Every decoration has two costs. The first is ink, which the census measures. The second is attention, which is what actually matters and which the census only approximates. There's a useful third question too. Can the people you're presenting to see the element at all? Geometry gives a clean answer to that one.
Can the back of the room see it?
Take a 16:9 slide, 13.333 inches wide, projected to a screen three metres across and rendered at 1920 pixels. That's a magnification of 8.86 times, so one pixel on the slide becomes 1.5625 millimetres on the screen. Whether an eye can resolve a mark depends on the visual angle it subtends. For a high-contrast line the practical acuity limit is about one arcminute.
| Mark | On screen | At 2 m | At 5 m | At 8 m | At 12 m |
|---|---|---|---|---|---|
| 1 px hairline: gridline, plot border | 1.56 mm | 2.686′ | 1.074′ | 0.671′ | 0.448′ |
| 2 px context line | 3.12 mm | 1.34′ | |||
| 3 px accent line | 4.69 mm | 2.01′ | |||
| 8 px marker | 12.50 mm | 5.37′ | |||
| 34 px bar | 53.12 mm | 22.83′ |
Advertisement
Read the top row. A one-pixel gridline is comfortably visible from two metres, marginal at five, and below the acuity limit at eight. Picture a room where the back row is eight metres from a three-metre screen. For those people, every hairline gridline on every chart is grey haze that nobody can resolve into a line. It costs ink and it costs contrast against the marks that matter. It gives back nothing. That's the strongest single argument for the delete order, and it comes from geometry. Taste has nothing to do with it.
Type behaves the same way. At eight metres, a 9 point label subtends 12.1 arcminutes. You can read it with effort, but not under time pressure. 12 point subtends 16.1 and reads comfortably, and 24 point subtends 32.2 and reads instantly. The type scale belongs to the mechanics branch, but here's what it means for charts. Say the axis labels are 9 point because that was the only way to fit twelve categories. You haven't solved a fitting problem. You've moved it onto the audience.
Which decorations should you delete?
Now the individual decorations, ranked by how often deleting them is correct.
Drop shadows and bevels are always wrong on a data mark. They add a second, softer copy of every edge. That's the very edge the eye uses to judge length, and length is the encoding a bar chart runs on. On the Nilaya chart they were 47.7% of all non-data ink by themselves.
Gradient fills inside a bar are wrong for the same reason, plus one more. A gradient makes the bar's colour vary along its length, so two bars of the same category read as two different colours at their tips. Put a gradient behind the plot and it also breaks the contrast guarantee, because the accent now stands on a background whose luminance changes across the slide.
Three-dimensional effects on a two-dimensional quantity are the worst case. The reader can't tell which part of the mark is the measurement. A 3-D pie makes the front slices larger in projected area than back slices of identical value. A 3-D column chart puts the top face of each column above the gridline it should meet.
Borders are cheap in ink and expensive in structure. A box around a plot competes with the axis for the role of frame, and a box around a legend advertises the legend. White space does the same job.
Are any gridlines worth keeping?
Gridlines are the one case where keeping some can be right. Keep them horizontal only if values are read horizontally. Keep three or four, never eleven. Put them behind the marks, in the lightest neutral that still resolves at your viewing distance. Better still, replace the whole grid with a single reference line at the number that matters, which is a target, a prior year or a zero.