Graphical Summaries

Graphical summaries are visual displays — histograms, boxplots, scatterplots, bar charts, and more — that condense a dataset into a picture revealing its shape, center, spread, and unusual features. Key result in words: a well-chosen graph communicates the structure of data faster and more honestly than a table of raw numbers. Where it applies: descriptive statistics, exploratory data analysis, science reporting, journalism, business dashboards. Who/when: developed from William Playfair (1780s) through John Tukey's Exploratory Data Analysis (1977).

A graphical summary is any diagram that represents the distribution or relationships within a dataset so that key features — center, spread, shape, outliers, and associations — become visible at a glance. They matter because the human eye detects patterns (clusters, gaps, trends, skew) far more rapidly than it parses lists of figures, and because a poorly summarized dataset can hide critical information. Each graph type is matched to a data type: bar charts and pie charts for categorical data, histograms and boxplots for a single quantitative variable, and scatterplots for two quantitative variables. A surprising fact: four datasets with identical means, variances, and correlations can look utterly different when plotted — this is Anscombe's quartet, a classic warning that numerical summaries alone can deceive.

The Core Insight

A dataset is a long list of numbers or labels that overwhelms working memory. A graphical summary maps quantities onto visual attributes — position, length, area, angle — that the visual system processes pre-attentively. The art lies in choosing a graph whose visual encoding faithfully preserves the feature you care about, while never distorting it.

Formal Statement

Types of Data

Graphical summaries depend on the measurement scale of the variable being displayed:

Categorical (qualitative) — values are labels in distinct groups (e.g., eye colour, blood type). Sub-types: nominal (no order) and ordinal (ordered, e.g., grades A–F).
Quantitative (numerical) — values are measured amounts. Sub-types: discrete (counts) and continuous (measurements on a scale).

The data type dictates which graphs are valid.

Frequency, Relative Frequency, and Density

For a categorical variable or a binned quantitative variable, define for each class $i$ :

f_{i} = count of observations in class i

⟳simply tally how many data points fall in each category or bin.

r_{i} = \frac{f _{i}}{n}, n = i \sum f_{i}

⟳divide each count by the total number of observations

n

; the

r_{i}

sum to 1.

For a histogram with bin width $w_{i}$ , the density (used when bins have unequal widths) is:

d_{i} = \frac{r _{i}}{w _{i}} = \frac{f _{i}}{n w _{i}}

⟳so that

area $= d_{i} \times w_{i} = r_{i}$ equals relative frequency — the bar's area, not height, represents proportion.

The Five-Number Summary

A boxplot is built from five order statistics of a sorted dataset $x_{(1)} \leq x_{(2)} \leq \dots \leq x_{(n)}$ :

{min, Q_{1}, median (Q_{2}), Q_{3}, max}

⟳the median splits the data in half;

Q_{1}

and

Q_{3}

are the medians of the lower and upper halves respectively.

The interquartile range measures spread of the middle 50%:

IQR = Q_{3} - Q_{1}

⟳subtract the first quartile from the third; it is resistant to outliers.

The 1.5 × IQR Outlier Rule

An observation $x$ is flagged as a potential outlier in a boxplot if it lies beyond the fences:

x < Q_{1} - 1.5 \cdot IQR or x > Q_{3} + 1.5 \cdot IQR

⟳Tukey chose the multiplier 1.5 empirically; for a normal distribution roughly 0.7% of data fall outside these fences, making true outliers conspicuous.

Why It Holds / Derivation

Why area encodes proportion in a histogram

We justify equation [3], that histogram bars should be drawn so area equals relative frequency.

We want the total area of all bars to equal 1, so that the histogram approximates a probability distribution.
The area of bar $i$ is height $\times$ width $= d_{i} \cdot w_{i}$ .
We require $\sum_{i} d_{i} w_{i} = 1$ .
Set $d_{i} w_{i} = r_{i}$ (the relative frequency of bin $i$ ). Then $\sum_{i} d_{i} w_{i} = \sum_{i} r_{i} = 1$ by equation [2]. ✓
Solving for height gives $d_{i} = r_{i} / w_{i}$ , which is equation [3].
Consequence: if all bins have equal width $w$ , then height $\propto r_{i}$ , so plotting raw counts or frequencies is fine. But with unequal widths, plotting frequency as height distorts the picture — a wide bin looks falsely dominant. Always use density. $■$

John Wilder Tukey | 1915–2000

An American mathematician and statistician at Princeton and Bell Labs, Tukey reshaped how data is explored. In his 1977 book Exploratory Data Analysis he invented the boxplot (box-and-whisker diagram) and the stem-and-leaf plot, championing the idea that one should look at data before fitting models. He coined the words "bit" and "software." His philosophy — let the data suggest hypotheses rather than only test them — underpins modern data science and the very notion of a graphical summary.

