Avoid changing global .Random.seed

[weiyaw]

Hi,

Would you consider using withr::local_seed in mice() rather than set.seed?

Currently set.seed sets .Random.seed in the users' global environment which can result in unintended results and bugs.

Regards,

Kenyon

[stefvanbuuren]

Thanks for your suggestion of a neater alternative.

I created a branch randomseed that isolates the seed streams used in the mice() and mice.mids() functions. The random number state is reset to the initial state on completion of mice(...) or mice.mids(...), so both functions do not interfere anymore with the calling code, e.g.,

context("mice: local random stream")

set.seed(1)
a <- runif(10)

set.seed(1)
b1 <- runif(6)

imp <- mice(nhanes, maxit = 0, seed = 82, print = FALSE)
imp2 <- mice.mids(imp, maxit = 1, print = FALSE)

b2 <- runif(4)

test_that("mice and mice.mids use a local random stream", {
  expect_equal(a, c(b1, b2))
})

As long as the branch exists, install by

devtools::install_github(repo = "amices/mice@randomseed")

I did not look at the parallel functionality in the package.

Does this work for you?

[stefvanbuuren]

PR #432 implements the local seed in the master version.