The three colours you cannot spend twice

Colour that already means something

This audience reads several palettes fluently, without a legend, because they are encoded in the standards they work to.

Palette	Means	Where it comes from
Green / yellow / red	Normal, moderate, severe	MUAC tape banding, triage
IPC phases 1–5	Minimal, stressed, crisis, emergency, catastrophe	IPC colour standard, fixed
JMP ladder	Safely managed → surface water	WHO/UNICEF joint monitoring
Blue	Water	WASH cluster convention

The IPC palette is not a suggestion. Phase 3 is a specific orange and a map that uses a different one for Phase 3 will be read as a different phase by someone who looks at IPC maps weekly. Use the published values or do not use phase colours.

The MUAC banding is why the signal palette is reserved on this platform. Green, amber and red are --color-signal-normal, -moderate and -severe, and they are for data state — indicator status, severity, quality flags — and nothing else.

What repurposing costs

# Three districts, three signal colours. Nothing here is a severity.
COLOURS = {"Nord-Ouest": "#2f8f5b", "Centre": "#d9a02b", "Sud-Est": "#c0442f"}

# A palette chosen because it is on the brand board, not because it means this.

A reader who sees red on Sud-Est reads “severe” before they read the axis. If Sud-Est happens to be the worst district the chart got lucky and taught a bad habit; if it is the best, the chart has said the opposite of its own data.

The cost is cumulative and it lands on the next chart. Once a reader has seen the signal band used decoratively, they stop reading it as a measurement — and the chart where it is a measurement has lost the shorthand it depended on.

So the rule is not aesthetic. Reserve the signal band for severity, quality flags and indicator status. Everything else gets the neutral palette.

INK, SPRUCE, SLATE = "#16211E", "#2F5D50", "#5A6B66"
ax.barh(labels, values, color=[SPRUCE if e else SLATE for e in emphasis])

scale_fill_manual(values = c("TRUE" = "#2F5D50", "FALSE" = "#5A6B66"))

Two colours is usually the whole palette a programme chart needs: one for the series, one for the thing you are pointing at. Every figure on this platform uses exactly that, and none has needed a third.

Colour as encoding, and what it can carry

Three uses, and only one of them is about which colour is prettier.

Categorical. Distinct hues, at most about six before a reader stops distinguishing them. Above six, use small multiples or direct labels instead of a legend.

Sequential. One hue, light to dark, for an ordered quantity. Coverage, rates, counts.

Diverging. Two hues from a meaningful midpoint — change from baseline, above and below a target. The midpoint has to be real. A diverging scale centred on the mean of the data invents a midpoint and colours half the districts red for being below average, which every dataset guarantees.

import matplotlib.pyplot as plt

fig, ax = plt.subplots()
ax.barh(districts, change, color=["#c0442f" if c < 0 else "#2f8f5b" for c in change])
ax.axvline(0, color="#16211E", linewidth=1)     # the midpoint, drawn

# If the midpoint is not drawn, the reader guesses where it is.

Here the signal colours are legitimate, because the encoding is a state: worse or better than the previous round, with zero as a real midpoint drawn on the chart.

Colour is the weakest channel; use position first

The ranking of visual encodings by how accurately they are read is not a matter of taste.

Rank	Encoding	Use for
1	Position on a common scale	The main comparison
2	Length	Bars
3	Angle, slope	Lines over time
4	Area	Rarely; a bubble is read as its radius
5	Colour intensity	An additional dimension, never the main one
6	Colour hue	Category labels

If the finding is only visible in the colour, the chart has encoded it in the weakest available channel. A choropleth is exactly this trap and the next lesson is about it.

Redundant encoding is not redundant. Marking the emphasised bars by colour and by a caption that names them means the chart still works in greyscale, at low resolution, and for a reader who cannot distinguish the two hues — which is lesson 5.

Nine ways sector charts get colour wrong

Each of these is common and each has a one-line fix.

Mistake	Fix
Signal colours on non-severity categories	Neutral palette
A rainbow scale for an ordered quantity	Single-hue sequential
IPC-like oranges that are not IPC values	Use the published values, or do not imply phases
Red for “female”, blue for “male”	Neutral pair; sex is not a severity
Diverging scale centred on the data mean	Centre on a real midpoint, and draw it
A legend for two categories	Label them directly on the chart
Colour carrying the finding	Move the finding to position or length
More than six categorical hues	Small multiples
Colour and only colour distinguishing two lines	Add a direct label at the line’s end

The last one is the most common in this sector, because a two-line chart looks finished. Print it in greyscale and it becomes one line and a mystery.

Report it whole

Figure conventions used in this report

  Charts use a neutral two-colour palette: one tone for the series, a darker
  tone for values the caption calls out.

  Green, amber and red are reserved for severity and data-quality status --
  MUAC banding, indicator status, completeness flags -- and are not used for
  geographic or demographic categories anywhere in this report.

  IPC phase maps use the published IPC colour values without modification.

  Where a chart uses a diverging scale, the midpoint is a stated reference
  (the previous round, or the national target) and is drawn on the chart.

Four lines in the methods annex, and they answer the question a reviewer asks about every map in the pack. A report whose colour conventions are stated is one whose charts can be trusted to mean the same thing twice.

What comes next

Colour intensity is the weakest encoding, and a choropleth puts the entire finding into it. The next lesson compares one map of twelve communes against twelve small charts, and finds the map cannot show the thing the last lesson established.