Colour wisely…

Colour – the attribute of an image that makes it acceptable or destined for the bin. Colour has a funny effect on us – it’s a double-edged sword that greatly strengthens, or weakens data representation in such a huge level. No one really talks about what’s a good way to colour an image or a graph, but it’s something that most can agree as being pleasing, or disgusting. There are two distinctive advantages to colouring a graph: it conveys both quantitative and categorical information very, very well. Thus, I will provide a brief overview (with code) on how colour can be used to display both quantitative and qualitative information. (*On the note of colours, Nick has previously discussed how colourblindness must be considered in visualising data…).

1. Colour conveys quantitative information.
A huge advantage of colour is that it can provide quantitative information, but this has to be done correctly. Here are three graphs showing the exact same information (the joint density of two normal distributions) and  we can see from the get-go which method is the best at representing the density of the two normal distributions:

Colouring the same graph using three different colour maps.

Colouring the same graph using three different colour maps.

If you thought the middle one was the best one, I’d agree too. Why would I say that, despite it being grayscale and seemingly being the least colourful of them all?

  • Colour is not limited to hues (i.e. whether it’s red/white/blue/green etc. etc.); ‘colour’ is also achieved by saturation and brightness (i.e., how vivid a colour is, or dark/light it is). In the case of the middle graph, we’re using brightness to indicate the variations in density and this is a more intuitive display of variations in density. Another advantage of using shades as the means to portray colour is that it will most likely be OK with colourblind users.
  • Why does the graph on the right not work for this example? This is a case where we use a “sequential” colour map to convey the differences in density. Although the colour legend clarifies what colour belongs to which density bin, without it, it’s very difficult to tell what “red” is with respect to “yellow”. Obviously by having a colour map we know that red means high density and yellow is lower, but without the legend, we can interpret the colours very differently, e.g. as categories, rather than quantities. Basically, when you decide on a sequential colour map, its use must be handled well, and a colour map/legend is critical. Otherwise, we risk putting colours as categories, rather than as continuous values.
  • Why is the left graph not working well? This is an example of a “diverging” colourmap.
    It’s somewhat clear that blue and red define two distinct quantities. Despite this, a major error of this colour map comes in the fact that there’s a white colour in the middle. If the white was used as a “zero crossing” — basically, where a white means the value is 0 — the diverging colour map would have been a more effective tool. However, we can see that matplotlib used white as the median value (by default); this sadly creates the false illusion of a 0 value, as our eyes tend to associate white with missing data, or ‘blanks’. Even if this isn’t your biggest beef with the divergent colour map, we run into the same colour as the sequential colour map, where blue and red don’t convey information (unless specified), and the darkness/lightness of the blue and red are not linked very well without the white in the middle. Thus, it doesn’t do either job very well in this graph. Basically, avoid using divergent colourmaps unless we have two different quantities of values (e.g. data spans from -1 to +1).

2. Colour displays categorical information.
An obvious use of colour is the ability to categorise our data. Anything as simple as a line chart with multiple lines will tell you that colour is terrific at distinguishing groups. This time, notice that the different colour schemes have very different effects:

Colour schemes can instantly differentiate groups.

Colour schemes can instantly differentiate groups.

Notice how this time around, the greyscale method (right) was clearly the losing choice. To begin with, it’s hard to pick out what’s the difference between persons A,B,C, but there’s almost a temptation to think that person A morphs into person C! However, on the left, with a distinct set of colours, there is a clear distinction of persons A, B, and C as the three separate colours. Although a set of distinct three colours is a good thing, bear in mind the following…

