lme4 allFit() giving confusing results when wrapped in a function - r

Question

I am using allFit() in lme4 to automatically scan through possible optimizers, since the default optimizer usually doesn't converge in this situation. My code works fine when I run it line-by-line, but when I run it wrapped in a simple function, it gives different results.
I've looked at the output of the allFit call and it seems that when it's NOT inside the function, it returns a list of lmerModLmerTest objects as desired.
However, inside the function, it returns a list with the values simpleError, error, and condition. Why is it doing this?
I'm using RStudio, R 3.6, lme4 1.1-21, lmerTest 3.1-0.
UPDATE: The problem is that the update() method used by allFit cannot find the 'tt' data frame when re-fitting the models. I have put breakpoints into the code and it seems that the 'test' data exists in the function environment, however, so I don't understand why it can't find it...
UPDATE 2: It appears that if I change the assignment of the test data to <<-, it works. This is dangerous, though, by breaking functional programming, and I think it may fail when I try to parallelize. I am testing further... still open to suggestions!
Here is the code that works, not inside the function:
library(lme4)
multi_arm_var_sim <- function(nsub = 20, nclust = 100, narm = 2, iccs = c(.01, .04), betas = c(0,.3)){
sig_b2 <- -1*iccs / (iccs - 1)
n <- nsub * nclust * narm
y <- rep_len(NA, n)
arm <- as.factor(rep(0:(narm-1), each = nsub*nclust))
clustid <- rep(1:(nclust*narm), each = nsub)
clustRElist <- rnorm(narm*nclust, mean = 0, sd = rep(sqrt(sig_b2), each = nclust))
clustRE <- rep(clustRElist, each = nsub)
sig_b2 <- rep(sig_b2, each = nclust*nsub)
error <- rnorm(n, mean = 0, sd = 1)
beta <- rep(betas, each = nclust*nsub)
linpred <- beta + clustRE + error
output <- cbind.data.frame(arm, clustid, sig_b2, clustRE, linpred)
return(output)
}
set.seed(2)
test_1 <- multi_arm_var_sim()
model_flex_1 <- lmer(linpred ~ arm + (1 + arm | clustid),
data = test_1)
diff_optims_1 <- allFit(model_flex_1, verbose = TRUE)
print(class(diff_optims_1[[1]]))
is.OK_1 <- sapply(diff_optims_1, is, "lmerMod")
print(is.OK_1)
And here is the code that doesn't work, same setup, wrapped in a function.
library(lme4)
multi_arm_var_sim <- function(nsub = 20, nclust = 100, narm = 2, iccs = c(.01, .04), betas = c(0,.3)){
sig_b2 <- -1*iccs / (iccs - 1)
n <- nsub * nclust * narm
y <- rep_len(NA, n)
arm <- as.factor(rep(0:(narm-1), each = nsub*nclust))
clustid <- rep(1:(nclust*narm), each = nsub)
clustRElist <- rnorm(narm*nclust, mean = 0, sd = rep(sqrt(sig_b2), each = nclust))
clustRE <- rep(clustRElist, each = nsub)
sig_b2 <- rep(sig_b2, each = nclust*nsub)
error <- rnorm(n, mean = 0, sd = 1)
beta <- rep(betas, each = nclust*nsub)
linpred <- beta + clustRE + error
output <- cbind.data.frame(arm, clustid, sig_b2, clustRE, linpred)
return(output)
}
get_pval <- function(){
tt <- multi_arm_var_sim()
model_flex <- lme4::lmer(linpred ~ arm + (1 + arm | clustid),
data = tt)
diff_optims <- lme4::allFit(model_flex, data = tt, verbose = TRUE)
print(class(diff_optims[[1]]))
is.OK <- sapply(diff_optims, is, "merMod")
print(is.OK)
}
set.seed(2)
get_pval()
Thanks!!