Chart design: perception-grounded choices

A chart that produces systematic misreadings of the data is editorially equivalent to a misleading number: it generates incorrect beliefs in the reader. Chart design in the Gauge Intelligence public archive is treated as an epistemic choice, not an aesthetic one.

Cleveland & McGill’s 1984 experiments established a hierarchy of visual elementary tasks by accuracy. The same data, encoded differently, produces judgements that differ by a factor of two in error. The hierarchy provides the empirical foundation for prescriptive chart selection rules.

Every chart type in the archive is assigned to a hierarchy tier; chart types that ask readers to perform low-accuracy perceptual tasks for judgements that matter are restricted or prohibited.

The perceptual accuracy hierarchy

From most accurate to least: (1) position on a common scale; (2) position on non-aligned scales; (3) length; (4) angle and slope; (5) area; (6) colour intensity; (7) hue.

A bar chart encodes values as position on a common scale — tier 1. A donut chart encodes values as angle and arc length — tier 4. For the same proportional comparison, bar beats donut by at least 2× in reader accuracy.

Trellis charts (small multiples) allow tier-1 position comparisons across many panels simultaneously. This maps directly to the four-lens architecture, where comparing operator performance across corridors is the canonical multi-panel task (Cleveland, The Elements of Graphing Data (1985); Cleveland & McGill (1984)).

Archive chart policy

Endorsed. Position-on-common-scale charts: bar, dot plot, connected dot plot. Tier-1 encodings for all primary quantitative comparisons. These are the default for any published figure where the reader is expected to extract a value or rank entities.

Restricted — context only. Line charts for time series where the temporal connection is the primary message. Not for cross-entity comparisons at a single time; for that, a dot plot at the chosen time slice is the correct form.

Restricted — disclosure required. Stacked bar charts for compositions where the top category’s absolute value is not the primary message. Only the bottom category is readable as position on a common scale; values for stacked categories above the bottom require the reader to subtract baselines, which is a tier-2 task. Use is permitted where the total is the primary message and individual categories are secondary.

Prohibited. Donut and pie charts for proportional data where individual categories are being compared. Angle and area are tier-4 and tier-5 encodings; bar charts encode the same information at tier 1 without loss.

Small multiples and the four lenses

The four analytical lenses (operator, destination, corridor, supply chain) map naturally to small-multiples layouts: the same chart repeated across panels for each entity in the lens.

Cleveland’s trellis principle requires that panels in a small-multiples grid share common scales. Shared scales preserve tier-1 comparisons across panels; separate scales degrade them to tier-2.

Gauge Intelligence applies this in cross-corridor and cross-operator views: shared y-axis scale enables honest position comparisons that a separate-scale layout does not. Where the data range across panels is too wide for a single shared scale to be informative, the panels are split into magnitude bands, each with its own internally-shared scale, rather than allowed to drift to per-panel scales.

The shrunk A2F league table on the operators index is the standing live example of the tier-1 encoding in published form. Every operator is read against a single common percentage scale, with the 90% credible interval as the second column on the same scale.

Rendered horizontal-bar plots that encode the same numbers as bar lengths against the same axis will land on operator and corridor reference pages as the chart layer is built out. Until then the ranked table is the published surface, and the encoding is identical (position on a common scale).

Version history

Version 1.0 — April 2026. Initial publication. Endorsed, restricted, and prohibited chart types codified against the Cleveland-McGill hierarchy. Applies to all charts in the public archive and all licensed analytical content.

Sources

Cleveland, W. S. (1985). The Elements of Graphing Data. Wadsworth.

Cleveland, W. S. & McGill, R. (1984). “Graphical Perception: Theory, Experimentation, and Application to the Development of Graphical Methods.” Journal of the American Statistical Association, 79(387), 531–554.