Definition
Data visualization is the encoding of quantitative or categorical information into visual properties — position, length, angle, area, color, shape — so that the human visual system can do work that reading numbers cannot. Its power comes from perceptual pre-attentive processing: we detect differences in position and length quickly and accurately, without deliberate comparison. That is why the choice of encoding is not a matter of taste. Research by William Cleveland and Robert McGill, published in 1984, ranked encodings by how accurately people judge them, finding position along a common scale most accurate, followed by length, then angle and area, with color saturation among the least accurate. A chart is therefore an argument about what should be easy to compare. Whatever you encode in position, you have asserted is the important comparison; whatever you push into color, you have made harder to read precisely.
Why It Exists
Data visualization exists because quantities beyond a handful exceed working memory. A table of forty numbers can be read but not seen; the same values as a chart reveal trend, outlier and cluster in a second. It also exists because decisions increasingly depend on volumes of data no one can read directly — epidemiological surveillance, financial risk, logistics, climate records, product analytics. The alternative to visualization is not careful reading of raw data; it is summary statistics, which conceal structure. Francis Anscombe's 1973 quartet makes this concrete: four datasets with nearly identical means, variances and regression lines that look entirely different when plotted. Visualization exists to keep the shape of data visible when the summary would flatten it, and to let a reader check the analyst's claim rather than accept it.
Examples
- →John Snow's 1854 map of cholera deaths around the Broad Street pump in London, plotting cases by address to reveal a spatial cluster.
- →Charles Minard's 1869 chart of Napoleon's 1812 Russian campaign, encoding army size, geography, direction and temperature in a single figure.
- →Florence Nightingale's polar-area diagrams of mortality causes in the Crimean War, used to argue for sanitary reform to a non-statistical audience.
- →A modern epidemic curve, where the y-axis scale choice — linear or logarithmic — substantially changes what a general reader concludes.
History
Statistical graphics were largely invented in the late eighteenth and nineteenth centuries. William Playfair introduced the line chart, bar chart and pie chart in his Commercial and Political Atlas of 1786 and later works, giving economic data visual form for the first time. The nineteenth century produced its most celebrated artifacts: Snow's cholera map, Nightingale's diagrams, Minard's campaign flow map. The field then went comparatively quiet until John Tukey's Exploratory Data Analysis in 1977 re-established plotting as a method of inquiry rather than presentation, and Jacques Bertin's Sémiologie graphique of 1967 supplied a systematic theory of visual variables. Edward Tufte's The Visual Display of Quantitative Information, published in 1983, brought design criticism to the subject with arguments about data-ink, chartjunk and graphical integrity, and remains the most widely read entry point into the discipline.
In Modern Design
Contemporary practice divides roughly into exploratory and explanatory work. Exploratory visualization is made by an analyst for themselves, quickly and often ugly, to find what is in the data. Explanatory visualization is made for a reader, carries a claim, and must handle annotation, sequence, accessibility and the reader's likely misreadings. Most published charts fail as explanation rather than as analysis. Current standards emphasize a few practical rules: label directly rather than forcing legend lookups, start bar charts at zero because length is the encoding, do not use a second y-axis to imply a correlation, and never rely on hue alone, both because roughly one in twelve men has a color vision deficiency and because color is a weak quantitative encoding regardless. Interactivity has become routine, which introduces a new failure mode — information hidden behind hover states is invisible in the screenshot that will actually circulate.
Real-World Example
The pandemic years produced a public argument about axis choice that illustrates the stakes. Charts of cumulative cases on a linear scale showed a flat line that suddenly became a wall, which communicated shock but made it hard to see whether growth was slowing. The same series on a logarithmic scale turned exponential growth into a straight line whose slope changed visibly when transmission slowed, which was the decision-relevant signal — but many readers unfamiliar with log axes substantially underestimated the magnitudes involved. Neither chart lied. Each made a different comparison easy and a different one hard, and the choice determined what a non-specialist took away. The episode is a reminder that graphical integrity is not only about avoiding truncated axes and misleading areas; it includes an obligation to consider what an ordinary reader, not an ideal one, will conclude from the encoding you chose.
Key Principles
- →Match the encoding to the judgment: position and length for precise comparison, color and area only for coarse distinctions.
- →Start bars at zero — the length is the quantity, so truncating the axis falsifies it. Line charts of a trending series need not start at zero.
- →Show the data, not the decoration: every visual element should carry information or aid reading.
- →Label directly where possible; a legend forces the reader to hold a mapping in memory.
- →Never encode meaning in hue alone; redundant encoding with position, shape or text keeps the chart readable for everyone.
- →State uncertainty. A point estimate drawn as a confident line implies precision the data may not support.
Why it matters
Charts carry unusual rhetorical authority. They look like measurement rather than argument, so readers apply less scrutiny to them than to the equivalent sentence, which means a misleading chart travels further and meets less resistance than a misleading paragraph. Designers are typically the last people to touch a chart before publication, and often the only ones asking how it will be read rather than whether it is technically correct. That places a real responsibility on the role. It also creates opportunity: the skill of turning a dataset into a defensible, legible claim is scarce, valued across sectors, and largely independent of visual fashion. Learning it well also improves ordinary design judgment, because it forces constant explicitness about what a reader is being asked to compare.
Then vs Now
Then
Charts were drawn by hand, published rarely, and each one represented a substantial investment of effort by someone who had studied the data closely. Playfair, Minard and Nightingale each made a small number of figures intended to carry a specific argument.
Now
Charts are generated automatically in dashboards and spreadsheets at effectively zero cost, frequently by people who have not examined the underlying data. Abundance has made the editorial judgment about what deserves a chart, and which comparison it should privilege, the scarce skill.
Try it yourself
Find a chart in a news article published this week and reverse-engineer it. Write down, in order: what is encoded in position, what in length or area, what in color, and what is only in the text. Then name the single comparison the chart makes easiest, and the comparison it makes hardest — there is always one of each. Next, redraw the same data twice by hand or in a spreadsheet: once to support the article's claim as strongly as honesty permits, and once to support the opposite reading, without changing a single number. Note precisely which levers you used — axis range, aggregation period, sorting, baseline, absolute versus per-capita. That list is the set of decisions you are making every time you make a chart, whether or not you notice them.