MisleadingCharts
Back to the gallery

The machinery deaths that grew by moving in from the edge

Showing the misleading chart

Be the first to star this exhibit

A safety-committee slide we drew puts HSE’s count of workers killed in Great Britain, by kind of accident, into two treemaps: 2024/25 beside 2025/26. Every area is exact and every tile carries its count. The machinery tile goes from a thin strip on the right edge to a solid block inside the chart, and the slide calls it a rise. Machinery killed 10 workers in both years. The layout redrew the tile, 60% wider and 36% shorter; the biggest actual change, struck by a moving vehicle at 19 to 24, is hard to see.

01The claim

Deaths from contact with moving machinery rose this year while falls from height held steady, so the committee should switch its toolbox talks from working at height to machine guarding.

02The trick

Nothing on the slide is wrong, tile by tile. Each treemap gives every kind of accident an area exactly proportional to its deaths, and each tile carries its count. The problem is the comparison the slide asks for. Two treemaps side by side are two separate layouts: each year’s tiles are fed in largest first and the squarified algorithm packs them into rows that keep each tile as close to square as it can, starting from scratch every time. A tile’s shape and position therefore depend on every other tile and on the proportions of the box. Between the two years, struck by a moving vehicle overtook struck by an object and trapped by something collapsing grew from 15 to 18, so the rows broke in different places, and the machinery tile came out 60% wider and 36% shorter and moved in from the right-hand edge to sit inside the right half of the chart. (In the slide’s 440 × 412 layout, a strip 73 wide and 194 tall became a block 116 by 124.) Its count was 10 both years; its area grew 1.6%, only because the year’s total fell from 128 to 126. A thin strip along an edge looks small and a squarish block inside the chart looks big, so the eye reads the reshaping as growth. The labels say 10 and 10, in large type; the slide’s author read the picture rather than the labels, which is exactly what a picture invites. The layout is also touchy about the box: draw both treemaps 40 layout pixels shorter and the 2025/26 machinery tile is a strip on the right edge again, with not one number changed. Meanwhile the real changes are hard to see, because a tile only shows its change if you compare two areas of different shapes in different places. Falls from height dropped from 36 to 31 and its tile came out slightly wider. Struck by a moving vehicle rose from 19 to 24, the biggest increase, but nothing on the slide draws attention to it.

03The fix

The redraw puts all fourteen kinds of accident on one zero-based axis, with a hollow dot for 2024/25, a filled dot for 2025/26 and the change printed beside each row. Machinery is 10 and 10: two dots on top of each other. The changes the treemaps hid line up in the same column: vehicles +5, falls −5 (still the most common kind by a distance), trapped +3, electricity −2. Grey ticks for the three years before add the context the slide left out, which is how much these counts move on their own. HSE’s notes warn that small groupings “are susceptible to considerable year-on-year variation”. Over five years machinery deaths ranged from 8 to 15 and falls from 27 to 49, so a single change of five is weak evidence of anything, and the five-year totals are the steadier guide: falls 184 of 651, struck by a vehicle 110, struck by an object 108, machinery 52. These are counts, not rates, which is reasonable for adjacent years in one country, where employment barely moves, but not for comparing industries of different sizes. If a treemap is the house style and has to be repeated year to year, freeze the layout instead of recomputing it: keep the same order and the same rows every year (d3’s treemapResquarify reuses the previous layout’s rows when values change, so tiles grow and shrink in place instead of jumping), and print the change on each tile. Better still, give comparisons across years to a chart that shares an axis.