This article was downloaded by: [130.220.181.

50] On: 28 February 2021, At: 20:50

Publisher: Institute for Operations Research and the Management Sciences (INFORMS)
INFORMS is located in Maryland, USA

INFORMS Journal on Applied Analytics

Publication details, including instructions for authors and subscription information:
http://pubsonline.informs.org

A Practitioner’s Guide to Best Practices in Data

Visualization
Jeffrey D. Camm, Michael J. Fry, Jeffrey Shaffer

To cite this article:

Jeffrey D. Camm, Michael J. Fry, Jeffrey Shaffer (2017) A Practitioner’s Guide to Best Practices in Data Visualization. INFORMS
Journal on Applied Analytics 47(6):473-488. https://doi.org/10.1287/inte.2017.0916

Full terms and conditions of use: https://pubsonline.informs.org/Publications/Librarians-Portal/PubsOnLine-Terms-and-

Conditions

This article may be used only for the purposes of research, teaching, and/or private study. Commercial use
or systematic downloading (by robots or other automatic processes) is prohibited without explicit Publisher
approval, unless otherwise noted. For more information, contact [email protected].

The Publisher does not warrant or guarantee the article’s accuracy, completeness, merchantability, fitness
for a particular purpose, or non-infringement. Descriptions of, or references to, products or publications, or
inclusion of an advertisement in this article, neither constitutes nor implies a guarantee, endorsement, or
support of claims made of that product, publication, or service.

Copyright © 2017, INFORMS

Please scroll down for article—it is on subsequent pages

With 12,500 members from nearly 90 countries, INFORMS is the largest international association of operations research (O.R.)
and analytics professionals and students. INFORMS provides unique networking and learning opportunities for individual
professionals, and organizations of all types and sizes, to better understand and use O.R. and analytics tools and methods to
transform strategic visions and achieve better outcomes.
For more information on INFORMS, its publications, membership, or meetings visit http://www.informs.org
 INTERFACES
Vol. 47, No. 6, November–December 2017, pp. 473–488
http://pubsonline.informs.org/journal/inte/ ISSN 0092-2102 (print), ISSN 1526-551X (online)

A Practitioner’s Guide to Best Practices in Data Visualization

Jeffrey D. Camm,a Michael J. Fry,b Jeffrey Shafferb
a School of Business, Wake Forest University, Winston-Salem, North Carolina 27109; b Carl H. Lindner College of Business, University of
Cincinnati, Cincinnati, Ohio 45221
Contact: [email protected] (JDC); [email protected] (MJF); [email protected] (JS)

Received: June 27, 2016 Abstract. Data visualization is the process of visualizing data through tables, charts,
Revised: November 8, 2016; April 7, 2017 graphs, maps, and other visual aids. Data visualization is often thought of as a descriptive
Accepted: April 14, 2017 tool, but it is also important for use in predictive and prescriptive analytics. It serves as a
tool for descriptive data exploration, but also for communicating insights from data and
https://doi.org/10.1287/inte.2017.0916
from analytical models. In this tutorial, we provide a concise discussion of best practices
Copyright: © 2017 INFORMS for data visualization for effective communication.

History: This paper was refereed.

Keywords: professional: OR-MS • professional: implementation

Much has been written about the importance of com- the effectiveness of the visualization in communicating
munication in the application of operations research to the audience.
(OR) and in ensuring that the results from OR studies In this tutorial, we discuss best practices for effec-
are accepted and implemented. Woolsey (1979, 1981), tive data visualization, give examples of good and bad
Levasseur (1991, 2013), and Keller and Kros (2000) pro- visualizations, and describe a variety of types of charts
vide examples. Data visualization is an important tool and graphs and the comparisons for which they are
for data exploration; however, it is also critical for com- most effective.
municating results from analysis and recommenda-
tions from analytical studies. The Goal and Three General Principles
From the point of view of good, available software, The goal of a table or chart should be to display data
data visualization has never been easier. A variety of in a manner that conveys the desired message to the
excellent data-visualization software packages are now intended audience. The input data to the visualiza-
available. A few of the leading specialized software tion can be raw data, descriptive statistics from the
packages include Tableau, SAS Visual Analytics, and raw data, or the output of analytical models. Hence,
Spotfire. The open source language R also provides the principles we discuss in this paper are applica-
excellent data-visualization capabilities, and Microsoft ble across the full spectrum of analytics: descriptive,
has recently updated the data-visualization capabili- predictive, and prescriptive. For example, in descrip-
ties in its ubiquitous Excel package to allow the cre- tive analytics, scatter plots are useful for exploring
ation of a wider range of charts. Excel also has a rec- and conveying possible relationships between pairs of
ommended-chart option, which provides the user with variables. Overlaid line charts are often an excellent
a list of suggested visualizations based on the selected choice for the output of predictive models when the
data. Comparisons of data-visualization tools can be goal is to show the model results and prediction inter-
found in recent ratings by PC Magazine (Baker 2017) vals to convey the uncertainty associated with those
and Predictive Analytics Today (2017), and in Gartner’s results. A network geographic map, with arcs show-
“Magic Quadrant for Business Intelligence and Analyt- ing distribution-center-to-customer-zone assignments,
ics Platforms” (Sallam et al. 2017). However, software is an excellent way to convey the efficient structure
availability does not relieve the user of the responsibil- of the solution from a supply-network optimization
ity of making intelligent choices—choices that impact model.
473
 Camm, Fry, and Shaffer: A Practitioner’s Guide to Best Practices in Data Visualization
474 Interfaces, 2017, vol. 47, no. 6, pp. 473–488, © 2017 INFORMS

