alpha and fill legends in ggplot2 boxplots?

Question

I'm trying to combine alpha and fill aesthetics. It works when I'm using geom_bar (or geom_points, for color), but the alpha legend doesn't work in when I'm using geom_boxplot.

library(data.table)
library(ggplot2)
dt = data.table(x = rep(1:5,6), y = rnorm(30),
               tag1 = rep(c('hey', 'what'), 15),
               tag2 = rep(c('yeah', 'yeah', 'so', 'so', 'so'), 6))

It works for bars:

ggplot(dt[, list(y = mean(y)), by = list(x, tag1, tag2)],
       aes(x = x, y = y, fill = tag1, alpha = tag2,
           group = interaction(x,tag1,tag2))) +
  geom_bar(stat = 'identity', position = 'dodge')

But not for boxplot - the alpha legend is empty.

ggplot(dt, aes(x = x, y = y, fill = tag1, alpha = tag2,
               group = interaction(x, tag1, tag2))) +
  geom_boxplot()

A simpler version can be done with no fill - it seems like bar defaults to gray/lightgray, and boxplot defaults to white/lightwhite:

ggplot(dt[, list(y = mean(y)), by = list(x, tag2)],
       aes(x = x, y = y, alpha = tag2,
           group = interaction(x,tag2))) +
 geom_bar(stat = 'identity')

ggplot(dt, aes(x = x, y = y, alpha = tag2,
               group = interaction(x, tag2))) +
  geom_boxplot()

But I'm not really sure how to fix this. Any thoughts?

Answer 1

I'm not sure why ggplot doesn't actually provide the alpha levels in the legend for boxplots, but you can hard code it using override.aes. (Editorial note: I find the alpha aesthetic a bit confusing for either the boxplot or the bar plot. It's hard to mentally separate the transparency from the fill color and the grey-scale alpha legend exacerbates the problem, because nothing is mapped to grey in the plot.)

In the code below, to improve visibility of the legend, I've removed the box lines from the alpha legend and increased the legend key height. I've also edited the aesthetics to remove the need for the group argument.

ggplot(dt, aes(x=factor(x), y=y, fill=tag1, alpha=tag2)) + 
  geom_boxplot() +
  scale_alpha_manual(values=c(0.2,0.7)) +
  guides(alpha=guide_legend(override.aes=list(fill=hcl(c(15,195),100,0,alpha=c(0.2,0.7)),
                                              colour=NA))) +
  theme(legend.key.height=unit(1,"cm"))

Another option would be to use interaction for both the fill and alpha aesthetics, but it turns out ggplot doesn't include any colors in that case:

ggplot(dt, aes(x=factor(x), y=y, alpha=interaction(tag1,tag2)), 
       fill=interaction(tag1,tag2)) + 
  geom_boxplot() +
  scale_fill_manual(values=rep(hcl(c(15,195),100,65), 2)) +
  scale_alpha_manual(values=rep(c(0.3, 1), each=2)) +
  theme(legend.key.height=unit(2,"cm")) 

So, instead you can do it all with the fill aesthetic, but include transparency in the colour specification. This works, but, once again, because transparency and color are somewhat intermingled in visual perception, it's probably better to just go with four different colors.

ggplot(dt, aes(x=factor(x), y=y, fill=interaction(tag1,tag2,sep="-"))) + 
  geom_boxplot() +
  scale_fill_manual(values=hcl(c(15,195,15,195),100,65, alpha=c(0.4,0.4,1,1))) +
  theme(legend.key.height=unit(1,"cm")) +
  labs(fill="Tag 1 - Tag 2")