Simplifying Statistics

for Graduate Students

 Simplifying Statistics
for Graduate Students
Making the Use of Data Simple and User-Friendly

Susan Rovezzi Carroll and David J. Carroll

Rowman & Littlefield

Lanham • Boulder • New York • London
Published by Rowman & Littlefield
An imprint of The Rowman & Littlefield Publishing Group, Inc.
4501 Forbes Boulevard, Suite 200, Lanham, Maryland 20706
www.rowman.com

86-90 Paul Street, London EC2A 4NE, United Kingdom

Copyright © 2023 by Susan Rovezzi Carroll and David J. Carroll

All rights reserved. No part of this book may be reproduced in any form or by
any electronic or mechanical means, including information storage and retrieval
systems, without written permission from the publisher, except by a reviewer who
may quote passages in a review.

British Library Cataloguing in Publication Information Available

Library of Congress Cataloging-in-Publication Data Available

ISBN 9781475868388 (cloth : alk. paper) | ISBN 9781475868395 (pbk. : alk. paper) |
ISBN 9781475868401 (ebook)

The paper used in this publication meets the minimum requirements of

American National Standard for Information Sciences—Permanence of Paper
for Printed Library Materials, ANSI/NISO Z39.48-1992.
 To all graduate students.

May this journey be one that you enjoy,

are proud of, and benefit from for many years.
 Contents

Introduction ix
Chapter 1 Variables and How to Measure Them 1
Chapter 2 Intelligently Managing Your Data 11
Chapter 3 Graphing Techniques to Support Your Data 17
Chapter 4 Means, Medians, and Modes and When to Use Each 27
Chapter 5 Measuring Variability: An Important Role in Data 35
Chapter 6 Random Sampling and Other Useful Sampling Strategies 47
Chapter 7 Stating Hypotheses and Hypothesis Testing 55
Chapter 8 t-Test Procedures 63
Chapter 9 ANOVA Procedures 75
Chapter 10 Chi-Square Procedures 85
Chapter 11 Correlation Procedures 93
Chapter 12 Regression Procedures 103
Chapter 13 Practical Tips for Graduate Students 113
About the Authors 121

vii
 Introduction

One of the greatest barriers to completing a graduate thesis or a doctoral dis-

sertation is research that includes statistics. Oftentimes, the journey through
graduate school is gratifying when the content courses in the chosen field
of study are undertaken. Conversely, the courses in research methods and
statistics are met with trepidation and anxiety. Many graduate students feel
lost when it comes to dealing with data. Some students even quit at the end
of their programs, which is why ABD (all but dissertation) applies to many
PhDs.
This book is intended to help you move through the barriers that seem
formidable but are not. This is not a statistics text and does not purport to
be such. Rather, it is a way to introduce you to basic statistical concepts in a
gentle, easy-to-comprehend manner so that you can become the success that
you set out to be when you signed on to your academic journey.
Chapter 1 presents the concept of variables: What are they and how do
you measure them? Knowing the differences in measurement scales, selecting
the correct measurement scales for variables, and identifying which ones you
use for different statistical procedures are significant steps in setting up data.
This understanding is vital; the use of statistical procedures depends on this
basic knowledge.
Next, chapter 2 discusses the steps needed to manage data sets. Often-
times, the steps are bypassed allowing for errors and misinformation to
emerge from data sets. Taking the time to develop frequency distributions

ix
x  •  Introduction

and more importantly, examining them in detail, are some of the most valu-
able steps in data management.
Chapter 3 presents graphing techniques that are most beneficial to gradu-
ate students’ research. Correctly choosing and then setting up graphs are
discussed in this chapter along with some mistakes to avoid. Graphing can
be an asset in the ability for graduate students to communicate their results
in a simple, informative manner.
Chapter 4 begins the statistics presentation with descriptive statistics. The
mean, median, and mode are used often every day. This chapter explains the
three measures and when to use each one. Chapter 5 is a complement as it
takes chapter 4 and explains why variability is critical as a partner to de-
scriptive statistics. The benefit of understanding, and then using, the normal
curve and percentiles is discussed in this chapter.
Chapter 6 shows graduate students the different sampling strategies. It
explains the process of random sampling and also includes several other
sampling strategies that might be of use in different situations. Especially
important is the presentation of nonrandom sampling options, which may
be more practical for graduate student research.
Chapter 7 reviews hypothesis testing as a prerequisite to running statisti-
cal analyses. It is imperative to state the research question and corresponding
null hypothesis that statistics will be used to test. Type I and Type II errors,
along with the Bonferroni correction, are examined.
Chapters 8 and 9 begin the introduction of inferential statistics. T-tests
and ANOVA (analysis of variance) procedures are examined in a straightfor-
ward manner so that graduate students can confidently execute them. These
parametric statistics have application in many graduate research designs
where two or more groups are compared on some measure.
In chapter 10 the nonparametric statistic chi-square analysis is presented.
Because this procedure examines categorical variables such as demographics,
its usage to graduate students is immense.
Correlation and multiple regression are presented in chapters 11 and 12.
Correlation is a descriptive statistic that tests relationships among variables.
This allows graduate students to avoid speculating about what is and is not
related. The correlation coefficients do the work. Multiple regression is a
complex, multivariate procedure presented in a basic and uncomplicated
manner. This statistic has many applications but may require external exper-
tise to help execute it.
Chapter 13 gives graduate students advice on determining their research
topic, using library references, selecting a research design, and writing up re-
 Introduction  •  xi

sults from their data. Strategies to do this with finesse, accuracy, and honesty
are offered to help graduate students make the most of their data.
Expertise with data is expected of graduate students. It is part of the rea-
son you are earning an advanced degree. Simplifying Statistics for Graduate
Students: Making the Use of Data Simple and User-Friendly is an antidote for
the research and statistics blues. It can provide graduate students with the
confidence they need when signing on to the research methods and statistics
courses required in their programs. Graduate students should enjoy the jour-
ney and this book can help.
 CHAPTER ONE

Variables and How to Measure Them

Few of us eagerly anticipate taking a research or statistics course. That in-

cluded me in my own graduate program. Many times, there is a knowledge
gap between high school mathematics and the quantitative courses of gradu-
ate study. For some, math was a subject in high school taken primarily as a
college application requirement. For others, the image of statistics was and
remains formula-driven, and thus intricately complex and intimidating.
Some of those brave souls who have taken a statistics course most likely had
their preconceptions confirmed by a professor who taught the subject from a
computational perspective.
Consequently, many graduate students and new faculty are wary of pursu-
ing research based on data. This is unfortunate; it can be so exciting! This
book is not a statistics text. It is a book for those who want to use data for
their thesis, dissertation, or faculty research with a sense of confidence.
Look at the data set in textbox 1.1 that represents pre- and posttest values
for only twenty-eight subjects. It can make your head spin. What can you tell
from “eyeballing” the data set?

• Are most subjects performing well or performing poorly?

• What is the typical performance score in the data?
• Are subjects making gains–if so, to what degree?
• Are segments of subjects (by gender, age, ethnicity) performing differently?
• How does performance compare with peers?

1
2  •  Chapter One

Textbox 1.1. Pre- and Postdata Set for 28 Subjects

001 2221222132111211201021202011102011111210122011221011
002 1311113131112300301113113111101333121321111101110110
003 32213221322122102020211030 2120200120021031200020011
004 3330333123300320002032203021303001200210103000300031
005 22112213232112112011221121 2311211122021 231201121111
006 1221122111111211211121212111111011110211111001211111
007 222122212331121020321131301131200210121 1212000310211
008 2211221112210211101122212012112101211211311010211122
009 33313321232102203100213030 21302031300330303000320030
010 33303330332003203031331030 22302000200310201000220021
011 3210322132210210201032202010212001100111111011111111
012 33313320332103203020323130 3330311110032 031300122012
013 33303231322003103130313030 2331100120021 031200122013
014 1201221102211202102011023013301103100310012001120111
015 2221222122200210213130102011312003310100313001220011
016 1310321232100101300031103013321103110110321001310101
017 2201221032021201211123123110101210013101111030203101
018 2201221032021201211123123110101210013101111030203101
019 2330232023312313302121203011202112301210303010210011
020 3200120133101100201021020010011003010100313001010131
021 2221332033210311212032202022103001211301313002330030
022 2320322112310320212131102011212002110212130001100312
023 3321312213310100300130002111212100310210321031332300
024 33202320223003203010323030 20303000100230302000110021
025 1211222122311211312111211112122111111212212101111211
026 22202320132013301111212011 2120101110111 0311001110131
027 23312230323013023021122211 2310212210211 0122011122210
028 23312221122012102011222120 2120210121121 1112001221121

Most of us cannot answer these questions by reviewing a set of data. Even

the small set of 28 pre- and postvalues displayed in textbox 1.1 is incompre-
hensible. Our human minds are limited in their capacity to absorb a set of
numbers, retain them, process them, and then make sense of them.
Statistics is an area of math that seeks to make order out of a diverse col-
lection of facts. With a single number, statistics can summarize the proper-
ties of large groups of numbers. They help us to obtain an understanding of
our data. Statistical techniques summarize data, making complex masses of
numbers simple. By using them, we are able to crunch large sets of numbers
into usable and actionable information.
 Variables and How to Measure Them   •  3

Basically, statistics should be thought of as numbers that summarize the

properties of large groups of numbers. If you have been given a massive assort-
ment of data and have no hope of understanding them, although you have
looked at them for hours, use statistics. Life will become less complicated.

General Classifications of Data: Quantitative and Qualitative

Statistics involves collecting information called data, analyzing it, and mak-
ing meaningful decisions based on the data. Collected data, which represent
observations or measurements of something of interest, can be classified into
two general types: qualitative and quantitative.

• The term qualitative data refers to observations that are descriptive.

These data represent categories. This might include demographic data
such as gender, ethnicity/race, town of residency, education, household
composition, and others.
• The term quantitative data represent various observations or measure-
ments that are numerical such as academic grades, blood pressure read-
ings, customer satisfaction ratings, attendance, and others.

The Essential Component in Data: Variables

Data are derived from characteristics about individuals, objects, or events.
These characteristics are called variables. Anything that varies and can be
measured is called a variable. Variables can be quantitative or qualitative.
We attach numbers to our variables in an effort to measure them and apply
statistics to them.
Variables that are qualitative are called categorical variables. They have
different categories, and each category takes on a whole number or integer
to represent it. For example, school type would be a categorical variable. For
some school systems it might have three categories:

1 = Elementary (Kindergarten through fifth grade)

2 = Middle (sixth through eighth grades)
3 = Secondary (ninth through twelfth grades)

Variables that are quantitative are classified as either discrete or continuous. If

they are discrete, they can take on only whole numbers or integers. For ex-
ample, discrete variables might include number of children in the home, to-
tal number of hospital admissions, SAT scores, number of DUIs, or monthly
4  •  Chapter One

deaths from COVID-19. These are represented by whole numbers. If they are
continuous variables, they can take on fractions or decimals. Body weight,
credit card balance, body temperature, height, or race time are examples of
continuous variables, where race time in the 100-meter dash might be 13.2
seconds. See table 1.1 for a classification of variables.

Table 1.1. Classification Chart for Data and Variables

Qualitative Data Categorical Whole Numbers Residential Locations
Variables Only Urban (1)
Rural (2)
Suburban (3)
Quantitative Data Discrete Whole Numbers Hospital Readmissions
Variables Only 3, 5, 8
Continuous Fractions and Grade Point Average
Variables Decimals 3.29

All variables are measurable, or they would not be called variables. We

have to assign numbers to all of our variables in order to apply statistics to
them. In order to do this, we have to understand the scales of measurement.
This is essential. Unfortunately, it is an area that is largely misunderstood. If
measurement scales are used incorrectly, the wrong statistical technique will
be applied to the data. This will yield erroneous information.

Four Measurement Scales

Whatever exists, exists in some amount and can be measured. Measure-
ment involves quantifying people, objects, or events on their characteristics.
When we collect information about people, objects, and events, we must
turn that information into numbers so that we can measure it and make
deductions about what we find out. We must express it in numbers not just
descriptive phrases.
Because measurement involves quantifying people, objects, or events on
their characteristics, you must assign numbers to variables. There are four
scales of measurement used to assign numbers to variables. Each is differenti-
ated according to their degree of precision. For example, a pediatrician might
measure a child’s general health by level of physical activity, nutritional
practices, and weight. The latter is the most precise measure although all
three can provide information about the same variable—health.
 Variables and How to Measure Them   •  5

Nominal Measurement Scales

The first measurement scale in the hierarchy is the nominal scale. The term
“nominal” means to name. The properties of nominal scales are:

• Data categories are mutually exclusive.

• Data categories have no logical order.

Observations are simply classified into categories and assigned a number

with no relationship existing between or among the categories. It classifies
without ordering. The variable Home Loan Application can be nominally
scaled with a number of either 1 (complete) or 0 (incomplete) categories.
Another example is the variable of Workshop Attendance; it would have
nominal scaling—present (1) or absent (0).
The numeric values could be any combination. Three options for nominal
coding of the same variable, such as Residential Location, are presented in
table 1.2.

Table 1.2. Three Nominal Scale Options for the Variable Residential Location
Option One Option Two Option Three
Numbers Numbers Numbers
Variable Assigned Variable Assigned Variable Assigned
Rural 1 Rural 2 Rural 3
Urban 2 Urban 3 Urban 2
Suburban 3 Suburban 1 Suburban 1

There is no logical ordering of the categories. Numbers are assigned to

the categories, but no quantitative meaning is assigned to the numbers. The
numbers mean absolutely nothing. We are simply using the numbers to classify
people, objects, and events. The numbers we assign to those categories have
no ordering, no ranking, no “higher than,” no “lower than,” no “more of,”’
or no “less than” associated with them. They are what their name suggests—
nominal, a name to identify. The numbers we assign to our variables in their
categories are not quantifiable except to count them up. The total number of
males we have will be how many we add up with the value we have assigned
to it. This is important to note when you are calculating statistics.
There are two basic requirements for nominal measurement:

1. All members of one category must be assigned the same numeral.

2. No two categories are assigned the same numeral.
6  •  Chapter One

Some nominally scaled variables have only two categories. These are
called dichotomous variables. This means that only two numbers can be as-
signed to each of the categories, respectively. Some examples of dichotomous
variables may include primary language spoken at home (English/Spanish),
parent (father/mother), staff (supervisor/trainee), student (special education/
general education), loan applicant (accepted/rejected), patient (inpatient/
outpatient), and many others. You may use two numbers (such as 1 and 2 or
0 and 100 or 200 and 300) as the nominal scaling applied to each category.
Nominal scaling is ideal to categorize schools. The type of school the child
attends might be categorized as Elementary (1), Middle (2), and Secondary
(3). Or, if we have five elementary schools, we could nominally scale all
seven of the schools that the children in a town attend as in table 1.3.