Regardless of the type of input data (data or model introduced what he calls the data-ink ratio—the ratio
output), the creation of the visualization must be of ink required to convey the intended meaning (data-
guided by the intended audience, the interests of that ink) to the total amount of ink used in the table or
audience, and the message that is to be conveyed. chart. According to Tufte, the data-ink ratio should be
This requires an understanding of the makeup of the as close to one as possible; that is, we should avoid ink
audience and an anticipation of questions audience that does not add information.
members will pose. In short, empathy for your audi- The third general principle, consistent with the
ence should be the main driver in designing your notion that simpler is better, is to use color purposely and
visualization. effectively. Although the use of color might be attractive
Before embarking on a discussion of the three guid- to make the data visualization seem “prettier,” it can be
ing principles for data visualization, it is instructive to a distraction to the audience and hinder your message.
know a bit about how we process information. Every As an example, color can be used effectively to draw
second our brains are busy processing the informa- attention to part of a visualization or to distinguish cat-
tion that our five senses are feeding us. The brain fil- egories in the data; however, using too much color can
ters stimuli from our environment to process what is distract the audience. Color should be used intention-
important in two main ways: unconscious and con- ally, that is, only if it assists in conveying your message.
scious, as Kahneman (2011) refers to as System 1 In the following sections, we provide examples and
and System 2, respectively. System 1 represents the more detail on these three general principles of good
uncontrolled functions, such as reading people’s facial data visualization. For a more detailed discussion on
expressions or the immediate reaction we have to an how the brain processes images for data visualiza-
event. System 2 represents the controlled functions tion, we refer the reader to Few (2009), particularly to
that require conscious effort, such as solving a math Chapter 3.
problem.
In data visualization, we leverage the pre-attentive An Example: Barring Pie Charts
attributes, which are System 1. We encode data in a cer- Let us consider a simple descriptive analytics example
tain way; for example, if we highlight something with to illustrate the use of the three general principles, as
color, the reader sees this and interprets it in 250 mil- well as the iterative nature of the process for creating
liseconds without even thinking about it. Hence, by a good visualization. Like prototyping in the modeling
utilizing methods that are attuned to System 1, data- process, visualization often benefits from incremental
visualization tools, when used properly, can help a improvements over multiple iterations. Consider the
reader gain a quick and correct understanding of the following example.
data being presented. The three guiding principles, We have data that have yielded the percentage of
which we discuss next, are all about making sure students declaring each of the 12 categories of majors
we capitalize on how we process information using (plus undesignated) in a college of business. Suppose
System 1. these data are to be used as part of the recruiting
The first general principle in data visualization is process for the college; that is, the audience includes
that design and layout matter. The design and layout of potential students, and in many cases, one or more
your visualization should facilitate ease of understand- parents of these potential students. Undoubtedly, the
ing your message. The design and layout, including the students and parents will have many questions about
type of chart or table used, should draw attention to the majors; however, we know that their questions are
the parts of the visualization that are important in con- likely to include the following:
veying your message to your intended audience. For • What is the most popular major?
example, axes should be labeled clearly and legends • What is the least popular major?
should be close enough to the display to facilitate easy • What are the top three majors?
comparisons. • How does the major I am considering compare to
The second general principle is to avoid clutter. In the other majors in terms of its popularity?
short, simpler is better. To make this point, Tufte (2001) • What percentage of students are undesignated?
Camm, Fry, and Shaffer: A Practitioner’s Guide to Best Practices in Data Visualization
Interfaces, 2017, vol. 47, no. 6, pp. 473–488, © 2017 INFORMS 475

The data that we have available are the percentage Consider Figure 2 in which we have eliminated the
of the total student population in each major, which unnecessary third dimension.
illustrates a part-to-whole relationship—the percent- Can we improve upon this chart? Studies have
age breakdown of the total for the variable under con- shown that humans are very good at measuring the
sideration. Pie charts are often touted as being useful length or height of an object on a common baseline. For
for showing this part-to-whole relationship (Figure 1). example, when two people stand next to each other,
What are the potential problems with this choice of we can easily tell which person is taller and by how
chart? One drawback is that it is difficult to see the much. Humans are also good at measuring the posi-
legend and the colors associated with each major. The tion of objects; for example, how far one person is
layout and design do not facilitate easy comparison standing from another. However, they are much less
because the viewer must move back and forth between adept at accurately measuring angles, arcs, and area,
the legend and the pie. With 13 categories, many of the and comparing them against each other. We can lever-
colors are very similar, which also makes comparison age this strength at measuring length in the design of
difficult. It might be better to label each slice of the
a chart by carefully choosing the type of chart to use.
pie with its major; however, doing so would make the
Bar charts are a good choice for comparing categories.
chart even more cluttered. Having the percentage next
Given the questions we believe our audience will pose,
to each slice is helpful.
that is, comparisons of the percentages of students in
An interesting question is, “What does the third di-
each major, a bar chart would be a better design for our
mension add to this chart to help us better understand
purposes.
the data?” The answer is nothing. In this example,
adding the third dimension makes comparing cate-
gories more difficult. Figure 2. Compared to the Three-Dimensional Pie Chart in
Figure 1, the Data-Ink Ratio in This Pie Chart Is Higher and
Figure 1. This Three-Dimensional Pie Chart Shows the the Chart Is Easier to Understand
Percentage of Undergraduates by Major for a College Undergraduate majors
of Business
6%
1%
Undergraduate majors 19%
8%
6% 19%
1%
8%

4%
2% 4%
4% 2%

26% 26%
4%

17%

5% 17%
5% 5%
2% 1%
5% 2% 1%
Accounting Business economics
Economics Entrepreneurship Accounting Business economics
Finance Hospitality management Economics Entrepreneurship
Industrial management International business Finance Hospitality management
Information systems Marketing Industrial management International business
Operations management Real estate Information systems Marketing
Undesignated Operations management Real estate
Note. The third dimension reduces the data-ink ratio and makes the Undesignated
chart more difficult to understand.
 Camm, Fry, and Shaffer: A Practitioner’s Guide to Best Practices in Data Visualization
476 Interfaces, 2017, vol. 47, no. 6, pp. 473–488, © 2017 INFORMS

