Dual axis

A second value axis is the standard answer to a real problem: two series that belong in the same story and share no unit. Revenue is in thousands, complaints are a count, and neither one drawn against the other’s scale says anything. So the plot grows a right-hand axis, one series is read against each, and the chart appears to have solved it. What it has actually done is take the single most important thing the reader will do with the picture, compare the two shapes, and hand the person who drew it a knob that decides the answer.

The specimen above never changes a number. Eight months of revenue, eight months of complaints, drawn as the same bars and the same line throughout. All that moves is the domain of the right axis. Set it wide and the complaint line lies low and almost flat under the bars, and the chart says complaints are a rounding error against a business that is growing. Set it tight and the identical eight counts climb most of the plot, steeper than the bars, and the chart says complaints are outrunning revenue. Both pictures are correctly labelled. Both are drawn from the same file. The crossing point where the line meets the bars, which is the thing every reader takes away, exists nowhere in the data: it is a property of the two domains, and whoever chose them chose it.

There are honest uses, and they are narrower than the pattern’s popularity suggests. When the second axis is a fixed transformation of the first, degrees Celsius against degrees Fahrenheit, a count against that same count as a percentage of a known total, there is no free knob at all, because the mapping is arithmetic rather than taste. The same goes for a second axis whose domain the subject matter already fixes, such as a rainfall scale a meteorological convention pins. Where the two units are genuinely unrelated, the usual repairs are to index both series to their own starting value and plot them on one axis in percent, or to stop overlaying them at all and use small multiples, which keeps the comparison honest by refusing to draw it as a single shape. Datawrapper makes the stronger version of the argument, which is that a reader cannot tell a real crossing from a manufactured one, so the chart cannot be trusted even when it is right.

The neighbours divide by which part of the drawing does the lying. A truncated axis lies about magnitude, moving one baseline so a difference looks bigger than it is; a dual axis lies about correlation, handing each series its own free knob so two shapes can be made to agree or disagree at will. Axis is the component both are choices about, the drawn line with its ticks and its labels, and chart is the whole drawing they serve. When the domains are picked to win an argument the numbers do not support, the word stops being a charting term and becomes a dark pattern: the evidence is all present, in small type, at the edge nobody reads.