Colour scale choice directly shapes the conclusion

What it is

A heatmap color scale comprises its numeric domain, transformation, clipping, palette type, reference point, and missing-value encoding. These choices determine which differences receive more visible contrast, which values share a bin, and how readers interpret positive versus negative or high versus low. Visual conclusions about hotspots, boundaries, and anomalies can therefore change under defensible scales. The scale is not decoration or a license to maximize drama; data semantics and tasks should determine it, and readers need the mapping and range.

Why it happens

Linear, logarithmic, or other transformations first alter data intervals; a palette then maps transformed values to lightness, chroma, and hue. A diverging scale assigns opposite directions around a reference, a sequential scale expresses one-way order, and binned or quantile scales compress continuous differences into ranks. A wide domain can suppress central variation, clipping gives several extremes the same endpoint color, and automatic per-panel scaling makes identical colors mean different values. Readers compare rendered color differences and may not reconstruct these operations.

Studying it

Prespecify read-off, ranking, anomaly, sign, or block-detection tasks and compare semantically defensible linear, transformed, clipped, and palette mappings on the same data. Record accuracy, time, confidence, and conclusion reversals; use simulated data to test detection and false alarms for known effects and null structure. For multiple charts, test shared and independent domains. Run scale-sensitivity analysis for important conclusions: if a result appears under only one defensible setting, report that setting and uncertainty rather than selecting the supportive image.

Where it stops holding

A logarithmic scale cannot directly include zero or negative values, and alternative transforms also change distance meaning. Quantile colors primarily encode rank, not magnitude difference. Diverging scales require a meaningful reference; a mean is not inherently “normal.” Clipping can protect central discrimination but must mark saturation and expose original values. Absolute cross-panel comparison generally requires a shared domain, transform, and reference. Independently scaled panels can compare shape only when cross-panel color-magnitude comparison is explicitly ruled out. Perceptual uniformity still varies with display and viewing conditions.

Applying it

• Choose a sequential or diverging scale from data semantics, explicitly defining domain, transform, reference, bins, clipping, and missing state.
• Put units, meaningful ticks, endpoints, and saturation markers in a visible legend; preserve unmapped values in details and exports.
• Declare shared versus independent scaling across heatmaps, and make rescaling after interactive filters visible, lockable, or reversible.
• Recheck the main conclusion under several reasonable scales before release; when it is unstable, show the sensitivity, add a numeric view, or weaken the claim.

Related

• Same group: U2.10.1 Heatmaps suit dense two-dimensional values · U2.10.2 Row and column ordering decides whether structure is visible
• Adjacent: U4.02.1 Sequential scales must increase perceptually monotonically · U4.03.1 Diverging scales suit data with a meaningful midpoint
• Search terms: color scale · scale sensitivity · diverging midpoint · quantile color map