Eight panels of the same chart: what changes for a reader when each panel fits its own vertical range instead of one shared range?
answer
- same distance, same number, every panel
- the largest group sets a shared range
- shapes are compared before ticks are read
- fitted to itself means comparable to nothing
- name the comparison, then the setting
basics
~20 sA range per panel makes each panel legible and the set incomparable: every panel is rescaled to its own rows, so a small group's bump draws like a large group's. A shared range keeps the comparison but flattens small panels.
solid answer
~50 sWith one shared range, a given distance up the page means the same number in every panel, so the panels can be compared - and the largest group sets the limits, which can flatten small panels into nearly straight lines. With a range per panel, each is fitted to its own rows, so every panel shows its detail and nothing about apparent height carries between them. The tick values are still printed, but that is a weak defence: readers compare shapes before they read numbers. Decide by the reader's question - comparing groups means one shared range; studying the shape inside each group means a range per panel, said loudly on the page. Quote no default: where the panel split is a clause of one specification with one scale, the range is shared by construction; where the panels are separate charts in a grid, each fits itself unless you make them agree.
go deeper
Recall that panels are only comparable when the same distance up the page means the same number in each, and that fitting every panel to its own rows removes that guarantee.
Explain both costs in the same breath: a shared range is set by the largest group and flattens the small panels, while a per-panel range makes every group look equally eventful.
Show judgment by starting from the comparison the reader will make, and by offering a third option - change the unit plotted, or split the set by magnitude - rather than picking a side quietly.
The standing call is what a chart must state about itself before anyone acts on it, and whether the ranges are shared belongs on that list alongside the origin and what was aggregated.
## The two arrangements The **axis range** is the lowest and highest value an axis actually shows, which may be narrower or wider than the data. Across a set of panels there are two arrangements of it: - **one shared range** - every panel is drawn against the same limits, chosen to hold all the rows in the whole set; - **a range per panel** - each panel's limits are fitted to that panel's own rows. Nothing else about the chart changes. The marks, the grouping column and the panels are identical. What changes is what a reader is able to conclude, and that is a very large change for a setting that looks cosmetic. ## Why the shared range is the comparable one A comparison across panels only means something if a given distance up the page stands for the same number in every panel. A shared range guarantees exactly that, and a per-panel range removes exactly that. The failure is easy to state and hard to notice. Suppose one group's values run from zero to fifty and another's from zero to five thousand, and each panel is fitted to its own rows. Both panels now fill their frame. A rise that looks the same size in the two panels differs by a factor of a hundred, and there is nothing in the *picture* to say so. The usual objection is that the numbers are printed on the ticks, so a careful reader can recover the comparison. That is true and it is not a defence: - readers compare shapes first and read tick values afterwards, if at all; - the entire value of a panel set is that it can be **scanned**, and scanning is precisely the mode in which nobody reads tick values; - the more panels there are, the less likely any individual axis is read at all. ## Why a shared range is nonetheless sometimes the wrong call A shared range is set by the largest group in the set. If the groups differ widely in magnitude, the small panels become flat lines pressed against the baseline - and the detail that justified drawing a panel for that group has been thrown away. If the reader's question is *what does this group's own behaviour look like*, a shared range answers it with an empty-looking panel. | | one shared range | a range per panel | |---|---|---| | preserves | comparisons of level and of size between panels | detail inside every panel, whatever its magnitude | | destroys | visible detail in the small panels | any reading of height or slope across panels | | right when | the reader ranks or compares the groups | the reader studies each group on its own | | how it misleads | small groups look as though nothing happened | every group looks equally eventful | ## Where the default comes from, and why you cannot quote one This is the point at which authors state a default that is only true of the tool they know: - Where the panel split is **a clause of one chart specification**, the scale is defined once for the whole chart, so every panel inherits the same limits by construction and freeing them is the explicit setting. - Where the panels are **separate charts laid into a grid**, each one is fitted to its own rows unless you ask them to share or set the limits yourself, so the comparable arrangement is the one that takes work. The empirical test costs a minute and settles it: put two groups of wildly different magnitude into the chart and look at whether the small panel flattens. If it does, the range is shared. ## Getting more of both When the two goals genuinely conflict, there are better answers than picking a side silently: 1. **Keep the shared range and change the unit.** Each group as a change from its own starting point, or as a share of its own total, is comparable across panels in a unit that is genuinely comparable - and the small group stops being invisible. 2. **Split the set in two.** Group the panels by magnitude into two panel sets, each internally sharing a range, and say on the page which set is which. 3. **Keep the per-panel range and make it loud.** State each panel's maximum in its label, or say plainly in a caption that the panels are not on a common scale. Never leave the tick values as the only warning. 4. **Share one axis only.** Where the comparison a reader makes runs along one axis - usually the one carrying time - share that axis and let the other be fitted per panel. ## What an interviewer is listening for That you name the comparison before you name the setting. A candidate who says *free the ranges, it looks better* has described an appearance; a candidate who says *the reader is comparing group sizes here, so the ranges must be shared, and the small panels will be flat, so I would plot change from baseline instead* has described a decision and its cost.
- A reader compares two panels and concludes one group is twice the other. What do you check first?Whether the two panels are drawn against the same limits, and then whether they start from the same baseline. If either differs, the conclusion came from the layout rather than the data, and the reader had no way to know. Only after both agree is the comparison of heights a statement about the numbers.
- When is a range per panel the right call rather than a concession?When the reader's question is the shape inside each group - a seasonal pattern, a step change, a spike - and not the level between groups. It is right only when the page says the ranges differ, in each panel's label or in a caption, because otherwise the reader will compare heights anyway and will be wrong.
- The groups span three orders of magnitude and the reader needs both detail and comparison. What do you do?Change what is plotted rather than how it is scaled. Each group as a change from its own starting point, or as a share of its own total, keeps one shared range while making every panel legible, because the unit itself is now comparable. Failing that, split the panels into two sets by magnitude and label the split.
Five photographs of five buildings, each framed so the building fills the shot. Every picture is sharp, and nothing in the set tells you which building is tallest. Fitting each panel to its own rows does exactly that to a chart.
saying these in an interview costs you the question
- Says free ranges look better, without noticing the panels stopped being comparable.
- Assumes every panelling interface shares one range unless told otherwise.
- Believes the printed tick values are warning enough that the panels differ.
- Treats a shared range as always correct, even when small panels flatten to nothing.
- Thinks changing the range changes the data rather than what a reader can conclude.