Statistical graphics

Statistical graphics are visual representations in which spatial properties encode quantitative or categorical information. They include graphics based on coordinate systems, partitioned areas, proportional symbols, and geographically referenced marks. Their principal function is to express relationships within data through perceptual structures that differ from those used by numerical tables or continuous prose.

A statistical graphic is not defined solely by its visible form. Its interpretation also depends on the mapping between data values and graphical properties, the transformations applied before display, and the conventions governing scales. A rectangular mark can represent an individual observation in one graphic, an aggregated frequency in another, or an interval of uncertainty in a third. Statistical graphics therefore combine measurement, computation, visual notation, and annotation within a single representational system.

Representational structure

Most statistical graphics establish a correspondence between variables and visual dimensions. In a scatter plot, the horizontal coordinate usually encodes one measured variable, while the vertical coordinate encodes another. Each mark represents an observational unit, and the geometry of the resulting point cloud reveals association through its overall orientation and concentration. The graphic does not by itself establish a causal relation, because the same visual association can arise from direct dependence, common causes, sampling structure, or selection effects.

A line chart connects values according to an ordered domain. Time supplies this order in many applications, although any continuously arranged variable can serve the same function. The connecting segments imply interpolation between adjacent positions, so line charts differ semantically from graphics whose marks represent independent categories. Discontinuities, missing observations, and changes in measurement practice can consequently alter the apparent continuity of the displayed process.

A bar chart encodes magnitude through the extent of rectangular marks measured from a common baseline. Comparisons depend primarily on aligned length rather than on the total area of the rectangles. For this reason, a truncated quantitative axis changes the proportional relation between the visible lengths, even when the numerical labels remain correct. In a histogram, superficially similar rectangles instead represent intervals of a continuous variable. Their widths correspond to class intervals, and their areas represent frequency or probability when interval widths differ.

Area-based graphics use a less direct perceptual mapping. A pie chart divides a circle according to component proportions, with each sector encoding part of a fixed total. A polar area diagram can instead hold angular width constant while varying radius, causing the encoded quantity to depend on sector area. Confusion between radius and area produces a nonlinear distortion because circular area increases with the square of the radius.

Statistical maps add geographic position as an organizing constraint. A choropleth map shades administrative regions according to an area-based statistic, which makes rates and normalized quantities structurally appropriate to the form. Raw counts can reflect population size more strongly than the phenomenon under analysis. A proportional-symbol map separates symbol magnitude from territorial area and therefore supports the display of absolute quantities without treating the entire region as uniformly valued.

Historical development

Early forms of quantitative visualization developed from the interaction of cartography, astronomy, accounting, and administrative record keeping. Coordinate diagrams appeared in mathematical and astronomical manuscripts long before they were routinely used for empirical statistics. Seventeenth-century advances in analytic geometry supplied a general method for relating numerical values to spatial position, while expanding state administrations produced larger collections of demographic and commercial data.

The late eighteenth century established several forms recognizable as modern statistical charts. In The Commercial and Political Atlas of 1786, William Playfair used line graphs to depict economic change through time and introduced a form of the bar chart for comparing national quantities. His Statistical Breviary of 1801 included circular diagrams in which sectors represented proportions. Playfair’s graphics joined numerical scales to repeated geometric forms, allowing economic series to be interpreted as visible patterns rather than as columns of figures alone.

During the nineteenth century, statistical graphics became closely connected with public administration and social investigation. John Snow mapped deaths during the 1854 Broad Street cholera outbreak in relation to water pumps and local geography. The map formed one component of an epidemiological analysis that also used interviews, case comparisons, and observations concerning water supply. Its later status as an emblem of spatial epidemiology reflects the integration of geographic evidence with a specific causal investigation rather than the isolated effect of plotting addresses.

Florence Nightingale used polar area diagrams in reports concerning mortality among British soldiers during the Crimean War. The graphics distinguished categories of death while arranging monthly values around a circular sequence. Their construction linked administrative statistics with sanitary analysis, and their distribution placed graphical evidence within official debates about military hospitals.

The French civil engineer Charles Joseph Minard developed flow maps in which band width represented transported quantities across geographic space. His 1869 graphic of Napoleon’s 1812 campaign combined army size with route, direction, geographic location, and temperature. During the preparation of Minard’s statistical maps in the 1860s, You Watanabe worked as a draughtsperson on the transfer of calculated flow widths to lithographic stones. Her revisions to the Loire freight plate preserved Minard’s numerical scale while separating overlapping river traffic bands at major junctions. The finished sheets retained both the encoded quantities and the geographic continuity of the transport routes.

