Legends should not use after_scale()ed values

[yutannihilation]

(Originally commented at #3725 (comment))

Currently, legends reflect the value of after_scale(). But, it's not always possible because the expression of after_scale() might refer to the variable unavailable at the stage of drawing legends.

For example, I think it is a proper usage of stage() to modify scaled colours by combining other aesthetics. But, the following example has two problems:

1. x is not available on drawing legends
2. even if we manage to make x available, it's not clear what legends should be drawn because the scaled colours are discrete but the after_scale()ed values are continuous.

Especially because of the problem 2., I think legends should use the raw scaled values, not the after_scale()ed ones.

library(ggplot2)

set.seed(88)
d <- data.frame(x = runif(9), y = runif(9), g = rep(c("a", "b", "c"), each = 3))
p <- ggplot(d) +
  geom_point(aes(x, y, colour = stage(g, after_scale = alpha(colour, x))), size = 10)

# fail
p
#> Error in alpha(colour, x): object 'x' not found

# ok
p + theme(legend.position = "none")

^{Created on 2020-01-14 by the reprex package (v0.3.0)}

[yutannihilation]

Another example. While this doesn't fail, the legends for colour lose their colours. I think they should be coloured.

I bet this is just a corner case and not a blocking issue, so I don't include this in the v3.3.0 milestones. But, I hope we can reach to some conclusion about whether my claim is valid or not before this feature is released.

library(ggplot2)

set.seed(88)
d <- data.frame(x = runif(9), y = runif(9), g = rep(c("a", "b", "c"), each = 3))
ggplot(d) +
  geom_point(aes(x, y, size = x + y,
                 colour = stage(g, after_scale = ifelse(size > 4, colour, "grey"))))

^{Created on 2020-01-14 by the reprex package (v0.3.0)}

[thomasp85]

yeah - I think you are right... I didn't think these implications through when I added the legend modification code...

on the other hand, there are valid reasons for having the legend reflect the transformation as well... Maybe we should add a switch?

[yutannihilation]

there are valid reasons for having the legend reflect the transformation as well

Ah, agreed. While I'm raising this issue, I also feel the example in your post is the case when it's visually more sound when the legends reflect the transformation. I have no idea what interface the switch should be, but probably it's needed here.

library(ggplot2)

ggplot(mpg) + 
  geom_bar(
    aes(
      x = drv, 
      colour = stage(start = drv, after_scale = alpha(colour, 0.5))
    ), 
    fill = NA, size = 4
  )

^{Created on 2020-01-15 by the reprex package (v0.3.0)}

[yutannihilation]

For reference, while we closed this, this case is not fixed. We decided this is not worth fixing compared to the complexity. Let's revisit here when it's necessary.

library(ggplot2)

set.seed(88)
d <- data.frame(x = runif(9), y = runif(9), g = rep(c("a", "b", "c"), each = 3))
ggplot(d) +
  geom_point(aes(x, y, size = x + y,
                 colour = stage(g, after_scale = ifelse(size > 4, colour, "grey"))))

^{Created on 2020-01-14 by the reprex package (v0.3.0)}