  • Make sure the colours don’t clash with respect to lightness! Try to pick something that’s distinct (blue/red/green), rather than two colours which can be interpreted as two shades of the same colour (red/pink, blue/cyan, etc.)
  • Pick a palette to choose from – a rainbow is typically the best choice just because it’s the most natural, but feel free to choose your own set of hues. Also include white and black as necessary, so long as it’s clear that they are also part of the palette. White in particular would only work if you have a black outline.
  • Keep in mind that colour blind readers can have trouble with certain colour combinations (red/yellow/green) and it’s best to steer toward colourblind-friendly palettes.
import numpy as np
import matplotlib.pyplot as plt
import scipy.stats as sp
from mpl_toolkits.axes_grid1 import make_axes_locatable

### Part 1
# Sample 250 points
np.random.seed(30)
x = np.random.normal(size = 250)
np.random.seed(71)
y = np.random.normal(size = 250)

# Assume the limits of a standard normal are at -3, 3
pmin, pmax = -3, 3

# Create a meshgrid that is 250x250
xgrid, ygrid = np.mgrid[pmin:pmax:250j, pmin:pmax:250j]
pts = np.vstack([xgrid.ravel(), ygrid.ravel()]) # ravel unwinds xgrid from a 250x250 matrix into a 62500x1 array

data = np.vstack([x,y])
kernel = sp.gaussian_kde(data)
density = np.reshape(kernel(pts).T, xgrid.shape) # Evaluate the density for each point in pts, then reshape back to a 250x250 matrix

greys = plt.cm.Greys
bwr = plt.cm.bwr
jet = plt.cm.jet

# Create 3 contour plots
fig, ax = plt.subplots(1,3)
g0 = ax[0].contourf(xgrid, ygrid, density, cmap = bwr)
c0 = ax[0].contour(xgrid, ygrid, density, colors = 'k') # Create contour lines, all black
g1 = ax[1].contourf(xgrid, ygrid, density, cmap = greys)
c1 = ax[1].contour(xgrid, ygrid, density, colors = 'k') # Create contour lines, all black
g2 = ax[2].contourf(xgrid, ygrid, density, cmap = jet)
c2 = ax[2].contour(xgrid, ygrid, density, colors = 'k') # Create contour lines, all black

# Divide each axis then place a colourbar next to it
div0 = make_axes_locatable(ax[0])
cax0 = div0.append_axes('right', size = '10%', pad = 0.1) # Append a new axes object
cb0  = plt.colorbar(g0, cax = cax0)

div1 = make_axes_locatable(ax[1])
cax1 = div1.append_axes('right', size = '10%', pad = 0.1)
cb1  = plt.colorbar(g1, cax = cax1)

div2 = make_axes_locatable(ax[2])
cax2 = div2.append_axes('right', size = '10%', pad = 0.1)
cb2  = plt.colorbar(g2, cax = cax2)

fig.set_size_inches((15,5))
plt.tight_layout()
plt.savefig('normals.png', dpi = 300)
plt.close('all')

### Part 2
years = np.arange(1999, 2017, 1)
np.random.seed(20)
progress1 = np.random.randint(low=500, high =600, size = len(years))
np.random.seed(30)
progress2 = np.random.randint(low=500, high =600, size = len(years))
np.random.seed(40)
progress3 = np.random.randint(low=500, high =600, size = len(years))

fig, ax = plt.subplots(1,2)
ax[0].plot(years, progress1, label = 'Person A', c = '#348ABD')
ax[0].plot(years, progress2, label = 'Person B', c = '#00de00')
ax[0].plot(years, progress3, label = 'Person C', c = '#A60628')
ax[0].set_xlabel("Years")
ax[0].set_ylabel("Progress")
ax[0].legend()

ax[1].plot(years, progress1, label = 'Person A', c = 'black')
ax[1].plot(years, progress2, label = 'Person B', c = 'gray')
ax[1].plot(years, progress3, label = 'Person C', c = '#3c3c3c')
ax[1].set_xlabel("Years")
ax[1].set_ylabel("Progress")
ax[1].legend()

fig.set_size_inches((10,5))
plt.tight_layout()
plt.savefig('colourgrps.png', dpi = 300)
plt.close('all')