Let’s play a quick game! Work through these 17 questions to see if you have an eye for good data viz design. Many (but not all) are related to color choices:
03:00
HSV and HSL are used commonly in color pickers (e.g. Google color picker). HCL underlies some default {ggplot2} parameters.
HSV
HCL
You don’t need to worry much about the underlying theory of color spaces, but know that changing any of the parameters (e.g. hue, saturation, etc.) can influence how we perceive information in a data visualization.
02:00
- mainly formed by selecting different hues
- hues assigned to each group must be distinct and ideally have different lightnesses
- limit to no more than 7 hues
- colors assigned to data values on a continuum, based on lightness, hue, or both
- lower values typically associated with lighter colors & higher values associated with darker colors (though not a hard and fast rule; make choices clear with legend)
- can use a single hue or two hues
- combination of two sequential palettes with a shared endpoint at the central value
- central value is assigned a light color (light gray is best)
- use a distinctive hue for each of the component palettes
We’ll be testing out different palettes throughout this lesson. Instead of having to retype the code for our plots each time, let’s create and save two versions of a penguin scatter plot. We can then call either of these plot objects to modify with different color scales:
Requires a categorical color scale
Color vision deficiency aka colorblindness is the decreased ability to see color or differences in color. It’s estimated that about 1 in 12 men (8%) and 1 in 200 women (0.5%) are affected (Wikipedia).
Color plate tests are used to help identify different forms of color blindness. Try using the Let’s get color blind Chrome extension to emulate different forms of colorblindness while looking at the above plates. Image source: American Optometric Association
colors don’t follow any natural perceived ordering (no innate sense of higher or lower)
perceptual changes in rainbow colors are not uniform (e.g. colors appear to change faster in yellow region than green region)
insensitive to color vision deficiencies
Try one of these improved versions (right), instead:
Problematic, perceptually nonuniform and unordered rainbow colormaps
Improved, perceptual uniform and diverging rainbow colormaps
The viridis color scales are perceptually-uniform (even when printed in gray scale) and colorblindness-friendly:
Continuous viridis scales
Binned viridis scales
There are a number of different ways to apply viridis color scales, but I often opt for scale_*_viridis_*() functions, which come as part of {ggplot2}.
Try out the palette options below, then check out the documentation and play around with some alternative options as well.
ColorBrewer offers a number of colorblind-friendly color schemes for maps and other graphics. Check them out using {RColorBrewer} or the web-based interface.
Colorblind-friendly palettes, viewed using display.brewer.all()
ColorBrewer’s web-based interface for exploring palettes
{RColorBrewer} comes with a couple useful functions for quickly viewing and assembling your palette’s HEX codes:
Use the right function (all part of {ggplot2}) for the type of data / palette:
Use scale_color_brewer() to apply qualitative palettes
Use scale_color_distiller() for unclassed continuous color scales
Check out the documentation and play around with some alternative options.
The {paletteer} package provides a common interface for accessing a near-comprehensive list of palettes (over 2,500!!) across various packages.
There are a number of ways to browse the list of supported palette packages, including in the {paletteer} documentation, on the r-color-palettes GitHub repo, on the R Color Palettes website, and the R-Graph Gallery’s Color Palette Finder.
- two groups of palettes: discrete and continuous
- discrete palettes can be fixed (have a set # of colors) or dynamic (adjustable # of colors based on your specifications)
For qualitative (categorical) data 
:
scale_*_manual()
For quantitative (numeric) data:
Unclassed palettes 
:
scale_*_gradient(): creates a two color gradient (low-high)scale_*_gradient2(): creates a diverging color gradient (low-mid-high)scale_*_gradientn(): creates a n-color gradient
Classed palettes 
:
scale_*steps(): creates a two color binned gradient (low-high)scale_*_steps2(): creates a diverging binned color gradient (low-mid-high)scale_*_stepsn(): creates a n-color binned gradient
Use the fill variant of the above functions for areas, bars, etc. and the color variant for points, lines, etc.
