Above about five series, a line chart stops being a chart and turns into a knot. Crossings make this concrete. Nine sector series over twelve months cross each other 31 times, an average of 0.86 crossings for each of the 36 pairs, and the busiest single pair crosses six times. Every crossing is a place where a reader tracking one line can lose it.
Small multiples are the same data drawn as a grid of small panels, one series per panel, all on the same scale. The idea is old. The argument for it is Edward Tufte's, in The Visual Display of Quantitative Information (1983), where the reader compares shapes instead of following lines. Each comparison is between whole shapes, and people compare shapes fast and in parallel in a way they can't follow lines.
Why do small multiples need a shared scale?
The shared scale is what makes it work, and it's also the rule software defaults break most often. Every panel gets the same vertical axis, chosen to cover the whole dataset. For the nine sectors that range is 96.57 to 126.30, a span of 29.73 index points. Now look at what free scaling would do. Health care's own range across the year is 3.57 index points. Give it its own axis and that 3.57-point wiggle fills the same panel height that a 29.73-point move fills elsewhere, an exaggeration of 8.32 times. Scanning the grid, you'd see health care as the most volatile sector in the set when it's the least. Nothing on the page would tell you the axes differed, unless you read nine sets of tick labels.
Free scaling is defensible in one case only: when the panels show different quantities in different units, so a shared scale means nothing. Even then the panels aren't really small multiples. They're a set of separate charts arranged in a grid, and the reader has to be told so. Label each panel's range, or index everything and go back to a shared scale.
Advertisement
What decisions make a small multiples grid work?
Three more decisions decide whether a grid works.
Panel order is an argument, the same way category order on a bar chart is. Sort by the value that matters to the message. For the nine sectors that's full-year return, so the grid runs from communication services at 26.30% down to health care at −1.85%. Alphabetical order is the choice that says nothing matters. Where an intrinsic order exists, it wins, as with geography on a map layout or the stages of a funnel.
Grid shape follows from the count and the aspect ratio of the space. Nine panels go three by three. Eight go four by two on a wide slide and two by four on a page. Panels need to be wide enough that the shape inside them is readable, which usually means fewer, larger panels beat more, smaller ones. If you can't fit them, you're showing too many series.
The grey-ghost technique puts back the context the split takes away. In each panel, draw all nine series in a very light grey and the panel's own series in the accent, so each small chart shows one sector against the whole set. It costs nothing in ink terms, since the grey lines are barely there, and it restores the comparison the grid was in danger of losing.
When do small multiples help, and when don't they?
So when shouldn't you use them? If the message is about the relationship between two particular series, keep them on one chart, because separating them destroys the comparison the message depends on. If there are three series, one chart with direct labels beats three panels. And if each panel would hold four data points, the small multiple is a table wearing a costume.
The nine-sector case shows what the rebuild buys you. On the crowded chart, health care is a line near 100 with eight other lines around it, the least eventful thing on the page. In the grid it's the one panel whose shape runs the other way. The correlation confirms it, at −0.84 against the average of the other eight, where the next lowest is utilities at +0.68. The message that only exists after the rebuild is that eight of these nine sectors are one bet and the ninth is a different one. No amount of decluttering the original would have produced it.