Changes in treatment assignment from 0.12 to 0.16

[lukesonnet]

See the following MWE:

install.packages("randomizr") # currently on version 0.16

library(randomizr)
prs <- rep(1/7, times = 7)
bls <- paste0(rep(letters[1:6], each = 100), rep(1:3, times = c(40, 10, 50)))
set.seed(42)
t1 <- block_ra(blocks = bls, prob_each = prs)

devtools::install_version("randomizr", "0.12.0")

library(randomizr)
set.seed(42)
t2 <- block_ra(blocks = bls, prob_each = prs)
table(t1==t2)

The result is that there are a few cases where the randomization is not equivalent. Indeed I believe this is due to a change in load balancing (Issue #35). This isn't ideal for those of us who use randomizr in old randomizations and rely on rerunning those scripts to ensure the randomization protocol doesn't change as data upstream from the randomization changes. Also, would like to know (for a friend) if the old version was "wrong" because then I've got a bigger problem.

[acoppock]

I believe that the new version is "better" but that the old version was not "wrong". I think you're seeing differences because of https://github.com/DeclareDesign/randomizr/issues/50.

So yes, the randomization distribution under 12 admits slightly more possibilities than under 16.