`fct_lump_prop()` has undocumented feature for one other level

[caneff]

There is nothing in documentation that says lumping into other does nothing if that is only one other category. I don't mind the behavior but I spent some time debugging my code because of it.

What happens:

library(forcats)
f <- factor(c(1,2,2,2))
fct_lump_prop(f, prop=0.5)
#> [1] 1 2 2 2
#> Levels: 1 2

^{Created on 2020-10-01 by the reprex package (v0.3.0)}

However I expect it to look like

#> [1] Other 2 2 2 2
#> Levels: 2 Other

Again I'm fine with the current behavior it is just unexpected for how the function reads. 1 is less than the 50% of the total so I expected it to become other. So either update documentation or behavior.

[thackl]

Just ran into the same issue, also with fct_lump_n(). To me, this feels quite counterintuitive even if it were documented...

library(forcats)
fct_lump_n(c("a","a","c"), n=1)
#> [1] a a c
#> Levels: a c
fct_lump_n(c("a","a","b","c"), n=1)
#> [1] a     a     Other Other
#> Levels: a Other

^{Created on 2021-01-30 by the reprex package (v0.3.0)}

[hadley]

I think the behaviour makes sense because what's the point in having an "other" level that contains only one level?

[huftis]

I think the behaviour makes sense because what's the point in having an "other" level that contains only one level?

If for nothing else, then for 1) consistency among the fct_lump_*() functions, 2) consistency with the documentation, and 3) the principle of least surprise. Regarding the first reason, fct_lump_min() does label the ‘other’ level "other" even if this doesn’t involve any merging:

> fct_lump_min(c("a","a","c"), min=2)
[1] a     a     Other
Levels: a Other

Any in case you want to change the fct_lump_min() to be consistent with fct_lump_n() instead of the other way around: The current fct_lump_min() behaviour is actually very useful. For example, I sometimes have to ‘anonymise’ levels, e.g., institution names, that have less than five observations (and then lump them into one level). Then I can write:

fct_lump_min(x, min = 5, other = "[NAME REDACTED]")

And this will work correctly even if there is only one such institution.

[hadley]

Ok; let's just hope this doesn't affect too much code in the wild.