Data visualisation

01Start from the question, not the chart type

A chart is an argument about which comparison matters. Comparing magnitudes across categories suggests a bar chart; change over an interval suggests a line; composition of a whole suggests a stacked bar with a small number of parts; relationships between two variables suggest a scatter plot; geographic variation may need a map, though a map often answers a different question than the one being asked.

The frequent failure is choosing a form first. A map of raw counts, for instance, mostly shows where people live, which is rarely the finding. If the question is about intensity rather than location, a rate on a bar chart may be both more accurate and more legible.

02Axes and the honesty of the baseline

For bar charts, where length encodes the value, the value axis starts at zero; a truncated baseline exaggerates differences by construction. For line charts, where the encoding is the slope of change, a non-zero baseline is acceptable and often necessary, but the range must be labelled clearly so the reader can see the scale.

Two further decisions carry the same weight. A logarithmic scale is legitimate for series spanning orders of magnitude and must be announced, because most readers will assume linearity. Dual axes should generally be avoided: they let the designer manufacture a visual correlation by choosing two scales.

03Colour with meaning and restraint

Colour should encode something. Sequential shades of one hue suit ordered quantities; two hues diverging from a neutral midpoint suit deviation from a reference; distinct hues suit unordered categories, and beyond about six categories they stop being distinguishable.

Colour must never be the only channel. Roughly one in twelve men has some form of colour vision deficiency, so categories need direct labels, patterns or ordering as well as hue, and text on coloured backgrounds needs a contrast ratio adequate for reading.

04Classification in maps

On a choropleth map the choice of class boundaries changes the picture more than the data does. Equal intervals, quantiles, natural breaks and a manual scheme built around a meaningful threshold will each produce a different map from identical values.

Because the choice is consequential it must be stated: the method, the number of classes, and the treatment of areas with no data, which should be visually distinct from areas with a value of zero. Normalising by population or another appropriate denominator is usually required before mapping at all.

05Uncertainty in a graphic

A clean line implies a precision that estimates do not possess. Where figures are sample-based, the graphic should show intervals, shaded bands or explicit notes. Where a series has been revised, the vintage belongs in the caption; where values are provisional, they should be marked as such rather than presented identically to final figures.

06Accessibility as a requirement

A graphic that only works as an image excludes readers using screen readers, readers on small screens and readers with the images disabled. The baseline is a text alternative that states the finding rather than describing decoration, a caption with source and period, and the underlying values available as a table.

Interactive graphics need keyboard operation, visible focus, and a static fallback. Interactivity should also not be the only route to the main point: the headline finding must be legible before anyone interacts with anything.

07Annotation is where the argument lives

A chart without annotation asks the reader to find the finding. Direct labelling of the series, a note on the one point that carries the story, a marked reference line where a threshold or a target exists, and a shaded period where a definitional break or an interruption in collection occurred: each of these moves the interpretation from the caption into the graphic where the reader is looking.

Annotation also prevents a specific dishonesty. A gap in a series drawn as a continuous line asserts values that were never collected, and the honest treatment is to break the line and label the gap. The same applies to a period affected by a change in recording practice: marking it lets a reader see the discontinuity that the shape of the line would otherwise hide.

08One crowded chart or several small ones

When a comparison involves more than about five series, a single chart becomes a tangle in which no series can be traced and the reader falls back on the legend. Repeating a small chart once per category, on identical axes, usually communicates more: the shared scale makes the comparison exact, the shapes are legible individually, and an outlier is visible as a shape that differs rather than as a colour to be hunted for.

The cost is space and the risk is inconsistency. Every panel must share the same axis range, the same units and the same period, because a grid of panels with individually fitted scales invites a comparison that the scales contradict. Where one panel needs a different range, it belongs outside the grid with its own labelling.

09What every published graphic needs

Title stating the finding, axis labels with units, source with publication date, period covered, a note on method where a choice was made, and an indication of missing or suppressed values. A graphic without these is an illustration, not evidence, and cannot be checked by a reader.

10Common distortions to check before publication

Truncated bar baselines; area used to encode a single value, which exaggerates differences because readers compare area rather than radius; three-dimensional effects that displace the visual comparison; inconsistent bin widths; time axes with uneven spacing; and cropping a period so that a trend appears to start where the story wants it to start.