Data visualization and infographics: misleading charts, chartjunk, and honest design

Intuition Beginner

Data visualization turns numbers into pictures. A good chart reveals patterns; a bad one hides them or invents false ones. Edward Tufte gave us tools to tell the difference: maximize the ink that shows data, minimize decoration he called "chartjunk," and check the "lie factor" — the size of the effect drawn divided by the size of the effect in the data.

Common traps: a y-axis that does not start at zero exaggerates differences; 3D pie charts distort angles; dual axes imply false correlation; area symbols sized by radius instead of area mislead; cherry-picked time windows reverse the trend. Alberto Cairo adds: show context, show uncertainty, and never let design choices do the arguing the data cannot support.

Visual Beginner

Technique	How it misleads
Truncated y-axis	Exaggerates small differences
3D pie chart	Distorts angle proportions
Dual y-axes	Implies false correlation
Area by radius	Quadruples instead of doubling
Cherry-picked range	Trend reverses over full data

These traps work because the eye reads the picture before the numbers.

Worked example Beginner

A news graphic shows two bars: unemployment last month 6.0%, this month 6.1%. The y-axis runs from 5.9 to 6.1, so the second bar is twice as tall as the first. The headline: "Unemployment spikes."

The data changed by 0.1 percentage points (about 1.7%). The visual changed by 100%. The lie factor — visual change divided by data change — is roughly 60. Starting the axis at zero would show two nearly equal bars. The chart is accurate in its labels but misleading in its impression. An honest design starts at zero, marks the break, or reports the 0.1-point figure in the headline.

Check your understanding Beginner

Formal definition Intermediate+

Data-ink ratio (Tufte). Data-ink is the non-erasable ink used to present data. The ratio is (data-ink) / (total ink). Maximize it: erase non-data ink and redundant data-ink wherever possible.

Lie Factor (Tufte). For a quantity changing from value v1 to v2, rendered by a graphic element whose measure (length, area) changes from m1 to m2:

LF = (percentage change in the graphic) / (percentage change in the data) = (|m2 - m1| / m1) / (|v2 - v1| / v1).

LF = 1 means the picture is honest; values far from 1 signal distortion in either direction.

Chartjunk = decorative graphics that convey no information: 3D extrusion, moiré vibration, gratuitous icons, heavy gridlines, gradient fills.

Small multiples = arrays of similar graphs sharing the same scale, varying one dimension, enabling direct comparison (Tufte, Envisioning Information, 1990).

Sparklines = word-sized graphics embedded in running text, giving a data-dense time series without breaking the line of reading.

Graphical integrity principles (Tufte). Show data variation, not design variation; label explicitly and in detail; reveal data at several levels of detail; keep the visual measure consistent with the quantity (area for area, length for length).

Cairo's truthfulness criteria (The Truthful Art, 2016). A visualization should be: (1) honest — no distortion; (2) functional — fit to purpose; (3) complete — reveal context and uncertainty; (4) insightful — reveal patterns; (5) illuminating — engage aesthetics in service of understanding.

Key result: Cleveland-McGill graphical perception Intermediate+

Cleveland and McGill (1984, "Graphical Perception") ranked elementary perceptual tasks by accuracy, from most to least accurate: position along a common scale; length; direction and angle; area; volume; curvature; shading and color saturation. This ranking is the empirical foundation for the chart-type recommendations: bar charts (position) outperform pie charts (angle), which outperform area bubbles.

Heer and Bostock (2010) replicated the ordering via crowdsourcing (Amazon Mechanical Turk) with larger samples, confirming the result. This places visualization critique on an experimental basis rather than aesthetic opinion.

For dual axes and truncated ranges, the danger is that the eye treats the visual slope as the effect size. Two series on independent axes invite spurious correlation; a single axis with a generous baseline anchors magnitude correctly. When the data range is narrow relative to its baseline (global temperature anomalies of ±1°C on a 288K baseline), starting at zero can erase the signal — so the rule is not "always zero" but "make the variation interpretable and disclose the baseline."

Bergstrom and West (Calling Bullshit, 2020) codify detection heuristics: check the axes, check the baseline, check whether proportional symbols scale by the right dimension, check the time window, check the denominator (rate vs. count), and check what is omitted. See 29.01.03 for statistical reasoning and 35.02.04 for epidemiology and confidence intervals.

Exercises Intermediate+

Advanced results Master

D3.js (Bostock, Ogievetsky, and Heer, 2011), Tableau, Power BI, Flourish, and Datawrapper let designers bind data to visual channels with interactivity — hover, zoom, filter, linked brushing. Interactivity lets the reader audit the chart: query points, rescale axes, switch from count to rate. This is honest visualization's strongest tool, provided the default view is itself honest. See 33.07.* for web technologies.

Generative AI changes the picture. ChatGPT, Claude, and Copilot produce charts from natural-language prompts. Risks: (1) hallucinated datasets — the model invents numbers; (2) inconsistent axis behavior across regenerations; (3) defaults that truncate axes or pick misleading chart types; (4) plausible but fabricated provenance ("source: WHO, 2023" with no such report). Always re-derive the chart from primary data. See 20.02.06 for AI ethics and 35.08.03 for precision medicine AI and healthcare visualization pitfalls.

