Fire Data Analysis

Handbook
Third Edition
FA-266/November 2021
Fire Data Analysis
Handbook
Third Edition
FA-266/November 2021
 Mission Statement
We support and strengthen fire and emergency medical
services and stakeholders to prepare for, prevent, mitigate
and respond to all hazards.
Foreword
The fire service exists today in an environment constantly
inundated with data, but data are seen of little use in the
everyday, real world in which first responders live and work.
This is no accident. By themselves, pieces of data are of
little use to anyone. Information, on the other hand, is very
useful indeed. What ’s the difference? At sporting events,
people in stadiums hold up individual, multicolored squares
of cardboard to form a giant image or text, which could be
recognized only from a distance. This is a good analogy for data
and information. The individual squares of cardboard are like
data. They are very numerous and they all look similar taken
by themselves. The big image formed from the organization
of thousands of those cards is like information. It is what can
be built from many pieces of data. Information then is an
organization of data that makes a point about something.

The fire service of today is changing. More and more, it is not

fighting fires as much as it is doing emergency medical services
(EMS), hazmat, inspections, investigations, prevention and
other nontraditional but important tasks, which are vital to the
community. Balancing limited resources and justifying daily
operations and finances in the face of tough economic times
is a scenario that is familiar to every department.

Turning data into information is neither simple nor easy. It

requires some knowledge of the tools and techniques used
for this purpose. Historically, the fire service has had few of
these tools at its disposal and none of them has been designed
with the fire service in mind. This book changes that. It was
designed solely for the use of the fire service. The examples
were developed from fire data collected from departments all
over the nation. This book also was designed to be modular
in form. Many departments’ information needs can be met by
using only the first few chapters. Others with a more analytical
and statistical background may want to go further. The point
is, it’s up to the reader to decide. This handbook is a tool, like
a pumper or a ladder, to help do the job.

The U.S. Fire Administration

Foreword iii
iv Fire Data Analysis Handbook
Table of Contents
Chapter 1: Introduction..................................................... 1
Why data analysis?.................................................................... 2
National Fire Incident Reporting System............................... 3
Data entry and data quality..................................................... 6
Statistical packages for computers......................................... 8
How to use this handbook....................................................... 9
Books on statistics and data analysis..................................... 10

Chapter 2: Histograms....................................................... 13
Data as a descriptive tool......................................................... 13
Types of variables...................................................................... 13
Developing a histogram............................................................ 20
Cumulative frequencies............................................................ 22
Summary.................................................................................... 23

Chapter 3: Charts................................................................ 25
Introduction............................................................................... 25
Bar charts................................................................................... 26
Column charts............................................................................ 28
Line charts.................................................................................. 31
Pie charts.................................................................................... 32
Dot charts................................................................................... 33
Pictograms.................................................................................. 35
Summary.................................................................................... 36

Chapter 4: Basic Statistics................................................. 37

Measures of central tendency................................................. 37
Properties and uses for measures of central tendency....... 38
Measures of dispersion............................................................ 39
Normal distribution and standard score............................... 41
Properties and uses for measures of dispersion.................. 42
Skewed distributions................................................................ 43
Central limit theorem................................................................ 44

Chapter 5: Analyses of Tables........................................... 45

Introduction............................................................................... 45
Describing categorical data...................................................... 45
The chi-square test.................................................................... 47
2-way contingency tables......................................................... 54

Table of Contents v
 Percentages for 2-way contingency tables............................ 55
Joint percentages....................................................................... 56
Row percentages....................................................................... 57
Column percentages................................................................. 58
Selecting a percentage table.................................................... 59
Testing for independence in a 2-way contingency table...... 60
Constructing a table of expected values................................ 61
Calculation of chi-square for a 2-way contingency table..... 62

Chapter 6: Correlation....................................................... 67
Introduction............................................................................... 67
Scatter diagram......................................................................... 67
Correlation coefficient.............................................................. 69
Calculating the correlation....................................................... 71
Other types of correlations...................................................... 74

Appendix: Critical Values of Chi-Square.......................... 77

Reference............................................................................. 79

vi Fire Data Analysis Handbook

Chapter 1: Introduction
The primar y objec tive of this handbook is to describe
statistical techniques for analyzing data typically collected by
fire departments. Motivation for the handbook stems from
the belief that fire departments collect an immense amount
of data and it should not go unused. With basic analytical
training and an understanding of statistical techniques, a fire
department can gain a better understanding of the nature of
fires in the area they serve and effectively present data and
information to help save lives and property.

Consider the incident reports that fire departments complete.

You document information such as the type of situation found,
action taken, time of alarm, time of arrival, time completed,
number of engines responding and number of personnel
responding. For fires, the list grows even longer, including area
of fire origin, form of heat of ignition, type of material involved
and other related factors. Additionally, if a civilian or firefighter
is injured, there are other reports to complete.

Fire departments have a legal requirement to document these

incidents. Victims, insurance companies, lawyers and many
others want copies of reports. Fire departments maintain files
to retrieve individual reports.

The reports can, however, be beneficial to fire departments by

providing insight into the nature of fires and casualties in their
jurisdiction. Basic information is probably already available.
Typically, the number of fires handled last year, the number of
fire-related injuries and the number of fire deaths are tracked. It
is another story, however, if more probing questions are asked:

ĵ How many fires took place on Sundays, Mondays, etc.?

ĵ How many fires took place each hour of the day or month
of the year?
ĵ What was the average response time to fires?
ĵ How much did response times vary by fire station areas?
ĵ What was the average time spent at the fire scene?
ĵ How much did the average time vary by type of fire?

Chapter 1: Introduction 1
 This handbook describes statistical techniques to turn data
into information to answer these types of questions and many
others. The techniques range from simple to complex.

For example, the next 2 chapters describe how to develop

charts to provide more effective presentations about fire
problems. These charts may be useful for city or county
of ficials to explain the activities and needs of your fire
department. “Chapter 4: Basic Statistics” discusses measures
of central tendency (mean, median and mode) and measures
of dispersion (range, variance and standard deviation).
“Chapter 5: Analyses of Tables” explains the chi-square statistic
and its use in analyzing table data. “Chapter 6: Correlation”
discusses the Pearson correlation coefficient and additional
correlations. These are all techniques that can tell you more
about the nature of fires and casualties.

One way to become more comfortable with data analysis is

to work with real data. For this handbook, we obtained data
from fire departments in several large metropolitan areas.
Working with real data makes it easier to understand the
different techniques.

Why data analysis?

You may still question why we should go to all this trouble to
analyze data. Many decisions do not require analysis, such as
decisions on personnel, grievance proceedings, promotions
and even decisions on how to handle a fire. It is certainly true
that fire departments can continue to operate in the same way
they always have without doing a lot of analysis.

On the other hand, there are 3 good reasons for looking closely
at the data:

1. Gain insights into fire problems.

2. Improve resource allocation for combating fires.
3. Identify training needs.

The most compelling reason is that analysis gives insight

into fire problems, which in turn can affect operations in the
department. For example, you may find that the average
time to fires in an area is 6 minutes, compared to less than

2 Fire Data Analysis Handbook

2 minutes overall. This may be helpful in requests for more
equipment, more personnel or justifying another fire station.

Another example of improved resource allocation, statistical

analysis of emergency medical calls can determine the impact
of providing another paramedic unit in the field. Increasing the
number of EMS units from 4 to 5 may, for example, decrease
average response times from 5 minutes to 3 minutes — a
change that may save lives.

The analysis can also be used to identify training needs. Most

firefighting training is based on a curriculum that has been
in place for many years. Analysis of your fire data can allow
you to see how training matches characteristics of fires in a
particular jurisdiction. This is not to say that other training
is not important. However, knowing more about the fires in
an area can lead to improvements in training. Additionally,
analysis of firefighter injuries may indicate a need for new
types of training.

In summary, this handbook will help you deal with the volume
of data collected on fire incidents. By using the techniques
presented here, you will be able to improve your skills in
collecting and analyzing data, as well as presenting the results.

National Fire Incident Reporting System

The National Fire Incident Reporting System (NFIRS) was
established more than 45 years ago to collect and analyze
data on fires from departments across the country. More
than 24,000 fire departments from all 50 states, the District
of Columbia and the Native American Tribal Authority report
their fires and losses to the NFIRS. This makes the NFIRS the
largest collector of fire-related incident data in the world. Fire
departments report over 1.2 million fire incident responses
each year to the NFIRS.

Incident data collection is not new. In 1963, the National Fire

Protection Association (NFPA) developed a dictionary of fire
terminology and associated numerical codes to encourage
fire departments to use a common set of definitions. This
dictionary is known as NFPA 901, Standard Classifications for

Chapter 1: Introduction 3
 Fire and Emergency Services Incident Reporting. The current
NFIRS 5.0 data standard represents the merging of the ideas
and definitions from NFPA 901 and the many suggested
improvements from the users of the NFIRS 4.1 coding system.

Version 5.0 of NFIRS consists of 11 separate modules that allow

fire departments to report any type of incident they respond to.

The Basic Module (Module 1) is required. It includes incident

number and type, incident date, alarm time, arrival time, time
in service, and type of action taken.

Modules 2 through 5 are required if applicable. If you respond

to a fire, you complete the Fire Module (Module 2). It includes
property details, cause of ignition, human factors, equipment
involved and other information.

If you respond to a structure fire, you complete Module 3, the

Structure Fire Module. It includes such things as structure type,
main floor size, area of fire origin, and presence of detectors
and automatic extinguishment equipment.

If there were civilian or fire service casualties, you complete

Modules 4 or 5, respectively.

The remaining modules are optional at the local level. They

include EMS (Module 6), Hazardous Materials (Module 7),
Wildland Fire (Module 8), Apparatus or Resources (Module 9),
Personnel (Module 10), and Arson (Module 11).

Usually, the state fire marshal’s office in each NFIRS state has
the responsibility for collecting data from its fire departments
or overseeing the NFIRS submissions. It also manages system
access for its fire departments. Some states maintain a state-
level incident database. These data files are combined with
data from other fire departments into a statewide database.
The state then submits the data files to the NFIRS national
database at the U.S. Fire Administration (USFA), National
Fire Data Center (NFDC), by using the USFA file upload
tool. Or, per state policy, local fire departments can submit
their files directly to the national database using the USFA
file upload tool. Departments and states that do not use a

4 Fire Data Analysis Handbook

vendor product may still participate in the NFIRS by using its
no-cost, web-based data entry tool. Today, there are very few
departments that have no electronic capabilities, but some
remain. Their paper reports are sent to their state office which
then enters the reports into the NFIRS.

All states and fire departments within them have been invited
to participate in the NFIRS on a voluntary basis. Most of the
data are collected electronically through third-party vendor
software. The NFDC maintains the NFIRS specification for
vendors so they may prepare products to meet the NFIRS 5.0
data standard. Data on individual incidents and casualties are
preserved incident by incident at local, state and national levels.

The NFDC, among other organizations, can analyze NFIRS

data at the national level to help develop public education
campaigns, make recommendations for national codes
and standards, guide allocations of federal funds, ascertain
consumer product failures, identify the focus for research
efforts, and support federal legislation.

Ever y f ire depar tment is responsible for managing its

operations in such a way that firefighters can do the most
effective job of fire control and fire prevention in the safest way
possible. Effective performance requires careful planning; this
can only happen if accurate information about fires and other
incidents is available. Patterns that emerge from the analysis
of incident data can help departments focus on current
problems, predict future problems in their communities and
measure their programs’ successes.

The same principle is also applicable at the state and national

levels. The NFIRS provides a mechanism for analyzing incident
data at each level to help meet fire protection management
and planning needs. In addition, NFIRS information is used by
labor organizations to analyze such matters as workloads and
firefighter injuries.

Chapter 1: Introduction 5
 Data entry and data quality
Data quality is an area of great importance. The following
criteria are used in monitoring data in the NFIRS during the year:

ĵ The data are complete.

ĵ The data are accurate.
ĵ The data are current.

These criteria are monitored by creating reports that show

the number of reporting fire departments, the number of
incidents by state, the number of invalid incidents and the
number of unreleased incidents. The USFA provides the
reports to the state NFIRS program managers and works
with them to resolve any data issues. USFA provides technical
assistance (e.g., telephone support) to states to help address
any data quality and data reporting needs.

Audits of the data are performed during the year to identify

any inconsistencies. The audits focus on 3 criteria: gaps in
reporting, critical errors in the data and outliers in the data.
In particular, the USFA works closely with states to monitor the
quality of data coming from third-party vendor software. The
USFA assists states in monitoring vendor data quality issues
or contacts vendors directly to discuss an issue at a state’s
request. Examples of data quality issues that are reviewed are
questionable, high dollar-loss incidents and questionable, high
numbers of fire deaths.

Quarterly, USFA staff queries the database for questionable

values (i.e., outliers) and verifies the values with state- and
local-level NFIRS program managers. These important steps
ensure that the data meet the USFA’s 3 criteria before the data
are released in the NFIRS Public Data Release format.

One assumption throughout the handbook is that data on

fire incidents and casualties are available for analysis. Manual
analysis is possible, but the tedious calculations quickly
overwhelm the ability to perform analysis in any meaningful
manner. Today, many analy tical and statistical software
packages exist to help process and analyze data quickly and
accurately. Even spreadsheet software, such as Microsoft

6 Fire Data Analysis Handbook

Excel, can be used to complete analysis on smaller sets of data
at the department level.

Most fire departments enter their data into either third-party

vendor software that is purchased by the state or local fire
department, or the free web-based tool supplied by the USFA
to states. If third-party vendor software is used, it must be
compatible with the NFIRS standard. A list of active-status
vendors is available from the USFA, but it is the responsibility
of the individual states to ensure that a vendor’s software
meets the qualifications. If the USFA web-based tool is used,
it must be supported by the state.

A word of caution: Any third-party vendor software should

contain an error-checking routine. Data quality is always a
concern, and in data science, the principle “garbage in, garbage
out” certainly applies to fire department reports. The software
should, for example, check each item to make sure a valid code
has been entered. Whenever the software encounters an error,
it should provide the user with the opportunity to correct the
error before it becomes part of the database. For example,
alarm times obviously cannot have hours greater than 23 and
minutes greater than 59. Data entry software should check
hours and minutes for valid numbers and allow corrections to
be made immediately.

