New ggplot2.discrete.fill option doesn't work with ordered factor columns

[rlh1994]

I have found what I think is an issue with the new ggplot2.discrete.fill options, although it may be my understanding of what exactly "discrete" means and this may be by design. When using the option it works correctly for character columns, or as in the example factor(char_column)'s, but does not appear to work for already existing factor columns as this uses the default colours instead.

As I said this may be by design in that factors are treated different in plotting and I haven't realised, but this isn't clear in documentation anywhere I looked. I am using ggplot2 version 3.3.2, and R version 3.6.1.

library(ggplot2)
okabe <- c("#E69F00", "#56B4E9", "#009E73", "#F0E442", "#0072B2", "#D55E00", "#CC79A7")

cty_by_var <- function(var) {
  ggplot(mpg, aes(cty, colour = {{var}}, fill = {{var}})) +
    geom_density()
}

price_by_var <- function(var) {
  ggplot(diamonds, aes(price, colour = {{var}}, fill = {{var}})) +
    geom_density()
}

price_by_var2 <- function(var) {
  ggplot(diamonds, aes(price, colour = as.character({{var}}), fill = as.character({{var}}))) +
    geom_density()
}

# Works correctly with okabe colours 
withr::with_options(
  list(ggplot2.discrete.fill = okabe),
  print(cty_by_var(class))
)

# Works incorrectly, using the default colours
withr::with_options(
  list(ggplot2.discrete.fill = okabe),
  print(price_by_var(cut))
)

# Works correctly with okabe colours 
withr::with_options(
  list(ggplot2.discrete.fill = okabe),
  print(price_by_var2(cut))
)

^{Created on 2020-07-27 by the reprex package (v0.3.0)}

[yutannihilation]

Thanks, confirmed. Curious... (I replaced your reprex with the rendered version).

[yutannihilation]

Just remembered. The default colour/fill scales for ordered factors are different from the other. But, I think the option should work for ordered factors.

library(ggplot2)
okabe <- c("#E69F00", "#56B4E9", "#009E73", "#F0E442", "#0072B2", "#D55E00", "#CC79A7")

withr::with_options(
  list(ggplot2.discrete.fill = okabe),
  print(
    ggplot(mpg, aes(cty, fill = factor(class))) +
      geom_density()
  )
)

withr::with_options(
  list(ggplot2.discrete.fill = okabe),
  print(
    ggplot(mpg, aes(cty, fill = factor(class, ordered = TRUE))) +
      geom_density()
  )
)

^{Created on 2020-07-27 by the reprex package (v0.3.0)}

[yutannihilation]

A bit more context.

First two of this type of options were ggplot2.continuous.{colour,fill}, which were introduced in #2178. These were dedicated options to choose viridis palette as default or not.

scale_{colour,fill}_ordinal(), which are the default scale for ordered factors, were introduced in #2179. The intent of this is to distinguish ordered factors from unordered ones.

ggplot2/R/zxx.r

Line 6 in b4bc293

scale_colour_ordinal <- scale_colour_viridis_d

ggplot2/R/zxx.r

Line 65 in b4bc293

scale_fill_ordinal <- scale_fill_viridis_d

In short, at first, the system considered only two choices; viridis or not. Now that we have much more flexibility, we probably need more considerations. My take is:

• ggplot2 should have ggplot2.ordinal.{colour,fill} for consistency.
• ggplot2.discrete.{fill,colour} should affect (with a warning?) the fill/colour of ordinal variables when no ggplot2.ordinal.{colour,fill} is set.

[rlh1994]

I agree with your suggestions, having one option for all possible scale types feels correct. Same with discrete being the default when no ordinal is specified as I suspect many users (me included) aren't even aware of scale_*_ordinal.

[thomasp85]

I'm not sure I agree that the discrete colour setting should be used for ordinal data. What is your rationale for that? ordinal data is not very common so my guess is that when used, it is used for a reason and the scale should be sequential

[yutannihilation]

Since ordinal scale is a discrete scale, I thought it's natural that ordinal scale is subject to discrete scale's default. But, this is convincing...

the scale should be sequential

[dgkf]

Thanks for tagging #4253, @yutannihilation. One added motivation in that issue is to allow for customization via thematic themes. Currently the thematic package doesn't touch ordinal scales, which causes the viridis defaults to stand out against some color pallets.

[cpsievert]

Since I feel partly responsible for the confusion/request, I took a quick stab at this in #4277. @yutannihilation / @thomasp85 let me know if you think this is worth taking on.