Displaying data with graphs

Before you calculate a single average, look at your data. A good graph shows the overall pattern, the unusual values and the questions you should be asking, often in a few seconds. A bad graph can hide the story or invent one. This lesson covers the standard graphs for categorical and quantitative variables, how to describe what a graph shows, and how graphs are used to mislead.

Distributions

The distribution of a variable tells you what values it takes and how often it takes each one. For a categorical variable that means the categories and the count or percentage in each. For a quantitative variable it means the range of values and where they cluster. Almost every graph in this lesson is a picture of a distribution, and almost every summary in the next lesson is a number that describes one feature of a distribution.

Graphs for categorical variables

The transit survey asked 400 residents how they usually get to work or school. A frequency table lists each category with its count, and a relative frequency table converts counts to proportions or percentages, which makes groups of different sizes comparable.

# Frequency and relative frequency table for travel mode (invented survey of 400).
from collections import Counter
modes = ["car"] * 252 + ["bus"] * 64 + ["walk"] * 36 + ["bicycle"] * 28 + ["other"] * 20
counts = Counter(modes)
n = len(modes)
print(f"{'mode':<10}{'count':>6}{'percent':>9}")
for mode, k in counts.most_common():
    print(f"{mode:<10}{k:>6}{k / n:>9.1%}")
print(f"{'total':<10}{n:>6}{1:>9.1%}")

mode       count  percent
car          252    63.0%
bus           64    16.0%
walk          36     9.0%
bicycle       28     7.0%
other         20     5.0%
total        400   100.0%

A bar graph shows each category as a bar whose height is the count or percentage. The bars are separated by gaps because the categories are distinct, and their order does not matter for nominal data, so sorting them from largest to smallest usually makes the graph easier to read. A pie chart shows each category as a slice of a circle. Pie charts work only when the categories make up a whole, and even then people judge angles and areas much less accurately than lengths, so a bar graph is almost always clearer. Pie charts are especially poor when there are many slices or when you want to compare two pies.

To compare a categorical variable across groups, use a side-by-side or segmented bar graph. For example, Lucas compared travel mode for residents under 30 and those 30 or older. A segmented bar graph shows, for each age group, a bar divided into the percentage using each mode, which makes it easy to see that younger residents are more likely to take the bus. Lesson 12 tests whether such differences could be due to chance.

Graphs for quantitative variables

A histogram groups a quantitative variable into intervals of equal width, called bins or classes, and draws a bar for each bin whose height is the count or percentage of values in it. Unlike a bar graph, the bars touch, because the bins cover a continuous range. The choice of bin width matters: too few bins hide the shape, too many make it jagged. Most software chooses a reasonable default, but it is worth trying a couple of widths.

# A text histogram of 400 invented commute times (minutes), with 10-minute bins.
import random
random.seed(141)
times = [round(random.gammavariate(3.2, 7.5)) for _ in range(400)]   # right-skewed, like real commutes
bins = range(0, 90, 10)
for lo in bins:
    k = sum(lo <= t < lo + 10 for t in times)
    print(f"{lo:>2}-{lo + 9:<3} min | {'#' * (k // 4)} {k}")
print("over 89 min:", sum(t >= 90 for t in times))

 0-9   min | ######### 36
10-19  min | ################################# 135
20-29  min | ########################## 104
30-39  min | ################## 74
40-49  min | ###### 26
50-59  min | #### 16
60-69  min | # 4
70-79  min | # 5
80-89  min |  0
over 89 min: 0

A stemplot, or stem-and-leaf plot, is a quick hand-made display for small data sets. Each value is split into a stem (all but the last digit) and a leaf (the last digit). The stems are written in a column and the leaves are written beside their stems, in order. A stemplot looks like a histogram turned on its side but keeps every individual value, so you can still read off the data. For example, the values 12, 15, 21, 23, 23, 28 and 34 become stems 1, 2 and 3 with leaves 2 5, 1 3 3 8 and 4.

A dotplot draws one dot for each observation above a number line, stacking dots for repeated values. Dotplots are excellent for small to moderate data sets because nothing is hidden. A boxplot, which we construct in lesson 4, summarizes a distribution with five numbers and flags unusual values; it is especially useful for comparing several groups side by side.

Describing a distribution: shape, centre, spread and outliers

When you look at a graph of a quantitative variable, describe four things. Shape: is the distribution roughly symmetric, or skewed, with a longer tail on one side? Does it have one peak (unimodal) or two (bimodal)? Centre: where is the middle of the data? Spread: how far do values range from the centre? Outliers: are there individual values that fall well outside the overall pattern?