The data collected to describe an incident are the foundation

of the system. Therefore, editing and correcting errors is
a system-wide activity involving local, state and federal
organizations. All errors resulting from the edit /update
process need to be reported to fire departments, and the
submission of corrections from fire departments is essential.
This is especially important for fatal errors, which prevent the
data from being included in the NFIRS database.

At the local level, fire departments need to establish data

quality procedures if they intend to take full advantage of
their data. There should be a system in place to double-check
the collection and data entry. Field edits and relational edits
can be built into the system that will reveal unacceptable and
unreasonable data. Data management personnel can use
these techniques to improve and validate the data.

Chapter 1: Introduction 7
 In summary, data entry software should include code-checking
routines to identify errors in individual items in the report and
errors reflected through inconsistencies between items. Because
data entry software cannot be expected to find all errors, fire
departments also need to implement data quality procedures to
ensure that correct data are entered into their systems.

Statistical packages for computers

In this handbook, we present many dif ferent t ypes of
analyses. “Chapter 3: Char ts,” for example, discusses
several types of charts, including bar charts, column charts,
histograms, line charts and dot charts. Other chapters show
how to calculate statistics, such as means and variances, and
how to do more advanced calculations, such as chi-square
tests and correlation coefficients.

For a good understanding of the analysis, it is important to

know what is involved in the statistical calculations, but it is not
recommended to do data analyses by hand. There are several
good analytical and statistical software packages available for
data analyses. If you intend to apply the techniques in this
handbook, you should acquire and learn how to use a full-
featured statistical analysis software application. Excel (and
Microsoft Access) may get you started but can quickly get
overwhelmed by larger datasets and complex analysis.

The following are a few examples of full-featured software

packages:

Software
Website
package
IBM® SPSS® www.ibm.com/products/spss-statistics
R www.r-project.org
Stata www.stata.com
SAS ®
www.sas.com
NCSS www.ncss.com
JMP ®
www.jmp.com

8 Fire Data Analysis Handbook

How to use this handbook
Data analysis is not an easy process. It requires careful data
collection, attention to detail, access to statistical programs
and skills in result interpretation. These are not impossible
tasks, but they require time and patience on your part for
success. Equally important, you need experience. In the long
run, you can only develop capabilities in analysis by applying
techniques from this handbook on actual data sets.

As a final note, one way of thinking about analysis is to

consider it a 4-stage process.

ĵ Stage 1: Collect the data, which is what the NFIRS does. In

and of themselves, the data are meaningless.

ĵ Stage 2: Organize and summarize into information that

can be analyzed.

ĵ Stage 3: Analyze according to whatever problem or issue

is being considered. This yields a better understanding of
the information.

ĵ Stage 4: Use the information to make decisions.

Our ultimate objective is to make better and more informed

decisions at the local fire department level. Data have no utility
in a vacuum, and fire reports stay as data if we do nothing
with them. Analysis turns data into information. We move,
for example, from knowing individual alarm and arrival times
to knowing average travel times. Our review of travel times
increases our knowledge about what is going on with fire
incidents, which results, in turn, in more informed decisions
within fire departments.

Chapters 2 and 3 are devoted to descriptions of different

t y pes of char t s and graphs. “ Chapter 2: His tograms ”
describes histograms, which are probably the easiest charts
to understand. Chapter 3 expands to other types of charts:
column charts, pie charts and dot charts.

Chapter 1: Introduction 9
 In Chapter 4, several basic statistics are introduced, including
means, medians, modes, variances and standard deviations.
Chapter 5 discusses analysis of tables, which is particularly
important since fire data often come to us as summaries in
the form of tables. In Chapter 6, correlations and variable
relationships are discussed. In these chapters, the goal is to
present how to perform the calculations associated with these
subjects as well as how to interpret the results.

In developing these chapters, we recognized that readers will

have varying backgrounds and capabilities. Therefore, while
a certain understanding of the principles behind the various
techniques is presented, in most cases a practical application
approach is used. The subject material becomes more difficult
as the handbook progresses. The first few chapters are easier
to understand. More technical subjects, such as chi-square
analysis and correlation, are more difficult and may require
knowledge of basic algebra. Even in these chapters, however,
emphasis has been placed on understanding results rather
than concentrating on theory.

We made every effort to simplify what can be a very complicated

topic. While there are many mathematical and statistical
symbols normally involved with the formulas and calculations
used in this handbook, none are used here. This is meant to be
a handbook, not a statistical textbook. It is written so anyone
can pick it up and be able to do basic statistical analysis of data.
For those who want more in-depth discussions of the subject
matter, a list of texts is included.

Books on statistics and data analysis

The following is a sampling of books on data analysis techniques
as well as some specific statistical topics handled or referred
to in this book. Most are basic or intermediate in scope, but all
have more detail than can be presented in this handbook.

The Art of Data Science: A Guide for Anyone Who Works with Data
by Roger Peng and Elizabeth Matsui (Skybrude Consulting, 2016).

Too Big to Ignore: The Business Case for Big Data (1st Edition) by
Phil Simon (John Wiley & Sons, Inc., 2013).

10 Fire Data Analysis Handbook

Practical Statistics for Data Scientists: 50+ Essential Concepts
Using R and Python (1st Edition) by Peter Bruce and Andrew
Bruce (O’Reilly Media, Inc., 2017).

Storytelling with Data: A Data Visualization Guide for Business

Professionals (1st Edition) by Cole Nussbaumer Knaflic ( John
Wiley & Sons, Inc., 2015).

The Data Detective: Ten Easy Rules to Make Sense of Statistics by

Tim Harford (Riverhead Books, 2021).

Envisioning Information (4th Edition) by Edward R. Tuf te

(Graphics Press, 1990).

Analyzing Tabular Data: Loglinear and Logistic Models for Social

Researchers by Nigel Gilbert (UCL Press, 1993).

Data Analysis: An Introduction by Michael S. Lewis-Beck (SAGE

Publications, Inc., 1995).

From Numbers to Words: Reporting Statistical Results for the

Social Sciences by Susan E. Morgan, Tom Reichert and Tyler R.
Harrison (Allyn and Bacon, 2002).

Misused Statistics (2nd Edition) by Herbert F. Spirer, Louise

Spirer and A. J. Jaffe (M. Dekker, 1998).

Say It With Charts: The Executive’s Guide to Visual Communication

(4th Edition) by Gene Zelazny (McGraw-Hill, 2001).

Schaum’s Outline of Theory and Problems of Beginning Statistics

by Larry J. Stephens (McGraw-Hill, 1998).

Sorting Data: Collection and Analysis by Anthony P. M. Coxon

(SAGE Publications, Inc., 1999).

Statistics (3rd Edition) by David Freedman, Robert Pisani and

Roger Purves (W. W. Norton & Co., Inc., 1997).

Statistics: Concepts and Applications by Amir D. Aczel (Irwin, 1995).

Statistics and Data Analysis: An Introduction (2nd Edition) by

Andrew F. Siegel and Charles J. Morgan ( John Wiley & Sons,
Inc., 1998).

Chapter 1: Introduction 11
 Statistics: The Exploration & Analysis of Data (4th Edition) by Jay
L. Devore and Roxy Peck (Brooks/Cole, 2001).

Your Statistical Consultant: Answers to Your Data Analysis

Questions by Rae R. Newton and Kjell Erik Rudestam (SAGE
Publications, Inc., 1999).

12 Fire Data Analysis Handbook

Chapter 2: Histograms
Data as a descriptive tool
“A picture is worth a thousand words” is an old saying that
applies to numbers as well as words. The task of reaching
conclusions from numbers is challenging, particularly when
we are looking for trends and patterns in the data. It is for this
reason that we turn our attention to histograms and other
charts in this chapter and Chapter 3. These tools will assist
you in understanding fire data, since the human mind seems
to comprehend pictures quicker than words and numbers.
However, before we delve into describing histograms and
other charts, we define the types of variables used to create
these charts.

Types of variables
For purposes of analysis, fire department variables can be
divided into 2 types: qualitative variables and quantitative
variables.

Qualitative variables are variables that are classified into

groups or categories. For example, fires can be broken into
structure fires, vehicle fires, refuse fires, explosions, etc.
Qualitative variables are also known as categorical variables
(data) since they are not measured in quantity but segregated
into groups. Examples of categorical data in the fire service
would include property use, cause of ignition, extent of flame
damage, etc. Most categorical variables that will be used in fire
data analysis will be found in the NFIRS modules.

Quantitative variables always take on numerical values that

reflect some type of measurement. Quantitative variables
can be discrete or exact or can be analog or continuous.
An example of a discrete variable would be the number of
days in the month or year (1 through 30 or 1 through 365),
but no fractions of days, whereas time, in hours, minutes,
seconds and infinite fractions of seconds, would be analog or
continuous. Other examples of quantitative variables would be
the number of fires in a district over a period of time (discrete),
the response time from alarm to arrival on the scene (analog),
and the dollar losses of fires (discrete).
Chapter 2: Histograms 13
 There is a distinction between a variable and data. A variable
is a characteristic that varies or changes. For example, days of
the week vary from Sunday through Saturday; months vary
from January through December; and types of fires vary, such
as structure fires, vehicle fires, residential fires, etc. A variable
is said to be independent if its variation does not depend on
another variable. Additionally, a variable is considered to be
dependent if its values depend on another variable.

Whenever observations are made on a variable, data are

created to be analyzed. Each time an NFIRS report is completed,
data for the variables listed are created. For example, by listing
the day of week, hour of day, month, type of situation found
and values for all other applicable variables in the NFIRS Basic
Module, data are created. The data then can be summarized in
a variety of ways, such as tables, graphs and charts.

The techniques used to summarize data found in Chapters 2

and 3 include:

Histograms Column Pie Pictograms

charts charts
Bar Line Dot
charts charts charts

This chapter describes histograms, while Chapter 3 is devoted

to the other techniques. With these graphic aids, we can
answer several basic questions. When are fires most likely to
occur? What are the primary causes of residential fires? Vehicle
fires? How many civilian injuries occurred last year by month?
What are the ages of civilian casualties? What percent of the
fire incidents have travel times less than 4 minutes? How many
structure fires resulted in dollar losses greater than $50,000?

A histogram is a column graph where the height of the

columns indicates the relative numbers, counts, frequencies
or values of a variable. The values may be numeric, such as
travel times, or nonnumeric, such as days of the week. The
column may also be used to show the relative frequency
(proportion) or percentage of counts in each category. The
relative frequency of a category is the count in a particular
category divided by the total count. The following examples
show how to organize and display fire data into histograms.

14 Fire Data Analysis Handbook

Example 1. One of the most fundamental ways to describe the
fire problem is to show how fires are distributed by month, day
of week and hour of day. Figure 2-1 shows a frequency list of
fires by hour of day for Canton, Ohio, for 1 year. A list or array
of numbers such as this is almost always the starting point for
a descriptive analysis, but the numbers by themselves are not
very useful. It is difficult to get a “feel” for what is happening
by scanning a list of numbers.

To grasp what the numbers say in Figure 2-1, we can develop a

frequency histogram, as shown in Figure 2-2. Similarly, Figures
2-3 and 2-4 show histograms by day of week and month of
year. Study these figures for a few minutes and draw your
own conclusions about what they represent. Don’t dwell on
individual numbers, but instead look for patterns. Ask yourself
3 questions:

1. Where are the low points and high points in the histogram?

2. What groups of times (hours, days or months) have similar

frequencies?

3. Is there anything in the histogram that runs counter to

your experience?

Figure 2-1. Fires by hour of day — Canton, Ohio

Time period Number Time period Number

Midnight - 1 a.m. 15 Noon - 1 p.m. 31
1 - 2 a.m. 15 1 - 2 p.m. 33
2 - 3 a.m. 13 2 - 3 p.m. 39
3 - 4 a.m. 13 3 - 4 p.m. 35
4 - 5 a.m. 13 4 - 5 p.m. 46
5 - 6 a.m. 11 5 - 6 p.m. 39
6 - 7 a.m. 16 6 - 7 p.m. 30
7 - 8 a.m. 11 7 - 8 p.m. 50
8 - 9 a.m. 17 8 - 9 p.m. 32
9 - 10 a.m. 17 9 - 10 p.m. 29
10 - 11 a.m. 19 10 - 11 p.m. 28
11 a.m. - Noon 19 11 p.m. - Midnight 24

Chapter 2: Histograms 15
 Answers to these questions provide the first insights into your
fire data and any conclusions drawn from it.
While these histograms suggest several conclusions, the key
ones are:
1. Canton has 2 distinct hourly patterns. The hours from noon
to midnight overall have almost twice the fires than the
hours from midnight to noon. The hours of 7 p.m. to 8 p.m.
and 4 p.m. to 5 p.m. have more fires than any other hours
in the day.
2. The fewest fires occur in the time period from 2 a.m. to
6 a.m. and from 7 a.m. to 8 a.m.
3. Sunday sees the most fires with a continuous decline until
Friday, which sees the fewest fires per day.
4. May has the most fires with June, July and November tied
for second. The fewest number of fires occur in February.
With these histograms, we begin to see a picture of the fire
problem in Canton. Histograms allow for an easy descriptive
and analytical procedure without having to think too much
about the numbers themselves. Graphical displays should
always strive to convey an immediate message describing a
particular aspect of the data.

Figure 2-2. Fires by hour of day — Canton, Ohio

60
Number of incidents

50
50 46
39 39
40 35
31 33 30 32
29 28
30 24
20 1515 16 17 17 1919
13 13 13 11 11
10
0
2 - 3 p.m.

5 - 6 p.m.
6 - 7 p.m.
7 - 8 p.m.
8 - 9 p.m.
9 - 10 p.m.
10 - 11 p.m.
11 p.m. - Midnight
Midnight - 1 a.m.
1 - 2 a.m.
2 - 3 a.m.
3 - 4 a.m.
4 - 5 a.m.
5 - 6 a.m.
6 - 7 a.m.
7 - 8 a.m.
8 - 9 a.m.
9 - 10 a.m.
10 - 11 a.m.
11 a.m. - Noon
Noon - 1 p.m.
1 - 2 p.m.

