drop = FALSE is not working properly (enforcing empty space for some missing levels in a boxplot)

[LuisLauM]

In the following example (using the iris dataset), I am creating a factor class variable in which one of the species does not contain values of level C. When I make the plot, I cannot find a way to make ggplot not drop the empty level (virginica-C). It suppose that the argument drop = FALSE do that, but it does not.

require(dplyr)
require(ggplot2)

iris %>% 
  
  mutate(fct_x = factor(x = sample(x = c("A", "B", "C"), size = nrow(.), replace = TRUE), 
                        levels = c("A", "B", "C"))) %>% 
  
  filter(!(Species == "virginica" & fct_x == "C")) %>% 
  
  ggplot(aes(x = Species, y = Sepal.Length, fill = fct_x)) +
  
  geom_boxplot() +
  
  scale_fill_discrete(drop = FALSE)

In other words, as you can see in the figure that I am attaching, the virginica group does NOT show an empty space for group C (because there are no elements of type virginica-C) and that is exactly what I want to achieve: to show that empty space in the figure.

[smouksassi]

the position scale is not related to the fill scale see how you can control position using the positioning:

require(dplyr)
#> Loading required package: dplyr
#> 
#> Attaching package: 'dplyr'
#> The following objects are masked from 'package:stats':
#> 
#>     filter, lag
#> The following objects are masked from 'package:base':
#> 
#>     intersect, setdiff, setequal, union
require(ggplot2)
#> Loading required package: ggplot2

set.seed( 123456)
iris %>%
  mutate(fct_x = factor(
    x = sample(
      x = c("A", "B", "C"),
      size = nrow(.),
      replace = TRUE
    ),
    levels = c("A", "B", "C")
  )) %>%
  filter(!(Species == "virginica" & fct_x == "C")) %>%
  ggplot(aes(x = Species, y = Sepal.Length, fill = fct_x)) +
  geom_boxplot(position = position_dodge(preserve = "single")) 

^{Created on 2022-06-16 by the reprex package (v2.0.1)}

[LuisLauM]

the position scale is not related to the fill scale see how you can control position using the positioning:

require(dplyr)
#> Loading required package: dplyr
#> 
#> Attaching package: 'dplyr'
#> The following objects are masked from 'package:stats':
#> 
#>     filter, lag
#> The following objects are masked from 'package:base':
#> 
#>     intersect, setdiff, setequal, union
require(ggplot2)
#> Loading required package: ggplot2

set.seed( 123456)
iris %>%
  mutate(fct_x = factor(
    x = sample(
      x = c("A", "B", "C"),
      size = nrow(.),
      replace = TRUE
    ),
    levels = c("A", "B", "C")
  )) %>%
  filter(!(Species == "virginica" & fct_x == "C")) %>%
  ggplot(aes(x = Species, y = Sepal.Length, fill = fct_x)) +
  geom_boxplot(position = position_dodge(preserve = "single")) 

Created on 2022-06-16 by the reprex package (v2.0.1)

Well, when you see the documentation (?scale_fill_hue), drop is explicitly shown as an available argument for scale_fill_hue (and scale_fill_discrete as well), so I think drop should be available as a parameter for position spacing. Your solution it is just useful for fixing the boxplot widths, but not for leave empty spaces for missing levels.

[smouksassi]

can you elaborate on what expected behavior you are looking at see below for how things differ when playing with
scale_x_discrete(drop=TRUE) vs scale_fill_discrete(drop=TRUE)

library(dplyr)
#> 
#> Attaching package: 'dplyr'
#> The following objects are masked from 'package:stats':
#> 
#>     filter, lag
#> The following objects are masked from 'package:base':
#> 
#>     intersect, setdiff, setequal, union
library(ggplot2)
library(patchwork)

set.seed( 123456)
a <- iris %>%
  mutate(fct_x = factor(
    x = sample(
      x = c("A", "B", "C",NA),
      size = nrow(.),
      replace = TRUE
    ),
    levels = c("A", "B", "C")
  )) %>%
  filter(!(Species == "virginica"),! fct_x=="C") %>%
  ggplot(aes(x = Species, y = Sepal.Length, fill = fct_x)) +
  geom_boxplot(position = position_dodge(preserve = "single"))+
  scale_x_discrete(drop=FALSE)+
  scale_fill_discrete(drop=FALSE)

b <- a +
  scale_x_discrete(drop=FALSE)+
  scale_fill_discrete(drop=TRUE)
#> Scale for 'x' is already present. Adding another scale for 'x', which will
#> replace the existing scale.
#> Scale for 'fill' is already present. Adding another scale for 'fill', which
#> will replace the existing scale.

c <- a +
  scale_x_discrete(drop=TRUE)+
  scale_fill_discrete(drop=FALSE)
#> Scale for 'x' is already present. Adding another scale for 'x', which will
#> replace the existing scale.
#> Scale for 'fill' is already present. Adding another scale for 'fill', which
#> will replace the existing scale.

d<- a +
  scale_x_discrete(drop=TRUE)+
  scale_fill_discrete(drop=TRUE)
#> Scale for 'x' is already present. Adding another scale for 'x', which will
#> replace the existing scale.
#> Scale for 'fill' is already present. Adding another scale for 'fill', which
#> will replace the existing scale.