Table 1.3. Nominal Scale for a School

System with Seven Schools (Five
Elementary)
Number
Schools in the School System Assigned
Peachtree Elementary School 1
Riverside Elementary School 2
Justin Elementary School 3
Harris Elementary School 4
Connor Elementary School 5
Maples Middle School 6
Maples High School 7

Being able to nominally scale our variables this way allows us to drill down
our data or inspect our data much more thoroughly.
One pitfall threatens. Many individuals get confused about what the ac-
tual variable is and what are the respective categories. Sometimes, you might
think the variable is female instead of gender with categories. This differen-
tiation is very important, especially when you are investigating statistical dif-
ferences with independent and dependent variables, which will be discussed
later in the book. Ask yourself: What variable are the individual categories
representing? The answer should be the variable itself and not its categories.

Ordinal Measurement Scales

The second type of scaling in the measurement hierarchy is called ordinal,
where there is relative ranking and ordering of an attribute in different cate-
gories. There is “more than” and “less than,” “higher than” and ‘lower than,”
“least of” and “most of.” There is a qualitative relationship among numbers
 Variables and How to Measure Them   •  7

in the ordinal scale. Unlike nominal scaling, the numbers in ordinal scales
have meaning. Ordinal scales give more information and more precise data
than nominal scales do. Here are three examples of ordinal-scaled variables:

Variable: Frequency
Never (1)
Rarely (2)
Sometimes (3)
Frequently (4)
Always (5)

Variable: Satisfaction
Dissatisfied (1)
Satisfied (2)
Very Satisfied (3)

Variable: Attainment
Below Expectations (1)
Meets Expectations (2)
Exceeds Expectations (3)

The properties of ordinal scales are:

• Data categories are mutually exclusive.

• Data categories have some logical order.
• Data categories are scaled according to the amount of a particular char-
acteristic they possess.

One of the most common uses of ordinal scaling is with ratings, preferences,
rankings, goal attainment, satisfaction, degrees of quality, and agreement
levels—typical of the Likert scale found on questionnaires. Here is an ex-
ample of ordinal scaling using a rubric to measure student performance:

Beginning (1)
Developing (2)
Proficient (3)
Advanced (4)

As a note, it is helpful to assign your numbers in ordinal scaling in ways

that make sense. A rating of “excellent” on a 5-point rating scale can have
8  •  Chapter One

a value of 5 or a value of 1. It is relatively arbitrary. However, it makes more

conceptual sense that a low rating with a value of 1 should be assigned to a
low rating—“poor,” whereas a high rating (5) should be assigned to a high
rating—“excellent.” After you tabulate your data, it is easier to interpret
what you have found out if you assign your ordinal numbers in gradations
that make sense.
Another example of ordinal scaling that is used often is applied to agree-
ment statements. There are five responses that could be assigned ordinal
values. Again, it makes more conceptual sense to assign the value “1” to the
least amount of agreement—Strongly Disagree and vice versa. Higher levels
of agreement should be reflected in the values assigned—5 = Strongly Agree.
Again, this is arbitrary.

Strongly Disagree (1)

Disagree (2)
Undecided/Uncertain (3)
Agree (4)
Strongly Agree (5)

Interval and Ratio Measurement Scales

There are two metric scales of measurement: interval and ratio scales. Be-
cause of their metric nature, these two measurement scales afford the most
precise data. With both there are equal intervals or units between any two
consecutive numbers. For interval and ratio scales, the distance between
numbers 2 and 3 and between 4 and 5 are exactly the same. This is differ-
ent from ordinal scales where we cannot claim for certain that the distance
between good (3) and fair (2) is exactly the same as between poor (1) and
fair (2). We assign numbers in ordinal scales, but the distance is not in exact
units. With interval and ratio scales, there is the ordering that is found in
ordinal scales, but now we have exact units of measurement.
The only difference between interval and ratio scales is the role of zero
(0). In interval scales zero is artificial rather than real. For ratio scales, the
role of zero is real; there is the absence of that variable being measured.
With these two scales we can compute high-level, sophisticated statistics
called parametric statistics (which will be discussed later in the book).
The properties of interval scales are:

• Data categories are mutually exclusive.

• Data categories have a definite logical order.
 Variables and How to Measure Them   •  9

• Data categories are scaled according to the amount of a particular char-

acteristic they possess.
• Equal differences in the characteristics are represented by equal differ-
ences in the numbers assigned to the categories.
• The point zero (0) is just another point on the scale.

Interval scaling is used often in graduate student research. When we measure

leadership, personality, aptitude, interests, attitudes, achievement, and other
variables, it is likely that they are measured on interval scaling.
Ratio scaling is similar to interval scaling in terms of equivalent values
between numbers. The only difference is that in ratio scales a zero (0) means
something important. It means the absence of whatever the scale is measur-
ing. This type of scaling is encountered more commonly in the physical
sciences where there can really be “none”—weight, time, height, calories, or
wherever zero is real, not artificially created. If you are measuring the level
of pain in postsurgical patients (where none is possible), you would be using
ratio scaling.
The properties of ratio scales are:

• Data categories are mutually exclusive.

• Data categories have a definite logical order.
• Data categories are scaled according to the amount of a particular char-
acteristic they possess.
• Equal differences in the characteristics are represented by equal differ-
ences in the numbers assigned to the categories.
• The point zero (0) reflects absence of the characteristic.

Chapter Summary
As a summary of the four measurement scales, they can be remembered by
the following characteristics:

• Nominal: numbers are assigned without order and meaning.

• Ordinal: numbers are assigned with order but without equal intervals
between them.
• Interval: numbers are assigned with order and equal units.
• Ratio: numbers are assigned with order, equal units, and a true zero
point.
10  •  Chapter One

When setting up data, graduate students must know the information

covered in this first chapter. It is the foundation of being data literate.
Knowing what the differences are in measurement scales, selecting the cor-
rect measurement scales for variables, and identifying which ones to use for
different statistical procedures are significant steps in setting up data. This
understanding is vital; the use of statistical procedures depends on this basic
knowledge.
 CHAPTER TWO

Intelligently Managing Your Data

Often, when we look at a set of data, it is basically a blur of numbers. This

is easily addressed by displaying data with a method to the madness. There
are several steps that graduate students should follow when they are working
with data sets.
The first step is simply organizing your data. Many graduate students skip
this step because it is elementary and seems too simple to be useful. It is
advised that you take the time to organize your data. Doing this will help
you to understand your data and make preliminary conclusions. Equally im-
portant, you will be familiarized with the data to the point where oversights
can be uncovered, avoiding embarassment to you and minimizing the value
of your work.

Frequency Distributions
One of the most useful expenditures of time is to produce a frequency distribu-
tion for each of the variables in your data set. This is a systematic arrange-
ment of numeric values from the highest to the lowest—with a count of the
number of times each value was obtained. It is a procedure for organizing and
summarizing data into a meaningful representation. It does not tell you ev-
erything about your data, but it provides a beginning and convenient way of
grouping data so that meaningful patterns can be found. Statistical packages
like SPSS can do this for you, but here is how you do it if you did it yourself.

11
12  •  Chapter Two

The first step in developing a frequency distribution is to put all your

values for one variable in order from the lowest value to the highest value in
a column format. Then, use slash marks or tally marks beside every value in
your column—each and every time it occurs. Some values may occur only
once, and some may occur more often. This is why tallies are helpful. Lastly,
count up the tally marks and place a real number beside them that shows the
frequency that each value occurred in your distribution. This information in
chart form can make sense of a set of numbers that appears to be a jumble
at first glance.
There are symbols that are typically used. When we set up our frequency
distributions, the lowercase letter “x” indicates our values, and the lowercase
letter “f” indicates the frequency of occurrence. For example, let’s say your
website click-through data look like the data set shown in table 2.1.

Table 2.1. Data Set for Daily Website Clicks for 50 Days
Week Monday Tuesday Wednesday Thursday Friday
1 699 459 450 450 445
2 420 420 420 420 420
3 430 420 445 435 467
4 420 320 420 320 420
5 689 445 479 450 467
6 435 430 430 430 420
7 420 467 467 450 435
8 320 320 320 320 420
9 430 445 430 435 467
10 420 420 320 420 467

The lowest number of clicks is 320 and the highest is 699. The frequency
distribution orders the 50-day click rates from lowest to highest and then
tally marks are placed to show the frequency each was obtained. Percentages
(the relative frequencies) are calculated to provide even more information to
you. Finally, cumulative frequencies and cumulative percentages are added to
complete the picture (review table 2.2).
From this frequency distribution, you can draw some important conclusions:

• The most frequently occurring click rate was 420—the most typical for
30% of the days.
• More than half (56%) of the click rates were 430 or lower.
 Intelligently Managing Your Data   •  13

Table 2.2. Frequency Distribution for Daily Website Clicks (N = 50 days)

Frequency Relative Cumulative Cumulative
Clicks Tallies (f) Frequency Frequency Frequency
320 /////// 7 14% 7 14%
420 /////////////// 15 30% 22 44%
430 ////// 6 12% 28 56%
435 //// 4 8% 32 64%
445 //// 4 8% 36 72%
450 //// 4 8% 40 80%
459 / 1 2% 41 82%
467 ////// 6 12% 47 94%
479 / 1 2% 48 96%
689 / 1 2% 49 98%
699 / 1 2% 50 100%

• There is a broad range of click rates from highest (699) to lowest (320).
• Two click rates (699 and 689) were very different from the other
rates—outliers.

Class Intervals
Let’s say that you have twice as many days in your sample: 100 instead of 50.
You might use a shorthand called class intervals. The number of classifications
of data (classes) can be reduced by combining several of the actual clicks
into an interval or band of values. You are essentially consolidating the data
points into bundles to make them more manageable and comprehensible.
A good rule of thumb is to have between 10 to 20 class intervals, alto-
gether. This is particularly true if you intend to graph your frequency distri-
bution data into a frequency polygon or histogram, which will be discussed in
the next chapter. Ten to 20 class intervals will summarize your data without
distorting the shape of your graph. Too few intervals compress the data and
thus conceal meaningful changes in the shape of your graph. Too many in-
tervals stretch out the data, so they are not summarized enough for a clear
visualization.
There are two general rules for using class intervals:

• The class interval should be of such size that between 10 and 20 inter-
vals will cover the total range of values. This provides for a manageable
number of intervals without losing the general shape of the distribution.
• Whenever possible, the width of the class interval should be an odd
rather than an even number. Under this rule, the midpoint of the
 14  •  Chapter Two

interval will be a whole number rather than a fraction. This will be-
come important when you graph your data.

How to Construct a Frequency Distribution

Using Class Intervals
1. First, determine the range in your frequency distribution from your
highest value to your lowest value. For the point of illustration, let’s
say we have a low of 320 and a high of 767. The difference or range is
(767 – 320) = 447.
2. Divide the range by either 10 or 20 to get the number of intervals that
makes the most sense and is most manageable. For 10, you would di-
vide 447 by 10, and it would yield 45, as the class interval. For 20, you
would divide 447 by 20 and get 22, as the class interval. For the sake
of illustration, we choose 10; so, our class interval is 45.
3. Begin your intervals a little below your lowest value; in our data it is
320. With 45-point increments, our first-class interval is 300–344.
4. Once you set up your class intervals, then you proceed the same way
you did with the frequency distribution in Table 2.2. Use tallies for the
interval in which your values fall. A tally would include a click rate of
396 in the interval of 391–436.

Table 2.3 displays what our frequency distribution with class intervals would
look like.

Table 2.3. Frequency Distribution of Daily Website Clicks with Class Intervals
(N = 100 days)
Class Intervals Frequency Relative Cumulative Cumulative
of 45 Tallies (f) Frequency Frequency Frequency
300–344 / 1 1% 1 1%
345–390 ////// 6 6% 7 7%
391–436 ///////////////////////// 25 25% 32 32%
437–482 ///////////////// 18 18% 50 50%
483–528 /////////// 11 11% 61 61%
529–574 //////////// 12 12% 73 73%
575–620 ///////// 9 9% 82 82%
621–666 ////// 6 6% 88 88%
667–712 ////// 6 6% 94 94%
713–758 ///// 5 5% 99 99%
759–804 / 1 1% 100 100%
 Intelligently Managing Your Data   •  15

As you can see, using class intervals to compose a frequency distribution

makes it easy to analyze as well as to present your data. Again, important
conclusions can be drawn:
• Most of the click rates were low.
• The most frequently occurring click rates were between 391 and 436.
• There is a broad range of click rates from below 345 to over 758.
• There were a few extremely high and extremely low click rates.

As a note, using class intervals is less cumbersome when your data are con-
tinuous and have a wide range of values. Think about the variable of age
and the adults in your community if you were conducting a survey of city
residents. Listing all possible ages might begin at 18 and go beyond 100 or
more years! Doing a tally of each and every age would require a lot of paper,
never mind time. This is a case when categorizing your variable into class
intervals is a good idea. You might use intervals of five years in this case so
those class intervals would look like this. (There would be less than 20 in-
tervals for adult ages.)

Class Intervals for Age for City Residents

18–22
23–27
28–32
33–37
38–42
43–47
48–52
53–57
58–62
63–67
68–72 and so forth.