Building Intuition

Think of a dataset as a crowd of people in a stadium. A table is like reading every ticket stub one by one — accurate but unmanageable. A histogram is like flying overhead: you instantly see where the crowd is dense (the peak), how spread out they are, and whether anyone is sitting far off alone (an outlier). A boxplot is a stripped-down aerial view showing only the central 50% (the box), the typical range (whiskers), and the loners (dots). A scatterplot seats two crowds on perpendicular axes so you can spot whether tall people also tend to be heavy — an association. The right graph turns invisible structure into something your eye catches in a fraction of a second.

A boxplot and its five-number summary, showing the IQR box, whiskers, and an outlier

<svg viewBox="0 0 640 280" xmlns="http://www.w3.org/2000/svg">
  <rect width="640" height="280" fill="#0a0e17"/>
  <!-- axis -->
  <line x1="60" y1="220" x2="600" y2="220" stroke="#94a3b8" stroke-width="2"/>
  <text x="320" y="255" fill="#e2e8f0" font-size="14" text-anchor="middle">value</text>
  <!-- ticks -->
  <line x1="100" y1="215" x2="100" y2="225" stroke="#94a3b8"/>
  <text x="100" y="240" fill="#e2e8f0" font-size="11" text-anchor="middle">0</text>
  <line x1="300" y1="215" x2="300" y2="225" stroke="#94a3b8"/>
  <text x="300" y="240" fill="#e2e8f0" font-size="11" text-anchor="middle">50</text>
  <line x1="500" y1="215" x2="500" y2="225" stroke="#94a3b8"/>
  <text x="500" y="240" fill="#e2e8f0" font-size="11" text-anchor="middle">100</text>
  <!-- whiskers -->
  <line x1="150" y1="110" x2="220" y2="110" stroke="#94a3b8" stroke-width="2"/>
  <line x1="150" y1="90" x2="150" y2="130" stroke="#94a3b8" stroke-width="2"/>
  <line x1="380" y1="110" x2="450" y2="110" stroke="#94a3b8" stroke-width="2"/>
  <line x1="450" y1="90" x2="450" y2="130" stroke="#94a3b8" stroke-width="2"/>
  <!-- box -->
  <rect x="220" y="80" width="160" height="60" fill="none" stroke="#d4a853" stroke-width="3"/>
  <!-- median -->
  <line x1="300" y1="80" x2="300" y2="140" stroke="#d4a853" stroke-width="3"/>
  <!-- outlier -->
  <circle cx="560" cy="110" r="5" fill="none" stroke="#d4a853" stroke-width="2"/>
  <!-- labels -->
  <text x="150" y="75" fill="#e2e8f0" font-size="11" text-anchor="middle">min</text>
  <text x="220" y="160" fill="#e2e8f0" font-size="11" text-anchor="middle">Q1</text>
  <text x="300" y="70" fill="#e2e8f0" font-size="11" text-anchor="middle">median</text>
  <text x="380" y="160" fill="#e2e8f0" font-size="11" text-anchor="middle">Q3</text>
  <text x="450" y="75" fill="#e2e8f0" font-size="11" text-anchor="middle">max (within fence)</text>
  <text x="560" y="95" fill="#e2e8f0" font-size="11" text-anchor="middle">outlier</text>
  <!-- IQR bracket -->
  <line x1="220" y1="180" x2="380" y2="180" stroke="#94a3b8" stroke-width="1" stroke-dasharray="4"/>
  <text x="300" y="200" fill="#e2e8f0" font-size="12" text-anchor="middle">IQR = Q3 − Q1</text>
</svg>

Worked Examples

Building a Five-Number Summary and Boxplot

Given: test scores of 11 students:

42, 55, 61, 61, 68, 70, 73, 78, 85, 91, 97

(already sorted,

n = 11

). Steps:

Median ( $Q_{2}$ ): with $n = 11$ , the median is the 6th value $= 70$ .
Lower half (below the median): $42, 55, 61, 61, 68$ — five values. Its median is the 3rd $= Q_{1} = 61$ .
Upper half: $73, 78, 85, 91, 97$ . Its median is the 3rd $= Q_{3} = 85$ .
Min $= 42$ , Max $= 97$ .
IQR $= 85 - 61 = 24$ (equation [5]).
Fences (equation [6]): lower $= 61 - 1.5 (24) = 25$ ; upper $= 85 + 1.5 (24) = 121$ . No value lies outside $[25, 121]$ , so no outliers.

Answer:

{min, Q_{1}, med, Q_{3}, max} = {42, 61, 70, 85, 97}, IQR = 24 points

Histogram with Unequal Bins Using Density

Given: ages of 100 visitors grouped into uneven bins:

| Age range | Count $f_{i}$ | Width $w_{i}$ | |-----------|-------------|-------------| | 0–10 | 20 | 10 | | 10–20 | 30 | 10 | | 20–40 | 40 | 20 | | 40–80 | 10 | 40 |

n = 100

. Steps: compute density

d_{i} = f_{i} / (n w_{i})

(equation [3]) for correct bar heights:

0–10: $d = 20/ (100 \cdot 10) = 0.020$
10–20: $d = 30/ (100 \cdot 10) = 0.030$
20–40: $d = 40/ (100 \cdot 20) = 0.020$
40–80: $d = 10/ (100 \cdot 40) = 0.0025$

Note the trap: the 20–40 bin has the highest count (40) but the same density as the 0–10 bin. Plotting raw counts would make 20–40 look tallest and most dominant; the density view shows it is no more concentrated than the youngest group. Answer:

densities = 0.020, 0.030, 0.020, 0.0025 per year

Each bar's area equals its relative frequency (0.20, 0.30, 0.40, 0.10), summing to 1.