3 - 4 p.m.
4 - 5 p.m.

Hour of day

16 Fire Data Analysis Handbook

 Figure 2-3. Fires by day of week — Canton, Ohio
100
Number of incidents 96
91 90
90 87
83
80 76

70 68

60
Sunday

Monday

Tuesday

Wednesday

Thursday

Friday

Saturday
Day of week

Figure 2-4. Fires by month — Canton, Ohio

80 74
Number of incidents

70
61 61 61
60
50 46 47 45 47 44
43
40 36
30 68
30
20
Jan Feb Mar Apr May Jun Jul Aug Sep Oct Nov Dec
Month of year

Example 2. Ages of civilian casualties. Suppose a fire chief

is interested in developing a fire prevention program aimed
at reducing civilian injuries and deaths. Descriptive data on
civilian casualties is available from the NFIRS reports, and
there are a number of different descriptions that could be
developed from the data. One of the most basic is descriptive
data on the ages of civilian casualties.

Figure 2-5 shows the ages of civilians injured or killed in fires in

Denver, Colorado, for 1 year. This distribution is considerably
different from the previous histograms primarily because it
does not have the same “smoothness.” However, the 5-year

Chapter 2: Histograms 17
 age groups show some interesting patterns. For example, the
age group 36 to 40 accounts for the most civilian casualties,
followed in frequency by 26 to 30 and 46 to 50, respectively.

Also of interest is how the frequency takes a rather sudden

drop for the 16 to 20 and 56 to 60 age groups. Spikes in the
data occur at the 26 to 30 and 36 to 40 age groups.

The figure also reveals several gaps in the data for ages 6 to 10
and 76 to 80 as a result of no reported casualties in these age
groups. Due to these gaps at either end of the distribution, 2
outliers are created in the under 5 and 81 to 85 age groups.

Figure 2-5. Ages of civilian casualties — Denver, Colorado

14
12
Number of casualties

12
10 9
8
8 7
6 6
6 5 5 5
4 3 3
2 2 2 2
2
0
36-40
41-45

56-60
61-65
66-70
71-75
81-85
46-50
51-55
1-5
11-15
16-20
21-25
26-30
31-35

Ages of civilian casualties

Notes: Age was not provided for 7 casualties. 52% of the casualties were between 26 and
50 years old.

Spikes are high or low points that stand out in a histogram.

Gaps are spaces in a histogram reflecting low frequency of data.
Outliers are extreme values isolated from the body of data.

In histograms and other charts, it is sometimes useful to

include comments and conclusions with the chart. In Figure
2-5, a note was provided that 7 casualty records did not
include age information and were therefore not included in
the histogram. Other notes provide summary information on
the data such as the percent of casualties between the ages 26
and 50 years old. Anyone studying the histogram could reach
the same conclusion, but the summary saves time and effort.

18 Fire Data Analysis Handbook

Example 3. Response times to f ires. Response times to
fires are one of the most important data sets to study in fire
departments. Many fire departments have objectives for average
response times to fires and try to allocate personnel to achieve
these response times. Figure 2-6 shows a frequency distribution
for response times to fires in Boston, Massachusetts.

Figure 2-6. Response times — Boston, Massachusetts

Response time Frequency

Less than 1 minute 129
1 to 2 minutes 206
2 to 3 minutes 759
3 to 4 minutes 1,406
4 to 5 minutes 1,312
5 to 6 minutes 747
6 to 7 minutes 384
7 to 8 minutes 206
8 to 9 minutes 110
9 to 10 minutes 62
10 to 11 minutes 18
11 to 12 minutes 15
12 to 13 minutes 15
13 to 14 minutes 10
14 to 15 minutes 5
15 to 16 minutes 6
16 to 17 minutes 2
17 to 18 minutes 0
18 to 19 minutes 1
19 to 20 minutes 1
Total fire calls 5,394

Notice in this example that the times are clustered at the low
end of the distribution as would be expected since response
times to fires are generally low for most fire departments.

Chapter 2: Histograms 19
 Figure 2-7 provides a frequency histogram for this distribution.
In this figure, we have combined the last few points into a
category of 10 minutes or more. A histogram with the same
shape as in this figure is said to be skewed to the right or
skewed toward high values. What is meant by these terms
is that the distribution is not symmetrical, but instead has
a single peak on the left side of the distribution with a long
tail toward the right. In fire departments, on-scene time data
(from time of arrival to time back in service) and fire dollar loss
data also reflect values skewed to the right.

Figure 2-7. Response times — Boston, Massachusetts

1,600
1,406
Number of incidents

1,400 1,312
1,200
1,000
800 759 747
600
384
400
206 206
200 129 110 62 73
0
<1 1-2 2-3 3-4 4-5 5-6 6-7 7-8 8-9 9-10 >10

Number of minutes

Developing a histogram
Making a histogram is relatively straightforward:

1. Choose the number of groups for classifying the data. In

most cases, 5 to 10 groups are sufficient, but there are
exceptions, such as histograms by hour of day. Sometimes
the groups are natural, as in our exhibits by day of week and
month. With other data, developing appropriate intervals
will be necessary as was done in Figure 2-5 with the ages
of civilian casualties.

2. Determine the number of events (fires, casualties, etc.) for

each of the groups.

3. For data such as ages and response times, intervals usually

need to be defined. For these intervals, convenient whole
numbers should be chosen. That is, try to avoid the use of

20 Fire Data Analysis Handbook

 fractions in the groups and always make the intervals the
same width. In Figure 2-5, intervals of 5 years were used
for grouping the data. Data such as day of week do not
require this step since their intervals are naturally defined.

4. Determine the number of observations in each group.

Statistical software packages are particularly useful in this
step since they usually include routines for tabulating data.

5. Choose appropriate scales for each axis to accommodate

the data. Again, most statistical packages will do this with
a default setting.

6. Display the frequencies with vertical bars.

Do not expect to get a histogram — or any other type of chart —

exactly right on the first try. You may need several tries before
you get a satisfactory histogram.

The histograms presented in the previous section offer good

examples of different characteristics for describing the data.
In Beginning Statistics with Data Analysis, a text by Mosteller et
al. (1983), the following definitions of histogram characteristics
are presented:

1. Peaks and valleys. The peaks and valleys in a histogram

indicate the values that appear most frequently (peaks) or
least frequently (valleys). Figure 2-2 shows clear peaks and
valleys for incidents by hour of day.

2. Spikes and holes. These are high and low points that stand
out in the histogram. In Figure 2-5, for example, there is a
spike for the 36 to 40 age group and a hole for the 16 to
20 age group.

3. Outliers. Extreme values are sometimes called outliers

and are points that are isolated from the body of the data.
In Figure 2-5, there are 2 outliers: the under 5 and the 81
to 85 age groups.

4. Gaps. Spaces may reflect important aspects of a histogram.

In Figure 2-5, there are gaps in the 6 to 10 and the 76 to 80
age groups.

Chapter 2: Histograms 21
 5. Symmetry. Sometimes a histogram will be balanced along
a central value. When this happens, the histogram is easier
to interpret. The central value is both the mean (average)
for the distribution and the median (half the data points
will be below this value and half above).

Cumulative frequencies
There are 2 other t ypes of distributions which will be
important in later chapters: the cumulative frequency
and the cumulative percentage frequency. A cumulative
frequency is the number of data points that are less than or
equal to a given value. A cumulative percentage frequency
converts the cumulative frequency into percentages.

Example 4. With the data in Figure 2-6, we can calculate the

cumulative frequency and cumulative percentages for the
response time data from Boston, Massachusetts, found in
Figure 2-8.

Figure 2-8. Cumulative response times — Boston,

Massachusetts

Cumulative Cumulative
Response Frequency
frequency percent
Less than 1 minute 129 129 2.4
1 to 2 minutes 206 335 6.2
2 to 3 minutes 759 1,094 20.3
3 to 4 minutes 1,406 2,500 46.3
4 to 5 minutes 1,312 3,812 70.7
5 to 6 minutes 747 4,559 84.5
6 to 7 minutes 384 4,943 91.6
7 to 8 minutes 206 5,149 95.5
8 to 9 minutes 110 5,259 97.5
9 to 10 minutes 62 5,321 98.6
10 or more minutes 73 5,394 100.0
Total 5,394 100.0

22 Fire Data Analysis Handbook

The first entry under the “cumulative frequency” column is
129, which is the same as in the “frequency” column. The
second entry shows 335, which is 129 + 206, the sum of the
first 2 entries in the “frequency” column. By adding these
2 numbers, we can say that 335 incidents have response
times less than 2 minutes. The next entry is 1,094 (129 + 206 +
759) and means that 1,094 incidents have response times less
than 3 minutes. The cumulative frequencies continue in this
manner with the last entry in the column always equal to the
total number of incidents in the analysis.

The last column, labeled “cumulative percent,” merely converts

the cumulative frequencies into percentages. This step is
accomplished by dividing each cumulative frequency by 5,394,
which is the total number of incidents. The column shows that
2.4% of the incidents have response times less than 1 minute,
6.2% less than 2 minutes, 20.3% less than 3 minutes, etc.

In general, cumulative percentages describe data in “more than”

and “less than” terms. We can conclude, for example, that about
half the calls have response times of less than 4 minutes and
about 95% have response times less than 8 minutes. Response
times exceed 10 minutes in only about 1% of the calls.

Summary
A list of numbers is frequently the starting point for analysis.
If the question of interest is for specific information, then the
list of numbers serves the purpose. For example, Figure 2-1 is
useful if we are asked about how many fires occurred between
2 and 3 a.m., or if we want to know the exact difference
between the busiest and the least busy hour. On the other
hand, Figure 2-1 is not very useful for determining the 6 busiest
hours of the day.

Histograms provide a much better method for getting the feel

of a list of numbers and answering several questions about
relationships. The patterns in a histogram are especially
important, such as high and low frequencies and trends
indicated by spikes, outliers and gaps. Histograms give quick
graphic representations of the data that otherwise would be
hidden and hard to dig out of a table of numbers.

Chapter 2: Histograms 23
24 Fire Data Analysis Handbook
Chapter 3: Charts
Introduction
In this chapter, we extend beyond histograms to other types
of charts. Histograms are only one of many different ways of
presenting data. As an analyst, you must decide which type
of chart best portrays the results you want to represent. A
histogram may serve as the best vehicle in some cases, but
other types of charts should be considered, such as bar charts,
line charts, pie charts, dot charts and pictograms. Each of
these types of charts are discussed in this chapter.

There are 2 questions to keep in mind throughout this process:

1. What are the main conclusions from your analysis?

2. What is the best way to display the conclusions?

As with the previous chapter, several sets of real fire data

are presented. You should study each example carefully and
draw your own conclusions about the results. You may, in fact,
disagree with what the book emphasizes, or you may identify
an aspect of the data that was overlooked. In either case, the
point is to think about how you would present your viewpoints
in a graphical format to a given audience. The audience may
be an internal group of managers, an outside association or
group of citizens, or even your own city or county council. The
audience itself influences the type of chart that is selected.

Therefore, the first step is to determine the key results from

the data. Once they have been identified, select the best type
of chart to convey them. Often it is helpful to try different
charts to determine the best presentation for a particular
audience and data set.

Each of the following sections describes a different type of

chart. At the end of the chapter, you will find guidelines on
selecting a type of chart suitable for different conclusions.

Chapter 3: Charts 25
 Bar charts
A bar chart is one of the simplest and most effective ways to
display data.

In a bar chart, a bar is drawn for each category of data allowing

for a visual comparison of the results. For example, the figures
in Figure 3-1 give the causes of ignition (from NFIRS 5.0 codes)
for 12,600 structure fires reported in Chicago, Illinois, for 1 year.

Interest in a list of this type usually centers on how the items

compare to each other. What is the leading cause of ignition
in structure fires? How do unintentional causes compare to
intentional ones? How many causes are never determined?

Some results can be determined relatively easily from the

list of numbers. For example, “cause undetermined after
investigation” is clearly the leading cause of ignition followed by
“intentional,” “equipment failure” and “unintentional,” all close
in number. The remaining 3 — “not reported,” “other” and “act
of nature” — account for less than 1% combined. While these
comparisons can be made from the list, they require mental
manipulations and are not easily made or retained in full.

Figure 3-1. Cause of ignition for structure fires —

Chicago, Illinois

Cause of ignition Number Percent

Intentional 2,771 22.0
Unintentional 2,583 20.5
Equipment failure 2,654 21.1
Act of nature 29 0.2
Cause, other 40 0.3
Not reported 55 0.4
Cause undetermined
4,468 35.5
after investigation
Total 12,600 100.0

26 Fire Data Analysis Handbook

A bar chart overcomes these problems by presenting the data
in frequency order as displayed in Figure 3-2. The horizontal
dimension gives the percent, while the vertical dimension
shows the category labels. The bars are presented in numerical
order, starting with “undetermined” as the most frequent.
Each bar also contains the number of fires for that cause of
ignition as additional information to the reader.

It should also be noted that the categor y “cause under

investigation” had no cases reported, but this fact is mentioned
in a footnote since it is a listed option in the NFIRS module.
Also in a footnote are the complete titles of 2 of the categories
that were abbreviated in the table listing.

As a general rule, the horizontal dimension in a bar chart is

numeric, such as percentages or other numbers, while the
vertical dimension shows the labels for the items in a category.
It is not always necessary to include numbers in each bar,
especially if there is an accompanying table or list, but they can
be useful to readers unfamiliar with the data. If the numbers
are omitted from the chart, a total number should be provided
either in the title or a footnote.

Figure 3-2. Cause of ignition for structure fires —

Chicago, Illinois
Undetermined
*
4,468
Intentional 2,771
Equipment failure
**
2,654
Unintentional 2,583
Not reported 55
Cause - other 40
Act of nature 29
0 10 20 30 40
Percent
*”Undetermined after investigation.”
**Or “heat source failure.”
Note: No cases reported under “cause under investigation” category.