Examples using our discrete color palette:
Examples using our continuous color palette:
Figure 4. Appropriate diverging and sequential colour schemes for the following climate data (a), absolute temperature (b), absolute precipitation (c), temperature anomaly (d), precipitation or runoff anomaly (e and f) other climate variables with no symbolic association . Schemes in this figure are 7 class ones designed by Cynthia Brewer, (Brewer et al. 2003)
Image source: htmlcolorcodes.com
blue-green & red-orange are complementary and therefore offer the strongest possible contrast
Find descriptions of blue-green & red-orange on htmlcolorcodes.com
We associate meaning with different hues (e.g. cold / sad = blue, hot / angry = red), and importantly, these associations can differ among cultures.
Source: Information is Beautiful
Source: Information is Beautiful
“lightness, brightness, and saturation can communicate the level of seriousness, intensity, and emotional weight in a visual work” -Cédric Scherer
“lightness, brightness, and saturation can communicate the level of seriousness, intensity, and emotional weight in a visual work” -Cédric Scherer
(Right) COVID-19 tracker by the Johns Hopkins University (screenshot from 2020-07-27, courtesy of Cédric Scherer). Red tends to elicit panic / fear. (Left) A map of confirmed COVID-19 cases by Datawrapper (screenshot from 2020-07-27, courtesy of Cédric Scherer). Blues and greens help to avoid such a strong fearful emotional response.
Though it may be temping to use bright / bold colors to grab attention, it can lead to eye strain and make it more challenging for your readers to focus on your chart.
Though it may be temping to use bright / bold colors to grab attention, it can lead to eye strain and make it more challenging for your readers to focus on your chart.
Be sure to consider what we’ve already discussed:
- ensure that you’re picking colorblind-friendly color combos
- use color wheels to identify color harmonies
- think carefully about what emotions / messages your color choices will convey
- avoid lots of pure / fully-saturated hues
And also consider some other important sources of inspiration:
- your company or organization’s brand / logo
- steal colors from your favorite / relevant images using tools like Color Thief
- use a randomized palette generator, like coolors.co
- find a color picker for generating HEX codes – my favorite it HTML Color Codes
I recommend saving your palette to a named vector outside of your ggplot – this prevents lengthy palettes from creating cluttered ggplot code and allows you to reuse your palette across multiple plots:
We should always be consistent with our colors. E.g. if Gentoo penguins are blue in one plot, they should be blue in all plots. Notice that our colors don’t “stick” with the species they represent, but rather they’re applied in the order that they appear in our palette:
Adelie, Chinstrap & Gentoo penguins
Setting the names of our vector elements (colors) ensures that they stick with those factor levels across all of our visualizations:
Adelie, Chinstrap & Gentoo penguins
The x-axis already identifies each bar, so adding color provides no additional information. Use color intentionally, and only when it communicates something meaningful!
Ensure consistent use of colors across multiple visualizations that display the same groups.
The more colors you use, the more difficult to becomes to distinguish between groups. Consider an entirely different chart type, or use color to highlight the only the group(s) of interest. | “Consider the color grey as the most important color in Data Vis.” -Lisa Charlotte Muth
Always include a color key, in the form of a traditional legend or otherwise.
Many newsrooms avoid pink/blue altogether, but choices are not consistent | Avoid steretypical skin colors, don’t use gray for “other” or “multiracial” categories, use less saturated colors (saturated colors have strong associations, e.g. green as positive or right, red as dangerous or important), and keep shuffling your colors!
In most cases, readers will associate bright colors with lower values and darker colors with higher values. Build gradients accordingly.
“humans perceive bright colors on elevation maps to represent a high altitude, with darker colors representing naturally low-lying and shady areas like valley” (Cédric Scherer, Colors and Emotions in Data Visualization)
Color gradients should transition from a bright color (e.g. white) to a dark color (e.g. dark blue) in a consistent way, and they should work in black and white. Readers are also generally better able to distinguish colors on a gradient better if they are encoded through both lightness and two (sometimes three) carefully-selected hues.