Figure 3 shows a vertical bar chart of the data Figure 4. Sorting a Bar Chart Allows a Reader to Easily and
on business college majors. The differences in college Quickly Compare Differences Among Variables
majors are much easier to visualize in this figure than Undergraduate majors
in either Figures 1 or 2. For example, compare Account- 30
26%
ing and Finance in Figures 3 and 2. Without the data 25
labels, looking at Figure 2 and determining which 20 19%
17%
major has a higher percentage of undergraduates is
% 15
much harder. In Figure 3, even without data labels,
10 8%
we can relatively easily see that the bar for Finance is 6% 5% 5%
4% 4%
shorter than the bar for Accounting; thus, the percent- 5 2% 2% 1% 1%
age of Finance majors is smaller than the percentage of 0

Marketing

Accounting

Finance

Operations management

Undesignated

International business

Information systems

Business economics

Entrepreneurship

Economics

Industrial management

Hospitality management

Real estate
Accounting majors.
Can we improve upon Figure 3? If we think about
how this chart will be used, perhaps sorting the data
from high to low would make the chart easier to
read and more useful for the reader. Figure 4 shows
the same vertical bar chart sorted. Sorting allows the
reader to more easily find the highest and lowest per-
centage majors and to compare majors that are close in
percentage. What are the three largest majors? Compare the ease
Finally, our adeptness at comparing lengths is not of answering this question by looking at Figure 5 ver-
impaired by the use of a horizontal display versus sus Figure 2. We can also quickly determine the bottom
a vertical display. Figure 5 displays a horizontal bar
three majors and the second highest; or we can com-
chart, which is a better design for the same data; the
pare the third highest with the fourth highest. With
categorical text labels are easier to read because they
this simple example in Figures 1–5, we have illustrated
are oriented horizontally. In addition, because we have
the following three general principles:
labels next to each bar, we removed the horizontal axis
• Design and layout matter: Use them effectively.
for simplicity.
We evolved from a pie chart to a bar chart to a sorted
Figure 3. In This Bar Chart, Different Colors Are No Longer
Necessary to Distinguish Between Majors; Thus, We Can
Eliminate the Legend Figure 5. The Categorical Text Labels Are Horizontal; Thus,
Undergraduate majors
the Chart Is Easier to Read Than the Vertical Bar Chart in
30
Figure 4
26%
Undergraduate majors
25
Marketing 26%
20 19% 17% Accounting 19%
% 15 Finance 17%

10 8% Operations management 8%
5% 5% 6%
4% 4% Undesignated 6%
5 2% 1% 2% 1% Information systems 5%
0 International business 5%
Accounting

Business economics

Economics

Entrepreneurship

Finance

Hospitality management

Industrial management

International business

Information systems

Marketing

Operations management

Real estate

Undesignated

Entrepreneurship 4%
Business economics 4%
Industrial management 2%
Economics 2%
Real estate 1%
Hospitality management 1%
Camm, Fry, and Shaffer: A Practitioner’s Guide to Best Practices in Data Visualization
Interfaces, 2017, vol. 47, no. 6, pp. 473–488, © 2017 INFORMS 477

bar chart, and we then rotated the bar chart for ease of Design and Layout Matter: Types of
reading. Comparisons and Chart Types
• Avoid clutter: Too many labels, values, lines, and A number of primary comparisons can made when
dimensions hinder the reader in interpreting the data. visualizing data. The type of comparison being made
We eliminated the distracting third dimension. is critical when determining the type of data visualiza-
• Use color purposely and effectively: Use it only tion or chart type.
when it is needed to distinguish categories or draw the A categorical comparison is one of the more com-
attention of the reader. By choosing a bar chart, we mon comparisons, and several good choices are avail-
were able to eliminate the use of different colors and able. As we previously discuss, because humans are
the associated legend. adept at measuring height and width, we can leverage
From this example, we can also learn the following: this strength in our choice of chart type. As we saw in
• Avoid using a pie chart—use a bar chart instead. the first example, a bar chart, either vertically or hor-
• Sorting the data often makes relevant compar- izontally, is a great choice for making quick and pre-
isons easier; in our example, it can help our audience cise categorical comparisons. Figure 6 shows a simple
to more easily answer the questions we anticipate. example with two good approaches for a simple com-
• Avoid rotated text. If the data include long cate- parison (in this case, units of fruit sold). If our audience
gory names, consider rotating the chart. has a sales target of 16 units for each of these fruits,
In a more general way, this example illustrates what adding a target line helps the audience to easily com-
often happens in practice. We can rarely achieve the pare how each fruit is doing relative to the target. In
best visual display on the first attempt. Rather, itera- addition to a target line, other examples of this include
tive prototyping, based on the intended message, leads using “average,” “estimate,” “threshold,” and “previ-
to a narrowing down of design options followed by ous year.” Which of these to use should be driven by
successive improvements in labeling, the use of color, the problem context and questions you anticipate the
titles, and other details that improve the communica- audience will ask.
tion of our message. Berinato (2016) provides an ex- In designing a bar chart, consider the following:
panded discussion of prototyping in the context of • Order the data to provide additional context.
chart construction. • For a small number of bars, consider data labels.
In the next three sections, we provide more detailed Position the labels such that the reader can easily see
discussions of the three general principles. them. Notice the “data table” that is created when the

Figure 6. (Color online) Bar Charts for Categorical Comparisons Can Be Augmented with a Target Line
Bar chart for categorical comparisons
23
Bananas 23
17 Target 16 units
Apples 17 14
12