On the other hand, there are times when using class intervals shortchanges
the picture; you lose valuable data. Think about the variable “years that
nurses have been employed by your hospital.” If you used class intervals, you
would lose key information. A nurse who is brand-new, one who has been
with you for only one year, and one who has been on staff for three years may
be very different from each other. If you collapsed the “years employed” into
intervals of five years as shown with the age data, you would lose important
insights. Much of this is common sense. What is the information you want
to uncover with your data?
16  •  Chapter Two

Why Frequency Distributions Are Helpful

Whenever you compile a set of data, it is highly recommended that you first
set up (or run-on SPSS) a frequency distribution and study it. This is a strong
recommendation. Do not skip this step. Knowing your data thoroughly is criti-
cal before you begin to dig deeper. Besides developing an expertise with your
data set, there are several practical aspects to the development of frequency
distributions:

• You can identify atypical or odd values like that in table 2.2 (689 and
699). You have to be able to identify outliers before you can determine
what you should do with them. And it is important to decide whether
to keep or discard outliers. They can skew your results!
• You can compare the spread the values, visually. If you set up two fre-
quency distributions representing two segments of your data (such as
Weeks 1–5 and Weeks 6–10), you might be able to infer why the click
rates varied.
• By indicating the frequency that each value occurred in the distribu-
tion, this information could help you to make decisions about statistical
analyses you will execute later. This is particularly true with categorical
variables. For example, if you wanted to categorize click rates into high,
moderate, and low, you might look at the percentages to classify them.
• Finally, the frequency distribution may reveal data entry errors that you
typed in accidentally. If you had a click rate of 9 or 9999 that showed
up in your frequency distribution, you would know that that was an
error. You would be able to make your correction and not run the data
blindly, compromising the purity of your data set. This happens often
when frequency distributions are not run on all the data before the sta-
tistical analyses are executed. Take the time to do the basic work before
you plunge into the analytics.

Chapter Summary
The process of becoming an expert with your data takes work. Setting up the
frequency distribution for your data set, graphing it, and then interpreting
what it conveys, are often-skipped steps in data management. The time it
takes to do the “grunt” work as a preliminary step is a meaningful invest-
ment of time. There are advantages to be gained in many areas. You will be
able to reduce errors, make wise decisions about statistical techniques, and
ultimately provide a better foundation for interpreting your results. Take
the time to know your data set. It will be well worth the hours spent. Don’t
shortchange yourself and compromise your research quality.
 CHAPTER THREE

Graphing Techniques
to Support Your Data

The Shapes of Graphs from Distributions

The frequency distribution is an excellent tool for displaying and inspecting
your dissertation or thesis data. Yet it does not convey the information as
quickly or as impressively as graphs do. Graphs allow us to see the shape of
a distribution. They are designed to help obtain an intuitive feeling for the
data at a glance. The message should be readily apparent.
An effective graph is simple and clean. It should not attempt to present so
much information that it is difficult to comprehend the message. It should be
complete within itself and require little if any explanation in the narrative.

Benefits of Using Graphs

Graphs are most beneficial because they can provide you with a summary
data sheet as well as an opportunity to eyeball data. By looking at the shape
of the curve (the hump) in your graph you can make some general conclu-
sions quickly about the values in your data set. The illustrations in figures 3.1
through 3.5 will help in interpreting shapes.

• If the curve has a hump in the middle, its symmetry means that you
have most values in the middle of the range—few high and few low.
(Normal curve, as in figure 3.1.)

17
18  •  Chapter Three

Figure 3.1. Normal Curve. Source: Author developed

• If the hump of the curve is inclined toward the left, you have more low
values in your data set. The tail or end of this type of curve indicates
atypical or odd values in your data. In this case, there was a minority of
high values. (Positively skewed, as in figure 3.2.)

Figure 3.2. Positively Skewed Curve. Source: Author developed

• If the hump of the curve is to the right, you have more high values in
your data set. The tail or end of this type of curve indicates atypical or
odd values in your data. In this case, there was a minority of low values.
(Negatively skewed, as in figure 3.3.)
 Graphing Techniques to Support Your Data   •  19

Figure 3.3. Negatively Skewed Curve. Source: Author developed

• If the hump of the curve is flat, you have values that are spread out a
great deal. Values are all over the map. (Platykurtic as in figure 3.4.)

Figure 3.4. Platykurtic Curve. Source: Author developed

• If the hump of the curve is peaked, your values are very, very similar.
There is not much difference among values. (Leptokurtic as in figure
3.5.)
20  •  Chapter Three

Figure 3.5. Leptokurtic Curve. Source: Author developed

The discussion about the shapes of graphs will have more meaning in subse-
quent chapters.

Types of Graphs
Frequency Polygons
To construct a graph from a set of data, we start with the frequency distribu-
tion. A frequency distribution can be graphed into a frequency polygon. This
is how you convert a frequency distribution into a frequency polygon.
Draw a vertical side or ordinate axis of your graph. This is called the Y axis.
On the Y axis it is an accepted practice for the frequencies to be plotted. The
horizontal axis, called the abscissa, is used to plot the variable that you are
displaying data for. The horizontal axis is the X axis. As a rule of thumb for
good visual presentation, the vertical axis should be roughly two-thirds the
length of the horizontal axis.
Connect the points with a dot that intersects the values on the abscissa
axis to the frequencies on the ordinate axis. If class intervals are used, con-
nect the midpoint of the interval on the horizontal axis with the frequency
on the vertical axis. (This is why it is helpful to use odd numbers for the
span of your class interval. The midpoint is easy to identify.) As a note, if
the range in frequencies is large, you may want to start the vertical axis with
a value that is not zero (0).
Plotting the data this way allows the graph to take on a shape. When
you look at that shape, you can make a conclusion about the data from the
 Graphing Techniques to Support Your Data   •  21

frequency distribution. When there are a lot of data from many values, the
.

frequency polygon looks like a smooth curve. When there are fewer values,
it is jagged. Either way, the picture tells a story. The graph is a picture that
gives us an immediate message about our data. We look at it and we can infer
something about our data.
For our daily website click graph with class intervals, the simple message
is that there are many low values in this distribution, and it is a positively
skewed curve (see figure 3.6).

Figure 3.6. Frequency Polygon of Website Clicks. Source: Author developed

Bar Graph
A bar graph is a very convenient graphing device that is particularly useful
for displaying nominal data such as gender, ethnicity, and other categorical
variables. The various categories are located along the horizontal axis. The
frequency or count, as always, is on the vertical axis. The height of each bar
is equal to the frequency for that category. This is very helpful for viewing
the differences in the data for individual groups on some variable.
If your dissertation or thesis were presenting demographics on households
with and without children, and there were 200 households in your sample,
the results in the bar graph would be obvious. Even without the actual num-
bers inside the bars, the two columns of the bar graph show that twice as
many respondents did not have children as compared with those who did.
Please refer to figure 3.7.
22  •  Chapter Three

Figure 3.7. Bar Graph of Household Composition. Source: Author developed

Histograms
A histogram is another pictorial representation of a frequency distribution.
While bar graphs are useful for nominal-scaled, categorical variables, the
histogram is preferred if there is a quantitative variable where categories have
gradations such as age or income. A general rule in laying out the histogram
is to make the height of the vertical axis equal to approximately two-thirds
the length of the horizontal axis. Otherwise, the histogram may appear to be
out of proportion.
A vertical bar is constructed above each respective category equal in
height to its frequency count. All the rectangles have equal width. This is
a very common way to display data. Figure 3.8 displays what a histogram
would look like if you were reporting age segments. The age differences in
the respondents can immediately be seen. There are many respondents who
were older.

The Difference Between a Histogram and Bar Graph

One difference between a histogram and a bar graph is that with the histogram
the bars are placed next to each other to show continuity between the grada-
tions; when one interval begins, the other one ends. The data bars in histograms
have values attached to them and are continuous, such as with the variable
of age segments in figure 3.8. However, because the data are nominal scaled
in bar graphs, and not related to each other, the bars must not be connected
so any implication of continuity is prevented. This is illustrated in figure 3.7.
 Graphing Techniques to Support Your Data   •  23

Figure 3.8. Histogram of Age Segements. Source: Author developed

Pictographs
A pictograph uses pictures or symbols in place of numbers. Figure 3.9 portrays
a pictograph of the number of books read by a fourth-grade class during the
holiday vacation week. It uses a book icon, representing five books read for
each day Monday through Friday.

Figure 3.9. Pictograph of Books Read by 4th Graders.

Source: Author developed
24  •  Chapter Three

Pie Charts
Another method to display data is by using a pie chart. This graph is very easy
to generate and is user-friendly. It shows the categories of a variable by divid-
ing up “the pie.” Use 100% of the area in the circle and divide it up propor-
tionally to the categories of interest. Figures 3.10 and 3.11 display pie charts
that reflect the same data on Household Composition and Age Segments for
respondents, as depicted on the histogram and bar graph, previously shown.

Figure 3.10. Pie Chart of Household Composition. Source: Author developed

Figure 3.11. Pie Chart of Age Segments. Source: Author developed

 Graphing Techniques to Support Your Data   •  25

Advice for Developing Graphs

Graphing is an excellent way to demonstrate your data to the untrained eye.
There are two recommendations to observe when constructing your graphs.

1. A graph is a picture. So the message should be readily apparent. Keep-

ing it simple is a good practice to follow when graphing (or even when
constructing your tables).
2. Avoid the temptation to emphasize your point or position by “adjust-
ing” the vertical axis. This creates a truncated graph. The vertical scale
is cut off or restricted so that the information becomes distorted. The
picture is misleading. This is a practice that the media often use in
reporting or sensationalizing stories for the public’s (untrained) eye.
The axes are exaggerated so that the graph presents dramatic results
that really were rather “ho hum.” The vertical axis is restricted to a
short range, causing overemphasis of the results in either a positive or
negative light.

The graph in figure 3.12 reports a “steep decline in SAT values.” But look
how the vertical scale has been adjusted to begin at 492, distorting the
picture. The drop over a three-year period is negligible—from 502 to 496!
An initial glance at the graph would cause board of education members to
faint—before they could find out the real story in the actual data.

Figure 3.12. “Steep” Decline in SAT Scores over Three Years. Source: Author developed
26  •  Chapter Three

Chapter Summary
Graphs are most beneficial to graduate students because they can provide a
summary data sheet when presenting data. However, it is important to keep
in mind that a graph is a picture that is helping to tell a story about the data.
Those who will be viewing the graph are not as involved in your data and
analysis as you are and have other thoughts competing for their attention.
Don’t waste their time with graphs that are either too simplistic or overly
complex. Use graphs judiciously and with deliberation.
 CHAPTER FOUR

Means, Medians, and Modes

and When to Use Each

The importance of getting to know your data has been discussed. The
frequency distributions and the graphing techniques both produce a basic
profile of your data set. Taking the time to construct them helps to indicate
the characteristics of your data. There are also certain statistics that are
generated for the purpose of describing the distributions of your data or the
relationships between your variables. They are called descriptive statistics.
These very useful statistics bring together large amounts of data so they can
be presented and comprehended with minimal effort.
Descriptive statistics are widely applied. A good example of real-life ap-
plications is the US Census. By using some of the popular descriptive statis-
tics, we get a sense of important characteristics of households in the United
States. For example, descriptive statistics that are available in census data
may indicate:

• Average household size

• Ethnic, racial, and gender breakdowns
• Employment rates
• Households in poverty
• Characteristics of homeowners and renters
• Per capita income
• Levels of educational attainment

27
28  •  Chapter Four

If we had to generate this type of information from looking at raw data, it

would be too complex to grasp. Descriptive statistics make it easy to get a
sense of what is typical or average.
Measures of central tendency represent an important collection of descrip-
tive statistics. These are designed to describe the central part of the distribu-
tion that you created with your frequency distribution. They tell where most
values appear to group, cluster, or fall together—their commonality. In other
words, they tell us the average. With one single number you can obtain an
accurate picture of your entire distribution. It suggests the profile of a group
as a whole and is a concise description. There are three measures of central
tendency: the mean, the median, and the mode.

The Mean
The mean represents a whole data set with one single number! The mean is
the arithmetic average in your distribution of values. To obtain the mean you
add up all of the values (Xs) and divide by the total number of values (n) in
your distribution. As a rule, your data set should tend to cluster together and
not be spread all over. The beauty of the mean is that you utilize each and
every value in your data set to calculate it. As a result, it is the most stable
measure of central tendency and the one used most often. Since it is an aver-
age, you probably have been calculating a mean often in your everyday life.
In professional journals, the mean is often noted as M.
Here is a simple example to illustrate the use of the mean. During this
past school year, the board of education was contemplating the creation
of additional elementary school classroom space. The three neighborhood
elementary schools were brimming with children. This was due to recent
out-migrations, where families were moving from cities to a more rural lo-
cation. The board members decided to find out what the average class size
was in order to arm themselves with supporting data for overcrowding and
ultimately for additional classroom space.
Although class sizes were compiled for all grade levels, the data set for ten
third-grade classrooms was reported as follows in table 4.1.
The mean was calculated at 27 students per classroom—larger than what
was desirable. Here was the evidence the board of education needed. With
just one number representing all 10 classrooms, the board had a solid piece
of information.
The mean is a very good measure of central tendency.
 Means, Medians, and Modes and When to Use Each   •  29

Table 4.1. Class Size Data for

Third-Grade Classrooms
Third-Grade Classrooms Class Size
Class 1 15
Class 2 19
Class 3 22
Class 4 29
Class 5 29
Class 6 31
Class 7 31
Class 8 32
Class 9 32
Class 10 33
Mean 27

• It is based on more precise measurement scales such as interval and

ratio. Sometimes the ordinal scale is used too. This may occur with rat-
ing scales, performance scales, or satisfaction scales where averages are
useful. Purists would object to the use of a mean with ordinal data sim-
ply because the units of measurement are not exact. However, in many
applications, ordinal data is used with the mean because it makes sense.
That is the determinant of when to apply the mean to your data—if it
is ordinal scaled. If it makes sense in producing good information, by
all means use it. (Conversely, no mean can be calculated with nominal
scaled variables since the numbers have no meaning.)
• As mentioned before, all the values in a data set are used to calculate
the mean. This is another advantage.
• Finally, many of the powerful statistical analyses rely on the mean to
calculate formulas for statistical significance. The mean has become
the queen of central tendency. But this “royal” designation should be
avoided if your data set possesses outliers.

