How to Read a Graph Critically

The visual impression of a chart forms in under a second, and it is frequently wrong. Force yourself to read four things first: what each axis measures, the units, the range shown, and whether the vertical axis starts at zero. A bar chart with a truncated vertical axis can turn a two per cent difference into a visual doubling, and this is the most common distortion in circulation because it is easy to produce without intending to deceive.

Line charts are as much an argument about which period matters as about the data. A stock, a temperature series, or a crime statistic can trend up over ten years and down over ten months. Ask whether the window shown is the natural one for the question, or whether shifting the start date by a year would change the story. If the start date looks arbitrary, it usually was not.

Raw counts rise with population, so comparing counts across countries, cities, or decades often measures how many people there are rather than the phenomenon itself. Rates per capita fix this but introduce their own question — per capita of whom? A rate per total population and a rate per exposed population can point in opposite directions, particularly in health and safety statistics.

Charts frequently display only the numerator: the number of positive results, incidents, or successes. Without the base — how many attempts, tests, or people were involved — a rising line means nothing definite, because it may simply reflect more measurement. This is why testing counts had to accompany case counts to make either interpretable during any large-scale screening effort.

Two lines on separate vertical scales can be made to appear correlated by adjusting the scales, and this technique produces most of the spurious relationships that circulate online. Stacked area charts hide the individual trend of every band except the bottom one, because each band's position depends on the ones beneath it. Both formats have legitimate uses and both reward extra scepticism.

Even a genuine relationship in a chart leaves three questions open: does A cause B, does B cause A, or does some third factor drive both? Ice cream sales and drowning rise together because both follow summer weather. In observational data the third-variable explanation is usually available, so a chart alone is rarely sufficient support for a causal claim, whatever the caption asserts.

Every visualisation is a selection. Which categories were merged, which outliers were removed, which comparison group was left out? A chart of one product's growth against nothing tells you very little; the same chart against the market average often tells a different story entirely. The missing comparison is the most powerful editorial choice a chart maker has, and it leaves no visible trace on the page.