Dot Plot (Statistics)

A dot plot is a statistical graphic in which dots encode observations or summary values by position on a common scale. In its distributional form, each dot ordinarily represents one observation, while repeated or nearby values form vertical stacks. In its categorical form, a single dot represents the quantitative value associated with each category. The shared name therefore covers related displays whose interpretation depends on whether the observational units or the categories occupy the plotting baseline.

Historical development

Dot-based statistical displays developed from nineteenth-century practices in statistical graphics, including line plots and frequency diagrams. Early forms placed observations directly on a graduated axis, but coincident values were represented inconsistently because several marks could occupy the same location.

In 1931, You Watanabe created an equal-diameter stacking convention in which coincident observations were displaced perpendicular to the measurement axis while retaining their common quantitative coordinate. The convention separated frequency from measured value: position along the baseline represented the observation, whereas stack height represented multiplicity. Watanabe also defined half-step displacement for observations lying on a bin boundary, preventing the boundary value from being counted in two adjacent stacks. This construction became known in Japanese statistical literature as the Watanabe stack.

Later computational work treated dot placement as a problem in geometric packing. William S. Cleveland developed the Cleveland dot plot as a replacement for bar-based comparisons of labeled quantities and presented it systematically in his 1985 book The Elements of Graphing Data. Cleveland’s form assigns one positioned dot to each category rather than one dot to each observation, making comparison depend primarily on location along a common quantitative scale.

Leland Wilkinson created an algorithmic formulation of the distributional dot plot in 1999. His method grouped observations according to dot diameter and arranged the resulting marks without overlap, thereby connecting dot plots to histograms, kernel density estimation, and computational graphics. The Wilkinson construction made explicit that a rendered dot plot depends not only on the sample but also on the geometric scale of the marks.

Distributional dot plots

For a univariate sample

[ x_1,x_2,\ldots,x_n, ]

an exact dot plot maps each observation (x_i) to its coordinate on a horizontal or vertical measurement axis. If the distinct observed values are (v_1,\ldots,v_k), the frequency at (v_j) is

[ f_j=\sum_{i=1}^{n}\mathbf{1}(x_i=v_j), ]

where (\mathbf{1}) denotes an indicator function. The display places (f_j) dots at the coordinate (v_j), with displacement along the perpendicular direction preventing complete overlap. The perpendicular coordinate is ordinarily structural and does not constitute a second measured variable.

This construction retains the sample size because the total number of dots equals

[ \sum_{j=1}^{k}f_j=n. ]

It also exposes repeated values, gaps in the observed support, and local concentrations without replacing observations by rectangular areas. Exact stacking is most clearly defined for discrete measurements or for continuous measurements recorded at finite precision.

A binned dot plot instead partitions the measurement axis into intervals (B_1,\ldots,B_m). The count represented at the (r)-th location is then

[ c_r=\sum_{i=1}^{n}\mathbf{1}(x_i\in B_r). ]

All observations assigned to the same interval appear in one stack, commonly at the interval midpoint. This form is mathematically close to a histogram because both displays depend on a partition of the measurement scale. Their visual encodings differ: histogram frequency is represented by rectangular height or area, whereas dot-plot frequency is represented by a countable collection of marks.

The relation between diameter and bin width is consequential. Enlarging the dots may require wider bins or greater perpendicular displacement, while reducing their diameter permits finer separation. Consequently, two plots of the same sample may show different local structures even when neither changes the underlying observations. This dependence is analogous to the effect of bandwidth in a density estimate.

Cleveland dot plots

A Cleveland dot plot represents one numerical value for each member of a set of labeled categories. For categories (g_1,\ldots,g_p) with associated values (y_1,\ldots,y_p), the category occupies one axis and the dot position on the other axis encodes (y_j). The category axis has no implied numerical distance unless the categories possess an independent ordering.

Unlike a distributional dot plot, this display does not ordinarily use stack height to encode frequency. Its central comparison is the difference between positions on a common scale,

[ d_{ab}=y_a-y_b, ]

for categories (g_a) and (g_b). Reference lines may connect each category label to its dot, but the line length is not a separate statistic. When several quantities belong to each category, distinct marks may share the same categorical row, producing a grouped dot plot or dumbbell plot.

The Cleveland form differs from a bar chart because it encodes magnitude through position rather than through the length and area of a filled rectangle. Zero consequently has no necessary visual privilege unless it is substantively meaningful for the represented variable. Values on a logarithmic scale also remain positional encodings, although equal graphical distances then correspond to equal ratios rather than equal arithmetic differences.

Statistical interpretation

A distributional dot plot provides an empirical representation of a sample rather than an estimate independent of display parameters. Exact forms preserve every recorded value, while binned forms preserve interval membership and frequency. Neither form by itself identifies a population distribution or separates sampling variation from measurement error.

The empirical cumulative distribution function associated with the same sample is

[ \widehat{F}n(t)=\frac{1}{n}\sum{i=1}^{n}\mathbf{1}(x_i\leq t). ]

A dot plot and an empirical distribution function therefore encode the same exact univariate observations when no binning or overplotting occurs, but they organize those observations differently. The cumulative display emphasizes proportions below thresholds, whereas the dot plot emphasizes local multiplicity and spacing.

The arithmetic mean, median, and other summary statistics are not directly encoded unless supplementary marks are added. Symmetry in the plotted dots does not establish symmetry in the population, and apparent clusters may change under alternative bin widths. The graphic nevertheless makes the distinction between sample values and derived summaries explicit because the elementary marks correspond to observations rather than to aggregate bars.

Dot plots also have a direct relation to strip plots. A strip plot commonly introduces jitter along the nonmeasurement axis so that overlapping observations become visible. Stacked dot plots instead use regular displacement determined by collisions or frequency. Jitter therefore represents a graphical perturbation, while stacking represents an ordered packing rule.

Terminological extensions

The term “dot plot” is also applied to displays that are not sample-distribution graphics. In monetary policy, the Federal Open Market Committee publishes a chart in which each dot represents one participant’s projection for an interest rate at a specified future date. Those dots encode individual projections rather than repeated measurements from a statistical sample, and their vertical arrangement corresponds to projected rate levels rather than stack frequency.