Outliers
An outlier is an extremely high or low value in your distribution. It is atypi-
cal and does not resemble most of the other numbers in your distribution. It
affects the calculation of the mean and makes the mean less representative.
While the mean is looking to create a typical value in your data set, the
outlier is atypical.
What happens is that outliers cause the mean to shift in the direction of
the outlier. If the outlier is a high value, the mean is calculated higher than
it should be. If it is a low value, the mean is lower than it should be. Outliers
30  •  Chapter Four

cause your distribution to become skewed. This problem is particularly true

if the number of values in your distribution is small.
You can tell if you have outliers in your data set by setting up a frequency
distribution and then graphing a frequency polygon. This is another advan-
tage of taking the time to produce frequency distributions and using the
graphing techniques. If you have an asymmetrical curve, outliers are hanging
around. Some frequency polygons in chapter 3 suffered from the existence of
outliers in the data set. The curves in the graphs were irregular as opposed to
being balanced in form.
Going back to the example of classroom size, let’s alter the data set with
three particularly small third-grade classrooms. The data set would look like
that in table 4.2.

Table 4.2. Class Size Data for Third-

Grade Classrooms with Outliers
Third-Grade Classrooms Class Size
Class 1 28
Class 2 28
Class 3 29
Class 4 29
Class 5 29
Class 6 31
Class 7 31
Class 8 outlier 9
Class 9 outlier 8
Class 10 outlier 10
Mean 23

Here, the outliers caused the mean to shift down to 23 children per class-
room. The presence of the outliers decreased the case for additional class-
room space, but it is not a good representation of average class size in your
school system. Most class sizes in your data set distribution are around 28 to
31. If you used the mean as the measure of central tendency in this case, you
would be presenting an underestimated picture of need.
What should you do? Examine your data through the basic work of devel-
oping a frequency distribution, and then use more than one measure of central
tendency. One of the best, when your data has outliers, is the median.
 Means, Medians, and Modes and When to Use Each   •  31

The Median
When there are extreme values or outliers in your distribution, the median
is the preferred measure of central tendency. The median is the counting
average. It is simply the midpoint in your distribution of ranked, ordered
values. The median can be used with ordinal, interval, and ratio-scaled data.
Nominal-scaled data are inappropriate.
To calculate the median, list all of the values in your distribution from the
lowest to the highest and then find the midpoint—the place where it divides
your distribution into equal halves. That is, 50% of all values are above and
50% are below.

The Median: With an Odd Number of Values in Your Distribution

If there is an odd number of values in your distribution, the median is easy to
identify. Find the midpoint in the range of high to low values. That midpoint
is the median, since half of the values are above it and half are below it. A
formula can be used to locate the position in an ordered set of data. It is the

[Median = Number of values plus 1 divided by 2]

If you had 11 values (an odd number), the formula would be (11 + 1) ÷ by 2
= 6. The median would be the sixth value in the set of data where the values
are listed in order from lowest to highest.
Here is an example where we have an odd number (11) of values. The
midpoint is the sixth value or the exact middle of the distribution. In this
distribution, the median is 29, and 29 is an actual value in the data set. There
are five values above and below it in table 4.3.

Table 4.3. Median with an Odd Number of Values

Order of 1st 2nd 3rd 4th 5th 6th 7th 8th 9th 10th 11th
values Value Value Value Value Value Value Value Value Value Value Value
lowest to
highest
Distribution 28 28 29 29 29 29 31 31 32 32 33
of
11 values
Midpoint 50%
 32  •  Chapter Four

The Median: With an Even Number of Values in Your Distribution

If you have an even number of values, find the two that make the center-
most point, and then average them. If you have an even number of values,
the median may or may not be an actual value, depending on what the two
midpoints are. If the midpoints are identical values, then this is the median.
If they need to be averaged, then the median will be an average and not an
actual value in your distribution. That is why we say the median is the mid-
point, a point not a value.
Here is an example of when this is true. In table 4.2 we had our class size
data with the outliers in the group. The outliers shifted our mean down to
23. So, we decide to calculate a median.
The data set consisted of 10 values altogether: an even number of values.
We have to find the two midpoints and average them. The 10 values are
divided in half—five values on the left and five values on the right. The
median is calculated by taking two values at the midpoint or center and
calculating a mean. In this case, there is a value of 29 on the left and a value
of 31 on the right (see table 4.4).
The formula becomes

[Center point 1 + Center point 2] ÷ 2 = Median

In this example, the application of the formula is [29 + 31] = 60 ÷ 2 = 30

(the Median).

Table 4.4. Median with an Even Number of Values

Order of 1st 2nd 3rd 4th 5th 6th 7th 8th 9th 10th
values Value Value Value Value Value Value Value Value Value Value
lowest to
highest
Distribution 28 28 29 29 29 31 31 9 8 10
of
10 values
Center points

The median and midpoint of the distribution is 30. However, note that
the median is not an actual value in the set of data. As you can see, the
median is not affected one bit by the outliers. While the mean was 23, the
median of 30 is much more representative of the data set. This is the greatest
benefit of using the median. Oddball values in the data set do not affect it.
Although the median is not calculated by using all the values in the distribu-
 Means, Medians, and Modes and When to Use Each   •  33

tion, as the mean is, the median plays an important role in stating the central
tendency of your data set—especially if there are outlier values.

The Mode
The mode is an unsophisticated measure of central tendency. It is the most
frequently occurring value in your distribution. It is a simple but rough statistic
to calculate the central tendency of your data. The mode does not need to
be calculated. If you look at your frequency distribution, a simple eyeball
inspection of your data can tell you which value occurred most often. Look
at the tallies or frequencies. It is quick and can be obtained with a glance.
We did that in chapter 2 with our first data set of values in table 2.2. We
looked at the frequency distribution and could see that the mode was 420
for click rates.
The mode provides little information beyond simply identifying the value
in your data set that appears with the greatest frequency. Therefore, it should
only be used when you have a large data set of values, and not just a few. A
small set of values would not have enough frequency of occurrence to de-
velop a mode. You need many values so that you can be certain which value
turned up most often.
When there is one value that occurs with the most frequency in a distri-
bution, we say the distribution is unimodal or when graphed it has one hump
due to the fact that there is one mode. Many frequency distributions have
more than one mode. That is, more than one value turns up at the same high
level of frequency. Two modes in a frequency distribution create a bimodal
frequency distribution. This might occur in a set of data where two groups
are very different on the variable measured. If graphed, there would be two
humps in the curve or shape. If there are more than two modes, the distribu-
tion is called multimodal and the graphing shows several humps or curves.

Chapter Summary
Each measure of central tendency has a role to play in descriptive statistics.
The mean is the preferred method of calculating the center of your data set.
But when outliers emerge, the median is an alternative that is not affected
by these unrepresentative values. If you want to get a ballpark feeling for the
average, the mode will work with a quick glance at the frequency distribution
table. Here is a summary of the three measures and when to use each:
34  •  Chapter Four

The Mean
• If you want the greatest reliability
• If you will be calculating variability and other statistical computations
• If your distribution has no outliers
• If your data are interval, ratio, or ordinal scaled

The Median
• If your distribution is skewed by outliers
• If your data are interval, ratio, or ordinal scaled

The Mode
• If you need a quick estimate
• If you have interval, ratio, or ordinal scaled data
• If you can eyeball the data from a frequency distribution
 CHAPTER FIVE

Measuring Variability
An Important Role in Data

So far, a case has been made about the importance of constructing a fre-
quency distribution and plotting the data into a graph for an immediate
snapshot of findings. Similarly, the value of the three measures of central
tendency was established. They can indicate to us with a single number the
general characteristics of our set of data. There is another essential step:
measuring the variability of data sets. Many times statistics are reported with-
out mention of the variance or spread of values. The media is guilty of this.
Doing so can misrepresent data. This is why.
The variance is the manner in which your data are spread in either direc-
tion from the center of your distribution. It is important to know whether the
values tend to be homogeneous or whether they vary from each other. The
measures of variability tell us how representative our mean (M) is. Are the
values similar or are they spread out?
For example, look at the glucose levels for patients exposed to prediabetes
interventions, either Treatment A or Treatment B, at a hospital. The mean
blood sugar level in both cases is 100, so both treatments seem to be equally
effective. Closer inspection of the data shows that, while those in Treatment
A were very much the same at the end of the treatment, Treatment B had
some very high and very low results. For Treatment A, the lowest result was
98 and the highest was 102, not too far apart. For Treatment B, the lowest
result was 79 while the highest was 140. The spread or dispersion was great in
Treatment B, where there was high variability. Conversely, the low variabil-
ity in Treatment A was evident from the similar values. There are practical

35
36  •  Chapter Five

implications for Treatment B patients that would be overlooked if just the

mean was used to profile outcome data. Please see table 5.1.

Table 5.1. Two Treatments with Similar

Means
Patients in Patients in
Treatment A Treatment B
98 140
102 80
99 79
100 100
99 100
100 79
102 80
98 140
Mean = 100 Mean = 100

There are several descriptive statistics that measure variability: the range,
the variance, and the standard deviation.

The Range
The range is a very simple statistic and the most unsophisticated measure of
variability. In this regard, it possesses characteristics analogous to the mode in
central tendency. It is a rough estimate, quickly computed, but not tremen-
dously stable. This is because it is computed with only two values: the highest
value and the lowest value in your distribution of values. You subtract one
from the other. The range for Treatment A would be 4 (102 – 98) while the
range for Treatment B would be 61 (140 – 79). Even this crude measure of
variability is very helpful in providing us with information that the mean
value for Treatment B is not stable.
The range is influenced by the size of your data set. The larger the data set,
the greater the likelihood of extreme values because you have more potential
for outliers. This is a limitation of the range; outliers affect it. Since you are
using only two values in your data set to calculate the range, if there is an
outlier at either end, it will influence the calculation.

Variability and Graphs

Graphs, discussed in chapter 3, are quite useful in initially determining the
variability in your distribution. The shape of the frequency polygons and
 Measuring Variability   •  37

histograms that are constructed from the frequency distribution can tell us
the story about the spread of values at a glance. This is another asset of the
frequency distribution (and their respective graphs), and a further reason to
use them in getting to know your data set.
If the hump or curve in your graph is markedly flat, the values are spread
out; there is a great deal of variability. Subjects are not alike. This type of
curve, called platykurtic, means that your values are spread out around the
mean value of the distribution. On the other hand, if the hump or curve in
your graph is peaked and tight, there is little spread in values; the variability
is low. Subjects are alike. This type of curve, called leptokurtic, means that
your values are very close to the mean value of the distribution. (Please refer
to the corresponding figures in chapter 3.)
Another example is illustrated in two graphs that display the number of
pounds that bank employees lost during a Get Healthy Competition. The
intervention group participated in a special exercise and nutrition program
and the comparison group did not.
For both groups the mean number of pounds lost was 6 pounds. As you can
see, employees in the intervention group performed similarly. The members
lost between 4.5 and 8.5 pounds each. On the other hand, the comparison
group lost between 1 pound and 11 pounds. In fact, the range was 10 (11 – 1)
for the comparison group and only 4 (8 – 4) for the intervention group! This
difference in spread of pounds lost is different between the two groups. But
the mean is 6 pounds for each group. Please refer to figures 5.1 and 5.2, which
present visuals that indicate the variability or spread of values.

Figure 5.1. Comparison Group. Source: Author developed

38  •  Chapter Five

Figure 5.2. Intervention Group. Source: Author developed

The shapes of the curves immediately indicate the dispersion or spread of

values. For the comparison group the shape of the curve is flat (platykurtic).
The message from the graph is that values are spread out around the mean.
To claim that most bank employees were performing as well without the
special program is false. For the intervention group the shape of the curve is
peaked (leptokurtic). The mean is an excellent measure of central tendency
because there is little spread. It would be safe to conclude that most interven-
tion group employees were more successful than the comparison group even
though mean values are the same. If a statistical procedure were applied to
these data, there likely would be a significant difference between the inter-
vention and comparison groups.

Variance and Standard Deviation

There are two additional measures of variability: the variance and the standard
deviation. Both are much better indicators of dispersion in your data set than
the range. The reason is that they use all of the values in the data set in their
calculation (just as the mean does). While the mode and the range have
similar traits (quick, unsophisticated), the mean and the variance/standard
deviation are very much alike (precise, reliable). They are the kings of vari-
ability as the mean is the queen of central tendency. In addition, each of
these statistics relies on more precise measurement scales such as interval
and ratio. (Some use the ordinal scale but there is a trade-off in precision.)
The formulas for calculating both the variance and the standard devia-
tion can be found in any math or statistics book. However, if you are using
a statistical software package like SPSS, these will be calculated for you.
 Measuring Variability   •  39

Conceptually, each of these statistics is based upon how much each value in
your data set deviates from the mean, and then putting the deviation val-
ues into a formula for computation. The variance is the standard deviation
squared, and the standard deviation is the square root of the variance. They
are more or less “siblings” in the realm of variance. The symbol for reporting
the variance is “sigma squared” (ơ²). The symbol for standard deviation is SD
or sigma (ơ). These are what you might see in professional journals.
Because both the variance and standard deviation are calculated by de-
viations from the mean in your data set, their value is tremendous. With
one single number, you can tell whether most of the values in your data set
cluster closely around the mean or are spread out. The larger the standard
deviation, the more spread out are your values in your data set.
For example, let’s consider customer satisfaction scores over a two-year
period for a restaurant chain where 900 was the highest possible score. Last
year they had a mean of 600 and a SD of 50. This year they had a mean of
610 and SD of 100. Our mean (M) customer satisfaction scores rose from
600 to 610. What appears to be good news for our restaurant chain is in
fact a hidden problem. The standard deviation of 100 shows outlier values
have shifted the mean upward and artificially created the image that all of
our values were rising. While we can celebrate the seemingly good news this
year, it puts a strain on next year when we have to explain to the CEO why
the values are back down to (or below) 600. If the standard deviation were
calculated in addition to the mean, there would be a more accurate presenta-
tion of our restaurant’s performance.
Here is sound advice: Whenever you report the mean, you should report the
standard deviation. They are the statistical couple. The standard deviation
tells you a great deal: how representative your mean is as a measure of central
tendency for your data set.
Look at the data for the number of motor vehicle accidents occurring in three
towns of about the same size and demographic profiles over a five-year period
(see table 5.2):

Table 5.2. Motor Vehicle Accidents by Town

Town Mean Per Year Standard Deviation
Litchfield 60 2
Torrington 60 1
Winsted 60 15
40  •  Chapter Five

At face value three towns have the exact same number of accidents
among residents: 60. So, you assume that this is typical or average. However,
the standard deviations tell a very different story. For Torrington and Litch-
field, the means are very representative of the average. The SD is small for
both. If you actually drew a frequency polygon from the frequency distribu-
tion data per month, these two towns would have leptokurtic curves. The
values would be close to the mean.
But this is not true for Winsted. This town has a greater spread of values
around its mean of 60, as signified by the standard deviation of 15. The fre-
quency polygon for Winsted data would be flat, not peaked. And actually,
the median might be the better measure of central tendency in this case.

