There are three palette families, and the data chooses among them. Your preference doesn't come into it. Getting it wrong is the second commonest colour error after using too many hues. It does more damage, though, because it makes the reader see structure that isn't there.
What are the three palette families?
A categorical palette encodes unordered categories: channels, regions, product lines, cohorts. Its hues must be distinguishable from one another and must imply no order, because any perceived order is a claim the data doesn't make. Cap it at five. If the message is about one category and not all of them, prefer the accent-plus-neutral formula.
A sequential palette encodes an ordered quantity running from a meaningful low to a meaningful high, such as a count, a rate or an intensity. It runs light to dark in a single hue, or across two neighbouring hues, and the darkness carries the magnitude. It has no midpoint. The error is using one where the data has a natural midpoint.
A diverging palette encodes an ordered quantity with a meaningful centre: growth against zero, performance against a target, a variance against budget. It runs from one hue through a light neutral to a second hue, and the light neutral must fall at the centre the data means.
Advertisement
Worked example: growth across twelve territories
Fairmount Retail's twelve territories give a clean test case. Year-on-year growth runs from Central at −11.0% to Pacific Northwest at +11.2%, a range of 22.2 points. Five territories are above zero and seven below, and the mean is −1.125%.
The data is ordered, so a categorical palette is out. It also has a meaningful centre at zero. A territory that grew and a territory that shrank differ in kind, and not only in degree. So a sequential ramp is out too. A sequential scale maps low to light and high to dark, and the zero crossing falls 49.55% of the way up the range with no visual mark on it at all. Look at that chart and you see a smooth gradient. To find out which territories are actually shrinking, you have to read the labels. That's the very work the colour was supposed to do.
The diverging scale is correct, and it has one parameter that decides whether it tells the truth. The centre goes at zero. The limits go at plus and minus 11.2, symmetric around the centre, so that a territory 5 points up and a territory 5 points down get equal ink in opposite directions.
| Territory | Growth | Normalised to ±11.2 |
|---|---|---|
| Pacific NW | +11.2% | +1.000 |
| Mountain | +7.9% | +0.705 |
| Southwest | +4.1% | +0.366 |
| Great Lakes | +1.3% | +0.116 |
| Mid-Atlantic | +0.4% | +0.036 |
| New England | −0.8% | −0.071 |
| Southeast | −2.6% | −0.232 |
| Plains | −3.5% | −0.313 |
| Gulf | −5.2% | −0.464 |
| Ohio Valley | −6.9% | −0.616 |
| Upper Midwest | −8.4% | −0.750 |
| Central | −11.0% | −0.982 |
What can go wrong with a diverging scale?
Now anchor the same scale at the mean of −1.125% instead. Spreadsheet software will do that by default if you let it. New England, at −0.8%, gets coloured as growth even though it shrank. The colour mislabels one territory out of twelve, and no reader will catch it. Catching it means reading the number, and the colour was meant to save them that job.
Classing is the other decision. Five equal-width classes over the full range put three territories in the bottom class, three in the second, three in the middle, one in the fourth and two in the top. That's a reasonable spread. Quantile classing would force equal counts and so hide the fact that the top class contains a real outlier. For a growth rate, equal intervals are the safe default, because the reader assumes equal steps of colour mean equal steps of quantity.
How do you pick a palette family?
The most useful habit here costs nothing. Before you choose a palette, write down which of the three families the data belongs to. Ordered or unordered? If ordered, centred or not? Two words, and they settle it.