Tufte Design
Six fundamental principles of analytical design, distilled from Edward Tufte's five books on evidence presentation (1983–2020). The principles are universal — they apply equally to a scatter plot, a Jupyter notebook, a one-pager, a slide deck, a scientific paper.
Grand principle: The principles of analytical design are derived from the principles of analytical thinking. Cognitive tasks are turned into principles of evidence presentation and design. (Beautiful Evidence, p. 137)
Meta-rule: "It is better to violate any principle than to place graceless or inelegant marks on paper." These are guides for judgment, not a checklist.
1. Show comparisons, contrasts, differences
Every well-designed display answers "Compared with what?" — the fundamental analytical question.
How to apply:
- Show enough surrounding data (before/after, neighbors, baseline, control group) that the viewer can judge the effect. A single number out of context says nothing — the same data point can mean opposite things depending on its surroundings.
- Design small multiples — identically-formatted panels indexed by a third variable — when comparisons are the point. They are inevitably comparative, deftly multivariate, shrunken, high-density.
- Use parallelism: identical visual structures for things to be compared make differences pop without re-explanation.
- Use range-frames (axes that extend only to actual min/max, not round numbers) and dot-dash plots (the axis frame doubles as a marginal-distribution display): the data itself participates in the comparison.
Watch-outs:
- Never use pie charts — angles cannot be compared across pies, and two pies side-by-side are the only worse design than one.
- Arbitrary axis ranges, broken scales without warning, and rebased indices invent or hide differences.
- Quoting a number out of context is the most common quiet lie.
See references/vdqi.md, references/envisioning-information.md.
2. Show causality, mechanism, explanation, systematic structure
The reason to examine evidence is usually to understand why, not just what.
How to apply:
- Diagram the causal model behind the data. Make arrows specific: different arrow styles for different relationships (composition, sequence, hypothesis, influence, feedback). Generic arrows under-specify the link.
- Plot cause-and-effect with cause on X, effect on Y — the longer horizontal axis elaborates the causal variable in more detail.
- When time is involved, label the dated events (interventions, regime changes, policy shifts) directly on the graphic — they are part of the causal mechanism, not chart furniture.
- For decisions, present evidence in the form the decision requires: side-by-side options, parallel structure, the comparisons that matter. The Challenger O-ring data presented one way killed seven astronauts; presented another way it would have grounded the launch.
Watch-outs:
- Effects without causes: passive voice in writing, bullet lists in slides, raw correlations without a posited mechanism. "The borders were not hardened" — by whom? Be specific.
- The simple passage of time is not a good explanatory variable. Descriptive chronology is not causal explanation.
See references/visual-explanations.md, references/beautiful-evidence.md.
3. Show multivariate data — more than 1 or 2 variables
The world is multivariate. Flatland is a constraint of the medium, not of the data.
How to apply:
- Escape flatland through layering and separation, transparency, small multiples, micro/macro design — multiple variables coexisting on a single visual field.
- Mobilize every graphical element to do double or triple duty. A range-frame is the min/max display. A dot-dash-plot frame tick is a marginal distribution mark. Data-built grids mark events in the data, not arbitrary intervals.
- Increase data density. The eye easily resolves ~625 points per square inch (100/cm²); most published charts use fewer than 10. The Shrink Principle: most graphics can be made smaller without loss of legibility — use recovered space for additional variables, context, or supporting data.
- Color sparingly and meaningfully. Reserve color for categorical distinctions or ordered intensity ramps. Gray creates order; color creates puzzles because the eye has no natural ordering for hues. Imhof's rules: pure bright colors only in tiny accent areas; large areas in muted, subdued tones. Design all color choices to be safe for color-deficient viewers (~5–10%).
Watch-outs:
- Fake 3D distorts comparison and hides values.
- Information-carrying dimensions must not exceed data dimensions — don't use area for a 1D quantity; don't use volume for a 2D quantity.
See references/vdqi.md, references/envisioning-information.md.
4. Completely integrate words, numbers, images, diagrams
Evidence doesn't care what mode it is in. The reader needs it together.
How to apply:
- Place graphics inline with the prose that discusses them — same typeface, no separator rules, no "Fig. 2" abbreviation. Captions are paragraphs, not labels.
- Label directly on the graphic. Legends force back-and-forth lookup; words placed at the data are read with the data.
- Use sparklines — small, intense, wordlike graphics — embedded inline with words and numbers when the comparison is between many things or over many time steps. Pair sparklines with their most-recent numeric value and range to add precision.
- Tables outperform charts for small datasets and exact values. ~20 numbers or fewer almost always belong in a table. A well-organized supertable beats a hundred little bar charts.
- For complex evidence, interleave words, numbers, images on a single page with a shared visual weight — the way woodblock printing originally united type and image.
Watch-outs:
- Reports where images are piled at the back and tables are exiled to an appendix.
- Slide decks that fragment a single argument across ten slides.
See references/beautiful-evidence.md, references/visual-explanations.md.
