Before the questions, make sure you can: describe colour in terms of luminance, saturation and hue and explain why perceptual colour spaces matter; choose a categorical, sequential, diverging or cyclic colour map for an attribute and explain why rainbow maps mislead; design palettes that work for colour-vision deficiencies and check them; compute a WCAG contrast ratio and apply the thresholds; write a useful alt text; recognize the common misleading techniques and compute their effect; and decide whether a chart (or an AI-generated caption) is honest.
Colour is the channel designers love most and use worst. It is powerful for categories and for overviews, weak for precise values, invisible to some readers, and easy to abuse. This chapter turns colour into a set of decisions you can justify — which map for which attribute type, which contrast, which palette for colour-blind readers — and then widens the lens to honesty: a chart can be technically correct and still mislead through its axis, its window, its scaling or its caption. AI tools reproduce popular defaults, including the dishonest ones; your job is to catch them.
Colour has three channels
Think of colour not as "red" or "blue" but as three channels:
- Luminance (lightness) — a magnitude channel: light vs. dark is read as less vs. more. It is also what remains when a chart is printed in grey or seen by many colour-blind readers.
- Saturation (chroma) — magnitude, but with few distinguishable steps, especially for small marks.
- Hue — an identity channel: good for categories (about 6–9 distinguishable), with no natural order.
RGB and HSL describe how screens make colours, not how people perceive them: two colours with the same HSL lightness can look very different in brightness (yellow vs. blue). Perceptual colour spaces such as CIELAB and HCL are designed so that equal distances look like equal differences; good colour maps are built in them.
Choosing a colour map by attribute type
| Attribute | Map | Example schemes | CityPulse use |
|---|---|---|---|
| Categorical | categorical: distinct hues, similar lightness | tableau10, Okabe–Ito, ColorBrewer Set2 | district, rider type |
| Ordered, one direction | sequential: luminance ramp, optionally with hue | viridis, blues, ColorBrewer YlOrRd | PM2.5, trips |
| Ordered, meaningful midpoint | diverging: two hues meeting at a light neutral midpoint | blueorange, RdBu, PuOr | change vs. last year (midpoint 0), temperature anomaly |
| Cyclic | cyclic: hue wheel that wraps | sinebow, twilight | hour of day, wind direction |
In Altair: alt.Color("pm25:Q", scale=alt.Scale(scheme="viridis")); for diverging data set the midpoint explicitly: alt.Scale(scheme="blueorange", domainMid=0).
Why the rainbow misleads
Rainbow ("jet") maps were defaults for decades (MATLAB until 2014). They are harmful (Borland & Taylor, 2007, "Rainbow color map (still) considered harmful"):
- their lightness is not monotonic — yellow and cyan look brighter than the colours around them, so readers see false boundaries and features that are not in the data, while real gradients in other parts are hidden;
- hue has no natural order — is green more than yellow?
- they fail for colour-blind readers and in greyscale.
viridis (created for matplotlib 2.0 in 2015) is perceptually uniform, ordered by lightness, colour-blind friendly and readable in greyscale. Use it (or ColorBrewer sequential maps) for quantitative data; reserve cyclic rainbows like sinebow for truly cyclic attributes.
Colour-vision deficiency
About 8 % of men and 0.5 % of women of Northern-European ancestry have a colour-vision deficiency, mostly red–green (protan/deutan types). In a lecture of 100 students, expect several. Consequences:
- Avoid red–green as the only distinction (the classic "good/bad" traffic-light palette). Simulated deuteranopia makes standard red (#d62728) and green (#2ca02c) nearly identical (Lab V6.2).
- Use colour-blind-safe palettes (Okabe–Ito, viridis, ColorBrewer "colour-blind safe" sets) and vary lightness as well as hue.
- Never rely on colour alone (WCAG success criterion 1.4.1): add labels, shapes, patterns or direct annotation.
- Check: simulators (Coblis, browser developer tools, Lab V6.2) show your chart as others see it.