Grapes 14

Oranges 12
Target 16 units

Bananas Apples Grapes Oranges

“The language of bar charts speaks both vertically and horizontally.” - Jeffrey Shaffer
Notes. Target lines are often used to give the reader a reference point and context. Target lines can represent values, such as sales goals,
previous year data, averages, and thresholds.
 Camm, Fry, and Shaffer: A Practitioner’s Guide to Best Practices in Data Visualization
478 Interfaces, 2017, vol. 47, no. 6, pp. 473–488, © 2017 INFORMS

labels are aligned near the vertical axis of the chart. A bar chart is recommended for these types of com-
This makes it very easy for the eye to follow. parisons. To emphasize comparisons among categories
• If the chart must include many categories, then and part-to-whole comparisons, the bar chart can be
balance the use of axis labels with data labels. Both are supplemented by a stacked bar chart (Figure 7).
unnecessary. In the example above, why use axis labels Another very common comparison in data visual-
to give the reader approximate values, when the data ization is showing a trend over time. Time should gen-
labels provide the actual values? erally be on the horizontal axis. Line charts are great
• Avoid rotated text. If some category names are for visualizing a trend, but a bar chart can also be used.
long, consider rotating the chart. Figure 8 shows two good approaches for times-series
• Mute or remove gridlines and remove chart data. Note that the line chart does not show all data
values because doing so would make the chart appear
borders.
too cluttered. Here we simply show the highest value
• Use a light font color on dark colors and dark font
and the most recent value. For example, if our audience
colors on light colors (e.g., white font color on the dark-
for Figure 8 is concerned about planning candy pur-
colored bars in Figure 6).
chases, then having actual numbers, as we show on the
• Always start the axis at zero on a bar chart or any
bar chart in the figure, with the actual number of Hal-
other chart that encodes data using length or height loween visitors, might be more useful. We could also
(e.g., bar chart, histogram, lollipop chart, area chart). enhance the figure by adding an average line, similar
Because these types of charts use height or length to to our target line in Figure 6.
encode the data for comparison, the axis should begin In designing a time-series chart, consider the
at zero; otherwise, the visual comparison is incorrect following:
and can be misleading. • Time should be on the horizontal axis; time should
In addition, the part-to-whole comparison is often progress from earlier to later and read from left to right.
necessary. Unfortunately, this is often shown as a pie • If using bars to represent time, do not break the
chart, which for the reasons previously discussed, most axis. Bars must begin at zero. It is acceptable to break
data-visualization experts consider to be bad practice. the axis on a line chart; however, it has a magnifying

Figure 7. (Color online) A Stacked Bar Chart Can Enhance a Bar Chart and Is Generally Preferable to a Pie Chart for Making
Part-to-Whole Comparisons
What makes up a credit score? How to improve a credit score:

Payment history 35% Pay accounts on-time month after month

• Each month of on-time payments will help improve
Amounts owed 30% Pay off negative items
Bring past-due accounts current
Length of credit history 15%
Reduce amounts owed
• Lower the percent of balance to available credit
Types of credit used 10%
Limit number of inquiries
• New inquires can lower a score by 5–10 points
New credit 10%
Pay off monthly charges prior to statement date
• Pay off part or all of the balance mid-cycle

Payment history and amounts owed represent 65% of a credit score

Payment history Amounts owed Length Type New credit
35% 30% 15% 10% 10%

Length of credit history

Source: FICO, 2010 Types of credit used
created by Jeffrey A. Shaffer
Camm, Fry, and Shaffer: A Practitioner’s Guide to Best Practices in Data Visualization
Interfaces, 2017, vol. 47, no. 6, pp. 473–488, © 2017 INFORMS 479

Figure 8. (Color online) Both Line Charts and Bar Charts Can Be Used Effectively for Time-Series Data
Line or bar chart for time series data (on x-axis)

Number of trick-or-treaters by year Number of trick-or-treaters by year

869

800

600
454
869
400
726 673
492 542
200 454
391

0
2008 2009 2010 2011 2012 2013 2014 2008 2009 2010 2011 2012 2013 2014
Note. The line chart emphasizes minimum and most recent values to the reader; the bar chart can more effectively convey information on
values for each year without appearing too cluttered.

effect on the visualization, making the visual slope of thus, it helps the reader to generate insights from the
the line appear much more dramatic as the magnifica- chart.
tion increases. Dots can also be used in a scatter plot (with or with-
• Plot consistent time periods using a line chart. If out jitter). Figure 10 combines a traditional scatter plot
time periods are missing or unequal, then note this
and deal with it in the visualization; for example, use a Figure 9. (Color online) Jitter Adds Some Randomness in
dotted line where years are missing. the Horizontal Position of the Dots and Performance Bands
The use of dots can be effective in various visual- Allow the Easy Designation of Different Grade Ranges
ization methods. This includes making comparisons, Final exam grades of data visualization students
showing time-series data when an interval is miss- University of Cincinnati—Spring 2013
ing, plotting a distribution, or showing a correlation Dot plot Using jitter
between variables. If possible, avoid using shapes other
100 100
than dots or circles. Circles and dots have a single cen-
ter point that is clear to readers. Triangles, squares,
asterisks, stars, or other shapes hinder the reader in 90 90
making a quick and accurate determination of the cen-
ter of the point.
Final exam (%)

Final exam (%)

Figure 9 shows an example of a dot plot, also known 80 80

as a strip plot, to visualize the grades in a data-
visualization class. A standard dot plot is on the left;
the same plot using jitter and background shading for 70 70
the performance bands (50 to 100 percent) is on the
right. Jitter is a random value, or for our purposes,
pseudo-random value, which is assigned to the dots to 60 60

separate them so they are not plotted directly on top of

each other. It is a great technique to use in dot plots, 50 50
box plots with dots, and scatter plots, because it allows
Notes. The use of jitter allows a viewer to easily differentiate the dots.
the reader to more easily observe differences in values We can see its value by comparing the chart on the right with the
that would otherwise be plotted on top of each other; chart on the left in which all dots are in the same horizontal position.
 Camm, Fry, and Shaffer: A Practitioner’s Guide to Best Practices in Data Visualization
480 Interfaces, 2017, vol. 47, no. 6, pp. 473–488, © 2017 INFORMS

Figure 10. (Color online) A Scatter Diagram and Box Plots Can Display the Relationship Between Two Variables and Are
Useful for Displaying the Distributions of Correlated Variables

-AJOR,EAGUE"ASEBALLPLAYERS(EIGHTANDWEIGHT

4EAM !LL 0OSITION !LL

 

 
(EIGHTINCHES

 

 

 
        
