How the Cleveland-McGill Hierarchy Ranks Visual Encodings for Accuracy

A designer building a public health dashboard might reach for a bubble chart or a color-coded map, arguing that visual variety holds a reader’s attention and makes the data memorable in a crowded information environment. A statistician looking at the exact same data will insist on a plain bar chart or a scatter plot, warning that human eyes are physically incapable of accurately comparing the area of two circles or the intensity of two shades of blue.[3]

This tension between engagement and precision is not merely a matter of aesthetic preference or design philosophy. It is governed by strict biological limits in human visual processing. In 1984, statisticians William Cleveland and Robert McGill set out to quantify exactly how badly the human eye fails at different geometric tasks, arguing that "the subject of graphical methods for data analysis and for data presentation needs a scientific foundation." Their work shifted visualization from an art to a measurable science.[1][6]

To build that empirical foundation, Cleveland and McGill ran controlled laboratory experiments with 54 subjects, asking them to estimate the percentage differences between various geometric shapes. They tested position along a common scale, length, angle, area, volume, and color saturation. By recording the exact margin of error for every single estimation, they created a mathematical baseline for how the human brain decodes visual information. The researchers presented subjects with pairs of marked elements within a chart and asked them to judge what percentage the smaller element was of the larger one, isolating the cognitive act of magnitude estimation.[1][6]

The results yielded a strict, ranked hierarchy of graphical perception that remains the gold standard in the field. At the very top—the most accurate visual encoding by a wide margin—is position along a common scale, the mechanism used in standard bar charts and scatter plots. When data is encoded this way, human error rates drop to their absolute minimum, allowing readers to extract point estimates with near-perfect reliability. This is because the visual cortex is highly evolved to judge the alignment of objects on a shared horizon.[1][3][5]

Below position, accuracy degrades rapidly and measurably. Length (used in unaligned bar charts) and angle (used in pie charts) occupy the middle tier of the hierarchy. Subjects attempting to decode angles consistently made larger estimation errors than those comparing aligned bars, explaining why data analysts routinely advise against using pie charts for complex datasets. The brain struggles to translate rotational geometry into linear quantities. The error rates for angle judgments were nearly double those of position judgments, a penalty that compounds as the number of slices increases.[1][3]

At the bottom of the hierarchy sit area, volume, and color saturation. When subjects were asked to compare the size of two circles or the darkness of two shaded regions, their error rates spiked dramatically. The human visual cortex simply does not process two-dimensional area or color intensity with mathematical linearity. A circle that is twice the mathematical area of another will rarely be perceived as exactly twice as large by a human reader. This perceptual distortion makes bubble charts and heat maps highly vulnerable to misinterpretation, especially when the underlying values are close in magnitude.[1][5]

For over two decades, the Cleveland-McGill hierarchy remained a laboratory-bound theory, accepted by statisticians but rarely tested in the wild. That changed in 2010, when computer scientists Jeffrey Heer and Michael Bostock at the University of Washington Interactive Data Lab rebuilt the 1984 experiments for the internet age. They wanted to know if the foundational rules of visual encoding survived outside the pristine conditions of a mid-century research laboratory, where modern readers consume data on varying screens. Their study became a watershed moment for modern data visualization.[2]

Heer and Bostock deployed the perceptual tests on Amazon’s Mechanical Turk platform, recruiting 50 participants per task at a cost of $0.05 per response. Their goal was to see if the pristine laboratory results held up when subjects were distracted, using uncalibrated monitors, and operating in uncontrolled environments. By crowdsourcing the graphical perception experiments, they introduced the exact kind of noise and friction that real-world dashboards face every day. The sample size and demographic diversity far exceeded the original 1984 cohort.[2]

The crowdsourced replication was definitive. The 2010 data perfectly mirrored the 1984 hierarchy. The proportional difference in log-error rates between judging a bar chart (position) and a bubble chart (area) remained mathematically constant, proving that the hierarchy is a hardwired cognitive constraint, not an artifact of laboratory conditions. Whether a subject is sitting in a quiet room with paper or clicking through a web browser in a noisy cafe, their brain fails at computing area just as reliably. The replication cemented the hierarchy as a universal law of human perception.[2][7]

Subsequent research has expanded this framework into modern cognitive psychology. A 2021 synthesis by Steven Franconeri and colleagues in Psychological Science in the Public Interest mapped these error rates directly to the mechanics of working memory and visual attention. Their evidence confirms that complex encodings force the brain to perform sequential, error-prone estimations rather than instant parallel processing. When a chart relies on color gradients, the reader must constantly look back and forth between the legend and the data, exhausting working memory. This cognitive load is what drives the spike in estimation errors.[4]

The hierarchy is now being tested beyond the screen entirely. A 2023 study presented at the IEEE VIS conference by Emily Wall and colleagues at Emory University extended perceptual accuracy tests into tactile encodings, measuring how well users can decode data through touch. Even across different sensory modalities, the fundamental rule holds: simple, aligned spatial comparisons outperform complex, multi-dimensional ones. The biological preference for a common scale appears to transcend vision. This finding has profound implications for designing accessible data interfaces for visually impaired users.

However, the evidence pack surrounding the hierarchy carries a crucial limitation: it optimizes entirely for point-estimate accuracy. If the goal of a chart is to allow a reader to extract an exact number, position along a common scale is mandatory. But if the goal is to reveal a broad geographic cluster or to draw a reader into a dataset they would otherwise ignore, strict adherence to the hierarchy can produce sterile, unengaging graphics. A choropleth map uses color—a low-accuracy encoding—but it excels at showing regional patterns that a bar chart would obscure.[3][4]

The data dictates a clear threshold for practitioners. When a medical dosage, a financial forecast, or a structural engineering tolerance is on the line, the visual encoding must sit at the top of the Cleveland-McGill ranking. Where precision is paramount, the biology of the human eye leaves no room for aesthetic compromise. The hierarchy proves that in the realm of data communication, what looks the most interesting is rarely what communicates the most truth. Designers must choose whether their primary objective is to entertain the eye or to inform the mind.[4][5][6]