How can I maintain consistent box width in a boxplot where factor*group combination has no observations?

Question

I am trying to create a boxplot for 2 groups across several factors along with labels for the number of observations. When there are no observations for one group a one factor level, the box for the group with observations takes up the space of both and looks odd.

Minimal example:

library(tidyverse)

mtcars %>%
  select(mpg, cyl,am) %>%
  filter(!(cyl == 8 & am == 0)) %>%
  ggplot(aes(factor(cyl),mpg,fill=factor(am))) + 
  stat_boxplot(geom = "errorbar") + ## Draw horizontal lines across ends of whiskers
  geom_boxplot(outlier.shape=1, outlier.size=3, 
               position =  position_dodge(width = 0.75)) +
  geom_text(data = mtcars %>%
              select(mpg, cyl, am) %>%
              filter(!(cyl == 8 & am == 0)) %>%
              group_by(cyl, am) %>%
              summarize(Count = n(),
                      q3 = quantile(mpg, 0.75),
                      iqr = IQR(mpg),
                      lab_pos = max(ifelse(mpg < q3+1.5*iqr, mpg, NA), na.rm = TRUE)),
                      aes(x= factor(cyl), y = lab_pos,label = paste0("n = ",Count, "\n")),
                  position = position_dodge(width = 0.75))

Which produces:

Is there a way to make the box for am(1) at cyl(8) half the width, so it's consistent with the other boxes on the plot? I have tried to use fake data, but that results in a count label for am(0) at cyl(8).

Answer 1

I was able to get a reasonable solution to this by installing the latest version of ggplot2 from GitHub and using position_dodge2 which uses preserve = "single" by default.

# Install devtools
install.packages('devtools')

# Install dependency of scales package
install.packages(c("RColorBrewer", "stringr", "dichromat", 
                   "munsell", "plyr", "colorspace"))

# Load devtools
library(devtools)

# Move to development mode
# This installed scales and ggplot2 in the "~/R-dev" directory, 
# so CRAN version of ggplot2 is not removed.
dev_mode(TRUE)

# Install scales
install_github("hadley/scales")

# Main branch of development
install_github("hadley/ggplot2", "hadley/develop")

# load development version of ggplot2
library(dplyr)
library(ggplot2)

mtcars %>%
  select(mpg, cyl,am) %>%
  filter(!(cyl == 8 & am == 0)) %>%
  ggplot(aes(factor(cyl),mpg,fill=factor(am))) + 
  stat_boxplot(geom = "errorbar",
               position =  position_dodge2(width = 0.75, preserve = "single")) + 
  geom_boxplot(outlier.shape=1, outlier.size=3, 
               position =  position_dodge2(width = 0.75, preserve = "single")) +
  geom_text(data = mtcars %>%
              select(mpg, cyl, am) %>%
              filter(!(cyl == 8 & am == 0)) %>%
              group_by(cyl, am) %>%
              summarize(Count = n(),
                        q3 = quantile(mpg, 0.75),
                        iqr = IQR(mpg),
                        lab_pos = max(mpg)),
            aes(x= factor(cyl), y = lab_pos,label = paste0("n = ",Count, "\n")),
            position = position_dodge2(width = 0.75, preserve = "single"))