Interactive uncertainty: fan charts, hypothetical outcome plots (Hullman, Resnick, and Adar, 2015), and Bayes-factor displays make uncertainty legible. Static error bars are routinely misread as confidence about the mean rather than about the spread. Animation (Hans Rosling's Gapminder) reveals a third dimension but can also manipulate temporal pacing to dramatize. See 29.01.03 for statistical reasoning and p-values, and 35.02.04 for epidemiology and confidence intervals.

Reproducible pipelines — Observable notebooks, Plotly, vegalite — encode the data-to-mark mapping as code, exposing every design decision as an auditable line. Reviewable pipelines are the master-level discipline.

Connections Master

Color theory and perception. Itten's and Albers's color interactions (see 34.03.02) translate to charts: sequential palettes for ordinal data, diverging palettes for deviations from a midpoint, never rainbow for ordinal data. Rainbow palettes create perceptual discontinuities and fail for color-blind viewers. Use viridis, cividis, or ColorBrewer palettes. Gestalt principles (proximity, similarity, closure, connectedness — see 29.03.*) explain why small multiples work and why dual-axis charts mislead: the eye groups by proximity regardless of axis.

Accessibility. Eight percent of men have some color-vision deficiency. Design with redundant encoding (color plus shape plus label), test with simulators, and provide alt text for screen readers. See 30.04.* for the sociology of disability and 34.03.* for design ethics.

Famous misleading charts. Fox News repeatedly used truncated y-axes on employment and welfare graphics. Climate-change-denier charts cherry-pick time windows or start axes at glaciation maxima. COVID dashboards conflated cases with tests, or used linear and log scales inconsistently. The Iraq WMD slides (Powell at the UN, 2003) used ambiguous imagery as if it were evidence. See 35.02.03 for viral pathogenesis and COVID data visualization, 27.07.* for climate change and the hockey stick, and 35.03.* for chronic disease and cancer statistics.

Data journalism. The NYT Upshot, FiveThirtyEight, ProPublica, and The Pudding set the standard for honest explanatory visualization. See 36.02.02 for data journalism.

Historical and philosophical context Master

William Playfair invented the bar chart, line chart, and pie chart in the late 18th century (The Commercial and Political Atlas, 1786). His bar chart of Scottish exports was the first — a deliberate choice to represent discrete quantities by length rather than area. Playfair's pie chart was later shown by Cleveland and McGill to be perceptually inferior, which is why Tufte and Few counsel against it.

Charles Joseph Minard's 1869 map of Napoleon's Russian campaign is the canonical masterpiece: a flow map combining geography, army size (line width), direction, temperature, and date in a single graphic. Tufte called it "the best statistical graphic ever drawn."

John Snow's 1854 London cholera map used spatial clustering to identify the Broad Street pump — a visualization that was simultaneously an epidemiological argument. Florence Nightingale's coxcomb diagrams showed that preventable disease, not combat, killed most Crimean soldiers.

Tufte's Visual Display (1983) launched the modern discipline by codifying data-ink and the lie factor against 1970s-80s chartjunk: 3D bar charts, moiré shading, clip-art icons. The graphical integrity movement is a reaction to desktop-publishing tools that made decoration free.

Every chart is a rhetorical act: it selects, frames, and emphasizes. Cairo's The Truthful Art frames honesty not as the absence of choice but as disclosure — show the baseline, show the uncertainty, name the denominator. Bergstrom and West's Calling Bullshit (2020) makes the same point: misleading charts flourish when readers cannot read the chart's grammar.

Bibliography Master

1. Tufte, E. R. The Visual Display of Quantitative Information (2nd ed., Graphics Press, 2001). Data-ink, chartjunk, lie factor, graphical integrity.

2. Tufte, E. R. Envisioning Information (Graphics Press, 1990). Small multiples, micro/macro readings.

3. Tufte, E. R. Beautiful Evidence (Graphics Press, 2006). Sparklines, truthfulness principles.

4. Cairo, A. The Truthful Art: Data, Charts, and Maps for Communication (New Riders, 2016). Honest visualization practice.

5. Cairo, A. How Charts Lie: Getting Smarter about Visual Information (Norton, 2019). Popular guide to misleading charts.

6. Few, S. Now You See It: Simple Visualization Techniques for Quantitative Analysis (Analytics Press, 2009). Visual perception and chart design.

7. Bergstrom, C. T. and West, J. D. Calling Bullshit: The Art of Skepticism in a Data-Driven World (Random House, 2020). Detecting misleading data visualizations.

8. Cleveland, W. S. and McGill, R. "Graphical Perception: Theory, Experimentation, and Application to the Development of Graphical Methods." Journal of the American Statistical Association 79 (1984) 531-554.

9. Heer, J. and Bostock, M. "Crowdsourcing Graphical Perception: Using Mechanical Turk to Assess Visualization Design." CHI 2010.

10. Bostock, M., Ogievetsky, V. and Heer, J. "D3: Data-Driven Documents." IEEE TVCG 17 (2011) 2301-2309.

11. Hullman, J., Resnick, P. and Adar, E. "Hypothetical Outcome Plots Outperform Error Bars and Violin Plots for Inferences about Reliability of Variable Ordering." PLOS ONE (2015).