(a | b) /
  (c | d)

^{Created on 2022-06-16 by the reprex package (v2.0.1)}

[LuisLauM]

can you elaborate on what expected behavior you are looking at see below for how things differ when playing with scale_x_discrete(drop=TRUE) vs scale_fill_discrete(drop=TRUE)

library(dplyr)
#> 
#> Attaching package: 'dplyr'
#> The following objects are masked from 'package:stats':
#> 
#>     filter, lag
#> The following objects are masked from 'package:base':
#> 
#>     intersect, setdiff, setequal, union
library(ggplot2)
library(patchwork)

set.seed( 123456)
a <- iris %>%
  mutate(fct_x = factor(
    x = sample(
      x = c("A", "B", "C",NA),
      size = nrow(.),
      replace = TRUE
    ),
    levels = c("A", "B", "C")
  )) %>%
  filter(!(Species == "virginica"),! fct_x=="C") %>%
  ggplot(aes(x = Species, y = Sepal.Length, fill = fct_x)) +
  geom_boxplot(position = position_dodge(preserve = "single"))+
  scale_x_discrete(drop=FALSE)+
  scale_fill_discrete(drop=FALSE)

b <- a +
  scale_x_discrete(drop=FALSE)+
  scale_fill_discrete(drop=TRUE)
#> Scale for 'x' is already present. Adding another scale for 'x', which will
#> replace the existing scale.
#> Scale for 'fill' is already present. Adding another scale for 'fill', which
#> will replace the existing scale.

c <- a +
  scale_x_discrete(drop=TRUE)+
  scale_fill_discrete(drop=FALSE)
#> Scale for 'x' is already present. Adding another scale for 'x', which will
#> replace the existing scale.
#> Scale for 'fill' is already present. Adding another scale for 'fill', which
#> will replace the existing scale.

d<- a +
  scale_x_discrete(drop=TRUE)+
  scale_fill_discrete(drop=TRUE)
#> Scale for 'x' is already present. Adding another scale for 'x', which will
#> replace the existing scale.
#> Scale for 'fill' is already present. Adding another scale for 'fill', which
#> will replace the existing scale.

(a | b) /
  (c | d)

Created on 2022-06-16 by the reprex package (v2.0.1)

I expected that scale_fill_discrete(drop = FALSE) will produce something like this:

As you can see, the (empty) space for veriginica-C is left, that is what I understand drop = FALSE means, as scale_x_discrete(drop = FALSE) does.

PS: This last figure was created in MS Paint (so I crop-paste some parts of the figure itself in order to show my point).

[smouksassi]

this is not how ggplot2 is designed to work scale_fill_discrete(drop = FALSE) will affect the fill scale so you keep or not the missing level color it won't affect the positioning on the x-axis

[LuisLauM]

this is not how ggplot2 is designed to work scale_fill_discrete(drop = FALSE) will affect the fill scale so you keep or not the missing level color it won't affect the positioning on the x-axis

So, there is no way to preserve empty space for a missing level, is that right?

[smouksassi]

my reply above with
geom_boxplot(position = position_dodge(preserve = "single"))

produce exactly what you edited in paint please reuse code and seed seed so things stay reproducible
I still don't get what you exactly want

[LuisLauM]

geom_boxplot(position = position_dodge(preserve = "single"))

I am talking about to a real leaving of space of a missing level.

If you run this piece of code, you will notice that in the output figure the groups A and C are smashed together, omiting the (empty) space for B:

require(dplyr)
require(ggplot2)

set.seed(666)

iris %>% 
  
  mutate(fct_x = factor(x = sample(x = c("A", "B", "C"), size = nrow(.), replace = TRUE), 
                        levels = c("A", "B", "C"))) %>% 
  
  filter(!(Species == "virginica" & fct_x == "B")) %>% 
  
  ggplot(aes(x = Species, y = Sepal.Length, fill = fct_x)) +
  
  geom_boxplot(position = position_dodge(preserve = "single"))

So it is not a matter of preserving widths, but spaces for missing levels.

[smouksassi]

A way to do it is below

require(dplyr)
require(ggplot2)
set.seed(666)
 plotdata <- iris %>% 
  mutate(fct_x = factor(x = sample(x = c("A", "B", "C"),
                                   size = nrow(.), replace = TRUE), 
                        levels = c("A", "B", "C"))) %>% 
  filter(!(Species == "virginica" & fct_x == "B")) 

  ggplot(plotdata,aes(x = fct_x, y = Sepal.Length, fill =fct_x   )) +
  geom_boxplot(position = position_dodge2(preserve = "single"))+
    facet_grid(~Species,switch = "x")+
    theme(strip.placement = "outside",
          axis.title.x.bottom = element_blank())

^{Created on 2022-06-16 by the reprex package (v2.0.1)}

[LuisLauM]

Yes, I got the same suggestion from people in Stackoverflow. I am using that way by now, but I hope ggplot team could end up with a more elegant solution. Now, I know that it is an #3345 but it seems it has a not-easy solution.

[danbebber]

It's 2 years since the last comment here. This functionality really should be implemented (maintaining correct position of second category boxes)

[teunbrand]

The issue is tracked in #3345, let's keep the discussion in one place.