Contrast and accessibility
The Web Content Accessibility Guidelines (WCAG 2.1) define a contrast ratio (L1 + 0.05) / (L2 + 0.05) from the relative luminances of two colours (from 1:1 to 21:1):
- normal text: at least 4.5:1; large text (18 pt, or 14 pt bold): 3:1 (criterion 1.4.3);
- graphical objects needed to understand the content (lines, bars, icons): 3:1 against adjacent colours (criterion 1.4.11).
Grey #777777 on white is 4.48:1 — just failing for body text; #767676 reaches 4.54:1. Bright yellow (#ffcc00) on white is only 1.51:1 — invisible as a thin line.
Accessible charts also need: a meaningful title and labelled axes with units; alt text that states the chart type, what is shown and the key message ("Line chart of monthly mean PM2.5 in six districts, 2026. All districts fell; Lucheng fell most, from 52 to 38 µg/m³."); the data table available for screen-reader users; keyboard-accessible interaction; and no information conveyed by colour alone.
Honest charts
Misleading techniques you will meet — often as AI or tool defaults:
| Technique | Effect | Fix |
|---|---|---|
| Truncated bar axis | exaggerates differences (lie factor, Chapter 5) | zero baseline for bars; position marks if you must zoom |
| Dual y axes | the two scales are arbitrary, so any "crossing" or "following" can be manufactured | two aligned panels, or index both series (e.g., 100 = January) |
| Cherry-picked window | a short window reverses or exaggerates the trend | show the long-term context; justify the window |
| Area / 3D / pictograms | area and depth distort magnitudes | 2D, length/position |
| Inconsistent scales in small multiples | panels look alike although magnitudes differ | shared scales (or a clear warning) |
| Missing normalization | totals reflect population or size, not rate | per capita, per km², per station |
| Hidden uncertainty | estimates look exact | error bars, intervals, ranges, sample sizes |
In Lab V6.4, a full year of PM2.5 falls by 23 % (−0.75 µg/m³ per month), but the June–September window rises by 12 %: a chart showing only that window would tell the opposite story.
Tufte's principles remain useful: maximize the data-ink ratio (remove ink that shows no data: heavy gridlines, backgrounds, decorations), avoid chartjunk, and show the data. But clarity, context and annotation are not junk — a reference line or a short label often helps more than minimalism.
Captions are claims. "Air quality improved significantly" must be supported by the data and, where appropriate, by a statistical comparison with normal variation (Chapter 14). AI-generated captions frequently overstate.
Accessibility is increasingly a legal requirement: public-sector websites in many countries must meet WCAG 2.1 AA, and companies face complaints when dashboards cannot be used by colour-blind or blind users. Organizations such as the BBC, the UK Office for National Statistics and Datawrapper publish accessible chart guidelines and tools that generate alt text and colour-blind checks automatically.
Research on chart accessibility studies how blind and low-vision users explore data through sonification, tactile charts and structured screen-reader navigation (e.g., Olli, MIT), and evaluates whether AI-generated chart descriptions are accurate and useful. Early results: descriptions help, but errors in AI descriptions are hard for blind users to detect — verification remains a human responsibility.
Key takeaways
- Colour = luminance (magnitude), saturation (magnitude, few steps), hue (identity). Design in perceptual spaces.
- Categorical → distinct hues; sequential → luminance ramp (viridis); diverging → two hues around a meaningful midpoint (
domainMid); cyclic → wrapping hues. - Rainbow/jet maps create false boundaries and lack order; avoid them for quantities.
- ~8 % of men have colour-vision deficiency: avoid red–green alone, use safe palettes, vary lightness, never colour alone.
- WCAG: contrast 4.5:1 text, 3:1 large text and graphics; alt text, titles, units, data tables.
- Honest charts: zero-based bars, no dual axes, context instead of cherry-picked windows, no area/3D tricks, shared scales, normalization, visible uncertainty — and captions supported by data.
Ready? Close the notes and practise.
30 questions. Predict the output before you check — that is the skill the exam measures.