Showing the raw point distribution is more reliable

What it is

A raw-point distribution display gives each displayable observation a mark, as in strip, jitter, beeswarm, and sina plots. Compared with bins, smoothing, or a few quantiles, point-level views preserve direct evidence of sample size, discrete values, gaps, heaps, and unusual records. They are “more reliable” when that evidence remains perceptible. The title is not a fixed chart ranking: overlap, jitter, sampling, weights, privacy transformations, and screen resolution can distort or defeat a point plot.

Why it happens

Points reduce irreversible compression from observations to summaries and can preserve a route to a record. They are not parameter-free. Random jitter alters a non-data coordinate; beeswarm packing depends on collision rules and point radius; opacity changes the salience of overlap. Fully coincident records can look like one mark, and one mark per weighted record does not make visible count equal weighted population. Reliability comes from disclosing these transformations and preserving provenance, not from drawing points alone.

Studying it

Compare points, histograms, density plots, box plots, and combinations for judgments of sample size, heaping, multimodality, extremes, group location, and distributional uncertainty. Manipulate sample size, radius, overlap, jitter, opacity, ordering, delivered dimensions, and data-generating shape; measure errors, time, confidence, and record lookup. Fix or vary the random seed to determine whether a finding depends on one layout. Separate sample-shape reading from population inference: showing every sampled observation does not automatically communicate sampling uncertainty.

Where it stops holding

When pixels and attention cannot separate marks, bins, density, contours, stratified samples, or multiresolution views may be more faithful; there is no device-independent point-count threshold. Sampling must expose rate, method, and full count while retaining aggregate context. Points for sensitive or rare groups can create re-identification risk and may require aggregation, thresholds, or privacy protection. Censored, missing, and zero values need distinct states. Duplicate values need stacking or counts rather than jitter that pretends continuity. Although grouped under “Histograms and distributions,” this leaf covers point-level evidence that complements a histogram, not a histogram variant.

Applying it

• When marks remain discriminable, layer or pair raw points with summaries or shape displays. Set a clear layer order so summaries do not erase minority or high-consequence records.
• Document jitter axis, amplitude, seed, radius, opacity, sampling, and weights. Tooltips, keyboard focus, and downloaded data should return true values rather than displaced coordinates.
• At high density, progressively switch to a binned or aggregate overview with filters, zoom, or controlled detail back to records; display total count and aggregation scope.
• Give screen readers group, count, range, duplicates, gaps, and anomaly summaries plus point navigation. Validate heaping, mode, and extreme-value tasks at delivered size.

Related

• Same group: U2.08.1 Histogram bin width changes the conclusion · U2.08.2 Box plots hide multimodal structure
• Nearby: U2.05.2 Many points overplot; use transparency or binning · U2.09.4 Sample size does not show in the box; annotate it separately
• Search terms: raw points · strip plot · beeswarm · jitter · overplotting