This suggests to me that the output is being bound into a matrix (which requires a single type) and then converted to a data frame later, which it should probably not be. Rather than potentially round-tripping through a string (with the possibility for information loss) I would suggest binding results together into a data frame directly, which should be more efficient anyway.

This suggests to me that the output is being bound into a matrix

Yes, this is indeed the case

It seems this is also caused by the behaviour of unlist, which results in a vector, and when performed on a list with different types inside, will convert them all to the same (in this case character). This unlisting is done to handle summary functions that return multiple elements (e.g. quantile2).

However, I don't know of a base R equivalent to purrr::flatten which would do what we want in this case (returning a list, but without nesting).

Would unlist(mylist, recursive = FALSE) already to the trick?

I don't think so, at least initial testing adding recursive = FALSE in the unlist calls doesn't seem to change anything

What would purrr::flatten do instead? Not sure I get how this would fix things.

I think purrr::flatten always returns a list, whereas unlist will return a vector even if recursive = F if there are no nested lists

x = list(1, 2,3)

unlist(x, recursive = F)
#> 1 2 3

purrr::flatten(x)
# [[1]] 
# [1] 1
#  [[2]] 
# [1] 2
# [[3]] 
# [1] 3 

I see. makes sense. and would the use of purrr::flatten fix the issue?

I think so. I can test tomorrow. Are you thinking to copy the function over to posterior?

I am not sure yet. If it works, perhaps you can make a PR using purrr::flatten and I will then edit the code to no longer require purrr.

Sure, I can edit the existing PR

Or any other approach like direct data.frame conversion, if this is also efficient.