Chapter 3: Charts 27
 A clustered bar chart shows 2 categories in the same chart.
In Figure 3-3, for example, the causes of ignition for structure
fires in Chicago are shown in a residential versus nonresidential
format. The figure shows that fires that are undetermined
after investigation comprise over 40% of the nonresidential
fires and only 16% of the residential ones. Interestingly, the
chart also shows an almost exact ratio of 40% and 15% for
unintentional causes of residential and nonresidential fires,
respectively. In addition, while the percentages are close for
residential and nonresidential fires under the “unintentional”
and “equipment failure” categories, the numbers differ by 3
to 4 times due to the large difference in total fires between
residential and nonresidential. The clustered or paired bar
chart clearly shows the differences in ignition causes for these
2 types of structure fires.

Figure 3-3. Cause of ignition residential versus

nonresidential fires — Chicago, Illinois
410
Undetermined 4,058
637
Intentional 2,134
488
Equipment failure 2,166
1,041
Unintentional 1,522
Not reported 18
37
Cause - other 9 Residential
31
16 Nonresidential
Act of nature 13
0 10 20 30 40 50
Percent

Column charts
In Chapter 2, several column charts were displayed. For
example, Figures 2-2, 2-3 and 2-4 showed Canton, Ohio fires
by hour of day, day of week and month, respectively. These
are all examples of time series presented as column charts.

28 Fire Data Analysis Handbook

 Column char t s of this t y pe are par ticularl y usef ul in
demons trating change over time. Where is the series
increasing, decreasing or staying about the same? If the
analysis shows change over time, then column charts are
particularly beneficial in presenting the changes.

As an example, the figure from Chapter 2 on fires by hour of

day is repeated in Figure 3-4. By looking from left to right, you
can visualize the change. The horizontal scale shows the hours,
but it is not really needed to be able to see the overall changes.
The numbers of reported fires are low in the early morning
hours, then increase in the afternoon and evening hours.

Figure 3-4. Fires by hour of day — Canton, Ohio

60
Number of incidents

50
50 46
39 39
40 35
31 33 30 32
29 28
30 24
1919
20 1515 16 17 17
13 13 13 11 11
10
0
2 - 3 p.m.

9 - 10 p.m.
10 - 11 p.m.
11 p.m. - Midnight
Midnight - 1 a.m.
1 - 2 a.m.
2 - 3 a.m.
3 - 4 a.m.
4 - 5 a.m.
5 - 6 a.m.
6 - 7 a.m.
7 - 8 a.m.
8 - 9 a.m.
9 - 10 a.m.
10 - 11 a.m.
11 a.m. - Noon
Noon - 1 p.m.
1 - 2 p.m.

3 - 4 p.m.
4 - 5 p.m.
5 - 6 p.m.
6 - 7 p.m.
7 - 8 p.m.
8 - 9 p.m.

Hour of day

Column charts show frequency distributions that allow for easy

identification of trends and other characteristics, particularly
with time series data. The horizontal scale defines the natural
groupings for the chart and the columns give the frequencies.

Another good application of column char ts is to show

comparisons across sets of data. Figure 3-5 lists the causes
of ignition from Figure 3-3. Due to their small numbers for
illustrative purposes, the “not reported,” “causes - other” and
“act of nature” categories have been combined into “other.”
Comparisons between the venues are not easy because the

Chapter 3: Charts 29
 totals differ so much. Nonresidential fires total just under
10,000 while residential fires have 2,619. A simple way to
overcome this problem is to develop percentages.

By converting the residential and nonresidential figures to

percentages, as shown at the bottom of the figure, you can
make a better comparison. The percentages for both add up to
100%. While there are many conclusions that could be drawn
from these percentages, the key ones are:

ĵ “Intentional,” “equipment failure” and “other” account

for about the same percentages in both residential and
nonresidential fires.

ĵ “Unintentional” fires account for 40% of the residential

fires, while 41% of the nonresidential fires fall into the
“undetermined ” category.

Figure 3-5. Comparison of causes of ignition in residential

versus nonresidential fires — Chicago, Illinois

Cause of ignition Residential Nonresidential

Intentional 637 2,134
Unintentional 1,041 1,542
Equipment failure 488 2,166
Cause undetermined
410 4,058
after investigation
Other 43 81
Total 2,619 9,981

Cause of ignition Residential Nonresidential

Intentional 24.3% 21.4%
Unintentional 39.7% 15.4%
Equipment failure 18.6% 21.7%
Cause undetermined
15.7% 40.7%
after investigation
Other 1.6% 0.8%
Total 100.0% 100.0%

30 Fire Data Analysis Handbook

To display this result, stacked column charts were developed,
as shown in Figure 3-6, using the percentages for each cause
of ignition. The columns have the same height since they both
total 100%. The colors highlight the differences among the
causes of ignition. The results just discussed should be clear
from the figure.

Figure 3-6. Comparison of causes of ignition by

percent — Chicago, Illinois
120
Undetermined
100 Other
80 Equipment
failure
60
Unintentional
40
Intentional
20
0
Residential Nonresidential
Note: “Other” causes include “not reported” and “act of nature.”

Line charts
Ef fec tive presentation of time series data also may be
developed from line charts. Figure 3-7 shows a line chart of
fires by hour of day for Canton, Ohio, previously displayed as
a histogram in Figure 2-2. The line chart immediately highlights
the jump in fires from a sharp rise in the early afternoon until a
peak at around 8:00 p.m. Many statisticians believe that a line
chart is the clearest way for showing increases, decreases and
fluctuations in a time series.

Chapter 3: Charts 31
 Figure 3-7. Fires by hour of day — Canton, Ohio
60
Number of incidents
50
40
30
20
10
0

6 - 7 p.m.
7 - 8 p.m.
8 - 9 p.m.
9 - 10 p.m.
10 - 11 p.m.
11 p.m. - Midnight
Midnight - 1 a.m.
1 - 2 a.m.
2 - 3 a.m.
3 - 4 a.m.
4 - 5 a.m.
5 - 6 a.m.
6 - 7 a.m.
7 - 8 a.m.
8 - 9 a.m.
9 - 10 a.m.
10 - 11 a.m.
11 a.m. - Noon
Noon - 1 p.m.
1 - 2 p.m.
2 - 3 p.m.
3 - 4 p.m.
4 - 5 p.m.
5 - 6 p.m.
Hour of day

Pie charts
A pie char t is an ef fec tive way of showing how each
component contributes to the whole. In a pie chart, each
wedge represents the amount for a given category. The entire
pie chart accounts for all of the categories.

For example, Figure 3-8 shows the causes of ignition for

structure fires in the Chicago Fire Department for 1 year
divided into “undetermined,” “equipment failure,” “intentional,”
“unintentional” and “other.” The percentages are included
with each wedge label. Although the percentage numbers are
not necessary, they aid in comparisons of the wedges. The
pie chart emphasizes the fact that the largest percentage
of fire causes is undetermined. In addition, “intentional,”
“unintentional” and “equipment failure” all account for about
the same percent of the causes.

32 Fire Data Analysis Handbook

 Figure 3-8. Cause of ignition for fires — Chicago, Illinois

Intentional
22.0%
Undetermined
35.5%

Unintentional
20.5%

Other
1.0%

Equipment failure
21.1%

In developing pie charts, you should follow these rules:

ĵ Convert data to percentages.
ĵ Keep the number of wedges to 6 or less. If there are more
than 6, keep the most important 5 and group the rest into
an “other” category.
ĵ Position the most important wedge starting at the 12 o’clock
position.
ĵ Maintain distinct color differences among the wedges.
While pie charts are popular, they are probably the least
effective way of displaying results. For example, it may be hard
to compare wedges within a pie chart to determine their rank.
Similarly, it takes time and effort to compare several pie charts
because they are separate figures.

Dot charts
Dot charts or scatter diagrams emphasize the relationship
between 2 variables. For example, the 10-year trend in other
residential fires over a decade was generally a decrease from
a high of 15,000 to a low of 11,000. During these years, a
decrease in fire deaths also occurred. You expect deaths to
decrease with a decrease in fires. This relationship is depicted
in Figure 3-9.

Chapter 3: Charts 33
 Figure 3-9. Fires and deaths
120
Outlier

100 Year 5 Year 1

80
Deaths

Year 3
Year 2
60 Year 7 Year 4

Year 8
40 Year 6
Year 9 Year 10
20
10 11 12 13 14 15 16
Fires (thousands)

The figure is a dot chart for fires versus deaths for the 10 years.
Fires are along the horizontal or x-axis, while deaths are along
the vertical or y-axis. The pattern is the important aspect of
a dot chart, rather than the individual dots. The horizontal
scale (x-axis) should reflect the independent variable, while
the vertical scale reflects the dependent variable. That is to say
that a decrease in fires (the independent variable) has a positive
correlation with a decrease in fire-related deaths (the dependent
variable). Recall from earlier in this text that a variable is said
to be independent if its variation does not depend on another
variable. Additionally, a variable is considered to be dependent
if its values depend on another variable.

Another useful application of scatter diagrams is to identify

outliers in the data. In Chapter 2, outliers were defined as
points that are isolated from the body of the data. In Figure
3-9, there is a general pattern showing a decrease in deaths
over time as fires decrease. While the decrease in fires pattern
is maintained for Year 5, however, deaths rise to the highest
count over the 10-year period. Therefore, Year 5 has many
more fire-related deaths than expected based on its amount
of fires. This outlier from the general pattern can be useful in
revealing an area of further data analysis that would account
for this discrepancy from the rest of the data.

34 Fire Data Analysis Handbook

Pictograms
The final type of chart takes advantage of pictures to display
data. Data by geographical areas, such as counties, census
tracts or fire districts can be presented on maps showing
the boundaries of the areas. Figure 3-10, for example,
shows firefighter deaths by region for 1 year. Each region is
broken down by career, volunteer, and, if applicable, wildland
department.

Figure 3-10. Firefighter deaths by region

West
Career - 8 North Central
Volunteer - 2 Career - 5
Wildland - 5 Volunteer - 15
Total - 15 Total - 20

Northeast
Career - 5
Volunteer - 25
Wildland - 2
Total - 32

South
Career - 14
Volunteer - 13
Total - 27

The key is that presentation in this manner is more effective

than any listing of the death rates. It can be easily seen that:
ĵ Career deaths in the South are 2 to 3 times more than in
other regions.
ĵ Volunteer deaths in the Western region are a fraction of
those in the rest of the country.
ĵ The Northeast has the most total deaths largely due to a
high number of volunteer deaths that is almost double the
next largest region.
You can easily imagine other pictograms for state and local
data. At the state level, data from individual counties may be
collected. A pictogram provides a good way of depicting the

Chapter 3: Charts 35
 county data by taking a state map showing county boundaries
and developing a figure similar to Figure 3-10. Similarly, for a
local jurisdiction, such as a city or a county, there may be data
for individual fire districts. A jurisdiction map with fire district
boundaries may be an effective way of presenting the data.

Summary
In this chapter, 6 t ypes of char ts were presented: bar
charts, column charts, line charts, pie charts, dot charts and
pictograms. The primary purpose of using any chart is to
indicate conclusions more quickly and clearly than is possible
with tables or numbers. It may be necessary to try several
types of charts before the most appropriate one is found,
but in a chart, simplicity is the key. The message is what is
important, so the chart form should not interfere with it.
As a quick reference guide on chart selection, the following is
recommended:
ĵ Use a bar chart with categorical data when the objective
is to show how the items in a category rank. Most fire data
are categorical, such as cause of ignition, property use,
area of origin, nature of injury, etc. These are reflected in
the NFIRS modules.
ĵ Use a column or line chart for data with a natural order,
such as hours, months or age groups. The chart will reflect
the general pattern and indicate points of special interest,
such as spikes, holes, gaps and outliers.
ĵ A pie chart is beneficial when the objective is to show how
the components relate to the whole. It is recommended that
the number of components be kept to 6 or less and that
the forming of several pie charts for comparison purposes
be avoided.
ĵ A dot chart or scatter plot depicts the relationship between
2 variables. Generally, these variables are continuous rather
than categorical. The pattern between the 2 variables is the
important aspect for a dot chart.

ĵ A pictogram is a pictorial representation of the data.

Breakdowns by geographic areas, for example, are
effectively shown by a pictogram.

36 Fire Data Analysis Handbook

Chapter 4: Basic Statistics
Data can be summarized in a variety of ways by using tables,
graphs and charts. In this chapter, ideas about summarizing
data will be ex tended by introducing basic descriptive
measures to include measures of central tendency (i.e., mode,
mean and median) as well as measures of dispersion (i.e.,
range, variance and standard deviation).

Measures of central tendency

Measures of central tendency provide a single summar y
figure that best describes the central location of an entire
distribution. The 3 most common measures of central
tendency are the mode, the mean and the median. Each of
the measures are defined, and the individual properties and
uses for each measure are also discussed.

The mode is the value that occurs most frequently in a

distribution. It is, therefore, easily recognized since no
calculations are necessary.

The mean is also known as the arithmetic mean or average.

However, since the term average is sometimes used
indiscriminately for any measure of central tendency, it
should be avoided. It is defined as the sum of all values in
a distribution divided by the total number of values. For
example, suppose that travel times to 9 incidents are 3
minutes, 2 minutes, 4 minutes, 1 minute, 2 minutes, 3 minutes,
3 minutes, 4 minutes and 3 minutes. Adding these travel times
gives 25 minutes in total, and dividing by 9 yields a mean travel
time of 2.78 minutes.

The third measure of central tendency is the median, which is

defined as the middle value (50th percentile) of a distribution.
To determine the median, the data must be ordered. Using
the 9 travel times from the above example, they would look as
follows if arranged in order: 1, 2, 2, 3, 3, 3, 3, 4, 4. The median
is the fifth, or middle, value, which is 3 minutes. There are 4
data values below and 4 data values above. In other words,
50% of the values lie on either side of the median, placing it at
the 50th percentile.

Chapter 4: Basic Statistics 37

 If there had been an even number of data values, then the
median would have been the mean of the 2 middle values.
For example, if the on-site times for 10 fire incidents were 12,
15, 17, 25, 27, 29, 32, 35, 37 and 42 minutes, then the 2 middle
values would be 27 and 29. Totaling them and dividing by 2
(calculating the mean value) results in a median value of 28.
Again, the median splits the values with 5 below and 5 above.

