`geom_boxplot` Whiskers never appear when IQR = 0 (`upper - lower = 0`)

[math-mcshane]

I found a problem with ...

geom_boxplot

I expected ...

Whiskers to appear. There's no convenient way to get whiskers to appear when IQR = 0 (upper - lower = 0) (that I am aware of*). Setting coef = 1000 or some other large value doesn't work (as it would in most other cases; it wouldn't be reasonable to expect coef*0 to be anything other than 0). In particular, it would make sense to let the whiskers extend to the min and max of the data (See the first bullet: https://en.wikipedia.org/wiki/Box_plot#Whiskers -- allowing boxplot whiskers to extend to the min and max should be a reasonable option). I debated whether to submit as bug or feature request, but since I'm discussing an edge case where IQR = 0 (rare), this seemed more like a bug.

Here is the code to reproduce the bug:

y = data.frame(x = c(0, 10, 10, 10, 100))

# Produces a boxplot with "box" of width 0 (IQR = 0)
ggplot(data = y, aes(x = x)) + geom_boxplot()
# No matter what we change coef to, we will never get whiskers
ggplot(data = y, aes(x = x)) + geom_boxplot(coef = 1000)
# Want: boxplot that looks like this: -|-

# Can hack this slightly with: 
ggplot(data = y, aes(x = x)) + geom_boxplot(xmin = min(y$x), xmax = max(y$x))
# And make the outliers disappear:
ggplot(data = y, aes(x = x)) + geom_boxplot(xmin = min(y$x), xmax = max(y$x), outlier.alpha = 0)

*Apologies if there is already a parameter that can be set to allow this.

[teunbrand]

You can arbitrarily parametrise the boxplot by using the boxplot geometry and summary stat.

This is somewhat more work than flipping a switch somewhere, but I think that one should only consciously deviate from what is popularly understood about boxplots. Making this harder makes you think longer about your decision, which makes the deviation more deliberate. Therefore, I think the current behaviour is appropriate and we should not add a flip for this behaviour.

library(ggplot2)

df <- data.frame(y = c(0, 10, 10, 10, 100))

ggplot(df, aes(x = 1, y)) +
  geom_boxplot(
    stat = "summary",
    fun.data = function(x) {
      data.frame(
        ymin   = min(x),
        lower  = quantile(x, 0.25),
        middle = median(x),
        upper  = quantile(x, 0.75),
        ymax   = max(x)
      )
    }
  )

^{Created on 2023-11-07 with reprex v2.0.2}

[math-mcshane]

Hi, @teunbrand , as you'll see in the wiki page I linked in my original comment, having whiskers extend to the min and max is not an uncommon approach. In fact, it was the original approach pioneered by Spear in 1950, as noted by @hadley himself:

-- in this 2011 paper by Wickham and Stryjewski: https://vita.had.co.nz/papers/boxplots.pdf

And you can see this on page 1 in Tufte's article/chapter:
https://www.jstor.org/stable/3545420?origin=crossref&seq=1

I'm not asking that the default be changed -- I agree that many people understand boxplots' whiskers' length defaults to 1.5*IQR. However, as my rare counterexample suggests, there are times where the default approach to making the whiskers extend to the min and max (picking arbitrarily large coef) fails under the current formulation of geom_boxplot.

As we can see from these two Stack Overflow posts (1 and 2), which have 11K and 26K views, it's not an uncommon desire. They also show that extending the whiskers to the min and max often leads to an unnecessary amount of code.

[teunbrand]

And the most upvoted answer also follows the summary stat approach, which I think is an entirely adequate solution with the appropriate amount of code overhead.