7EIGHTPOUNDS
#REATEDBY
*EFFREY!3HAFFER
&OLLOW (IGH6IZ!BILITY

3OURCE3/#2DATA
-,"(EIGHTS7EIGHTS

with a box plot on the horizontal and vertical axes to deliveries are in the ranges spanning three to six days,
show the distribution of the height and weight of Major but can also see that one delivery took over 10 days.
League Baseball players. Another common comparison measure is rank order.
Scatter plots are also suitable in comparing many A slope graph is useful for this type of comparison.
variables using a scatter matrix—a matrix that dis- Figure 13 shows an example of the change in car pro-
plays correlations among multiple variables. Figure 11 duction from Year 1 to Year 5 for car manufacturers.
shows the correlation among several variables related to The axis on the left displays the production volume in
aspects of rental populations in different New York City Year 1; the axis on the right displays the production
boroughs. We see that the percentage of college grad- volume in Year 5. The line connecting the dots indi-
uates correlates positively with median monthly rent, cates the change (slope) between Years 1 and 5 for each
but correlates inversely with poverty rate. Scatter matri- car manufacturer.
ces can be beneficial for displaying correlation among Although many other chart types are appropriate in
multiple variables and can often be helpful in designing data visualization, listing them all and giving an exam-
more sophisticated descriptive and predictive models. ple of each would be impossible. We suggest avoiding
A histogram is a basic chart type that is used to dis- the following chart types, because better visualization
play the distribution of data and compare frequencies tools, such as bar charts, stacked bar charts, or side-by-
across bins of ranges. Figure 12 shows a histogram of side bar charts, are usually available: pie chart, donut
delivery times (in days) for a supplier. Operations man- chart, funnel chart, gauges, pyramid chart, and radar
agers can use this chart to quickly determine that most chart.
Camm, Fry, and Shaffer: A Practitioner’s Guide to Best Practices in Data Visualization
Interfaces, 2017, vol. 47, no. 6, pp. 473–488, © 2017 INFORMS 481

Figure 11. (Color online) A Scatter Matrix Is Useful for Showing the Relationship Between Several Pairs of Variables and
Visualizing Correlations Between Variables
Scatter matrix comparing 4 variables
2K
Median monthly
rent ($)

1K

80
graduates
% college

60

40

20

40
rate (%)

30
Poverty

20

10

50
Travel time

40
(mins)

30
20
10

1K 2K 20 40 60 80 10 20 30 40 10 20 30 40 50
Median monthly Percentage college Poverty rate (%) Travel time (min)
rent ($) graduates (%)

In addition, word clouds and bubble charts may be useful as a high-level comparison of the frequency of
useful in certain instances, but they are not ideal for words occurring in a text source, but they are not useful
precise quantitative comparisons. Word clouds may be when trying to quantify the exact number of words in
text analysis. Bubble charts can be useful when the size
Figure 12. This Frequency Histogram of Supplier of the bubble is a secondary factor in the analysis. Color
Performance Shows That Most Supplies Are Delivered and shapes used in conjunction with scatter charts can
Within Three to Six Days
also convey additional dimensions in a visualization.
Supplier delivery performance However, as we discuss in the following sections, we
25 23 must be careful to avoid too much clutter and overuse
21
20 of color.
Frequency

15
Avoid Clutter
10 8 Any element that does not contribute to the intended
6 message of a data visualization is clutter. Clutter dis-
5
2
1
tracts from the message of the visualization; therefore,
0 it requires more mental effort on the part of the viewer.
1–2 3–4 5–6 7–8 9–10 11–12 As we previously mention, Tufte’s data-ink ratio is a
Delivery days way to measure clutter, and the elimination of clutter
 Camm, Fry, and Shaffer: A Practitioner’s Guide to Best Practices in Data Visualization
482 Interfaces, 2017, vol. 47, no. 6, pp. 473–488, © 2017 INFORMS

Figure 13. (Color online) A Slopegraph Compares Changes in Ranking; Change in Values Is Indicated by the Line Connecting
the Dots Between Year 1 and Year 5
3LOPEGRAPHOFCARPRODUCTIONINYEARVSYEAR

'ENERAL-OTORS
4OYOTA
4OYOTA 'ENERAL-OTORS


6OLKSWAGEN
0RODUCTIONINMILLIONS


6OLKSWAGEN (YUNDAI

(YUNDAI

 
9EAR

will result in a data-ink ratio close to one (i.e., ink is Understanding these principles often allows us to
used only to convey the message). eliminate clutter and increase our data-ink ratio. The
Knaflic (2015) discusses the Gestalt Principles of Vis- following example illustrates the principles of continu-
ual Perception and how they can aid in eliminating ity, connection, and enclosure.
clutter. These six principles, which allow us to create Assume we have 24 years (Periods 1–24) of budget
order from visual stimuli are: data (measured in billions of dollars) and wish to pre-
• Proximity: We place objects that are close together dict the budget for the next four years (Periods 25–28).
in a group; Using linear regression, we obtain a predictive model
• Similarity: We place objects with the same shape budgett  0.2529 + 3.0135t, where t  the period. Our
and (or) color in a group; objectives in constructing a graphic are to convey to our
• Enclosure: We place objects that are enclosed in a audience of budget analysts how well the predictive
group; model fits historical data, to show the predictions from
• Closure: If enough of a shape is shown, we map the model, and to emphasize the uncertainty associ-
the shape to something familiar to us; ated with those predictions.
• Continuity: We expect objects that follow a path Figure 14 shows the results of this regression with an
or curve to be more related than those that do not; overlaid line chart. We use continuity with the darker
• Connection: We place objects that are connected (black) line giving the observed budget over time and
in a group. the lighter (blue) line the estimated or fitted budget
Camm, Fry, and Shaffer: A Practitioner’s Guide to Best Practices in Data Visualization
Interfaces, 2017, vol. 47, no. 6, pp. 473–488, © 2017 INFORMS 483

Figure 14. (Color online) This Overlay Line Chart Illustrates the Results of a Predictive Model of Budget





