scale_color_gradient2 doesn't entirely respect limits argument

[rpruim]

If midpoint is not half-way between the two values in limits, then the gradient scale extends beyond the limit that is closer to midpoint but the guide stops where limits specifies. Especially in conjunction with na.value, this can mislead the reader.

In this plot the color scale goes below what the guide displays, and values below 2 are not indicated as NA.

library(scales)
library(ggplot2)
library(dplyr)
ggplot(mapping = aes(x = Sepal.Length, y = Sepal.Width) ) +
  geom_point(data = iris, aes(color = Petal.Length), size = 2) +
  geom_point(data = iris %>% filter(Petal.Length < 2), size = 4, shape = 5, color = "red") +
  geom_point(data = iris %>% filter(Petal.Length > 6), size = 4, shape = 0, color = "red") +
  scale_color_gradient2(low = muted("green"), high = muted("navy"), 
                        mid = "gray80",
                        midpoint = 3, limits = c(2, 6), 
                        na.value = "orange")

If we change midpoint to 3, then the scale truly runs from 2 to 6 and values below 2 are
flagged.

ggplot(mapping = aes(x = Sepal.Length, y = Sepal.Width) ) +
  geom_point(data = iris, aes(color = Petal.Length), size = 2) +
  geom_point(data = iris %>% filter(Petal.Length < 2), size = 4, shape = 5, color = "red") +
  geom_point(data = iris %>% filter(Petal.Length > 6), size = 4, shape = 0, color = "red") +
  scale_color_gradient2(low = muted("green"), high = muted("blue"), 
                        mid = "gray80",
                        midpoint = 4, limits = c(2, 6), 
                        na.value = "orange")

[karawoo]

Thanks for opening this issue and including an example. Would you mind including the plot images as well? The reprex package can help streamline this.

[rpruim]

Here you go. Notice the color change for NA's when we switch the value of midpoint from plot 1 to plot 2. Things in diamonds or squares should be marked as NA (orange) if limits is being respected tightly.

[foo-bar-baz-qux]

I think the issue is to do with the mid_rescaler which calls scales::rescale_mid. If the midpoint is set correctly to be the middle of the limits, the scaled limits will be [0, 1] such that all points outside the original limits, would also be outside [0, 1]. This leads to the expected result of those points outside the range being marked as NA.

However, the midpoint parameter might not do what is expected here if it is not in the middle of the data; if you look at the scales::rescale_mid function, it effectively changes the range of [0, 1] to correspond to twice the distance from the midpoint to either the minimum or maximum limit (whichever distance is greater). So using a midpoint of 3 instead of 4 in the original example, effectively scales the range [0, 6] down to [0, 1], rather than [2, 6] down to [0, 1] as one might expect. This leads to all points with Petal.Length < 2 to still be displayed since all original petal lengths are > 0, and therefore will still be greater than 0 after the scaling.

If you set the midpoint to 10, you get the opposite effect since now 10 is further from the minimum of the range (2) compared to the maximum of the range (6), so all points with petal length < 2 are considered NA since they will be < 0 in the scaled range.

Some possible solutions (from least to most conservative):

1. Set the default mid-point as being the middle of the from parameter in the function returned by mid_rescaler.
   • The default of midpoint = 0 can definitely lead to strange results depending on the range of the data. E.g. if the range of the data is 100 - 200, then a 0 midpoint effectively renders the low parameter unused.
2. Throw a warning in mid_rescaler when mid-point is not in the middle of the range.
3. Documentation change to reflect behavior when not actually set to the middle of the data range.
   • Perhaps there's a use-case for manually setting the mid-point, but the current behaviour may not be obvious from the documentation