Many points overplot; use transparency or binning

What it is

Overplotting occurs when observations land on the same or neighboring pixels, allowing occlusion to hide count, density, or class composition. Alpha blending can leave a darkness cue where translucent marks overlap. Hexagonal or rectangular binning and two-dimensional density displays instead aggregate local counts or estimated density into cells or contours. The choice depends on whether the task is to retrieve records, detect outliers, compare local density, or inspect class mixing—not on a universal row-count threshold.

Why it happens

Once opaque pixels saturate, one observation and many observations look alike. Lower alpha delays saturation but does not make darkness a reliably linear count scale: background, color compositing, draw order, display, and print all affect the result, while isolated points may become invisible. Binning explicitly summarizes after quantizing position, reducing mark count and stabilizing density reading, but cell width, origin, and shape change boundaries and peaks; kernel density adds a bandwidth decision. Both approaches trade detail for aggregation, and neither has a universal setting.

Studying it

Use distributions with known modes, sparse tails, outliers, and class overlap. Vary mark size, alpha, color, bin origin and width, density bandwidth, canvas size, and output medium. Separate tasks such as record retrieval, local-density comparison, multimodality detection, class-proportion estimation, and outlier discovery. Measure accuracy, miss type, and confidence, then repeat with realistic distributions. For sampling, repeat draws and check whether substantive conclusions change with the random seed.

Where it stops holding

Alpha still saturates in dense regions and can lose gradations on projectors, dark themes, low-contrast displays, or grayscale print. Blending multiple class colors may create colors absent from the legend. Bins are unsuitable for record-level audit and may absorb isolated cases or manufacture grid boundaries. Random samples reduce rendering load but can omit rare groups and outliers. Jitter is appropriate only for overlap caused by discrete or quantized values and deliberately changes displayed position. None of these methods repairs undisclosed missingness, weighting, or filtering.

Applying it

• Preserve a raw-point reference, then compare translucent points, hexbin, and 2D density against the actual task; do not switch on sample size alone.
• Tune alpha for both visible sparse points and unsaturated dense areas. Check light and dark themes, grayscale print, target display size, and a non-color-only density cue.
• Label the aggregate, bin width or bandwidth, and color scale. Move the grid origin or vary smoothing to test sensitivity.
• When record lookup matters, add zoom, filtering, and an accessible data table. If sampling is necessary, disclose population, method, fraction, randomness, and any weighted summary.
• Test predefined peak, sparse-group, and outlier questions. If reasonable parameter choices reverse the answer, expose that sensitivity or change the representation.

Related

• Same group: U2.05.1 Scatterplots suit the relation between two continuous variables · U2.05.3 Correlation is not causation
• Nearby: U2.08.1 Histogram bin width changes the conclusion · U2.14.4 More records, more overplotting, less structure
• Search terms: overplotting · alpha blending · hexbin · 2D density · bandwidth sensitivity