-ILLIONS

"UDGETACTUALBILLIONS
 -ODEL
PREDICTIONINTERVAL


                           
0ERIOD

from the model. We use connectedness and the break shades of one color to show intensity. For example, the
in the lines between Periods 24 and 25 to distinguish quantities of parts in a warehouse could be encoded
between the budgets of past periods and those in the by shades of grey, where a darker grey indicates
future. Finally, we use enclosure to emphasize, based higher quantities and a lighter grey indicates lower
on the 95 percent prediction interval, the uncertainty quantities.
of the predictions in Periods 25–28. A diverging color palette is similar to sequential,
but has a natural midpoint from which two sequential
Effective and Sparing Use of Color colors diverge. For example, a map that uses diverg-
Color should not be used simply to “spice up” a visu- ing colors to show profit and loss by state could rep-
alization or to avoid a “boring” visualization. It should resent profit by blue colors and loss by orange col-
be used purposefully. For example, it can effectively ors. Darker blue means higher profit and lighter blue
draw the attention of the reader, highlight a portion of means lower profit; a lighter shade of orange indi-
data, or to distinguish between different categories. For cates smaller losses and a darker shade larger losses.
a discussion of the use of color in data visualizations, Thus, divergence is really a combination of the first two
see Wexler et al. (2017, pp. 15–18). types of color usage, two categories of sequential color
Color can be used purposefully in data visualiza- diverging from a midpoint or zero.
tion in three primary ways: categorical, sequential, and Let us consider a simple example of the use of cat-
diverging. egorical color by extending our previous example of
Categorical color uses different color hues to distin- undergraduate majors (Figure 5). If we are preparing
guish between different categories. Examples include to use Figure 5 for an operations management fac-
categories that (1) involve apparel, such as shoes, socks, ulty meeting and our goal is to convey the percentage
shirts, hats, and coats, and (2) vehicle types, such as of undergraduate operations management majors rel-
cars, minivans, trucks, and SUVs. This color scheme ative to other majors, we can easily use color to draw
can be used in many different chart types. attention to operations management (Figure 15).
Sequential color can be used to encode a single mea- Figure 16 shows an example of sequential use
sure from low to high, for example, using light to dark of color—a map showing the population density
 Camm, Fry, and Shaffer: A Practitioner’s Guide to Best Practices in Data Visualization
484 Interfaces, 2017, vol. 47, no. 6, pp. 473–488, © 2017 INFORMS

Figure 15. A Sorted Horizontal Bar Chart of the Percentage states and that the states in the Eastern part of the United
of Undergraduates by Major for a College of Business States tend to have higher population density than those
Emphasizes Operations Management in the West.
Undergraduate majors Figure 17 shows a diverging use of color. The figure
Marketing 26% uses a blue-orange scale to show monthly same-store
Accounting 19%
sales for 10 retail outlets.
Finance 17%
As we have seen, color can be used effectively in deliv-
Operations management 8%
Undesignated 6%
ering the message about data. However, when commu-
Information systems 5% nicating with color, it is important to consider what we
International business 5% know about the presence of color blindness or color
Entrepreneurship 4%
vision deficiency (CVD). Birch (1993) found that one in
Business economics 4%
Industrial management 2%
12 men (eight percent) and one in 200 women (0.5 per-
Economics 2% cent) have CVD, and their primary issue is distinguish-
Real estate 1% ing between red and green. Consequently, we advise
Hospitality management 1% against the use of red and green together in a data
0 5 10 15 20 25 30
visualization.
%
One common solution among data-visualization
Note. This is an example of an effective use of categorical color to
draw the reader’s attention to a particular variable.
practitioners is to rotate around the color wheel to blue-
orange. Using blue instead of green to indicate “good”
(population per square mile) for each state in the United and orange instead of red to indicate “bad” works
States. From this map, a reader can quickly surmise that well; almost everyone (with very rare exceptions) can
the population density is greatest in the Northeastern distinguish blue and orange from each other. Thus, the

Figure 16. A Map Showing Population Density (Population per Square Mile) for Each State of the United States Illustrates the
Sequential Use of Color

0OPULATIONPERSQUARE
MILEBYSTATEFROM
CENSUS


n
n
n
n
n
n

Note. Darker shades of green indicate higher population density.

Camm, Fry, and Shaffer: A Practitioner’s Guide to Best Practices in Data Visualization
Interfaces, 2017, vol. 47, no. 6, pp. 473–488, © 2017 INFORMS 485

Figure 17. A Heat Map Shows Percentage Changes in Same-Store Sales for 10 Retail Outlets by Month
Percentage change in same-store sales
Month
Jan Feb Mar Apr May Jun Jul Aug Sep Oct Nov Dec
Anderson 5 3 2 7 8 7 8 5 8 10 9 8
Beechmont 3 4 5 9 4 2 5 3 2 0 1 2
Bellevue 7 7 7 14 12 8 5 9 10 9 7 6
Covington 12 14 13 17 12 11 9 8 7 8 5 3
Store

Dayton –2 –1 –1 0 2 4 3 5 6 7 8 8
Delhi –6 –6 –8 –4 –5 –5 –4 –7 –4 –2 0 –2
Florence –8 –5 –6 –2 4 7 9 12 11 12 14 12
Ludlow 5 3 0 6 4 4 5 6 10 11 8 7
Newport –1 –2 –1 0 2 3 3 4 7 7 8 8
Walnut Hills –4 –5 –3 –4 –1 –4 –5 –3 0 –2 0 –1

Notes. This heat map illustrates divergent use of color using blue and orange. Shades of blue indicate nonnegative percentage change in
same-store sales (more intense shades of blue are higher positive same-store sales), whereas shades of orange indicate negative (more intense
shades of orange are higher negative same-store sales).

