Evaluate scatterplot correlation readability with dispersion-based visual features

Evaluate scatterplot revisions against the visual features people use to judge correlation, not just the underlying Pearson r. For example, inspect whether perpendicular spread around the implicit trend, prediction-ellipse area, minor-axis length, and confidence-box width remain clearly separated across versions before accepting a redesign for correlation reading.

Correlation judgment in scatterplots follows a small set of visible point-cloud cues rather than the abstract statistic alone. When a redesign preserves separation in those cues, readers are more likely to distinguish nearby correlation levels as intended.

Mechanism: Readers appear to use visible dispersion around the regression direction as a proxy for association strength, so designs that preserve those cues support more accurate correlation judgments.

Evidence: Across 44 tested visual features, four dispersion-based features predicted judgment correctness better than correlation itself, and models built from those features matched or exceeded existing correlation-based perceptual models for scatterplots (Yang et al., 2019).

Notes: The strongest features all describe dispersion around the regression direction.

• User Goal: Improve or compare scatterplot designs for accurate reading of association strength.
• Task: Judge which of two scatterplots is more correlated, or audit a redesign for that task.
• Data: Bivariate quantitative data rendered as a point cloud.
• Chart Setting: Scatterplots, including side-by-side comparisons, where viewers infer correlation from the plotted points.
• Audience: Designers or reviewers evaluating correlation readability.
• Success Criterion: Readers can discriminate nearby correlation levels more reliably.

Break it when: The chart is not a scatterplot or the task is not judging correlation. Why: The evidence only establishes these cues for correlation discrimination in scatterplots, not for other chart families or other analytic tasks.

Sacrifice: A single-number review based only on Pearson r is no longer enough.
Risk: Checking only one cue can miss how some viewers judge correlation, because the results suggest several visual features may be used.
Mitigation: Inspect the small set of top-performing dispersion cues together instead of relying on one proxy.

Mistake: Accept a scatterplot revision because the data have the intended r values without checking the visible point-cloud features. Why it fails: Judgment correctness tracked dispersion-based features more closely than correlation alone.

Failure Sign: Two versions with different intended correlation readability still look similarly wide or similarly compact around the trend.
Quick Check: Compare candidate versions on perpendicular spread, prediction-ellipse area, minor-axis length, and confidence-box width.
Stronger Test: Model observed judgments with those features and verify that the feature-based fit matches or exceeds the correlation-only baseline.

• Add the four top-performing dispersion cues to the review checklist for correlation-focused scatterplots.
• Recompute or reinspect those cues after every encoding change that alters the point cloud.
• Reject a revision when it improves the numeric correlation contrast but weakens visible separation in those dispersion cues.