Properties and uses for measures of central

tendency
The mode is the only measure of central tendency that can be
used for qualitative data. This is really its only redeeming quality
other than to serve as an additional qualifier for a distribution.
The mode by itself is an unstable measure of central tendency.
Equal size samples taken from a distribution are likely to have
different modes. Further, on many occasions, distributions have
more than 1 mode (bimodal), which adds to the confusion.

The median is a better choice than the mode for a measure

of central tendency. Unlike the mode, it cannot be used with
qualitative data but with quantitative variables. The median
on-scene time for fires or the median dollar loss for fires can be
determined. However, the “median type of fire” or the “median
cause of ignition” has no meaning since these are qualitative
variables. Responding to how many values lie above and below,
but not to how far away, the median is less sensitive than the
mean to the presence of a few extreme values.

Generally, the mean is the most frequently used measure of

central tendency. Unlike the mode and the median, the mean is
responsive to the exact position of each value in a distribution.
It serves as a fulcrum point, balancing all of the values in a
distribution. Consequently, the mean is very sensitive to
extreme values (outliers) in a distribution.

For data from skewed distributions, the use of the median is

a better choice than the mean because it is not influenced by
large outliers. When a measure of central tendency needs to
reflect the total of the values, the mean is the best choice since
it is the only measure based on this quantity. Another of the
more important characteristics of the mean is its stability over

38 Fire Data Analysis Handbook

samples drawn from a distribution. This becomes especially
important when further statistical computation is done.

Measures of dispersion
While measures of central tendency provide a summary of
the values in a distribution, measures of dispersion provide
a summary of the variability or spread of the values in a
distribution. Measures of dispersion express quantitatively
the extent to which the values in a distribution scatter about
or cluster together. The 3 main measures of dispersion are the
range, variance and standard deviation. As with the measures
of central tendency, they will first be defined and then their
properties and uses will be discussed.

The range is the most basic measure of dispersion. Its

definition is simply the difference between the lowest and
highest value in a distribution. For example, with the 10 on-site
times used in the median discussion, the lowest value is 12
minutes and the highest is 42 minutes. Therefore, the range
is 30 minutes.

Another measure of the variability of a distribution is the

variance. To calculate the variance, it is necessary to first
obtain what is known as the deviation values of a distribution.
The deviation values are the difference between the values
in a distribution and its mean. Since the mean is the balance
point of the values in a distribution, the total of the deviation
values would be 0. Therefore, to calculate the variance, it is
necessary to square the deviation values to eliminate the
negative values.

To illustrate the calculation of a variance, the 9 travel times

used in the example for the mean are used. In Figure 4-1, the
mean of 2.78 has been subtracted from each individual travel
time and the result squared.

Chapter 4: Basic Statistics 39

 Figure 4-1. Calculation of variation

Travel time - mean

Travel time Squared
(2.78)
1 -1.78 3.17
2 -0.78 0.61
2 -0.78 0.61
3 0.22 0.05
3 0.22 0.05
3 0.22 0.05
3 0.22 0.05
4 1.22 1.49
4 1.22 1.49
Total 0.00 7.57
Variance .95

The middle column displays the amount of deviation from

the mean for each point. The first deviation is -1.78 (1 minute
minus 2.78 minutes), indicating that this travel time is 1.78
units from the mean and is to the left of the mean (since the
sign is negative). Note that the sum of the middle column is 0;
that is, the sum of the deviations from the mean is 0. In fact, an
alternative definition for the mean is that it is the only number
with this property.

In the right column is the square of each deviation. The sum

of the squared deviations is 7.57, and the variance is obtained
by dividing this sum by 8, which is 1 less than the total number
of values. The reason for subtracting 1 from the total number
of values will be discussed shortly. The variance from this
calculation is then 0.95. Since the variance is small compared
to the mean, it indicates that the values are close to the mean.

In discussing the calculation of the variance, the sum of the

squared deviations was divided by the number of values
minus 1. This was done to correct for a statistical error that
results when using inferential statistics. If the distribution is
the entire amount of values being considered, then dividing by
that number is perfectly legitimate. However, if the distribution
is merely a sample of a larger distribution, which it usually is,

40 Fire Data Analysis Handbook

then a better representation of the entire population of values
can be obtained by subtracting 1 from the sample distribution.

The final measure of dispersion is the standard deviation. It

is obtained by taking the square root of the variance. In the
current example, the standard deviation is 0.97 since this is
the square root of 0.95. This means that the spread (variability)
around the mean is not very large (in this case less than 1.0
compared to a mean travel time of 2.78 minutes). Therefore,
the mean is a good descriptor of the data in this example.

Normal distribution and standard score

Unless there is a compelling reason otherwise, statisticians
usually assume a normal distribution for any given set of
values. As shown in Figure 4-2, a normal distribution is equally
spread out in the general shape of a bell. In fact, it is known
as the bell curve. In a normal distribution, the mean, the
median and the mode are the same. Half the values are above
the mean and half below. Most of the values, 68%, fall within
1 standard deviation on either side of the mean; within 2
standard deviations, 95%; and within 3 standard deviations,
99.7% of the distribution is represented.

Figure 4-2. Normal distribution

High Mean
Relative frequency

-1 standard +1 standard
deviation deviation

Low
-4 -3 -2 -1 0 1 2 3 4

68.26%
95.44%
99.74%

Chapter 4: Basic Statistics 41

 By using a standard score, it is possible to compare values
from different distributions on an equal basis. A standard
score is a derived score that describes how far a given value
in a distribution is from some reference point, typically the
mean, in terms of standard deviation units. One of the most
commonly used standard scores is the z-score. Transforming
the values of a distribution to z-scores changes the mean to
0 and the standard deviation to 1, but does not change the
shape of the distribution. For example, in the travel times used
in Figure 4-1, a z-score of 1 would be equivalent to a score of
3.75 minutes. That would be calculated by adding the mean of
2.78 to the standard deviation of 0.973. In another distribution
of travel times with a different mean and standard deviation,
a z-score of 1 would be totally different. However, using the
z-scores, they could be compared equally without distorting
the original distributions.

Properties and uses for measures of dispersion

The range is ideal for preliminary work or in circumstances
where precision is not an important requirement. However, it is
not sensitive to the total condition of the distribution since only
the 2 outermost values determine its calculation. Therefore, the
range is of little use beyond the descriptive level.

Since the variance is the mean of the squares of the deviation

values of a distribution, it is responsive to the exact position
of each value in a distribution. It can, therefore, be ver y
important in inferential statistics because of its resistance to
sampling variation. However, it is of little use in descriptive
statistics because it is expressed in squared units.

The standard deviation, like the mean and the variance (from
which it is derived), is responsive to the exact position of
every value in a distribution. Because it is calculated by using
deviations from the mean, the standard deviation increases or
decreases as the individual values shift away from or toward
the mean. Like the mean, it is influenced by extreme scores,
especially with distributions that have a small amount of values.
As the number of values increase, each individual value has
less ability to shift the mean and the standard deviation. If the
mean and the standard deviation of a distribution are known,
a fairly accurate picture of the distribution can be obtained.

42 Fire Data Analysis Handbook

Once again, using the travel time example from Figure 4-1, the
mean is 2.78 and the standard deviation is 0.973. Assuming a
normal distribution, 1 standard deviation from the mean in
both directions should cover 68% of the values. In this case,
values between 1.807 and 3.753 include 2, 2, 3, 3, 3 and 3. Since
there are 9 values in the distribution, the 6 values that fall
within 1 standard deviation from the mean account for 67%.
Considering its small size, that is an extremely accurate picture
of the distribution. It is also a good example of how powerful
the combination of the mean and the standard deviation
can be. Each are the best measures of their type (central
tendency and dispersion) and both are used extensively in
more sophisticated statistical calculations.

Skewed distributions
Even though statisticians assume a normal distribution based
on a first impression, not all distributions are normal or
symmetrical. As stated before, in a normal distribution, the
mean, median and mode are all the same. However, this is
not the case with skewed distributions. As shown in Figure
4-3, distributions can be skewed positively or negatively. In a
positive skew, the extreme scores are at the positive end of
the distribution. This exhibits the “tail” on the right side and
pulls the mean to the right. Since the median and the mode
are less responsive to extreme scores, they remain to the left
of the mean. So, in a positively skewed distribution, the mean
has the highest value with the median in the middle, and the
mode is the smallest value. Conversely, in a distribution with
a negative skew, the extreme values and the “tail” are at the
negative end, the mean is the smallest value with the median
in the middle, and the mode is the largest value.

Chapter 4: Basic Statistics 43

 Figure 4-3. Skewed distributions
Mean
Median
Median Median
Mode
Mode Mean Mean Mode

Symmetrical Positive Negative

distribution skew skew

Central limit theorem

To perform statistical tests and analyses, statisticians rely on
their assumption of a normal distribution. However, as we
have seen, this is not always the case. Fortunately, there is a
rule which allows them to make this assumption even when the
distribution is not normal. The central limit theorem states that
the sampling distribution of means increasingly approximates
a normal distribution as the sample size increases. That is a
distribution whose individual values are the means of samples
drawn from the main distribution (population). The central limit
theorem allows inferential statistics to be applied to skewed and
otherwise normal distributions.

The central limit theorem is ver y power ful, and in most

situations it works reasonably well with a sample size greater
than 30. Therefore, it is possible to closely approximate what
the distribution of sample means looks like, even with relatively
small sample sizes. The importance of the central limit theorem
to statistical thinking cannot be overstated. Most hypothesis
testing and sampling theory is based on this theorem.

While there is a mathematical proof for the central limit

theorem, it goes beyond the scope of this text. It is discussed
here to show that there is a solid statistical base for assuming
a normal distribution for the statistical tests used in inferential
analysis of fire data. With the proper sample size, the results
will be valid even if the population distribution is not normal.

44 Fire Data Analysis Handbook

Chapter 5: Analyses of
Tables
Introduction
As was discussed previously, most fire data are qualitative
(categorical) in nature. Examples of categorical data in the fire
service would include property use, cause of ignition, extent of
flame damage, etc. Since this type of data cannot be expressed in
terms of the mean, median and standard deviation, the number
of each category can be used and listed in a table format. It
might be found, for example, that arson fires account for 56% of
all structure fires, equipment failure for 23%, and so on.

This chapter provides techniques for analy zing tables

d e v e l o p e d f ro m c a te go r i c a l da t a . T h is in c l u d e s t h e
development and interpretation of percentages for categorical
data and the use of a nonparametric statistical test known as
the chi-square. The chi-square is used to determine whether
the percentage distribution from a table differs significantly
from a distribution of hypothetical or expected percentages.

A nonparametric statistical test is one that makes little or

no assumptions about the distribution. As stated previously,
statisticians assume a normal distribution in their calculations.
However, categorical data by nature are not described in
this manner, i.e., mean, standard deviation, etc. Therefore,
statistical tests that have certain parameters to their use
would not be appropriate for this type of data. The chi-square
was designed to be used without these parameters, and as
such, is ideal for categorical data.

Describing categorical data

Summarizing a categorical variable is usually done by
reporting the number of observations in each category and
its percentage of the total. For example, consider Figure 5-1
for types of situations found in the fires of Lincoln, Nebraska.
These percentages are simple to calculate and easy to
understand: 24.9% of the fires are structure fires, 26.9% are
vehicle fires, and so on. As described in Chapter 4, the mode

Chapter 5: Analyses of Tables 45

 is the category with the largest number of data values. In this
example, the mode is vehicle fires, totaling 175 fires.

Figure 5-1. Type of situations found — Lincoln,

Nebraska, fires

Type of fire Number Percent

Structure fires 162 24.9
Outside of structure fires 44 6.8
Vehicle fires 175 26.9
Tree, brush, grass fires 166 25.6
Refuse fires 88 13.5
Other fires 15 2.3
Total 650 100.0

By way of comparison, Figure 5-2 shows the nationwide

picture of types of situations found for fires. From a national
perspective, structure fires accounted for 28.7% of the total,
closely followed by tree, brush and grass fires at 27.3% and
vehicle fires at 20.2%.

Figure 5-2. Types of situations found — nationwide fires

Type of fire Number Percent

Structure fires 523,000 28.7
Outside of structure fires 64,000 3.5
Vehicle fires 368,500 20.2
Tree, brush, grass fires 498,000 27.3
Refuse fires 226,500 12.5
Other fires 143,000 7.8
Total 1,823,000 100.0

Looking at these figures would prompt the question of

whether the distribution of fires in Lincoln dif fers from
the national picture. Some differences can be noticed by
comparing percentages. For example, 26.9% of the Lincoln fires
were vehicle fires, compared to 20.2% nationwide. Similarly,
2.3% of the Lincoln fires were other fires, compared to 7.8%

46 Fire Data Analysis Handbook

nationwide. Therefore, it would seem that the distribution of
fires in Lincoln deviates from the national picture. However,
a statistical test can be made to test this difference more
precisely. The next section provides such a test.

The chi-square test

The chi-square test (pronounced kī; rhymes with pie) is a
statistical test designed to be used with categorical data. Like
most statistical tests, it is stated in precise statistical language
by defining a hypothesis to be tested. The use of the term null
hypothesis is commonly seen. The null hypothesis merely
states that there is no difference between the 2 distributions
being compared. In this case, the null hypothesis would be
that there is no statistical difference between Lincoln and
the national percentages in the categories of fires in Figures
5-1 and 5-2. This is usually the way it is stated, that there is
no difference. If a difference is found, the null hypothesis is
rejected. It is sort of like innocent until proven guilty!

Although the chi-square test is conduc ted in terms of

frequencies, it is best viewed conceptually as a test about
proportions. To illustrate these ideas, it will be easier at this
point to use a format that does not include fire data. Instead,
consider a simple experiment where a die is thrown over
and over again. The resulting data values are the number of
dots showing after each throw. The number of dots varies
between 1 and 6; that is, there are 6 possible outcomes. If a
“fair” die is thrown a large number of times, one would expect
each number of dots to show up one-sixth of the time. The
chi-square test can be used with a certain degree of assurance
to determine if, in fact, the die is “fair.”