blue-orange palette is often referred to as a “color-blind give insights on why the solution makes sense. For the
friendly” palette. curious, the full linear programming model is given in
the appendix.
A Final Example: Using the The first author, seduced by Excel’s three-dimen-
Three Principles sional graphing capability, originally used Figure 18 to
We provide one final example that illustrates the use of convey the recommendation and explain the solution.
the three data visualization principles and the use of This figure clearly violates all three of the guiding
a visualization for communicating visually the output principles we presented. The design and layout do not
and (or) recommendation from a prescriptive model, facilitate a clear understanding of the recommenda-
that is, a linear program. tion. The third dimension adds no value in terms of
The Calhoun Textile Mill Case (Camm et al. 1987), conveying the message and, because of the use of three
based on a real project, discusses a make-versus-buy dimensions, some values are not even visible. Is any of
decision that can be modeled as a linear program. For Fabric 9 made on a dobby loom? This is impossible to
the upcoming quarter, Calhoun has known demand
for each of its 15 fabrics, which can be made on a Figure 18. The Solution to the Calhoun Textile Mill Case Is
dobby loom or a regular loom. The number of each Displayed As a Three-Dimensional Bar Chart
type of loom is given; therefore, the number of avail- Calhoun mills
able loom hours of each type, ignoring changeovers
and downtime, can be calculated easily. For a given
400,000
fabric, although the production rate on a dobby or reg-
350,000
ular loom is the same, Fabrics 1–4 cannot be made on a
300,000
Volume (yds.)

regular loom, and total loom capacity is insufficient to

250,000
satisfy demand. Therefore, some demand must be out-
200,000
sourced. The cost per yard to produce each fabric and
150,000
the outsourced cost per yard for each fabric are given.
100,000
Management wants to know the amount of each fabric 50,000
to produce on each type of loom and the amount to out- 15 14
0
13 12
source. Hence, once we have solved the problem, our 11 10 Dobby
9 8 7 6 Regular
objective is to develop a graphical display to communi- Fabric
5 4 Subcontract
3 2
cate to management the model’s recommendation and 1
 Camm, Fry, and Shaffer: A Practitioner’s Guide to Best Practices in Data Visualization
486 Interfaces, 2017, vol. 47, no. 6, pp. 473–488, © 2017 INFORMS

Figure 19. In a More Thoughtful Visualization of the Calhoun Textile Case Solution, We Use a Stacked Bar Chart to Show
Production Amounts on the Dobby and Regular Looms, and on Subcontracted Amounts
#ALHOUNMILLSMAKEVSBUY


$OBBY 2EGULAR 3UBCONTRACT



 /UTSOURCEPENALTY(R




 

 
Yards

 




&ABRICHASTHELOWESTSUBCONTRACT 
               
PENALTYINPERLOOMHOUR
&ABRIC



 

 
 
  
 
  
  

              
&ABRIC
Notes. We use color to draw attention to Fabric 6, which is the only fabric outsourced. The embedded bar chart explains why Fabric 6 is
outsourced by displaying the outsource penalty per yard for each fabric.

determine from Figure 18. In addition to the unneces- convey complicated information in a visualization in
sary dimension, the unnecessary “wall” shading clut- which we attempt to avoid overwhelming or confusing
ters the space and dramatically decreases the data-ink the reader. Knaflic (2015) includes an extended discus-
ratio. The label for the vertical axis is too far from the sion of data visualization to reduce cognitive load for
axis scale. While color is used somewhat sparingly, the reader.
using color rectangles to indicate zero yards is unnec-
essary and distracting. Summary
Compare Figures 18 and 19. In Figure 19, we use Knowing your audience and the message you intend
color and a stacked bar for each fabric. We can eas- to deliver are of critical importance in creating an effec-
ily visually compare volumes by type of loom for each tive visualization. In this tutorial, we discussed best
fabric. Nothing is hidden. We anticipate management’s practices for data visualization, including three general
first question, “Why Fabric 6?” by drawing more atten- principles: (1) design and layout matter: choose a design
tion to the fact that Fabric 6 is the only fabric out- and layout that facilitate your message to your audi-
sourced, and we explain that it is the best choice be- ence; (2) avoid clutter: the use of ink that does not con-
cause it has the lowest outsource penalty per loom hour tribute to your story will likely distract your audience
of any of the fabrics. This final point is illustrated using and confuse the message; and (3) use color purposely and
an embedded bar chart showing the outsource penalty effectively: do not overuse color, but use it purposely
per loom hour for each fabric type. If this appears too to convey your message. On the surface, these princi-
cluttered, this last point could also be made in a sepa- ples might seem like common sense; however, in our
rate graph. Figure 19 provides an example of trying to experience, they are much like the rule that says “add
Camm, Fry, and Shaffer: A Practitioner’s Guide to Best Practices in Data Visualization
Interfaces, 2017, vol. 47, no. 6, pp. 473–488, © 2017 INFORMS 487