Normal Curves or Normal Distributions

The standard deviation is one of the most valuable tools that graduate stu-
dents have at their disposal. One of its primary uses is in helping to interpret
data. To fully understand its value, a discussion of the normal curve is necessary.
Most of the data that we use is thought to be normally distributed. This
means that if we constructed a frequency distribution and then graphed a
frequency polygon, the hump or curve would be symmetrical. When graphed,
the frequency distribution resembles a bell, which means that there are a few
values on either end of the hump or curve; most values are in the middle.

Area under the Normal Curve

A normal distribution creates a bell-shaped curve or frequency polygon. These
are some characteristics of the normal curve:

1. It has one mode (unimodal)

2. The mean, median, and mode are exactly the same number
3. It is symmetrical and bell-shaped

When you draw a normal curve, you observe these characteristics. The
normal curve, when drawn in most statistics books, has ordinates or vertical
lines from the top of the curve to the baseline. When you look at the picture
of a normal curve, visualize thousands of people standing under the curve;
imagine faces looking at you. This is what the normal curve is representing—
data of individuals on whatever variables you are examining.
In the normal curve, the vertical lines are drawn from the top of the curve
to the baseline at 0 and at numbers 1, 2, and 3 to the left and to the right of 0.
Zero (0) represents the mean value or dead center. These vertical lines mark
 Measuring Variability   •  41

off areas under the curve and represent standard deviation units (1, 2, and 3)
or distance from the mean. The standard deviation (SD) units act like a ruler
and divide up the area under the normal curve as illustrated in figure 5.3.

Figure 5.3. Normal Curve, Standard Deviations, and Areas Under the
Curve. Source: Author developed

The base of the normal curve is divided into six units with three SD
above and three SD below the mean. Entries within the graph indicate the
proportion of the total area (or number of values) that falls into each of the
six segments or demarcations. These are the proportions of area under the
SD cutoffs:

• between the mean (0) and +1 SD units are 34.1% of all values
• between the mean (0) and –1 SD units are 34.1% of all values
• between ±1 SD are 68% of all values

• between +1 SD and +2 SD units are 13.5% of all values

• between –1 SD and –2 SD units are 13.5% of all values
• between ±2 SD are 95% of all values

• between +2 SD and +3 SD units are 2.5% of all values

• between –2 SD and –3 SD units are 2.5% of all values
• between ±3 SD are 99% of all values

Since we know that the mean is dead center, it is at the 50th percentile
mark. The next bit of information is important in interpreting values:
42  •  Chapter Five

• One standard deviation above the mean is an area of 34%. This indicates
that if your value is 1 SD above the mean, you are higher than 84% of
all other values in the distribution (50% plus 34%).
• One standard deviation below the mean is an area of 34%. This indicates
that if your value is 1 SD below the mean, you are higher than only
16% (50% – 34%).
• One standard deviation above and below the mean is an area of 68%. Over
two-thirds are within one standard deviation of the mean (±1 #SD). If
you look at the normal curve and pretend again that people are stand-
ing underneath it, then this is very logical. It is where most values fall,
creating the large hump.
• Two standard deviations above the mean cover an area of 48%. Between
the center (0) and +1 SD there is 34% of the area under the curve.
Between +1 SD and +2 SD there is 14% of the area under the curve.
As you can see, the further you get from the center, the less area there
is under the curve. This is because most values are in the center where
the hump is. Again, think of people standing underneath the curve. If
you obtain a value that is two standard deviations above the mean, that
is a select few. This value is better than 98% of all others (50% + 48%).
• Two standard deviations below the mean cover an area of 48%. Between
the center and –1 SD there is 34% of the area under the curve. Between
–1 SD and –2 SD there is 14% of the area under the curve. If you get
a value that is two standard deviations below the mean, this is in the
minority. The value is better than only 2% of all others (50% – 48%).
• Two standard deviations above and below the mean is an area of 95%. If
you look at the normal curve and pretend again that people are stand-
ing underneath it to create the hump in it, then the percentage of 95%
is very logical. Almost all people perform within two standard deviation
units of the mean (±2 SD).

The standard deviation units, together with the percentage of area under the
normal curve, can be used to determine the relative position of values above
and below the mean. The information is defined as the “distance from the
mean.” Extremely useful, this is why.
All standardized measurement tools have means and standard deviations
reported. You can compare your data to the national norms. The normal
curve gives you the ability to make a call on where your sample of subjects
stands in relation to all others. This is because the percentiles under the
normal curve indicate the relative position of those falling above and below
a certain designation.
 Measuring Variability   •  43

Standard or Z Scores
A standard or z score is another useful tool. Its symbol is lowercase z. It tells
you how many standard deviations a raw score is from the mean. You cal-
culate a z score by taking a subject’s raw score (X) minus the mean (M) and
then dividing it by the standard deviation (SD) of your data. This is your z
score. If the z value is negative, then your z score is below the mean. If it is
positive, it is above the mean. If it is zero, it is the mean.
For example, the IQ test has a mean of 100 and a standard deviation of
15. Your IQ was 130. If you apply the formula, your standard score would be
+2.00 or two standard deviations above the mean. Referring back to area
under the normal curve, your score would be higher than 98% of those who
took the IQ test.
Most statistics books have a table that helps you convert your raw data to
z values, and some statistical programs will even do the conversion for you.

Skewed Distributions
Before closing this chapter, we need to address the fact that often our distri-
butions or data sets are not normally distributed. Don’t panic. This is the case
more often than not with thesis or dissertation data. Just as a review, normal
distribution curves have the following characteristics:

• They are unimodal—having one mode only.

• The maximum height is at the mean where the zero (0) or vertical axis
touches the top of the curve.
• The mean, median, and mode are equal.
• They are symmetrical and bell-shaped. One half approximates the other
half, a mirror image.
• There are one-half of the values on one side and one-half on the op-
posite side of the mean.
• There are three standard deviations above and below the mean.
• They assume an infinite number of values underneath the curve, so the
two tails at either end do not touch the abscissa.
• They can be either platykurtic (flat) or leptokurtic (peaked), as well as
in between.

On the other hand, we could have positively or negatively skewed distributions.

How do we know? If values are clustered near the middle in a frequency dis-
tribution with a few values at the high and low ends, then it is pretty much a
44  •  Chapter Five

normal distribution. If there are concentrations of values on either the high

or low ends, then the distribution is skewed. Here are the hints.

• Our frequency distributions will confirm this because they will have
outliers in them.
• The mean, median, and mode will not be equal.
• The frequency polygon will extend outward either left of right. This
extension is called a tail.

(Please refer to the figures in chapter 3.)

Positively Skewed Distributions

The tail of the curve tells you what kind of skew you have. If the tail is to the
right, you have a positively skewed curve. With a positively skewed curve, the
hump of the curve is toward the left. Remember when graphing frequency
polygons, the left side of the horizontal axis begins with the lowest values.
Because the hump is in this vicinity, your distribution has many low values.
The few high values, represented in the tail area, skew it. These “outlier”
high values pull the mean upward; the average is overrepresented. The mean
shifts toward the outlier. In a positively skewed distribution, the mode is at
the peak of the distribution. The mean is off to the right and the median in
between.
An example of a positively skewed distribution would be results of testing
intended to tap an elite few. The test would be extremely difficult, yielding
very few high values. Testing to become a Navy Seal or a Rhodes Scholar
might be examples of such tests. The few that did well would be the values
on the tail and those who did not do well would be the large hump. Those in
the tail of the curve would qualify or be selected because the program targets
those with extremely high results.

Negatively Skewed Distributions

If the tail is to the left, you have a negatively skewed curve. With a negatively
skewed curve, the hump or curve shows where most values are falling. In this
case they are at the higher end of the horizontal axis: the hump or curve is
on the right side of the frequency polygon. The few low values, represented
in the tail area, skew it. These “outlier” low values pull the mean downward;
the average is underrepresented. The mean shifts toward the outlier. In a
negatively skewed distribution, the mode remains at the peak, the mean is
off to the left and the median is in between.
 Measuring Variability   •  45

An example of a negatively skewed distribution would be values on an

easy test where most did well. A driver’s education exam should have a
negatively skewed distribution. Most people pass it. Negatively skewed distri-
butions may be used to identify areas of need: academic, physical, psychologi-
cal, nutritional. You would be looking at those values in the tail to identify
those with high levels of need. This is not the majority; it is the minority
who have some kind of need that you want to identify from those who do
not have that need.

Chapter Summary
Reporting only part of the story about your data set occurs when measures
for variance are missing. You are looking at part of the information puzzle
but leaving off an important dimension. If your thesis or dissertation research
hinges on descriptive statistics, then using both central tendency measures
and measures of variation will be considered vital.
 CHAPTER SIX

Random Sampling and

Other Useful Sampling Strategies

Time is at a premium for graduate students. Although it probably was never

conceived in this fashion, statistics is a tool that saves time. It allows you to
conduct your research without spending many extra hours. Why is this so?
Statistics are based on samples. We have often heard the term random sample
or scientific sample. Yet few of us really know what this designation means or
understand how valuable it is.

Populations and Samples

The term population makes us think of all of the people in our town, region,
state, or country, and their respective characteristics such as gender, age,
marital status, ethnicity, race, religion, and so forth. In statistics, the term
population takes on a slightly different meaning.
The population in statistics includes all members of a defined group on
which we are collecting information. The operative descriptor is “all”: all
children under the age of five, senior executives, first offenders, hospital pa-
tients, or the entire community of households in whatever geographic circle
we are focused. The population in our statistical study is defined by the who
(target group) and the where (the geographic boundary in which this group
exists).
A part of the population is called a sample. Samples are studied to obtain
valuable information about the larger group called the population. Once we
define our population, we can take a sample of the population to conduct our

47
48  •  Chapter Six

statistics. A sample is a subset or subgroup of our population. It is a propor-

tion of the population, a slice of it and all its characteristics. A sample is a
group that actually possesses the same characteristics as the population—if it is
drawn randomly. This may be hard for you to believe, but it is true.
Randomly drawn samples must have two characteristics:

1. Every person has an equal chance to be selected for the sample.

2. Selection of one person is independent of the selection of another
person.

We use samples all the time to represent the whole. We take a sample
of exotic food, maybe a bite or two to see if it pleases our palette. We use a
sample of new software that comes in the mail to see if it really works as well
as the manufacturer purports it does. We take a sample of client references
to see if the consultant is as good as the proposal suggests. We visit a sample
of classrooms to determine if behavioral issues are increasing. We sample
some of the households in our community to see what languages are spoken
at home.
Why not use the entire population to draw our conclusions? The short
answer is it is a waste of time. For most purposes we can obtain suitable ac-
curacy, quickly and inexpensively, from a sample. Assessing all individuals
may be impossible, impractical, expensive, or even inaccurate. It is usually
not feasible to include an entire population in a study except if the number
of those in our population is small and manageable. Furthermore, statistics
make it unnecessary. A sign of ineptitude in data management is using an
entire population when a sample will provide the same results. You save
money, time, make fewer mistakes, and achieve the same end.
What is great about random samples is that you can generalize to the
population that you are interested in. If you randomly sample 300 households
in your community, you can generalize to the 50,000 households that live
there. If you match some of the demographic characteristics of the 300 with
the 50,000, you will see that they are surprisingly similar. Technically speak-
ing, if you calculated a mean for these 300 randomly selected households
and then drew another sample of 300 different households from the same
population as the first, your mean values (M) would be the same. This is the
beauty of random samples. If drawn scientifically, they represent the entire
population in which you are interested.
Many a doubting Thomas will still be skeptical about whether a sample
can truly represent the population. This is understandable. Yet if properly
conducted, sampling does work. Think about all the polls reported in the
 Random Sampling and Other Useful Sampling Strategies   •  49

media. They do not survey all of the nation’s households. Instead, a random
sample is drawn and surveyed. The results are pretty accurate. If you would
like to get a feeling for the size of samples that represent entire populations,
please refer to Krejcie and Morgan (1970; see source for Table 6.1). Table 6.1
suggests the great benefit that randomly selected samples afford. The larger
the population size, the smaller the sample. However, for smaller samples,
you must use almost the entire population. (Population size is noted by up-
percase “N” and sample size by lower case “n.”)

Table 6.1. Random Sample Sizes

(n) Required for Population (N)
Representation
Population Size (N) Sample Size (n)
50 44
100 80
500 217
1,000 278
1,500 306
3,000 341
5,000 357
10,000 375
50,000 381
100,000 384
Source: R. V. Krejcie and D. W. Morgan 1970.
“Determining Sample Size for Research Ac-
tivities.” Educational and Psychological Mea-
surements, 30, 607–10.

Samples must be drawn according to scientific principles—with precision

and accuracy. There is a price to having a truly representative sample. But the
process is quite simple and methodological. If you follow the steps, you will
find that the characteristics of your population are the same as those of your
sample. This is a good check on the validity of your sample. If the percent-
ages on some of the demographics are known about your population, then
the same demographics will appear in the sample. This is a good test and a
very convincing piece of information for those skeptics.