Suppose, for example, that a die is tossed 90 times and the

results are as shown in Figure 5-3.

Chapter 5: Analyses of Tables 47

 Figure 5-3. Results of die throws

Dots visible Number Percent

1 16 17.8
2 17 18.9
3 12 13.3
4 14 15.6
5 17 18.9
6 14 15.6
Total 90 100.0

If the die is a “fair” die, one would expect to have 1 dot turn
up exactly 15 times (one-sixth of the total), 2 dots visible
exactly 15 times, and so on. The actual results differ from these
expected results as shown in Figure 5-4.

Figure 5-4. Actual and expected results

Expected
Dots visible Actual number
number
1 16 15
2 17 15
3 12 15
4 14 15
5 17 15
6 14 15
Total 90 90

To summarize, a die has been tossed 90 times and obtained

the results shown in Figure 5-3. The null hypothesis is that the
die is “fair,” which means that there is no difference between
the actual and the expected number of times each number of
visible dots appears. The actual results are not the same as
the expected, either because the die is not “fair” or because
of variations inherent in throwing a die only 90 times. The
chi-square test will determine whether the actual results differ
significantly from the expected results.

48 Fire Data Analysis Handbook

In statistical terms, significance does not mean something
meaningful or important; rather, statistical significance refers
to the claim that a result from data generated by testing or
experimentation is not likely to occur randomly or by chance
but is instead likely to be attributable to a specific cause. In
other words, significance means that the results of a test are
likely real and not caused by luck or chance.

The following are the steps in performing the chi-square test:

1. Calculate the expected number for each category by

multiplying the expected or population percentages by
the total sample size. This calculation has already been
performed as shown in Figure 5-4 with the “Expected
number” column.

2. For each category, subtract the expected number from the

actual number and then square the result.

3. Divide the results from step 2 by the expected number.

4. Sum the results from step 3. This is the calculated chi-square

statistic. The larger this number, the more likely there is a
significant difference between the actual and expected values.

5. Find the degrees of freedom, which are defined as the

number of categories minus 1. In the die example, there
are 5 degrees of freedom.

6. Obtain the critical chi-square value from the table in

the Appendix by selecting the entry associated with the
appropriate degree of freedom. Note: The table includes
levels of significance from 0.05 to 0.001. Commonly, the
0.05 level is used for most determinations. This indicates
that results exceeding the critical value will be statistically
significant 95% of the time. The other levels are used
depending on how critical the results may be. For example,
the more stringent 0.001 level is used in drug testing where
lives may depend on the results.

7. If the computed chi-square statistic is greater than the

critical value obtained from the table, the null hypothesis
is rejected. Otherwise, the null hypothesis is accepted.

Chapter 5: Analyses of Tables 49

 Rejecting the null hypothesis means there is a significant
difference between the 2 distributions. Conversely, accepting
it means that the 2 distributions are essentially the same
with differences due to sampling or random variations.

Figure 5-5 summarizes these steps for the die example.

The “difference” column shows the difference between the
expected and actual numbers. The “squared difference” is the
square of the difference obtained by multiplying the number
by itself. The right-most column is the squared difference
divided by the expected number; for example, the first figure
is 0.067, obtained from 1 divided by 15. The chi-square value is
1.34, which is the sum of the values in the last column.

Figure 5-5. Actual and expected results — die-tossing

experiment

Divided
Dots Actual Expected Squared by
Difference
visible number number difference expected
number
1 16 15 1 1 0.067
2 17 15 2 4 0.267
3 12 15 -3 9 0.600
4 14 15 -1 1 0.067
5 17 15 2 4 0.267
6 14 15 -1 1 0.067
Total 90 90 1.340
Chi-square 1.34
value
Degrees of 5.00
freedom
Critical 11.07
value

From the Appendix, the critical chi-square value for 5 degrees

of freedom at the 0.05 level of significance is 11.07. Since the
calculated chi-square value of 1.34 is less, the null hypothesis
is accepted. Therefore, the results from the 90 throws do not
provide evidence that the die is unfair.

50 Fire Data Analysis Handbook

Degrees of freedom have been defined as the number of
categories minus 1. The rationale for determining degrees
of freedom is that each categor y may be considered as
contributing 1 piece of data to the chi-square statistic. These
data are free to vary except for the last category, since it is
determined already by what is left. It is, therefore, not free to
vary. Thus, the values in all categories except 1 are free to vary.
An illustration of this may be more helpful than an explanation.

Suppose you were asked to name any 5 numbers. In response,

you chose 25, 44, 62, 82 and 2. In this case, there were no
restrictions on the choices. There were 5 choices and 5
degrees of freedom. Now suppose you were asked to name
any 5 numbers again. This time you chose 1, 2, 3 and 4, but
were stopped at that point and told that the mean of the 5
numbers must be equal to 4. Now you have no choice for the
last number, because it must be 10 (1 + 2 + 3 + 4 + 10 = 20, and
20 divided by 5 equals 4). The restriction caused you to lose 1
degree of freedom in your choice. Instead of having 5 degrees
of freedom as in the first example, you now have 5 minus 1, or
4, degrees of freedom. Each statistical test of significance has
its own built-in degrees of freedom based on the number and
type of restrictions it makes. The chi-square has 1.

At this time, the question on whether the distribution of fires

in Lincoln differs from the nationwide distribution of fires can
be dealt with. It was noted that there were differences in some
categories; for example, Figure 5-1 shows that vehicle fires
account for 26.9% of the fires in Lincoln compared to 20.2%
nationwide. Similarly, other fires account for 2.3% of the fires
in Lincoln compared to 7.8% nationwide.

However, these are individual comparisons. The chi-square

test allows all categories to be tested simultaneously. The
null hypothesis is that “The percentage distribution of fires
in Lincoln does not differ significantly from the nationwide
picture.” If the calculated chi-square value is larger than
the appropriate critical value in the Appendix, then the null
hypothesis is rejected, which indicates that there is a significant
dif ference. Other wise, the null hypothesis is accepted,
indicating no significant difference in the 2 distributions.

Chapter 5: Analyses of Tables 51

 Figure 5-6 shows the calculations using the information in
Figures 5-1 and 5-2. The “actual number” column comes
directly from Figure 5-1. To obtain the expected number, the
percentages from Figure 5-2 are applied to the 650 Lincoln
f ires. For example, 28.7% of the nationwide f ires were
structure fires, which means we expect 28.7% of the 650 fires
in Lincoln to be structure fires. This calculation yields 186.6
fires (28.7% times 650 fires).

The “difference” column gives the difference between the

actual and expected numbers, and the next column is the
squared difference (the difference multiplied by itself). The
last column is the squared difference divided by the expected
value. The calculated chi-square value is the sum of the
column, which is 63.8.

In this example, there are 6 categories of fires, which means

there are 5 degrees of freedom. From the Appendix, the critical
chi-square value at the 0.05 level of significance is 11.07. Since
the calculated chi-square value of 63.8 is greater than the
critical value, the null hypothesis is rejected. The conclusion is
that the distribution of fires in Lincoln differs significantly from
those nationwide. As stated before, the table in the Appendix
lists the critical values for chi-square at various levels. For
the purposes of this type, the 0.05 level is sufficient, which
means that the difference will be significant 95% of the time
or at the 95% confidence level. In this particular example, the
obtained chi-square value far exceeds the critical value for
even the 0.001 level of significance, which is 20.51. This means
that it is significant 99.9% of the time with a chance of error of
only one-tenth of a percent! In most comparisons, this level of
confidence is rarely obtained.

52 Fire Data Analysis Handbook

 Figure 5-6. Actual and expected results — Lincoln fires

Divided
Type of Actual Expected Squared by
Difference
fire number number difference expected
number
Structure 162 186.6 -24.6 605.16 3.2
Outside 44 22.8 21.2 449.44 19.7
Vehicle 175 131.3 43.7 1,909.69 14.5
Grass 166 177.4 -11.4 129.96 0.7
Refuse 88 81.3 6.7 44.89 0.6
Other 15 50.7 -35.7 1,274.49 25.1
Total 650 650.0 63.8
Chi-square 63.8
value
Degrees of 5.0
freedom
Critical 11.07
value

Some of the rationale behind the chi-square statistic may be

helpful in understanding what it is actually reporting. The
dynamics of what contributes to the chi-square value are
evident in Figure 5-6. For example, the largest difference
(regardless of sign) between the actual and expected numbers
is 43.7 for vehicle fires. Squaring the difference and dividing
by the expected number gives 14.5, as shown in the last
column. As can be seen, vehicle fires are only the third largest
contributor to the chi-square value, even though this type
has the largest difference between the actual and expected
number of fires. The reason for this is that larger categories
have greater leeway for numerical variations, since it requires
more to account for the same amount of actual change than
smaller categories with fewer numbers to begin with. This
can readily be seen by looking at the top 2 categories in
contribution weight to the chi-square value. Outside fires with
a difference of 21.2 and other fires with a difference of -35.7
contribute 19.7 and 25.1, respectively, for a total of 44.8 toward
the 63.8 chi-square value. That is, 70% of the chi-square value
is made up of the 2 smallest categories! While the numerical
difference is less than that of vehicle fires, the actual amount

Chapter 5: Analyses of Tables 53

 of change in those categories is greater, because the numerical
difference is greater proportionally to the number of fires in
those categories. This is why it was stated earlier that “although
the chi-square test is conducted in terms of frequencies, it is
best viewed conceptually as a test about proportions.”

2-way contingency tables

Up to this point, chi-square has been applied in cases with
only 1 variable. It also has important application to the analysis
of bivariate frequency distributions. By studying bivariate
distributions with 2 categorical variables, the statistical
association between the 2 variables can be measured.
A ssociation allows the gaining of information about 1
variable by knowing the value of the other. The strength of
the association may run from no association to weak to quite
strong. The chi-square measures its existence and strength.

F igure 5 -7 is used as the s t ar ting point to introduce

contingency tables, statistical variable association and the
chi-square statistic’s role in measuring it. The NFPA’s “Survey
of Fire Departments for U.S. Fire Experience” for 1 year was
used to develop the figure. To facilitate the example, 5 of the
10 categories under “nature of injuries” were eliminated. The
“type of duty” category is as it appears in the original table.

There are 5 categories for location or “type of duty.” The first is

“responding to or returning from incident.” The next category,
“fireground,” covers injuries while on site at a fire. Similarly,
the third category, “nonfire emergency,” covers injuries while
on site at all nonfire incidents. The “training” category includes
any injuries sustained while the firefighter was training. The
last category covers all injuries not under the other categories
but while still on duty.

The nature of the injury also is divided into 5 categories. As

mentioned above, there were originally 10 categories of injuries,
but for simplicity’s sake, only the top 5 were used. They are:

1. Burns.
2. Smoke inhalation.
3. Wounds/cuts.
4. Strains/sprains.
5. Other.

54 Fire Data Analysis Handbook

 Figure 5-7. Firefighter injuries — type of duty and nature
of injuries

Nature of injuries
Type of
Smoke Wounds/ Strains/ Row
duty Burns Other
inhalation cuts sprains totals
Responding 65 115 960 2,250 710 4,100
to or
returning
from
incident
Fireground 3,255 2,580 9,210 16,410 3,635 35,090
Nonfire 185 185 2,440 8,025 2,725 13,560
emergency
Training 345 40 1,380 3,860 625 6,250
Other on 245 105 2,780 8,185 2,495 13,810
duty
Column 4,095 3,025 16,770 38,730 10,190 72,810
totals

Figure 5-7 shows that there were a total of 72,810 injured

firefighters. The top lef t number means there were 65
firefighters who were burned either responding to or returning
from an incident. Similarly, the number in the second row and
fourth column indicates that there were 16,410 firefighters who
suffered strains or sprains while on a fire incident. Further, this
number is the mode of the contingency table.

Outside of identifying the mode and showing the relative

position of each category within its variable, the numbers
in the table do not relay much information. Next, various
percentages are calculated from the table to provide more
insight. Finally, a chi-square value is calculated to measure the
strength of the relationship between the 2 variables.

Percentages for 2-way contingency tables

T here are 3 way s to c alculate percent ages for 2- way
contingency tables of frequencies. Each way highlights a
different feature of the table. More importantly, each provides
a different interpretation of the data and leads to different

Chapter 5: Analyses of Tables 55

 conclusions about the relationship between the 2 variables.
The 3 ways of calculating percentages are:

1. Joint percentages.
2. Row percentages.
3. Column percentages.

The t ype of percentage used depends upon where the

emphasis needs to be placed. Joint percentages allow the
direct comparison of table entries with each other. Row
percentages concentrate on the individual rows of the table
with percentages along the row totaling 100%. Similarly,
column percentages deal with the individual columns of the
table with column totals equaling 100%.

Joint percentages
To calculate joint percentages, each entr y in the table is
divided by the overall total. Figure 5-8 shows the percentage
calculation for the counts from Figure 5-7. The lower-left entry
is simply 4,095 divided by 72,810, which equals 5.6%. This
means that 5.6% of the total firefighter injuries sustained were
burns. The sum of all the entries in the table is 100.0%.

More logical comparisons can be made with joint percentages

than with just the raw counts. For example, the table shows
that 22.5% of all injuries were sprains or strains that occurred
while the firefighter was on the fireground. In a similar manner,
only 1.3% of all injuries were wounds or cuts suffered by
firefighters responding to or returning from an incident.

Figure 5-8 also provides important information from the

row and column totals. For example, from the second row
it is apparent that nearly half (48.2%) of all the injuries were
sustained at a fireground. There are 2 methods that can be
used to derive this percent. The first is to add the 5 percentages
across the row (4.5 + 3.5 + 12.6 + 22.5 + 5.0 = 48.1). The second
is to divide the row total of 35,090 (from Figure 5-7) by 72,810
to yield the 48.2%. (Note: Due to rounding, the percentages are
not always exactly the same using these methods.)

56 Fire Data Analysis Handbook

Similarly, column percentages provide information about
the nature of the injuries involved. For example, only 5.6%
of firefighters injured suffered from burns, 4.2% from smoke
inhalation, 23% from wounds or cuts, 14% from other injuries,
and over half (53.2%) from strains or sprains.