comments to your code.” Everyone agrees that follow- r13 + d13 + o13  10,000
ing them is a good idea; however, doing so requires r14 + d14 + o14  380,000
focus and a willingness to take the time to follow sim-
r15 + d15 + o15  62,000
ple practices. In that good data visualization is a way
0.1925r5 + 0.2625r6 + 0.2389r7 + 0.1911r8 + 0.1911r9
to quickly and effectively communicate quantitative
information, the principles apply to all types of analyt- + 0.1911r10 + 0.2679r11 + 0.2389r12 + 0.2253r13
ics: descriptive, predictive, and prescriptive. + 0.1911r14 + 0.2389r15 ≤ 196,560
We recommend several excellent texts for further 0.2149d1 + 0.2149d2 + 0.2149d3 + 0.2149d4 + 0.1925d5
reading on best practices in data visualization. These + 0.2625d6 + 0.2389d7 + 0.1911d8 + 0.1911d9 + 0.1911d10
include two classic texts, Tufte (1997, 2001), and the + 0.2679d11 + 0.2389d12 + 0.2253d13 + 0.1911d14
more recent texts, Cairo (2012) and Few (2009, 2012).
+ 0.2389d15 ≤ 32,760
Few (2006) also discusses data visualization in the
context of data dashboards. Finally, for a treatment di , r i , o i ≥ 0 for all i.
of data visualization and storytelling, we recommend
References
Knaflic (2015).
Baker P (2017) The best data visualization tools of 2017. Accessed
October 17, 2017, http://www.pcmag.com/roundup/346417/
Appendix the-best-data-visualization-tools.
Let d i  number of yds. of fabric i to produce on dobby looms Berinato S (2016) Good Charts: The HBR Guide to Making Smarter, More
i  1 . . . 15 Persuasive Data Visualizations (Harvard Business School Publish-
ing Corporation, Boston).
r i  number of yds. of fabric i to produce on regular looms Birch J (1993) Diagnosis of Defective Colour Vision (Oxford University
i  5 . . . 15 Press, New York).
o i  number of yds. of fabric i to purchase outside i  Cairo A (2012) The Functional Art: An Introduction to Information
1 . . . 15 Graphics and Visualization (New Riders, Berkeley, CA).
Camm JD, Dearing PM, Tadisina SK (1987) The Calhoun Textile Mill
Minimize Case: An exercise on the significance of linear programming
0.6573d1 +0.555d2 +0.655d3 +0.5542d4 +0.6097d5 +0.6153d6 model formulation. IIE Trans. 19(1):23–28.
Few S (2006) Information Dashboard Design (O’Reilly Media,
+0.6477d7 +0.488d8 +0.5029d9 +0.4351d10 +0.6417d11 Sebastopol, CA).
Few S (2009) Now You See It (Analytics Press, Oakland, CA).
+0.5675d12 +0.4952d13 +0.3128d14 +0.5029d15 +0.6097r5 Few S (2012) Show Me the Numbers, 2nd ed. (Analytics Press, Oak-
+0.6153r6 +0.6477r7 +0.488r8 +0.5029r9 +0.4351r10 land, CA).
Kahneman D (2011) Thinking, Fast and Slow (Farrar, Straus and
+0.6417r11 +0.5675r12 +0.4952r13 +0.3128r14 +0.5029r15 +0.8o1 Giroux, New York).
Keller C, Kros J (2000) Teaching communication in an MBA oper-
+0.7o2 +0.85o3 +0.7o4 +0.75o5 +0.75o6 +0.8o7 +0.6o8 +0.7o9 ations research/management science course. J. Oper. Res. Soc.
+0.6o10 +0.8o11 +0.75o12 +0.65o13 +0.45o14 +0.7o15 51(12):1433–1439.
Knaflic CN (2015) Storytelling with Data (John Wiley & Sons, New
subject to York).
Levasseur RE (1991) Effective communication—A critical skill for
d1 + o1  16,500 MS/OR professionals. Interfaces 21(2):22–24.
Levasseur RE (2013) People skills: Developing soft skills—A change
d2 + o2  52,000
management perspective. Interfaces 43(6):566–571.
d3 + o3  45,000 Predictive Analytics Today (2017) Top 74 data visualization software.
Accessed April 21, 2017, http://www.predictiveanalyticstoday
d4 + o4  22,000 .com/top-data-visualization-software/.
Sallam RL, Howson C, Idoine CJ, Oestreich TW, Richardson JL,
r5 + d5 + o5  76,500 Tapadinhas J (2017) Magic quadrant for business intelligence
and analytics platforms. Report, Gartner, Stamford, CT.
r6 + d6 + o6  110,000
Tufte ER (1997) Visual Explanations (Graphics Press, Cheshire, CT).
r7 + d7 + o7  122,000 Tufte ER (2001) The Visual Display of Quantitative Information, 2nd ed.
(Graphics Press, Cheshire, CT).
r8 + d8 + o8  62,000 Wexler S, Shaffer J, Cotgreave A (2017) The Big Book of Dashboards
(John Wiley & Sons, New York).
r9 + d9 + o9  7,500 Woolsey G (1979) The fifth column—An exemplary essay on com-
munication or corporate style, corporate substance, & the sting.
r10 + d10 + o10  69,000
Interfaces 9(3):10–12.
r11 + d11 + o11  70,000 Woolsey G (1981) The fifth column—Yet another essay on effective
communication or on learning where to begin, how to listen,
r12 + d12 + o12  82,000 and when to stop. Interfaces 11(1):10–12.
 Camm, Fry, and Shaffer: A Practitioner’s Guide to Best Practices in Data Visualization
488 Interfaces, 2017, vol. 47, no. 6, pp. 473–488, © 2017 INFORMS

Jeffrey D. Camm is Associate Dean of Business Analytics of British Columbia. He earned his PhD and MS degrees in
and the Inmar Presidential Chair in Business Analytics at the industrial and operations engineering from the University
Wake Forest University School of Business. He received his of Michigan, and a BS degree from Texas A&M University.
PhD in management science from Clemson University and He has consulted for a variety of organizations including
a BS in mathematics from Xavier University. Prior to join- Great American Insurance Group, Procter & Gamble, Cardi-
ing Wake Forest, he was a faculty member at the University nal Health, Boeing, Copeland Corporation, Coleman Com-
of Cincinnati. Professor Camm’s scholarship is on the appli- pany, Vancouver City Savings Credit Union, and others.
cation of optimization modeling to difficult decision prob- Jeffrey Shaffer is Vice President of Information Technol-
lems in a diverse set of application areas including, opera- ogy and Analytics at Unifund. He is also adjunct professor
tions planning and scheduling, supply chain optimization, at the University of Cincinnati teaching data visualization
product design, and conservation. He has received several where he was the 2016 Outstanding Adjunct Professor of the
teaching awards including the 2006 INFORMS Prize for the Year in Operations, Business Analytics and Information Sys-
Teaching of OR/MS practice. tems. He is the co-author of The Big Book of Dashboards and a
Michael J. Fry is Professor and Head of the Department of regular speaker on the topic of data visualization, data min-
Operations, Business Analytics and Information Systems and ing, and Tableau training at conferences, symposiums, work-
has been named a Lindner Research Scholar in the Carl H. shops, universities, and corporate training programs. He is
Lindner College of Business at the University of Cincinnati. one of 27 Tableau Zen Masters in the world and the win-
Professor Fry has served as Interim Director of the University ner of the 2014 Tableau Quantified Self Visualization Con-
of Cincinnati Center for Business Analytics, and he has been test which led him to compete in the 2014 Tableau Iron Viz
a visiting professor at Cornell University and the University Contest.