Steps for Drawing Simple Random Samples

Here are the steps for drawing a random sample, if you choose to use a ran-
dom sample in your graduate research. It may not be feasible to do this, but
it is important to know the steps in the process if you opt for this sampling
procedure.
50  •  Chapter Six

1. The first step in drawing a random sample is to identify all the mem-
bers in your population. You must be able to list them in what is called
a sampling frame. The frame should have the names without order to
them and be nonoverlapping (no duplicates). Alphabetizing the list by
surname is a way to ensure a random order in the sampling frame. Your
computer can sort alphabetically based on the last name, if surname is
entered into a database as a separate field. One reason that you need
to ensure random order of names is that some lists cluster them by
neighborhood location, housing type, income, or some other grouping.
This precludes their “equal chance” of being selected.
2. Give each name an identification number. Start with “1” and
continue.
3. Decide the size for your sample. You can use the table (see table 6.1),
or whatever you believe is right, in order to believe in the results you
obtain. As a rule of thumb, use as large a sample as possible. Whenever
you are calculating means, percentages, or other statistics, the popula-
tion is being estimated. Statistics calculated from large samples are
more accurate than those from small samples. Large samples give the
principle of randomness a chance to work.
4. Get a Table of Random Numbers (table 6.3 displays a partial one). Many
are located at the end of statistical or mathematical textbooks. The
Table of Random Numbers contains numbers generated mechanically
so that there is no discernable order or system to them. Each digit gets
an equal representation. The Table of Random Numbers consists of
rows and columns of numbers arranged at random so that they can be
used at any point by reading in any direction left or right, up or down.

We are now ready to draw our random sample. Here is our sampling frame
of 20 members of our population who are behavioral health counselors. The
director of the behavioral health clinic wants to select five counselors for an
overnight stay in the clinic with teens who are inpatients. In order to be fair
in the selection process, the director uses random sampling and lists the 20
counselors’ names in alphabetical order and assigns an ID number to each
from 1 through 20. Table 6.2 has the sampling frame.
The director knows that the largest ID number has two digits (20). So,
they are going to need a two-digit column in the Table of Random Numbers.
They put their finger down at any place in the Table of Random numbers.
This is their starting point. They have to decide in advance whether they
would use two digits going up, down, left, or right.
 Random Sampling and Other Useful Sampling Strategies   •  51

Table 6.2. Sampling Frame

ID Name
1 Annie
2 Ellen
3 Emily
4 Estelle
5 James
6 Jessica
7 Jose
8 Juan
9 Kaleen
10 Lia
11 Loren
12 Maria
13 Mary
14 McKenzie
15 Michelle
16 Miguel
17 Olivia
18 Raphaelle
19 Rebecca
20 Tanya

Table 6.3 is a Table of Random Numbers—just for the point of illustra-

tion. In boldface is the director’s starting point of 37. They do not have an ID
number that is 37 and so they proceed down: 37, 81, 89, and come to 06. ID
#6 is the first ID number that is within the band of between 1 and 20. This
is the first member of the random sample, and it is Jessica. They continue 82,
56, 96, 66, 46 until they come up with the next ID #13, Mary. Their three
last members of the sample are ID #8 (Juan), ID #5 (James), and ID #4 (Es-
telle). They now have five randomly selected members: Jessica, Mary, Juan,
James, and Estelle who will be staying overnight with the teens.

The Old Way Works Too

For such a small sampling frame, we could have done it the old way by draw-
ing five names out of a bowl that had all 20 names in it. And you can draw
a random sample that way instead of using the Table of Random Numbers.
You simply put all names in a container and thoroughly mix them up. If you
have 1,000 names in the bowl and want a sample of 100, draw the first name
out. That is the first person in the sample. Mix up the names. If you don’t
do this, then the name at the bottom will have less of a chance of being
selected. After mixing, draw the next name and continue until you get 100.
52  •  Chapter Six

Table 6.3. Table of Random Numbers (Partial)

53 74 23 99 67 61 32 28 69 84 94 62 67 86 24
63 38 06 86 54 99 00 65 26 94 02 82 90 23 07
35 30 58 21 46 06 72 17 10 94 25 21 31 75 96
63 43 36 82 69 65 51 18 37 88 61 38 44 12 45
98 25 37 55 26 01 91 82 81 46 74 71 12 94 97
02 63 21 17 69 71 50 80 89 56 38 15 70 11 48
64 55 22 21 82 48 22 28 06 00 61 54 13 43 91
85 07 26 13 89 01 10 07 82 04 59 63 69 36 03
58 54 16 24 15 51 54 44 82 00 62 61 65 04 69
34 85 27 84 87 61 48 64 56 26 90 18 48 13 26
03 92 18 27 46 57 99 16 96 56 30 33 72 85 22
62 95 30 27 59 37 75 41 66 48 86 97 80 61 45
08 45 93 15 22 60 21 75 46 91 98 77 27 85 42
07 08 55 18 40 45 44 75 13 90 24 94 96 61 02
01 85 89 95 66 51 10 19 34 88 15 84 97 19 75
72 84 71 14 35 19 11 58 49 26 50 11 17 17 76
88 78 28 16 84 13 52 53 94 53 75 45 69 30 96
45 17 75 65 57 28 40 19 72 12 25 12 74 75 67
96 76 28 12 54 22 01 11 94 25 71 96 16 16 88
43 31 67 72 30 24 02 94 08 63 38 32 36 66 02
50 44 66 44 21 66 06 58 05 62 68 15 54 35 02
22 66 22 15 86 26 63 75 41 99 58 42 36 72 24
96 24 40 14 51 23 22 30 88 57 95 67 47 29 83
31 73 91 61 19 60 20 72 93 48 98 57 07 23 69
78 60 73 99 84 43 89 94 36 45 56 69 47 07 41

This works the exact same way as using the Table of Random Numbers and
is useful if your sample is small in size.
There are also free random number generators on the internet that you
can use: two are www.random.org and www.randomizer.org.

Other Useful Sampling Strategies

Simple random sampling is the premier way to draw your sample. There
are other sampling strategies that are very useful as well. Many books on
sampling provide more in-depth information about the processes and the
benefits. Here is a summary of a few procedures that you can consider.
Systematic sampling is a cost-effective and often-used sampling strategy.
Again, you must have a population sampling frame listing names in random
order that is nonoverlapping. Determine both the size of the population and
the size of the sample you want to work with. Then, divide the sample size
(n) into the population (N) size to get your key number, symbolized as “k.”
 Random Sampling and Other Useful Sampling Strategies   •  53

• For example, if you had a population of 1,500 (N) and needed to have
a sample of 306 (n), the key number (k) would be 5.
• Then, you would randomly pick any number between 1 and 5. Let’s say
you picked “4.” That is your first ID number. ID #4 is the first member
of your sample from the list of 1,500.
• Then, systematically add 5 (your key number) to the first ID number of
4 and you get #9. ID #9 is the second member of your sample. Continue
adding your key number.
• The next ID number is ID #14, then ID #19, ID #24, ID #29 until you
get your sample size up to the 306 you intended.

You might use systematic sampling to select a sample of households in your

community for a housing survey. Town government might have the listing
already available, in random order and with ID numbers in sequence. This
saves you time and labor. If the town will give you a set of mailing labels or
email addresses, then it will be all the better.
Cluster sampling is exactly what its title implies. You randomly select clus-
ters, or groups, in a population instead of individuals. This would work if a
state wanted to sample all third graders on their writing skills. They would
randomly select third grade classrooms from all third-grade classrooms in the
state. Each of those classrooms selected would have 100% of the students in
that classroom in the sample. The sampling unit or cluster is the third-grade
classroom not the individual student. Sometimes, this is a more practical
effort than selecting individuals. Each member of the cluster has an equal
chance of being selected. The random selection of the clusters provides esti-
mates of the population. This is a good cost reduction technique.
Stratified sampling is used when the population is heterogeneous, and it
is important to represent the different strata or subpopulations. There is a
proportional representation of strata in the sample—proportional to the
population strata. We divide the entire population into strata (groups) to
obtain groups of people that are more or less equal in some respect. Then,
select a random sample from each stratum. This ensures that no group is
missed and improves the precision of our estimates. This might be used with
different racial/ethnic groups if we wanted to ensure that our sample included
a proportional representation of African Americans, Asians, and Latinos in
addition to the Caucasians that might predominate our demographic pool.
There are three nonrandom sampling strategies that graduate students
should consider using, especially if you have descriptive or nonexperimental
designs. These sampling strategies are convenience, purposive, and snowball
54  •  Chapter Six

sampling. Many times, random sampling strategies are impossible to under-

take at the graduate level. These three strategies are options to consider.
Convenience samples, exactly what the name suggests, are often what we
have to use because of reality. It is a nonrandom sampling technique. This
is a useful sampling strategy because you cannot draw a sample, yet you
have a group that is accessible, is representative of the target population
of interest, and available. Use what you have with the honest acknowledg-
ment that there are limitations. If you proceed to collect data with respect
to some systematic, thoughtful process, it is better than predicting with some
impractical power analysis what your sample size will be—and then failing
to achieve it.
Purposive sampling, another nonrandom sampling technique, is used when
the researcher seeks out subjects with specific characteristics to participate
in the research study. This might be done if your study involved parents of
daycare children. You might seek permission from daycare centers to ask
their parents to participate in the research.
Snowball sampling, another nonrandom sampling technique, asks each
research participant who volunteers for your study to identify one or more
others who might be willing to participate. This might be used if you were
studying African American female executives. The sample’s network might
produce more volunteers to add to your sample. There might be just a few
whom you identify initially but over the course of the research, those few
might help to identify like members of hard-to-find populations.

Chapter Summary
Except when the population of interest is small, it is imprudent to survey
an entire population. Sampling strategies are available for your use. By us-
ing simple random, systematic, cluster, stratified, convenience, purposive or
snowball sampling procedures we can get a handle on what we need to know,
save time, and get the thesis or dissertation completed expeditiously. One
of the most important uses of samples is that they afford us the opportunity
to make inferences without spending a fortune in time and money to do so.
 CHAPTER SEVEN

Stating Hypotheses
and Hypothesis Testing

Descriptive statistics, which describe our data, are very useful. Other statistics
are also very useful but take a little more comprehension in implementation.
They are called inferential statistics. This is because we can draw conclusions
or inferences from them.
In the course of your graduate school research, you pose guesses or
hunches that you believe to be true. In order to tell whether these hypotheses
are true or false, you have to subject them to a test. You may draw a randomly
selected sample from the population of interest, compute statistics on the
sample, and test whether the hunch is true or false. Central to the discussion
of inferential statistics is the concept of probability. When your statistical
analysis reveals that the probability is rare that your statistical result is due
to chance alone, we call this a statistically significant result. It means that our
observed outcome is so unique that it is unlikely to have occurred by luck.
The role of the sample, discussed in the previous chapter, is essential
to the process and makes the task of inferring a possibility. The inferential
statistics most commonly used by graduate students to draw these conclu-
sions include t-test, one-way analysis of variance, chi-square analysis, and
correlational techniques including regression, among many others. These
will be discussed in this book since they are considered to be user-friendly,
pragmatic, and applicable in research conducted by most graduate students.

55
56  •  Chapter Seven

Hypothesis Testing
Hypotheses are suppositions presumed to be true. Because of the importance of
posing these hypotheses, it is crucial to accept those that are true, and reject
those that are false. But how do we do this?
All research, whether it is in the graduate program or in the laboratory,
begins with the research question or hypothesis. It is always stated in the null
(negative) and is called the null hypothesis. This is because it is incumbent
on you, the researcher, to prove something is indeed true. It is similar to
the philosophy in the courtroom: innocent until proven guilty. The onus of
proof is on you to prove that the null is false. Technically speaking, the word
hypothesis is derived from a Greek word that means “an assumption subject
to verification.” The symbol for the null hypothesis is H0. Basically, there are
two types of null hypotheses: one that tests for differences and one that tests
for relationships between or among variables.

Null Hypotheses That Test for Differences

The first type tests for differences. Two null hypotheses that test for differ-
ences are presented below as an illustration of how to state them correctly.
Corresponding examples follow each.

• There is no difference between two groups on some variable, as repre-

sented by their mean scores. Example: there is no difference between
newly hired and long-term employees with respect to their levels of job
satisfaction.
• There is no difference among three or more groups on some variable,
as represented by their mean scores. Example: there is no difference
among Asian, African American, Latino, and Caucasian lawyers with
respect to their annual salaries.

Directional Null Hypotheses

Sometimes a null hypothesis that tests for differences takes a courageous step
and predicts the direction of the difference. This is called a directional null
hypothesis. Instead of stating a nondirectional null like this, “There is no dif-
ference between newly hired and long-term employees with respect to their
level of job satisfaction,” you would state a directional null like this, “Newly
hired employees do not have higher levels of job satisfaction than long-term
employees do.” The basis for this directional guess should be your knowledge
base, evidence in the professional literature or your own experience, and not
 Stating Hypotheses and Hypothesis Testing   •  57

be a superficial guess. You are actually stating which mean will be greater
when you calculate your statistics.
In this case, you really believe that newly hired employees have higher
levels of job satisfaction than long-term employees do. Remember that the
hypothesis must still be stated in the opposite of what you think is true. It
has to be stated in the null (which you really think is false). Then you have
to prove it to be false. When we accept the null (or fail to reject the null),
we are saying that our results are not statistically significant. When we reject
the null (or the null is false) we are saying that our results are statistically
significant and likely due to factors or conditions other than chance.

Null Hypotheses That Test for Relationships

The second type of null hypothesis tests for relationships between or among
variables.

• There is no relationship between one variable and another variable.

Example: there is no relationship between hours of sleep and work
productivity.
• There is no relationship between one variable and several other vari-
ables. Example: there is no relationship between work productivity and
hours of sleep, age and coffee consumption.

For practical utility this book will focus on these two types of null hypoth-
eses: those that test for differences and those that test for relationships.