A distribution is skewed to the right when its right tail is much longer than its left. Incomes, house prices, commute times, waiting times and the number of purchases per customer are typically right-skewed: most values are modest, but a few are very large. A distribution is skewed to the left when the left tail is longer, as with scores on an easy test or age at retirement. A symmetric distribution looks roughly the same on both sides of its centre, as many measurement errors and test scores on well-designed exams do.

Bimodal shapes often signal that two different groups have been combined. If the café's hourly sales show peaks at 8 a.m. and 3 p.m., the data mix morning commuters with afternoon students, and it might be more useful to study the two groups separately.

An outlier is a value that falls far from the rest. Outliers deserve attention, not automatic deletion. Some are errors: a commute recorded as 400 minutes may be a typing mistake for 40. Others are genuine and interesting: a customer who visited 90 times in a month may be an employee using the app, or a very loyal regular. Investigate first; then decide, and report what you decided.

Graphs over time

Data recorded at regular intervals, such as monthly sales, daily temperatures or the yearly unemployment rate, form a time series. A time series plot puts time on the horizontal axis and the variable on the vertical axis, connecting the points with lines. Look for a long-term trend (an overall rise or fall), seasonal variation (a pattern that repeats each year, week or day) and unusual movements. The café's monthly visits rise over two years, with a low point each June, when many students leave the city, and a peak each December: a trend plus seasonality.

# Monthly cafe visits (thousands) with trend and a seasonal pattern, and a 12-month moving average.
import math
visits = []
for m in range(24):
    trend = 40 + 0.6 * m
    season = 4 * math.cos(2 * math.pi * (m - 11) / 12)   # peak in December (month 11, 23)
    visits.append(round(trend + season, 1))
months = ["Jan", "Feb", "Mar", "Apr", "May", "Jun", "Jul", "Aug", "Sep", "Oct", "Nov", "Dec"]
for m in (0, 5, 11, 12, 17, 23):
    print(f"{months[m % 12]} year {m // 12 + 1}: {visits[m]:5.1f} thousand")
ma = sum(visits[12:24]) / 12
print(f"12-month average, year 2: {ma:.1f} thousand (year 1: {sum(visits[:12]) / 12:.1f})")

Jan year 1:  43.5 thousand
Jun year 1:  39.0 thousand
Dec year 1:  50.6 thousand
Jan year 2:  50.7 thousand
Jun year 2:  46.2 thousand
Dec year 2:  57.8 thousand
12-month average, year 2: 50.5 thousand (year 1: 43.3)

Seasonal patterns are why official statistics are often reported seasonally adjusted. Retail sales always jump in December and employment in construction always rises in summer, so agencies remove the usual seasonal swing to show whether the underlying trend is changing. When a news report says the unemployment rate rose, it usually means the seasonally adjusted rate.

Graphs that mislead

Graphs persuade because people trust what they see, and that makes them easy to abuse. The most common trick is a truncated vertical axis. If a bar graph of two interest rates, 4.1 and 4.3 per cent, starts its axis at 4.0, the second bar looks three times as tall as the first, although the rates differ by only 5 per cent of their value. For bar graphs, whose message is carried by bar length, the axis should start at zero. For line graphs that show change over time, starting away from zero can be reasonable, but the scale should be clearly labelled.

Other tricks include picture graphs in which an icon is scaled in both height and width, so that doubling a value quadruples the area; three-dimensional effects that distort the size of slices and bars; cherry-picked time windows that start at an unusual low or high; dual vertical axes with different scales that make two unrelated series appear to move together; and graphs with no labels or sources at all. When you see a graph, check the axes, the scale, the time window and the source before you accept its message.

Choosing a graph

Match the graph to the variable and the question. One categorical variable: a bar graph (or a pie chart if the parts form a whole and there are few of them). One quantitative variable: a histogram, dotplot, stemplot or boxplot. A quantitative variable across groups: side-by-side boxplots or histograms on the same scale. Two categorical variables: a side-by-side or segmented bar graph. Two quantitative variables: a scatterplot, covered in lesson 6. A variable over time: a time series plot. Label axes with the variable name and units, give the graph a clear title, state the source and the sample size, and keep decoration to a minimum.

Colour deserves thought. Roughly one in twelve men and one in two hundred women have some form of colour vision deficiency, most often difficulty telling red from green. Using colour palettes that remain distinguishable, adding direct labels and using patterns or shapes as well as colour makes graphs readable for everyone.