Figure 5-8. Firefighter injuries — joint percentages

Nature of injuries
Type of
Smoke Wounds/ Strains/ Row
duty Burns Other
inhalation cuts sprains totals
Responding 0.1 0.2 1.32 3.1 1.0 5.6
to or
returning
from
incident
Fireground 4.5 3.5 12.6 22.5 5.0 48.2
Nonfire 0.3 0.3 3.4 11.0 3.7 18.6
emergency
Training 0.5 0.1 1.9 5.3 0.9 8.6
Other on 0.3 0.1 3.8 11.2 3.4 19.0
duty
Column 5.6 4.2 23.0 53.2 14.0 100.0
totals

While Figure 5-8 provides more insight into these 2 variables, it

does not directly address other questions. For example, direct
comparisons between burns and smoke inhalation injuries
for any particular type of duty cannot be made. Similarly,
comparisons between types of duty for any particular injuries
cannot be made. To make these types of comparisons, row and
column percentage calculations must be made.

Row percentages
To convert table counts into row percentages, each entry in the
table must be divided by its row total. Therefore, the top-right
entry is calculated by dividing 710 by 4,100. This indicates that
17.3% of the total firefighters responding to or returning from
an incident sustained other types of injuries.

Chapter 5: Analyses of Tables 57

 Figure 5-9. Firefighter injuries — row percentages

Nature of injuries
Type of
Smoke Wounds/ Strains/ Row
duty Burns Other
inhalation cuts sprains totals
Responding 1.6 2.8 23.4 54.9 17.3 100.0
to or
returning
from
incident
Fireground 9.3 7.4 26.2 46.8 10.4 100.0
Nonfire 1.4 1.4 18.0 59.2 20.1 100.0
emergency
Training 5.5 0.6 22.1 61.8 10.0 100.0
Other on 1.8 0.8 20.1 59.3 18.1 100.0
duty

A table of row percentages allows for comparisons among the

categories represented by the rows. The total for each row is
100%, and these figures appear on the right of the table as a
reminder that row percentages are represented. (Note: Due
to rounding, the row percentages may not add up to 100%.)

As indicated, 17.3% suffered other types of injuries when they

were responding to or returning from an incident. A total of
1.6% had burn injuries, 2.8% had smoke inhalation injuries,
23.4% sustained wounds or cuts, and the vast majority, 54.9%,
had sprains or strains.

Looking at the second row, which is for firefighters injured at

the fireground, a somewhat different picture emerges. Burns
and smoke inhalation injuries accounted for 9.3% and 7.4%,
respectively. These are followed by wounds and cuts at 26.2%,
sprains and strains at 46.8%, and 10.4% for the other category.
Once again, these percentages total 100, accounting for all
firefighters injured while at the fireground.

Column percentages
To convert table counts into column percentages, each entry in
the table must be divided by the total for its column. The top-left

58 Fire Data Analysis Handbook

entry would be calculated by dividing 65 by 4,095, yielding 1.6%.
This indicates that only 1.6% of the firefighters who received
burns were responding to or returning from an incident.

Figure 5-10. Firefighter injuries — column percentages

Nature of injuries
Type of
Smoke Wounds/ Strains/
duty Burns Other
inhalation cuts sprains
Responding 1.6 3.8 5.7 5.8 7.0
to or
returning
from
incident
Fireground 79.5 85.3 54.9 42.4 35.7
Nonfire 4.5 6.1 14.5 20.7 26.7
emergency
Training 8.4 1.3 8.2 10.0 6.1
Other on 6.0 3.5 16.6 21.1 24.5
duty
Total 100.0 100.0 100.0 100.0 100.0

The table of column percentages looks at a particular nature

of injury across the 5 types of duty. With burn injuries, you
can see that most (79.5%) occurred at the fireground, 8.4%
during training, 6% on other types of duty, 4.5% at nonfire
emergencies, and only 1.6% while responding to or returning
from an incident. The “other” injury category shows a very
different breakdown. A total of 7% of the injuries occurred
while responding to or returning from an incident, while
35.7% were sustained at the fireground. Nonfire emergencies
accounted for 26.7%, followed by 24.5% for other on duties,
and lastly 6.1% during training.

Selecting a percentage table

The choice of a percentage table depends on the uses of the
data. Joint percentage tables are beneficial when the emphasis
is on the interrelationship between the 2 variables in the table.
For example, Figure 5-8 reveals that the combination of burns

Chapter 5: Analyses of Tables 59

 at the fireground accounted for 4.5% of the total. This figure
can be compared to other combinations in the table.

The row percentage table provides a way of emphasizing the

nature of injury for each type of duty. When a firefighter was
responding to or returning from an incident, Figure 5-9 shows
54.9% of the injuries were from strains or sprains, 23.4% from
wounds or cuts, 17.3% from other types of injuries, 2.8% from
smoke inhalation, and 1.6% from burns. These are useful
results by themselves and can be compared to distributions
in other rows.

The column percentage table emphasizes the type of duty

for each nature of injury. For burns only, Figure 5-10 shows
that 79.5% were sustained at the fireground, 8.4% occurred
during training, 6% were other on duty, 4.5% were on a nonfire
emergency, and 1.6% were responding to or returning from an
incident. Interestingly, the percentage of those burned while
responding to or returning from an incident is the same for
both the row and the column percentages.

Testing for independence in a 2-way contingency

table
This section uses the chi-square test to determine whether the
2 variables in a 2-way contingency table are independent of
each other. As before, a step-by-step procedure for calculating
the chi-square value is provided. It should be noted that, with
the chi-square calculations, as with the other calculations
that have been performed, virtually all statistical packages
automatically calculate the values. You can see that manual
calculation is arduous and time-consuming. Additionally,
manual calculations are more subject to error. Therefore,
a statistical software package should be used whenever
possible. However, the details of the computations are
shown here to enhance the understanding of the underlying
principles that are involved.

Before calculating the chi-square, however, a discussion of what

is meant by independence is needed. 2 variables are said to be
independent if knowledge about 1 variable cannot be used in
predicting the outcome of the other variable. In general, the
null hypothesis of independence for a 2-way contingency

60 Fire Data Analysis Handbook

table is equivalent to hypothesizing that in the population,
the relative frequencies for any row (across the categories
of the column variable) are the same for all rows, or that the
population of the relative frequencies for any column (across
the categories of the row variable) are the same for all columns.
So once again, the hypothesis to be tested by chi-square can
be seen as one concerning proportions. For example, there are
almost 9 times as many injuries sustained on the fireground as
there are responding to or returning from an incident, but if the
type of duty is unrelated to the number of injuries sustained,
then on a proportional basis, the number of injuries should
be the same for each type of duty.

Constructing a table of expected values

To calculate the chi-square value, the expected values for each
cell must be determined. The expected values are the counts
that would occur if the 2 variables were independent. The first
step in developing a table of expected values is to calculate the
proportion of cases in each cell. This can be done by column
or row.

Using the column, divide each column total by the grand

total. The proportion for the first column, “burns,” would be
calculated as follows: 4,095 divided by 72,810 equals 0.056.
Subsequent column proportions would be “smoke inhalation”
(0.042), “wounds/cuts” (0.23), “strains/sprains” (0.532), and
“other” (0.14). Note that the proportions are the same as the
column total percentages calculated for the joint percentages
in Figure 5-8.

It is easy to calculate the expected cell frequencies from the

expected cell proportions. For each cell, multiply the expected
column proportion for that cell by the row total for that cell.
For example, the cell representing firefighters responding to
or returning from an incident who sustained burns would be:
0.05624 (column proportion) times 4,100 (row total) equals 230.6.
Figure 5-11 shows the results of the remaining expected values.

Chapter 5: Analyses of Tables 61

 Figure 5-11. Firefighter injuries — table of expected values

Nature of injuries
Type of
Smoke Wounds/ Strains/ Row
duty Burns Other
inhalation cuts sprains totals
Responding 230.6 170.3 944.3 2,180.9 573.8 4,100
to or
returning
from
incident
Fireground 1,973.5 1,457.9 8,082.1 18,665.5 4,910.9 35,090
Nonfire 762.6 563.4 3,123.2 7,213.0 1,897.8 13,560
emergency
Training 351.5 259.7 1,439.5 3,324.6 874.7 6,250
Other on 776.7 573.8 3,180.8 7,346.0 1,932.8 13,810
duty
Column 4,095 3,025 16,770 38,730 10,190 72,810
totals

The table of expected values is the distribution of proportions in

each row (or column) that would be expected in the absence of
a dependent relationship between the 2 variables. In this case,
it would mean that the expected values are those that reflect
no relationship between the nature of injuries sustained and
the type of duty performed. As stated before, the same results
could have been obtained by calculating the row proportions
and multiplying them by the column totals. It should also be
noted that the row and column totals are exactly the same as
the original table of counts. That is, the development of the
expected value table preserves these totals. However, slight
discrepancies may exist due to rounding of decimals.

Calculation of chi-square for a 2-way contingency

table
The chi-square value for a 2-way contingency table is calculated
similarly to the one done for a single categorical variable.

1. Develop the table of expected values, as shown in Figure

5-11, using the method discussed in the previous section.

62 Fire Data Analysis Handbook

2. For each table entry, subtract the expected value from the
corresponding entry in the original table of counts and then
square the result. This difference measures the discrepancy
between the actual counts and what would be expected if
the variables were independent.

3. Divide the results from step 2 by the expected value. This

adjustment allows for the larger expected numbers which
are usually associated with larger deviations.

4. Sum the results from step 3. This is the chi-square statistic.

The larger the chi-square statistic, the more likely that
there is a significant statistical association between the 2
variables. However, the chi-square statistic also depends
on the number of categories, which must be taken into
account in the following steps.

5. Find the degrees of freedom, which are calculated for a

2-way contingency table by multiplying the number of rows
minus 1 times the number of columns minus 1. In the current
example, there are 5 rows and 5 columns. Therefore, the
number of degrees of freedom is (5-1) x (5-1) = 16.

6. Compare the computed chi-square statistic from step 4 to

the value in the chi-square table in the Appendix using the
appropriate degrees of freedom. The table value is called
the critical chi-square value.

7. If the computed chi-square statistic is greater than the

critical value in the table, then the null hypothesis of
independence is rejected and the variables are related. If
the computed chi-square statistic is less than the critical
value, the null hypothesis of independence is accepted and
the variables are not related.

It is important to keep in mind that in a 2-way contingency

table, the 2 variables are independent. If the null hypothesis is
accepted, it means that knowing the value of 1 of the variables
does not help in predicting the value of the other variable. In
the current example, the null hypothesis is that the type of
duty engaged in is independent of the nature of the injuries
sustained.

Chapter 5: Analyses of Tables 63

 Figure 5-12. Firefighter injuries — table of chi-square
entries

Nature of injuries
Type of duty Smoke Wounds/ Strains/
Burns Other
inhalation cuts sprains
Responding to 118.92 17.96 .26 2.19 32.33
or returning
from incident
Fireground 832.15 863.65 157.40 272.55 331.54
Nonfire 437.48 254.15 149.45 91.41 360.55
emergency
Training 0.12 185.86 2.46 86.22 71.28
Other on duty 363.98 383.01 50.50 95.82 163.53
Total Critical
chi-square value =
value = 26.3
5,324.77

Figure 5-12 shows the chi-square entries for the 2-way

contingency table. These entries are the results after step 3
above. The top-left entry was calculated as follows:

1. Figure 5-7 gave an actual count of 65 for this entry, and

Figure 5-11 gave an expected value of 230.6.

2. Subtracting the expected value from the actual count yields

-165.6 (65 minus 230.6), and squaring that figure results in
27,423.36.

3. Dividing this number by the expected value, 230.6, provides

the chi-square value of 118.92.

4. This value is then entered in Figure 5-12, and the procedure

is repeated for each of the other entries.

5. When all of the entries are calculated, they are all totaled.
This total is the total chi-square value.

6. In Figure 5-12, this total is 5,324.77. It is entered at the

bottom of the table.

64 Fire Data Analysis Handbook

Now, to test the hypothesis about the independence of the
2 variables, type of duty and nature of injury, compare the
total chi-square value to the critical chi-square value from
the Appendix. The critical chi-square value for 16 degrees of
freedom at the 0.05 level of significance is 26.3. Since the total
chi-square value greatly exceeds this value, the null hypothesis
is rejected. Therefore, there is a statistical association between
type of duty and nature of injury.

As a cautionary note, remember that a significant outcome of

the chi-square test is directly applicable only to the data taken
as a whole. The chi-square obtained is inseparably a function
of the, in this case, 25 contributions composing it — 1 from each
cell. Therefore, it cannot be said whether 1 group is responsible
for the finding of significance or whether all are involved.

Chapter 5: Analyses of Tables 65

66 Fire Data Analysis Handbook
Chapter 6: Correlation
Introduction
This chapter deals with the concept of correlation for
continuous (quantitative) data. Correlation is a statistical
measure which indicates the degree to which 1 variable
changes with another variable. For example, calls for EMS
generally increase with population grow th. That is, as
population increases, more medical service calls would be
expected. This would indicate a positive correlation between
population and EMS calls. The correlation measures the
strength of the association between the 2 variables.

If there is a correlation between 2 variables, then predictions

better than chance can be made from an individual score (or
whatever is being measured) on 1 variable to its predicted
score on the correlate variable. Any problem in correlation
requires 2 pairs of corresponding scores, 1 for each variable.
Generally, the greater the association (correlation) between 2
variables, the more accurately a prediction can be made on the
standing in 1 variable from the standing in the other.

This chapter starts with the scatter diagram illustrated in

Chapter 3 and proceeds with a discussion of the correlation
coefficient. Next, a typical calculation of a correlation is
presented for demonstration purposes. The chapter concludes
with a discussion of the applicability and uses of a correlation
and mention of other types of correlation.

Scatter diagram
Figure 6-1 shows a scatter diagram presented in Chapter 3 on
the number of fire deaths and the number of other residential
fires for a 10-year period. The horizontal axis gives the number
of fires (in thousands), and the vertical axis gives the number
of deaths. You can see from the figure that deaths are higher
with greater numbers of fires. The general trend is clear even
though the pattern is not perfect. The term “not perfect” refers
to the fact that the points do not fall on a straight line.