Alternative Hypotheses
If hypothesis testing shows that the null hypothesis is false, and is rejected,
the alternative hypothesis is stated. The symbol for the alternative hypothesis
is Ha. It asserts that the opposite of what the null hypothesis stated is true.
Almost always, graduate students believe that the alternative hypothesis is
true and that the null is false. But statistics are used to prove it. For the pre-
vious null hypotheses, these are the corresponding alternative hypotheses:

• There is a difference between newly hired and long-term employees

with respect to their levels of job satisfaction.
• There is a difference among Asian, African American, Latino, and
Caucasian lawyers with respect to their annual salaries.
• There is a relationship between hours of sleep and work productivity.
• There is a relationship between work productivity and hours of sleep,
age, and coffee consumption.
58  •  Chapter Seven

Probability Levels
How do graduate students determine whether a null hypothesis should be
accepted or rejected? Data are collected and statistics are executed. But the
key is the probability value associated with the statistical procedure that allows
the decision to be made. The probability value is also called the p value. It is
obtained once you execute your statistical analysis. You compare the signifi-
cance level you chose with the p value you obtained to determine whether
there are statistically significant results.
Probability applies exclusively to the likely occurrence of a future event.
In statistics, probability provides a quantitative measurement of the likeli-
hood that event will occur. In other words, probability is used to measure
the certainty or uncertainty of the outcome of an event. If it is sure to occur,
the probability is 100%. If it will never occur, the probability is 0%. A prob-
ability of .50 means the event should occur 50% of the time. Any event that
may or may not occur has a probability of between 0 and 1.00.
Professional journals often state with the results from a statistical pro-
cedure that “the probability level is less than .05” (p < .05). This means
that observed difference is likely to be real rather than having occurred by
chance. Or the statistical results may have occurred by chance only five out
of 100 times. Although somewhat arbitrary, the probability level is the mag-
nitude of error that one is willing to accept in making the decision to reject
the null hypothesis.
The conventional levels for rejecting the null hypothesis are either .05 or
.01. One (.01) is more conservative than the other (.05) because with .01
there is less willingness to have results due to chance alone; only one time in
100 will results be due to chance. A .05 significance level is more generous in
accepting a statement as true. A very conservative significance level is .001
or one time in 1,000. This may be used in scientific studies such as drug ef-
ficacy. For graduate students, the commonly accepted significance levels for
rejecting the null hypothesis are .05 and .01.
As a note, the significance levels must be chosen a priori. This is a Latin
term meaning “in advance.” Decide the significance level early on, before
collecting data, and accept it when the statistical testing is completed. It is
unethical to choose the conservative .01 significance level and then find out
that statistical significance would have occurred if the more liberal choice
of .05 was made.

Type I and Type II Errors

It is true that the selection of significance levels (.05, .01) comes with the
chance of making errors. These are called Type I and Type II errors.
 Stating Hypotheses and Hypothesis Testing   •  59

Type I Errors
A Type I error is when the null is rejected when it was actually true, con-
cluding falsely that there were differences or relationships when there were
none. This can be problematic when someone else replicates the study and
the results do not hold up.
Making a Type I error is based on the level of statistical significance. If a
.05 significance level was selected, then five times out of 100 the results will
be due to chance. Unfortunately, one of the five times occurred. A more con-
servative significance level is .01 where only one time out of 100 the results
will be due to chance alone. A liberal significance level is .10 where 10 times
out of 100 chance results are OK. This level is not practiced in most graduate
student research and can largely contribute to Type I errors.

The Bonferroni Technique

When performing multiple significance tests in graduate research studies,
there is the risk of making a Type I error. A way to reduce this risk is by
using the Bonferroni correction to adjust the probability level. To do this,
simply divide the significance level by the number of significance tests you
are performing. If using a significance level of p ≤ .05 and there are two sig-
nificance tests in the study, the significance level becomes .025 (.05 ÷ 2 =
.025) needed to reject the null hypotheses. It is more conservative, but it will
reduce the likelihood of making a Type 1 error. Your calculated p value must
be less than .025 to reject the null.
For example, if you measured the levels of stress (pre- and post-) and anxi-
ety (pre- and post-) for college freshmen, you would have two hypotheses
with two corresponding significance tests. The probability level would be
adjusted by the Bonferroni technique from .05 to .025. This would make it
harder both to reject the null, and concurrently avoid a Type I error.

Type II Errors
A Type II error occurs when accepting the null when it was in fact false,
concluding that there were no differences or relationships when in fact there
were. If the null should have been rejected, but it was not, the maintenance
of the status quo may be a mistake. For example, you use conservative sig-
nificance level such as .01 to determine if a pre-K program attendance had
an effect on academic performance for first graders. The statistical result sug-
gests no impact when in fact there was one. Not rejecting the null (accepting
the status quo) might limit funding for a program with important benefits
because of the choice of a conservative significance level.
60  •  Chapter Seven

Power to Reject the Null Hypothesis

One-Tailed and Two-Tailed Tests of Statistical Significance
With nondirectional null hypotheses, a graduate student is only suggesting
that there will be a difference in mean score results. Which mean (M) will
be higher is not speculated. In the case of a nondirectional null hypothesis, a
two-tailed test of significance is used. If there is a prediction of which mean
will be higher and which will be lower, then a one-tailed test of significance
is used. This might be the case if you are predicting that managers who have
an MBA will have higher employee retention than managers who do not
have an MBA. When stating a directional hypothesis, a one-tailed test of
significance is used.
This may seem like a lot of mumbo jumbo, but all it means is simply this:
a directional hypothesis is a serious commitment for someone to make; they
are predicting which mean score (M) of the groups they are comparing will
be higher or lower. When choosing the direction of the difference, the re-
ward is a one-tailed test of statistical significance. What this means is that
the ability to reject the null is boosted or heightened. The ability to reject
the null is called power—power to reject the null.

Factors Contributing to Power

There are certain conditions where it is easier to reject the null hypothesis:

• If using parametric statistics: Parametric statistics are sort of the elite

statistics. Some discussed in this book include the t-test, ANOVAs,
and correlational procedures. They are classified as parametric statistics
because they make assumptions about population parameters. They
work best if the distribution is not skewed but rather a normal curve,
bell shaped, and symmetrical. The groups that are compared should
have equal variances or spread. This is called homogeneity of variance.
Parametric statistics are more powerful in rejecting the null hypoth-
esis. Also, they are described as robust, a statistical term which means
that they can hold up even when the assumptions are violated. Para-
metric statistics rely on the computation of means and standard devia-
tions which use interval and ratio scaling.
Nonparametric statistics are classified as such because they require
meeting very few assumptions. They do not require normal distributions
or equal variances. They are oftentimes based on ordinal or nominal
measurement and are easier to compute. They use frequency counts
instead of complicated calculations. Because nonparametric statistics
 Stating Hypotheses and Hypothesis Testing   •  61

do not always rely on means and standard deviations, they lack the
precision that the parametric statistics possess.
Generally speaking, if the data come from a population that is nor-
mally distributed, use the parametric test. If not, use the nonparametric
test. In a later chapter, chi-square analysis, an essential nonparametric
statistic for use by graduate students, will be discussed.
• If using a directional hypothesis: One-tailed tests (used with directional
null hypotheses) are more powerful than two-tailed tests (used with
nondirectional hypotheses).
• If using large samples: These are more powerful than small sample sizes.
• If using a more liberal significance level: If choosing .05 instead of .01, the
chances to reject the null are greater.

Steps in Hypothesis Testing

The following four steps constitute the basis for hypothesis testing. This
framework may be more conceptual since computer programs have taken
some of the work out of the steps. Yet it is risky for graduate students to use
high-powered statistical packages with very little understanding of the pro-
cesses involved. Familiarity and awareness are essential for conducting any
inferential statistical procedure. This book is based on that premise and will
present the inferential statistics in that spirit in the next few chapters that
discuss statistical techniques.

1. State the hypothesis in the null form. The null can be stated for either
differences or relationships. If for differences, the null is either non-
directional or directional, but awareness of the type is a prerequisite.
2. Select your significance level, either .05 (p ≤ .05) or .01 (p ≤ .01). A sig-
nificance level of .05 establishes a 95% confidence level and is more
liberal but may result in a Type 1 error. A significance level of .01
establishes a 99% confidence level and is more conservative but may
result in a Type II error.
3. Conduct the statistical analysis and obtain the p value from the statistical
procedure. Compare the p value to the significance level you chose to
determine whether there is a statistically significant result. The p value
obtained must be less than or equal to the significance level you set in
Step 2. If you chose p ≤ .05, the p value obtained must be .05 or less.
4. Accept or reject the null.
No statistically significant result: accept the null as true.
Yes, a statistically significant result: reject the null as false.
62  •  Chapter Seven

5. Determine the practical value of the results. A result that is statistically

significant may be too small to have any value in a real-world setting.
If your mean difference is miniscule but statistically significant, what
is the practical value? Don’t let ego overtake common sense. You have
to make a judgment call about your results, and an honest one at that.
An objective way to test the practical significance is by using an effect
size index like Cohen’s d where a small effect size is <.20, a medium
effect size is .20 to .79, and a large effect size is >.79. Determining the
practical value of your results is an important last step.

A final thought is in order regarding statistical significance. Many are led

to believe that getting a statistically significant result is the gold ring. This
is not always the case. It depends on what research question is being asked.
For example, if the question is whether males and females have different
attitudes toward science, it would be positive if there were no significant
differences. In this case, a statistically significant result may be meaningful,
but not desirable.

Chapter Summary
The use of inferential statistics allows graduate students to make sound con-
clusions based on data. There is no guesswork. However, a method to the
madness must be observed. There is a scientific nature to hypothesis testing:
its principles must be respected. But it is a small price to pay for the increase
in knowledge that we get in exchange.
 CHAPTER EIGHT

t-Test Procedures

For graduate students, many decisions are focused on comparisons such as the
following examples:

• To determine which method of instruction has a greater impact on

student performance;
• To compare online and print advertisements and their impact on cus-
tomer purchasing;
• To assess hospital satisfaction between inpatients and outpatients;
• To assess leadership skills before and after manager training; or,
• To evaluate which treatment is more effective in reducing teen sub-
stance abuse.

When two groups are compared, the statistic that is most useful is the t-test.
It is both an inferential and a parametric statistic.
The purpose of a t-test is to determine if there is a statistically significant
difference between the mean scores of two groups. This is where the mean,
the queen of central tendency, is vital. The mean scores of two groups are
compared via the formula for a t-test. Because the t-test is a parametric sta-
tistic, it is powerful. If there are differences, even slight ones, the t-test will
uncover them. There are a few basic facts about t-tests:

1. A t-test is used if there are only two groups or two testing times (pre/
post or before/after) to compare.

63
64  •  Chapter Eight

2. This statistical technique answers the null hypothesis: there is no dif-

ference between two groups or testing times on respective mean scores.
3. There is one independent variable with two (2) categories and there is
one dependent variable.

Independent and Dependent Variables

For many graduate students, there is much confusion around what consti-
tutes independent and dependent variables. To understand the difference
is fundamental to executing statistics properly. This is where the use of a
statistical software package can be dangerous if you are not aware of even
the basic content presented in this book. The simplicity of “click, click” su-
perficially erases the need to know. But understanding which variable is the
dependent and which one is the independent variable is vital to producing
correct statistical analysis.
The independent variable is independent of the outcome but presumed to
cause, affect, or influence the outcome. The dependent variable is dependent
on the independent variable; the outcome depends on how the independent
variable is used. An independent variable might be patient status (inpatient
and outpatient) and the dependent variable might be attitudes toward nurs-
ing care. You might administer a survey to determine whether there are
different attitudes.
Basically, the independent variable for a t-test is nominally scaled. There are
two categories with one variable. It is that simple. Many think that because
there are two categories, there are two independent variables. There is only
one independent variable with a t-test, and it has two categories. Table 8.1
presents a few examples of independent variables and their two categories.

Table 8.1. Independent Variables and Two Possible Categories

Independent Variable 1st Category 2nd Category
College students 4-year 2-year
Employees C-suite managers Middle managers
Musicians String Woodwind
Customers First time Repeat
Patients Maternity Emergency care
Latinos Puerto Ricans Latin Americans
Educators Teachers Paraprofessionals
Community Full-time residents Summer-only residents
 t-Test Procedures   •  65

To nominally scale these independent variables for statistical analysis in a

t-test, you can assign a “1” to one group and a “2” to the other. Remember,
with nominal variables the numbers are only labels. They mean nothing in
terms of measurement but everything in differentiating the categories of your
independent variable.
With a t-test there is also only one dependent variable. It is continuous
in its numeric range and uses ordinal, interval, or ratio measurement scales.
Common examples of dependent variables are test scores, health measures,
success rates, satisfaction levels, organizational climate, attitudes, and many
others. See table 8.2 for some dependent variables that correspond to the
independent variables mentioned previously.

Table 8.2. Dependent Variables and Independent Variables

Independent Variable Dependent Variable
College students Grade point averages
Employees Job satisfaction
Musicians Age
Customers Dollars spent
Patients Nurse empathy
Latinos Political orientation
Educators Educational level
Community Attitude toward local schools

Preliminary Assumptions
In the last chapter parametric statistics were described. This group of statistics
is more powerful than nonparametric statistics in rejecting the null hypoth-
esis. But they depend on your meeting a few assumptions before you use them.
Since the t-test is a parametric statistic, there are a few preliminary steps that
you must try to abide by before you implement this statistical procedure:

• The two groups should have equal variances on the dependent variable. The
variability of the individual groups’ mean scores should be equivalent.
They should have the same degree of variability. This is called a test of
homogeneity of variance and done as a preliminary step, a kind of insur-
ance program for your data. This is particularly true when your groups
are of different sizes (n)—common in real-world settings versus labora-
tory settings. The test of homogeneity of variances will be done for you
with statistical program packages like SPSS, as a preliminary step for
conducting the t-test.
66  •  Chapter Eight

• The two groups should have an equal number of subjects. If the two groups
are unequal—like 20% more subjects in one group than in the other—
look at the standard deviation. If they are similar, go ahead and use
the t-test. If it is not, use the Mann-Whitney U test, a nonparametric
counterpart, discussed at the end of this chapter.
• Groups should be equivalent on all other variables except the dependent vari-
ables. If you are examining training approaches in two leadership devel-
opment programs by comparing scores on a leadership test, make sure
that the two groups are similar on other characteristics. These might
include gender breakdown, years working at the company, position in
the company, educational levels, age, and other variables. These factors
could account for the differences and not the leadership development
programs.

In your effort to meet the assumptions of the t-test, keep in mind that t-tests
are robust. They can hold up even when the assumptions are violated. Al-
though this robustness lets us off the hook to some extent, it is a good idea
to approximate to the greatest extent possible the rules of the game, the as-
sumptions. This will ensure excellent data management, avoid errors, and ulti-
mately produce sound data. As a note, many graduate research data sets do not
meet these assumptions. The operative word is to try to meet them. Your thesis
or dissertation is a learning opportunity as a researcher. If your data set violates
these assumptions, do not quit. Acknowledge the shortcomings and proceed.

