Foundations

The through-line: Bertin (theory, 1967) → Cleveland & McGill (experiment, 1984) → Mackinlay (automation, 1986) → Heer & Bostock (web-scale replication, 2010) → Munzner (synthesis + task framing, 2014). When one citation must carry the encoding ranking, cite Munzner’s marks-and-channels chapter — it’s the validated modern synthesis.

Bertin — the visual variables (Semiology of Graphics, 1967)

Two planar dimensions (position x, y) plus six retinal variables: size, value (lightness), texture, color (hue), orientation, shape. Each variable has “levels of organization”: selective (pick a group at a glance), associative (group across variation), ordered (read ranking without a legend), quantitative (read ratios).

• Only position and size are quantitative. Value is ordered but not quantitative. Hue and shape are selective/associative only — this is the origin of “don’t use hue or rainbows for quantitative data.”
• Bertin split color into hue and value as separate variables; popular lists that say just “color” lose the distinction that matters.
• Derived from cartographic theory, not experiments — Cleveland & McGill is the experimental successor.

Cleveland & McGill — graphical perception (JASA 1984)

The experimentally derived accuracy ranking of elementary perceptual tasks for proportional judgments of quantitative data, most → least accurate (grouped items are ties):

1. Position along a common scale
2. Positions along nonaligned scales
3. Length, direction, angle
4. Area
5. Volume, curvature
6. Shading, color saturation

Corollaries from the paper itself: dot charts beat pies; grouped bars beat stacked (stacked segments off the baseline become nonaligned-position judgments); framed-rectangle charts beat shaded choropleths.

Common misreadings to avoid: hue is not in this ranking at all (it lacks a perceptual ordering for quantity — restatements putting “hue” at the bottom are importing Mackinlay/Munzner); and only the position/length/angle portions were tested experimentally — area/volume/shading placements rest on psychophysical theory (Weber, Stevens).

Heer & Bostock — crowdsourced replication (CHI 2010)

Replicated Cleveland–McGill for web-rendered charts: the ranking held. Additional validated findings worth using directly:

• position > length ≳ angle > area for proportion judgments.
• Rectangular ≈ circular area judgment accuracy; 1:1-aspect rectangles were the worst (viewers proxy area with side length) — so squarified treemaps are not perceptually optimal.
• Charts under ~40px tall degrade comparisons; little accuracy benefit beyond ~80px height on a 0–100 scale.
• Gridlines improve accuracy — keep them ≥ ~8px apart; overly dense gridlines steeply increase error.

Tufte — graphical integrity and economy (Visual Display, 1983)

Verified definitions:

• Data-ink ratio = data-ink ÷ total ink; principles: maximize it, erase non-data-ink, erase redundant data-ink — “within reason“ (Tufte’s own qualifier, usually dropped; this is a redesign heuristic, not an absolutist rule).
• Chartjunk = non-data or redundant ink that tells the viewer nothing new (moiré vibration, heavy grids, the self-promoting “duck”).
• Lie factor = (effect size shown) ÷ (effect size in data); integrity requires ≈ 1 (0.95–1.05).
• Small multiples — “for a wide range of problems… the best design solution.”
• Six principles of graphical integrity: proportional representation; thorough labeling; show data variation, not design variation; deflated/standardized monetary units; dimensions in the graphic ≤ dimensions in the data; don’t quote data out of context.

What later research modifies: Bateman 2010 and Borkin 2013/2015 show relevant embellishment can aid recall without harming simple-chart comprehension (see infographics.md for exact scope and limits). What stands uncontested: the lie factor and proportional integrity. Fair synthesis: minimalism is a strong default for analytical work; decoration is a scoped tool for narrative work; distortion is never licensed.

Shneiderman — the information-seeking mantra (1996)

Verbatim: “Overview first, zoom and filter, then details-on-demand.“ A design sequence for interactive interfaces, not a rule of static composition. The same paper’s taxonomy is usually dropped: 7 data types × 7 tasks — the last three tasks being relate (view relationships), history (undo/replay), and extract (export selections); dashboards routinely forget all three.

Refinements worth knowing: Shneiderman offered it as distilled practice wisdom, not validated law (Craft & Cairns 2005 documented the validation gap). For very large/dense data, a details-first entry can beat overview-first (van den Elzen & van Wijk 2014); Keim’s visual-analytics variant is “analyze first, show the important, zoom/filter, analyze further, details on demand.”

Munzner — nested model and what-why-how (2009; VAD 2014)

Nested model — four levels, each feeding the next: (1) domain situation → (2) data/task abstraction → (3) visual encoding & interaction idiom → (4) algorithm. Errors cascade outward-in, and each level needs its own kind of validation: a lab study can’t rescue a wrong abstraction; benchmarks can’t validate an encoding. Most bad dashboards are level-2 failures — pin the data/task abstraction before debating chart types.

What-why-how:

• What: dataset types (tables, networks/trees, fields, geometry); attribute types (categorical, ordinal, quantitative).
• Why: task = action + target. Actions: analyze (discover/present/enjoy; produce), search (lookup/browse/locate/explore), query (identify/compare/summarize). Targets: trends, outliers, features, distributions, correlations, topology, shape.
• How: encode (marks & channels), manipulate (change/select/navigate), facet (juxtapose/partition/superimpose — small multiples live here), reduce (filter/aggregate/embed).

The modern channel rankings (cite these rather than raw Cleveland–McGill):

• Magnitude channels (ordered data): position on common scale → position on unaligned scale → length → tilt/angle → area → depth → color luminance/saturation → curvature → volume.
• Identity channels (categorical data): spatial region → color hue → motion → shape.

The two-list structure resolves the perennial confusion: hue is a poor magnitude channel but the second-best identity channel.

Mackinlay — effectiveness by data type (ACM TOG 1986)

Extended Cleveland–McGill into separate effectiveness rankings for quantitative, ordinal, and nominal data (the APT system) — the ancestor of automated chart recommendation and the formal grounding for “match the encoding to the attribute type.”