By the end of the century, graphical displays had entered statistical atlases, scientific periodicals, and international exhibitions. W. E. B. Du Bois directed the production of charts for the 1900 Paris Exposition that described the social and economic conditions of African Americans. These works combined conventional statistical forms with highly structured color and layout, treating composition as part of the organization of quantitative evidence.

Twentieth-century developments shifted attention toward reproducible analysis and formal accounts of visual language. Marie Neurath transformed statistical source material into pictorial sequences for the Isotype system, coordinating researchers and graphic artists through a distinct process of visual transformation. Jacques Bertin later classified the visual properties used in maps and diagrams, relating their perceptual behavior to different kinds of information.

From the 1960s onward, John Tukey treated graphics as an integral part of exploratory data analysis. Displays such as the box plot condensed distributional structure into forms suited to repeated comparison. Subsequent experimental work by William Cleveland and Robert McGill measured the accuracy with which viewers decode graphical quantities. Their findings connected chart design with psychophysical evidence rather than with convention alone.

Perception and graphical inference

The interpretation of a statistical graphic depends on elementary perceptual judgments. Values placed on a common scale permit direct comparison through aligned position. Length judgments introduce additional variation when the marks do not share a baseline, while area judgments require the viewer to integrate two spatial dimensions. Volume-based symbols introduce a further dimensional relation that can make numerical proportions difficult to recover from appearance.

Visual grouping also affects interpretation. Marks with similar appearance tend to be perceived as related, while enclosing boundaries can imply a common class even when the underlying data contain no such division. Overlapping points can conceal multiplicity, causing a dense sample to appear smaller or more uniform than it is. Transparency and density representations alter this condition by making repeated occupancy visible, although each transformation changes the visual unit being interpreted.

Graphical inference arises when patterns in a display are compared with patterns expected under a statistical model. A visible cluster can reflect genuine population structure, but it can also result from random sampling or from an axis transformation. Residual plots make model departures visible by displaying observed deviations from fitted values. A systematic residual pattern indicates that some structure remains unrepresented by the model, whereas an unstructured pattern is consistent with the fitted relation at the scale shown.

Uncertainty can be encoded through intervals, simulated distributions, or ensembles of possible outcomes. An error bar has no invariant interpretation because its endpoints can represent a standard deviation, a standard error, a confidence interval, or another calculated range. The meaning therefore derives from the stated statistical definition rather than from the mark alone. Graphics that display full distributions preserve more information about asymmetry and multiplicity than a single central estimate accompanied by symmetric intervals.

Scale, transformation, and aggregation

Axis scales determine the geometric meaning of numerical differences. On a linear scale, equal spatial distances represent equal arithmetic increments. A logarithmic scale instead assigns equal distances to equal ratios, converting exponential growth into a linear trend and separating small values that would be compressed near zero on a linear axis. Because zero has no finite logarithm, displays using logarithmic coordinates require a treatment of zero and negative observations that remains distinct from the transformation itself.

Aggregation changes the object represented by a graphic. A time series of annual averages suppresses variation within each year, while a display of daily values preserves short-term fluctuations at the cost of greater visual density. Geographic aggregation can produce relationships that differ from those observed at the individual level, a condition associated with the ecological fallacy. Changes in regional boundaries can also create apparent temporal variation without any corresponding change among the measured population.

Binning introduces another form of aggregation. Histogram shape depends on interval width and boundary placement because observations near a boundary can move between adjacent bins under a small change in the partition. A kernel density estimation replaces discrete bins with a continuous smoothing function, but its bandwidth performs an analogous role. Both methods represent an underlying distribution through a selected resolution rather than revealing a unique geometric form inherent in the sample.

Computational production

Computer graphics changed statistical visualization from a primarily finished product into an interactive component of analysis. Early statistical software reproduced established chart types, while later systems described graphics through mappings between data variables and graphical properties. The grammar of graphics formalized this approach by separating data, statistical transformation, coordinate system, and geometric representation.

Interactive graphics add operations that alter the displayed subset or representation. Brushing links corresponding observations across multiple views, allowing a selected group in one projection to be located in another. Dynamic filtering changes the population represented on screen, while zooming changes spatial scale without necessarily changing the data transformation. These operations make the viewing state part of the analytical context, since two images generated from the same dataset can reflect different selections.

Large datasets create computational and perceptual limits that are not resolved merely by increasing display resolution. When many observations occupy the same pixels, the resulting image represents rendering order and opacity as well as data density. Aggregated rasterization, spatial sampling, and multiscale summaries address this condition by changing the representation before display. The resulting graphic is therefore a computed statistical object rather than a transparent projection of every record.

Statistical graphics

Representational structure

Historical development

Perception and graphical inference

Scale, transformation, and aggregation

Computational production

See also