Common Misconceptions

A boxplot does not show shape or modality. Two datasets — one symmetric and unimodal, one bimodal (two peaks) — can produce identical boxplots, because a boxplot only records five summary numbers and hides the density between them. Always pair a boxplot with a histogram or dotplot if shape matters. Likewise, a longer whisker does not mean more data — it merely means the extreme values reach farther; the box, not the whiskers, holds the central 50%. Truncated and manipulated axes lie. Starting a bar chart's vertical axis at a value other than zero exaggerates differences; a 2% change can be made to look like a doubling. Bar charts must begin at zero. (Line graphs of time series may use a non-zero baseline, since they emphasize change, not magnitude — but this must be clearly labelled.)

Anscombe's Quartet

In 1973 statistician Francis Anscombe constructed four datasets, each with eleven $(x, y)$ points, that share — to two decimal places — the same mean of $x$ (9.0), mean of $y$ (7.5), variance, correlation ( $r \approx 0.816$ ), and regression line ( $y = 3.00 + 0.50 x$ ). Yet when plotted:

Set I is a noisy linear cloud (the "honest" case);
Set II is a perfect curve (relationship is non-linear, not linear);
Set III is perfectly linear except for one outlier dragging the line;
Set IV has all $x$ identical except one influential point.

Identical numerical summaries; four completely different stories. The lesson: always plot your data. Numbers alone can be summarized into agreement while concealing the truth.

Connections

Graphical summaries sit at the heart of Descriptive Statistics, turning the numerical Measures of Central Tendency (mean, median, mode) and Measures of Spread (range, IQR, standard deviation) into pictures. Histograms are empirical estimates of an underlying Probability Distribution, and a smooth, bell-shaped histogram hints at the Normal Distribution. Scatterplots are the visual face of Correlation and Linear Regression, where the strength of a linear pattern is quantified by the correlation coefficient $r$ . The boxplot's resistance to outliers connects to the broader idea of Robust Statistics.

Real-World Applications

Epidemiology: epidemic curves (histograms of case counts over time) reveal whether an outbreak is point-source or propagating — used daily during COVID-19 reporting.
Finance: candlestick charts (a specialized boxplot variant) summarize each trading day's open, high, low, and close prices for stocks.
Climate science: time-series line graphs of global temperature anomalies (the "hockey stick" graph) communicate long-term warming trends.
Quality control: control charts in manufacturing plot measurements against control limits to detect when a process drifts out of specification.
Public health / education: growth charts — percentile curves of height and weight by age — let pediatricians compare a child against a reference population.
Journalism: newspapers like The New York Times and FiveThirtyEight use scatterplots, choropleth maps, and bar charts to make data-driven arguments visible to the public.
Sports analytics: shot charts (scatterplots over a court or pitch) summarize where a player scores or misses.

Historical Development

1765 — Joseph Priestley publishes timeline charts, an early use of position to encode time.
1786 — William Playfair, in The Commercial and Political Atlas, invents the line graph and bar chart to display trade data.
1801 — Playfair introduces the pie chart in his Statistical Breviary.
1858 — Florence Nightingale uses her "coxcomb" (polar-area) diagrams to argue that disease, not battle, killed most soldiers in the Crimean War — pioneering data visualization for persuasion.
1869 — Charles Joseph Minard publishes his famous flow map of Napoleon's 1812 Russian campaign, hailed as one of the best statistical graphics ever drawn.
1900s — Karl Pearson formalizes the histogram and coins the term.
1973 — Francis Anscombe publishes his quartet, dramatizing the danger of summary statistics without plots.
1977 — John Tukey publishes Exploratory Data Analysis, introducing the boxplot and stem-and-leaf plot.
1983 — Edward Tufte publishes The Visual Display of Quantitative Information, establishing principles of graphical integrity ("data-ink ratio," avoiding "chartjunk").

Subtleties & Limits

The choice of bin width for a histogram has no universally correct answer — too narrow and the picture is noisy spikes; too wide and real structure is smoothed away. Rules like Sturges' formula ( $k = 1 + lo g_{2} n$ bins) and the Freedman–Diaconis rule (width $= 2 IQR / n^{1/3}$ ) give guidelines, but the "best" binning remains partly subjective and dataset-dependent. More deeply, how to honestly visualize high-dimensional data (many variables at once) is still an active research frontier; techniques like parallel coordinates, t-SNE, and UMAP attempt it, but each distorts some relationships while preserving others — there is no perfect projection of many dimensions onto a flat page.