5. Thoroughly document the evidence
Undocumented displays are inherently suspect. Documentation is the quality-control mechanism for trust.
How to apply:
- Detailed title that says what the display is about — not just "Figure 2."
- Authors and sponsors named. Disclose conflicts of interest. People do things; agencies and departments do not.
- Data sources cited — where the data came from, when it was collected, what's in it, what's excluded.
- Complete measurement scales — units, range, baseline, linear vs. log, inflation-adjusted vs. nominal.
- Point out relevant issues: missing data, exclusion criteria, what the display can and cannot show.
- For analyses, document the evidence-reduction pipeline — what was filtered, excluded, transformed, and why. This is what separates evidence from rhetoric.
Watch-outs:
- Anonymous corporate reports. Undated charts. Charts without units. "Adjusted" with no adjustment formula.
- Cherry-picking: a report that shows only supportive evidence cannot be trusted — ask what was selected out.
See references/beautiful-evidence.md.
6. Content counts most of all
Analytical presentations ultimately stand or fall depending on the quality, relevance, and integrity of their content.
How to apply:
- The first question is never about color or technology. It is "What are the content-reasoning tasks this display is supposed to help with?" Answer that, and design choices follow.
- Better content is the best design improvement. No tool, no template, no clever encoding will rescue thin or wrong material.
- Lie Factor ≤ 1.05. The visual size of an effect must be directly proportional to the numerical effect:
lie_factor = (size of effect shown) / (size of effect in data). Factors > 1.05 or < 0.95 are substantial distortions.
- Refuse decorative additions that don't reveal content. Chartjunk (moiré, heavy grids, ducks, fake 3D) is content-evasion.
- For serious analytical work, use whichever tool lets you reason most clearly. PowerPoint's bullet-list cognitive style fragments evidence and suppresses sentences. Replace it with concise written reports — sentences, tables, sparklines, images — when the stakes are high.
Watch-outs:
- Designs that prioritize "looking professional" over revealing content.
- Decks with no underlying paper. If you can't write the argument as prose, the argument isn't ready.
See references/beautiful-evidence.md, references/seeing-with-fresh-eyes.md.
Choosing the medium
Pick the format that fits the data volume and the decision:
| Format | Use when |
|---|
| Sentence | 1–2 numbers in running text |
| Table | ≤ ~20 numbers, exact values needed. Order by content or value, not alphabetically |
| Supertable | Large highly-labeled table organized for many local comparisons |
| Sparkline | Wordlike graphic embedded inline; shape + pattern alongside numbers |
| Graphic | Large or complex datasets; comparisons; patterns; mechanisms |
| Map | Spatial location is part of the data |
| Labeled diagram | Claiming causality, mechanism, or system structure |
| Pie chart | Never |
| PowerPoint deck | Never, for serious analytical work — write a report |
Time-series caveat: time-series are best for large datasets with real variability. A simple linear change can be summarized in one or two numbers; don't waste a graphic on it.
Review pass
Run when reviewing any analytical display:
- Comparison. Is the comparison the display invites the one that matters? Are the relevant baselines/controls/neighbors shown?
- Causality. Is the underlying mechanism named, sketched, or visible? If the display claims causation, where is the causal logic?
- Multivariate. Is the display flatlandy when the data is rich? Could a second or third variable be layered in without crowding?
- Integration. Inline with the prose? Direct labels rather than legends? Tables where tables belong? Sparklines where sparklines belong?
- Documentation. Title, author, source, scales, units, caveats — all present? Conflicts disclosed?
- Content. Is the content worth showing? Lie factor ≤ 1.05? Free of chartjunk? Would better content beat any design improvement?
Deep dives
Per-book elaboration in references/:
references/vdqi.md — The Visual Display of Quantitative Information (1983, 2001). Lie factor, data-ink ratio, chartjunk, redesigns (range-frame, dot-dash plot, quartile plot), small multiples, friendly typography, line weight contrast, horizontal proportions.references/envisioning-information.md — Envisioning Information (1990). Escaping flatland, micro/macro readings, layering and separation, color (four uses, Imhof rules), narratives of space and time.references/visual-explanations.md — Visual Explanations (1997). Smallest effective difference, parallelism, visual confections, evidence for decisions (Challenger, cholera), disinformation design, the PGP method.references/beautiful-evidence.md — Beautiful Evidence (2006). The six principles (full elaboration), mapped pictures, sparklines, links and causal arrows, words/numbers/images together, corruption patterns, the cognitive style of PowerPoint.references/seeing-with-fresh-eyes.md — Seeing with Fresh Eyes (2020). See / reason / act, content-responsive typography, graphical sentences, annotations, instructions at point of need, smarter meetings.
The revelation of the complex. (VDQI, p. 191)
Design is choice. Principles generate design options and guide choices among them — they don't decide. If a display is more truthful, more revealing, or more elegant for breaking a rule, break it — but know which one you broke and why.