Keeping Colours Constant

A note to self

In the emissions shiny app from Saturday there is the option to remove some of the sectors or fuels from the graph. Unfortunately when you did this the graphs originally failed to remember the colours assigned to the lines in the previous version and reassigned them among the remaining ones. This was a problem, but one I managed to fix. For future self-reference, here is a tutorial how to do it.

For example, here is the same problem but with a scatter plot with the famous mpg data-set:

library(ggplot2)
library(RColorBrewer)
data(mpg)
kable(tail(mpg))
ggplot(mpg, aes(cty, hwy, color=fl)) + geom_point(size = 2.5) + 
  scale_color_brewer(type="qual", palette = "Paired") + theme_linedraw() +
  ggtitle("cty by hwy by fuel type")

Note that there is a fair amount of over plotting here, but that’s not important. cty is the mpg in the city; hwy is for highways; and fl is the fuel type: CNG, diesel, ethanol, premium, and regular.

Suppose we remove the premium fuel cars—the green dots—from the data set:

library(dplyr)
mpg <- filter(mpg, fl != "p")
ggplot(mpg, aes(cty, hwy, color=fl)) + geom_point(size=2.5) + 
  scale_color_brewer(type="qual", palette = "Paired") + theme_linedraw() +
  ggtitle("cty by hwy by fuel type, removing premium fuel")

Premium was green, but now regular is. This is a problem, but one with a solution: manual specification of colours from the palette. Lets start again:

data(mpg) # reset dataset
fl.types <- unique(mpg$fl)
fl.cols <- brewer.pal(length(fl.types), "Paired")
names(fl.cols) <- fl.types

There are several ways to do this, but in the course of writing this post I actually discovered that the way I was doing it in the emissions app was not ideal. In that case I was creating fl.cols as a data frame, left_joining it to the dataset, and then applying the colours with scale_color_manual(values=unique(mpg$cols)). This works fine in the app, but relies on the hidden assumption of the plotting order: because the emissions algorithm involves a tidy::gather it has it’s equivalent of the fl column as a factor, with the same order of its levels as the colours are assigned, this is not a problem. Unfortunately this is does not work immediately with the mpg dataset we have here, something which caused me no end of trouble.

Instead, fl.cols is a named vector, with the names corresponding to the fl levels. We can make the original graph in the new way like so:

ggplot(mpg, aes(cty, hwy, color=fl)) + geom_point(size = 2.5) + 
  scale_color_manual(values=fl.cols) + theme_linedraw() +
  ggtitle("cty by hwy by fuel type, v2")

…except this has a problem: the colours are in a new order, and for the same order-of-plotting reason as before. This is because scale_color_brewer() was assigning based on alphabetical order, while we assigned by order of appearence in unique(). This is easily fixed:

names(fl.cols) <- sort(fl.types)
ggplot(mpg, aes(cty, hwy, color=fl)) + geom_point(size = 2.5) + 
  scale_color_manual(values=fl.cols) + theme_linedraw() +
  ggtitle("cty by hwy by fuel type, v2.1")

But now, finally, filtering the mpg dataset will produce the colours we want, in the order we want!

mpg <- filter(mpg, fl != "p")
ggplot(mpg, aes(cty, hwy, color=fl)) + geom_point(size=2.5) + 
  scale_color_manual(values=fl.cols) + theme_linedraw() +
  ggtitle("cty by hwy by fuel type, removing premium fuel, v2")

Voilà!