRNG) are both necessary.

Following up on our offline discussion:

Your approach using {doFuture} and {doRNG} is valid. It produces reproducible results:

library("foreach")
library("doFuture")
library("doRNG")

registerDoFuture()
registerDoRNG()
plan("multisession", workers = 4)

set.seed(2025)
ans1 <- foreach(
  i = 1:10,
  .combine = "c",
  .errorhandling = "stop"
) %dorng% {
  rnorm(1)
}
head(ans1)
## [1] -0.46387557 -0.05150034 -0.22384683 -0.60535060  1.52634253  0.31122695

set.seed(2025)
ans2 <- foreach(
  i = 1:10,
  .combine = "c",
  .errorhandling = "stop"
) %dorng% {
  rnorm(1)
}
head(ans2)
## [1] -0.46387557 -0.05150034 -0.22384683 -0.60535060  1.52634253  0.31122695

stopifnot(identical(ans1, ans2))

However, in {simtrial}, we've standardized on using {future} itself to handle the parallel random number generation. Using {doRNG} is an alternative to {doParallel}, but isn't required when using {doFuture} as the backend.

library("foreach")
library("doFuture")

registerDoFuture()
plan("multisession", workers = 4)

set.seed(2025)
ans1 <- foreach(
  i = 1:10,
  .combine = "c",
  .errorhandling = "stop",
  .options.future = list(seed = TRUE)
) %dofuture% {
  rnorm(1)
}
head(ans1)
## [1] -1.0860108  1.3469334  0.1314693 -1.5942347  1.3797186  0.2381956

set.seed(2025)
ans2 <- foreach(
  i = 1:10,
  .combine = "c",
  .errorhandling = "stop",
  .options.future = list(seed = TRUE)
) %dofuture% {
  rnorm(1)
}
head(ans2)
## [1] -1.0860108  1.3469334  0.1314693 -1.5942347  1.3797186  0.2381956

stopifnot(identical(ans1, ans2))

[jdblischak]

If this code takes a while to run, you can consider only publishing it to the pkgdown site but not including it in the distributed package. See the docs for info at https://r-pkgs.org/vignettes.html#sec-vignettes-article

[jdblischak]

How informative are the timings in the vignettes? Is it worth adding another dependency to be installed during each CI run?

If the simulations are small, and they are running on minimal hardware like a CRAN or GitHub machine, it's not clear to me if this will be that informative to an end user that will want to run many simulations on more robust hardware.