Chapter 6: Correlation 67
 Figure 6-1. Fires and deaths — other residential
120

100

80
Deaths

60

40

20
10 11 12 13 14 15 16
Fires (thousands)

With relationships depic ted in this manner, the usual

terminology is to label 1 variable as the independent variable,
and the other as the dependent variable. In the case of Figure
6-1, “fires” serves as the independent variable and “deaths”
as the dependent variable. Obviously, the number of fires
influences the number of fire-related deaths. Generally
speaking, the more fires there are, the greater number of
fire deaths. This represents a positive correlation since the
increase in the independent variable is accompanied by an
increase in the dependent variable.

It is important to emphasize 2 points about correlations. First,

correlations assume an underlying linear relationship; that is,
a relationship that can be best represented by a straight line. It
should be noted, however, that not all relationships are linear.
There are, for example, curvilinear relationships where the
points on a scatter diagram cluster about a curved line.

Secondly, while correlation can be used for prediction, it does

not imply causation. The fact that 2 variables vary together is
a necessary — but not a sufficient — condition to conclude
that there is a cause-and-effect connection between them. A
strong correlation between variables is often the starting point
for further research.

68 Fire Data Analysis Handbook

Correlation coefficient
The correlation coef f icient measures the s treng th of
association bet ween 2 variables. The term “correlation
coefficient” is used by most statisticians but is the same as
the more commonly used correlation. The correlation is
always between -1 and +1. A correlation of exactly -1 or +1 is
called a perfect correlation and means that all the points fall
on a straight line. A correlation of 0 indicates no relationship
between the variables and would be represented on a scatter
diagram as random points with no discernible direction. As a
correlation coefficient moves from 0 in either direction, the
strength of the association between the 2 variables increases.

As stated before, a positive correlation means that, as the

independent variable increases, so does the dependent variable.
In a negative correlation, as the independent variable increases,
the dependent variable decreases. The sign of the correlation
indicates direction, not magnitude. Magnitude is indicated
by the size of the number regardless of the sign. Therefore, a
correlation of -0.82 is greater than a correlation of +0.63.

To summarize the relationship between a scatter diagram

and the correlation coefficient, the correlation coefficient is
a number that indicates how well the data points in a scatter
diagram “hug ” the straight line of best fit. With per fect
correlations, all the data points fall exactly on a straight
line that summarizes the relationship, and the value of the
coefficient is +1 or -1. When the association between the
2 variables is less than perfect, the data points show some
scatter about the straight line that summarizes the relationship
as in Figure 6-1, and the absolute value (regardless of sign)
of the correlation coefficient is less than 1. The weaker the
relationship, the more scatter and the lower the absolute value
of the correlation coefficient.

A nother impor t ant point is that correlations are not

arithmetically related to each other. For example, a correlation
of 0.6 is not twice as strong as a correlation of 0.3. Although it is
obvious that a correlation of 0.6 reflects a stronger association
than a correlation of 0.3, there is no exact specification of the
difference. Subsequently, there is no relationship between
correlations and percentages.

Chapter 6: Correlation 69
 To make direct comparisons bet ween correlations, the
correlation coefficient must be converted to a coefficient
of determination. The coefficient of determination is the
square of the correlation multiplied by 100. This yields the
percentage of association bet ween the 2 variables. For
example, a correlation of 0.50 would indicate a 25% (0.50 times
0.50 equals 0.25 times 100 equals 25) association between
variables. A perfect correlation of 1.00 would be equal to a
100% coefficient of determination. So a correlation of 1.00 is
4 times as strong as a correlation of 0.50, not twice as strong,
as might appear from comparing the correlations directly.

Additionally, the differences between successive correlation

coefficient values do not represent equal differences in
degree of relationship. For example, the difference between
a correlation of 0.40 and 0.50 does not represent the same
difference as that between correlations of 0.90 and 1.00.
This can be seen more clearly by examining the coefficients
of determination and their corresponding correlations in
Figure 6-2. There is more than double the difference between
correlations of 0.90 and 1.00 than between 0.40 and 0.50 when
the corresponding coefficients of determination are compared.

Figure 6-2. Relationship between correlations and

coefficients of determination

Correlation coefficient Coefficient of determination

1.00 100%
0.90 81%
0.80 64%
0.70 49%
0.60 36%
0.50 25%
0.40 16%
0.30 9%
0.20 4%
0.10 1%
0.00 0%

70 Fire Data Analysis Handbook

Calculating the correlation
Today, most calculators include a program to calculate the
correlation coefficient. Additionally, virtually all statistical
software packages calculate the various types of correlations.
However, for those who must calculate a correlation by
hand and show what factors make the coefficient positive or
negative and what factors result in a high or low value, the
deviation-score method is used.
Figure 6-3 shows the number of fires and civilian fire deaths
for a 10-year period. The correlation between these 2 variables
will be computed using the deviation-score method. The most
widely used correlation formula is the Pearson. Its full name is
the Pearson product-moment correlation coefficient. There
are other types of correlations suited for special situations,
but the Pearson is by far the most common. In fact, when
researchers speak of a correlation coefficient without being
specific about which one they mean, it may safely be assumed
they are referring to the Pearson product-moment correlation
coefficient. The term moment is borrowed from physics and
refers to a function of the distance of an object from the
center of gravity. With a frequency distribution, the mean is
the center of gravity and, therefore, deviation scores are the
moments. As shown, the Pearson correlation is calculated by
taking the products of the paired moments.

Figure 6-3. Total United States fires and civilian fire deaths

Year Fires (thousands) Deaths

1 2,041.5 4,465
2 1,964.5 4,730
3 1,952.5 4,635
4 2,054.5 4,275
5 1,965.5 4,585
6 1,975.0 4,990
7 1,795.0 4,050
8 1,755.0 4,035
9 1,823.0 3,570
10 1,708.0 4,045
Sum 19,034.5 43,380

Chapter 6: Correlation 71
 As shown in Figure 6-3, fires tended to decrease over the 10-year
period, while civilian fire deaths seem to have no obvious pattern
overall (though the last 4 years have an apparent decrease).
From this, it would seem that there is little association between
the variables that should result in a low correlation.

The computation of the Pearson correlation using the deviation-

score method is illustrated in Figure 6-4 and summarized in the
following steps:

1. List the pairs of scores in 2 columns. The order in which

the pairs are listed makes no difference in the value of the
correlation. However, if 1 raw score is shifted, the one it is
paired with must be shifted as well.

2. Find the mean for the raw scores of each variable.

3. Convert each score in both variables to a deviation score

by subtracting the respective mean from each.

4. Calculate the standard deviation (S.D.) for both variables.

Since the deviation scores are already done, they need only
to be squared and summed. Divide each of these totals by
the number of pairs (in this case 10) and take the square
root of each.

5. Multiply each pair of deviation scores, known as the cross-

product, and total the results.

6. Next, multiply the 2 standard deviations by each other and

multiply that result by the number of pairs (10).

7. Divide the results of step 5 by the results of step 6. The result

is the Pearson product-moment correlation coefficient.

8. Square this for the coefficient of determination.

72 Fire Data Analysis Handbook

 Figure 6-4. Deviation score calculation for Pearson
correlation coefficient

Fires — Deaths —
Fires — Deaths — Cross
Year mean mean
mean mean product
squared squared
1 +138.05 +127 19,057.8 16,129 +17,532.35
2 +61.05 +392 3,727.1 153,664 +23,931.60
3 +49.05 +297 2,405.9 88,209 +14,567.85
4 +151.05 -63 22,816.1 3,969 -9,516.15
5 +62.05 +247 3,850.2 61,009 +15,326.35
6 +71.55 +652 5,119.4 425,104 +46,650.60
7 -108.45 -288 11,761.4 82,944 +31,233.60
8 -148.45 -303 22,037.4 91,809 +44,980.35
9 -80.45 -768 6,472.2 589,824 +61,785.60
10 -195.45 -293 38,200.7 85,849 +57,266.85
Sum 0 0 135,448.2 1,598,510 +303,759.00

Mean Fires Deaths Correlation Coefficient of

1,903.45 4,338 coefficient determination
S.D. 116.382 399.814 +.653 42.6%

The correlation obtained in Figure 6-4 is relatively high, as

demonstrated by the coefficient of determination of 42.6%.
This indicates a measure of relationship between the variables.
It does not mean that the relationship is necessarily causal. For
example, a high positive correlation probably exists between
the amount of beer consumed and the amount of automobile
accidents over each year from 1900 to the present. Rather
than believe that beer consumption and the number of auto
accidents are causally related, however, it is more reasonable
to sug gest that some condition such as an increase in
population accounts for the increase in both beer consumption
and automobile accidents.

Since the correlation is positive, it means that as the amount

of fires increase/decrease, the number of deaths increases/
decreases as well. While on the surface this would seem intuitive,
as with the beer/accident example, there can be other conditions

Chapter 6: Correlation 73
 that would account for the common variance. For example,
an increase in fires would be expected as the population and
buildings and residences increased. On the other hand, as
knowledge and use of fire safety programs and procedures
increased over time, the number of fire deaths would be
expected to go down. The point is that there are usually many
alternate and rational explanations for changes other than a
causal one between 2 simultaneously changing variables.
The next step after obtaining a correlation that shows there
is a relationship is to use it as a predictor. This is done by
defining the straight line that the data points cluster around,
known as the regression line. The regression line is defined
algebraically, and the formula is used to make the predictions.
The predictions become more reliable as the correlation
increases. A discussion of the regression method is beyond
the scope of this handbook but is mentioned here to give a
fuller meaning to the correlation coefficient.

Other types of correlations

While the Pearson correlation is by far the most commonly
used, there are other types of correlations derived directly
or indirectly from the Pearson. These correlations are used
with data that are not continuous and quantitative as with
the Pearson. Several of them are presented here with a brief
description of their use. Details of their computation and use
can be found in some of the texts cited earlier.
ĵ Rank-order correlation. Sometimes it is useful to categorize
data by ranking. The largest gets a rank of 1, the second
largest a rank of 2, and so on. When both variables consist of
ranks, a rank-order correlation coefficient is calculated. It is
sometimes called the Spearman rank-order correlation. It is
found merely by applying Pearson’s procedure to the ranks.
ĵ Biserial correlation. The biserial correlation is suited to
cases in which 1 variable is continuous and quantitative and
the other would be, except that it has been reduced to just
2 categories (for example, if the correlation between the
number of fires and whether or not the number of civilian
fire deaths was above or below the median). This would
require the use of the biserial technique since 1 variable
is continuous and the other is expressed dichotomously.

74 Fire Data Analysis Handbook

ĵ Point biserial correlation. This is used as in the biserial,
except that the second variable is qualitative and
dichotomous and cannot be expressed as continuous and
quantitative. For example, a correlation between the number
of fires and the number of male and female civilian deaths.

ĵ Phi coefficient. This is the Pearson correlation coefficient

for 2 variables that are both qualitative and dichotomous.

ĵ Partial correlation. The partial correlation shows the

Pearson correlation coefficient between 2 variables in the
absence of 1 or more other variables. For example, with
the correlation of fires to deaths, the relationship each
has to the passage of time may account for the change in
each rather than a relationship to each other. By doing a
partial correlation between fires and deaths for each month
within a given year, time can be held constant. The resulting
correlations reflect a truer picture of the relationship
between fires and deaths.

There are other variations of correlations used for determining

variable relationships with different circumstances, but these
cover most of what is likely required. As stated before, these
tools along with the ones discussed in the previous chapters
are readily available in various statistical packages. Most of
them guide the user through the process with clear, concise
directions. The purpose of manually calculating these statistics
is to provide the user with a more complete understanding
of how the data are analyzed. This should make it easier
to interpret the results from using a statistical package. It
also serves as a good foundation for any further study with
statistical texts and course work.

Chapter 6: Correlation 75
76 Fire Data Analysis Handbook
Appendix: Critical Values of
Chi-Square
Level of significance

Degrees of
0.05 0.025 0.01 0.005 0.001
freedom
1 3.84 5.02 6.63 7.88 10.83
2 5.99 7.38 9.21 10.60 13.82
3 7.81 9.35 11.34 12.84 16.27
4 9.49 11.14 13.28 14.86 18.47
5 11.07 12.83 15.09 16.75 20.51
6 12.59 14.45 16.81 18.55 22.46
7 14.07 16.01 18.48 20.28 24.32
8 15.51 17.53 20.09 21.95 26.12
9 16.92 19.02 21.67 23.59 27.88
10 18.31 20.48 23.21 25.19 29.59
11 19.68 21.92 24.73 26.76 31.26
12 21.03 23.34 26.22 28.30 32.91
13 22.36 24.74 27.69 29.82 34.53
14 23.68 26.12 29.14 31.32 36.12
15 25.00 27.49 30.58 32.80 37.70
16 26.30 28.85 32.00 34.27 39.25
17 27.59 30.19 33.41 35.72 40.79
18 28.87 31.53 34.81 37.16 42.31
19 30.14 32.85 36.19 38.58 43.82
20 31.41 34.17 37.57 40.00 45.31
21 32.67 35.48 38.93 41.40 46.80
22 33.92 36.78 40.29 42.80 48.27
23 35.17 38.08 41.64 44.18 49.73
24 36.42 39.36 42.98 45.56 51.18
25 37.65 40.65 44.31 46.93 52.62
26 38.89 41.92 45.64 48.29 54.05
27 40.11 43.19 46.96 49.65 55.48
28 41.34 44.46 48.28 50.99 56.89
29 42.56 45.72 49.59 52.34 58.30
30 43.77 46.98 50.89 53.67 59.70
Appendix: Critical Values of Chi-Square 77
78 Fire Data Analysis Handbook
Reference
Mosteller, F., Fienberg, S. E., & Rourke, R. E. K. (1983). Beginning
statistics with data analysis. Addison-Wesley Publishing
Company.

Reference 79
16825 South Seton Ave.
Emmitsburg, MD 21727
usfa.fema.gov

FA-266/November 2021