Steps for Conducting a t-Test

This example will help to explain the basic purpose of the t-test. Even
though there are software programs to click on for a t-test, you should be
aware of the process so that you can ensure quality control over the results.
You are a department head of a school with 30 math teachers. Your Kin-
dergarten through eighth-grade (K–8) teachers and those teaching ninth
through twelfth (9–12) grades were compared on an instrument used to
assess teachers’ content knowledge in math. You wondered if there was a
difference between the two groups. You decide to compare the math scores
between the two groups of teachers.

1. State your null hypothesis. It is nondirectional.

There is no difference in mathematics knowledge between K–8 and
9–12 teachers.
 t-Test Procedures   •  67

2. Identify your independent variable and your dependent variable.

“Teacher group” was your independent variable with two categories:
K–8 teachers and 9–12 teachers. You took the listing of all of your
teachers and separated them into two groups. Group One was the K–8
group and Group Two was the 9–12 group. Using nominal scaling, you
assign each group a numeric label so that your t-test could be calcu-
lated. You used “1” and “2” as your numeric values, which represented
the two categories of teachers.
Your dependent variable was mathematics knowledge, as measured
by the score on the math test. You listed the score for each teacher
on the mathematics test. You also calculated the mean scores for each
of the two groups. The mean scores appeared different, but you could
not really tell just from eyeballing the data. A t-test formula had to
be applied to the data to determine whether there was a statistically
significant difference.
3. Set up your data for data entry into a database.
The data for your t-test was recorded in a statistics database such as
that in table 8.3. As a note, you used only the numeric values in your
database, that is, the teachers’ ID number (1 through 30), the Group
codes (1 or 2), and the mathematics test score for each of the 30 teach-
ers. You have data on fifteen K–8 and fifteen 9–12 teachers.
4. Execute the t-test statistical procedure to obtain a calculated t-value and a
probability level.
After you have entered your data from your database into the com-
puter software program, you execute the t-test procedure. The statistic
obtained is called your calculated t-value. The t-value is a continuous
number usually with two decimal places. For the sample data, your cal-
culated t-value actually is –4.97. It is critical to your decision whether
to reject of accept the null hypothesis.
5. Examine the p-value on the computer output that is associated with your
calculated t value.
Is it higher or lower than the significance level you established a priori?
6. Accept or reject the null hypothesis.
Your calculated t-value and its probability level was statistically sig-
nificant at the significance level you set, p ≤ .05. You reject the null
hypothesis. There is a difference in mathematics knowledge between K–8
and 9–12 teachers. You found that your 9–12 teachers had statistically
higher mathematics scores.
68  •  Chapter Eight

Table 8.3. Data Set for Two Groups of Teachers and Their
Mathematics Scores
ID Teachers Category Group Math
1 Mr. Ramirez 9–12 1 88
2 Ms. DeJohn 9–12 1 89
3 Ms. Santos 9–12 1 88
4 Ms. Considine 9–12 1 89
5 Ms. Ponte 9–12 1 76
6 Ms. Marchand 9–12 1 89
7 Mr. Ross 9–12 1 88
8 Ms. Beyer 9–12 1 89
9 Ms. Gallagher 9–12 1 88
10 Ms. Magistrali 9–12 1 89
11 Ms. Strand 9–12 1 89
12 Mr. Fischer 9–12 1 88
13 Ms. Bowen 9–12 1 89
14 Ms. Peak 9–12 1 88
15 Mrs. Pacheco 9–12 1 89
Mean for Group 1 87.73
16 Ms. French K–8 2 89
17 Mr. Dickinson K–8 2 78
18 Ms. Addazio K–8 2 75
19 Ms. Saidel K–8 2 78
20 Mr. Dwyer K–8 2 79
21 Ms. Latavia K–8 2 89
22 Ms. Matarese K–8 2 78
23 Ms. Kozlak K–8 2 75
24 Mr. Fitz K–8 2 78
25 Ms. Hernandez K–8 2 79
26 Mr. Guzman K–8 2 78
27 Ms. Santo K–8 2 75
28 Ms. Strawson K–8 2 78
29 Ms. Vega K–8 2 79
30 Ms. Hebert K–8 2 70
Mean for Group 2 78.53

Independent and Correlated t-Tests

Although this book is intended to keep things as simple as possible, there is a
critical addendum to the t-test discussion. There are two forms of the t-test:
one is called the Independent samples t-test and the other is called the Corre-
lated samples t-test, also referred to as a Paired, Matched, or Dependent Samples
t-test. As graduate students, you need to know whether you are conducting
an Independent or Correlated samples t-test.
 t-Test Procedures   •  69

The Independent samples t-test is quite simple to understand and identify.

The two groups, such as the K–8 and 9–12 teacher groups, have no relation-
ship to each other. They are independent of one another, hence the name
“Independent samples” t-test.
Conversely, Correlated samples t-tests use two groups that have a connec-
tion or relationship to each other. With this built-in relationship, it is more
likely that the mean scores of the two groups have a relationship too. So, the
Correlated samples formula takes the inherent relationship into account and
helps to find a statistical significance—if there indeed is one. Here are the
occasions when you might be using Correlated t-tests:

1. When you are measuring the same individual twice

Two groups of data are really the same individual with two scores on
them. Table 8.4 displays an example where the patients in a weight
loss program are measured twice on weight pre- and postprogram. For
each patient, the pair, or dual weights, compose the two groups of data
needed for a t-test.

Table 8.4. Patients and Their Pre-

and Postweight
Preweight Postweight
Olivia 120 110
Joey 120 120
Luke 130 125
Mia 150 145

Correlated t-tests are used most often for this purpose—if you want
to determine impact, efficacy, and persistence over time. Some ex-
amples of when a correlated t-test might be used include the following:
• When assessing the efficacy of a health care intervention, treat-
ment, or drug over two time periods
• When tracking community involvement from last year to this
year
• When monitoring the effect of professional development training
on knowledge acquisition, pre and post
2. If you have matched your sample on some other variable so that the two
groups are alike.
One of the primary tenets of the t-test is that the two groups you are
studying are alike except for your dependent variable. You want to
ensure that something else called extraneous variance or error does not
70  •  Chapter Eight

account for differences between the two groups instead of your depen-
dent variable.
For example, look at the data set of teachers, discussed previously.
What might account for the difference in mathematics knowledge
besides whether a teacher is K–8 or 9–12? The answers might be the
years in teaching, the amount of professional development, confidence
with the subject area, attitude toward the profession, or other factors.
If there was one particular factor that you felt might compromise your
data, you might try to match the groups on that variable so that they
are equivalent except for the dependent variable you were exploring.

Matched Samples
Here is an example of a matched sample. Using the same set of data that we
used in the t-test, the two teacher groups were matched on the number of
professional development courses they had taken. As you can see in table
8.5, each teacher in Group One is matched with another teacher in Group
Two on the number of professional development courses that they took. For
each ID number in Group One there is a member of Group Two that has
the exact same number of professional development courses. Ramirez with
French, DeJohn with Dickinson, and all the way through, there are pairs of
teachers with the same number of professional development courses in each
of the two groups.

Table 8.5. Teachers Matched on Professional Development (PD) for the Correlated
t-Test
Pair 9–12 Teachers PD Group Math K–8 Teachers PD Group Math
1 Mr. Ramirez 1 1 88 Ms. French 1 2 89
2 Ms. DeJohn 1 1 89 Mr. Dickinson 1 2 78
3 Ms. Santos 2 1 88 Ms. Addazio 2 2 75
4 Ms. Considine 3 1 89 Ms. Saidel 3 2 78
5 Mr. Ponte 4 1 70 Mr. Dwyer 4 2 79
6 Ms. Marchand 1 1 89 Ms. Latavia 1 2 89
7 Mr. Ross 4 1 88 Ms. Matarese 4 2 78
8 Ms. Beyer 5 1 89 Ms. Kozlak 5 2 75
9 Ms. Gallagher 3 1 88 Mr. Fitz 3 2 78
10 Ms. Magistrali 4 1 89 Ms. Hernandez 4 2 79
11 Ms. Strand 5 1 89 Ms. Guzman 5 2 78
12 Mr. Fischer 1 1 88 Ms. Santo 1 2 75
13 Ms. Bowen 2 1 89 Ms. Strawson 2 2 78
14 Ms. Peak 2 1 88 Ms. Vega 2 2 79
15 Ms. Pacheco 3 1 89 Ms. Hebert 3 2 70
 t-Test Procedures   •  71

The number of professional development courses would not be entered

into the t-test that you were calculating, but it would ensure that your
groups of teachers were equivalent on this important variable (number of
professional development courses) before you begin to compare mathematics
knowledge. Your null hypothesis is still focused on grade levels and on math-
ematics knowledge, not the number of professional development courses.

Degrees of Freedom
Independent sample and Correlated sample t-tests have different degrees
of freedom associated with each t-test form. Degrees of freedom are noted in
professional journals as df. You need to know how the degrees of freedom are
attained so that you can report your statistical results.

• If you use an Independent samples t-test, the degrees of freedom (df) are
equal to the number in your total sample minus 2 or [N – 2]. In the first
example of K–8 and 9–12 teachers, the degrees of freedom would be 28
[or 30 teachers – 2 = 28].
• If you are using a Correlated samples t-test, the degrees of freedom (df)
are equal to the number of pairs minus 1 [N of pairs – 1]. In the second
example of teachers, you used matched pairs on professional develop-
ment courses; the degrees of freedom would be 14 [or 15 pairs – 1].

Correlated samples t-tests make it more difficult for you to reject the null.
Since the data for the groups are related, you have to ensure that the differ-
ence you find is real and not due to the relationship between the sets of data.
Therefore, the bar is raised with Correlated t-tests in the computation for
statistical significance with respective degrees of freedom and critical cutoff
points for rejecting the null. (Statistical textbooks will have a deep discus-
sion of degrees of freedom along with statistical computations of t-tests for
graduate students who are interested.)

Reporting t-Test Results in a Table Format

This is how you might report your t-test findings in a table format. For an In-
dependent samples t-test, there are two groups of teachers with 15 (N) in each
group. The two mean scores on the math test and the respective standard de-
viations are reported. The t-value that is calculated with the SPSS statistical
package is 4.97, there are 28 degrees of freedom, and the calculated p-value is
.00. This is below the cut off level of significance you chose, which was .05.
You have statistically significant results; the null is rejected (see table 8.6).
72  •  Chapter Eight

Table 8.6. Reporting Independent Samples t-Test Results

Teacher Groups N M (SD) t value df p
9–12 teachers 15 87.33 (4.82) 4.97 28 .00*
K–8 teachers 15 78.53 (4.87)
*p ≤ .05

This is how you might report your t-test findings in a table format. For
a Correlated samples t-test, there are 25 workshop participants tested twice,
pre and post. The two mean scores on a leadership test and the respective
standard deviations are reported. The t-value that is calculated with the
SPSS statistical package is 4.69, there are 24 degrees of freedom, and the
calculated p-value is .00. This is below the cut off level of significance you
chose, which was .05. You have statistically significant results; the null is
rejected (see table 8.7).

Table 8.7. Reporting Correlated Samples t-Test Results

N = 25 M (SD) t value df p
Preworkshop 84.72 (6.07) 4.69 24 .00*
Postworkshop 90.56 (3.72)
*p ≤ .05

Nonparametric Counterparts
There are two nonparametric statistical procedures that answer the same
question that the t-test does. If the assumptions mentioned in the begin-
ning of this chapter are violated, then you may want to consider using a
nonparametric alternative. The Mann-Whitney U test and the Wilcoxon
matched-pairs signed ranks test are the counterparts to the Independent and
Correlated samples t-tests, respectively. They are not bound by normal dis-
tributions or equal variances. They can be used with ordinal data or ranked
data instead of interval/ratio scaling for the dependent measure. These non-
parametric counterparts can be executed in a statistical program like SPSS.
The downside is that they are not as powerful as the parametric t-tests in
rejecting the null hypothesis.
 t-Test Procedures   •  73

Chapter Summary
T-tests are very useful tools. Often, you are comparing two groups of students,
managers, patients, employees, customers and so forth. The t-test is a simple
and straightforward statistic that allows you to get beyond a frequency count.
One group’s mean score may look different from another’s mean score, or it
may look the same. The litmus test for statistically significant two group com-
parisons is called the t-test.
 CHAPTER NINE

ANOVA Procedures

The t-test is very useful when you have two groups or testing times (pre/post
or before/after) to compare. But what happens if you have more than two?
The statistical technique that is analogous to the t-test is called Analysis of
Variance or ANOVA. The purpose of ANOVA is understood quickly if it is
thought of as an extension of the t-test.
ANOVA is an inferential statistic. It is also a parametric statistic and as
such it is very powerful. It can reject the null hypothesis or find differences if
they exist among groups. The assumptions of homogeneity of variance, equal
group sizes, and normal distribution of scores should be adhered to, just as the
t-test should meet these assumptions. Yet as a robust statistic, ANOVA can
sustain having the assumptions violated and still perform its function. There
is a nonparametric version of ANOVA called the Kruskal-Wallis H test. If
you have serious violations of the assumptions, use it.
ANOVA is a little more complicated than a t-test even when using a
statistical software package. Although there are many versions of ANOVA
(which will be described briefly), this book will focus on a simple version,
One-Way ANOVA. For the purpose of discussion, the main point is that you
use ANOVA as you would a t-test but when there are more than two groups.
There is one dependent and one independent variable but with more than
two levels or categories.
Here are some facts about One-Way ANOVAs:

75
76  •  Chapter Nine

• This statistical technique answers the null hypothesis: there is no differ-

ence among three or more (3+) groups on their respective mean scores.
• There is one independent variable with three or more (3+) categories.
These levels are nominally scaled. As a note, you can use an ANOVA
to compare two groups, but the t-test is intended for that purpose.
• There is one dependent variable. It is continuous in its numeric range
and uses ordinal, interval, or ratio measurement scales, similar to the
t-test.
• The statistic you obtain to determine statistical significance is the F
ratio or F statistic.

Steps for Conducting a One-Way ANOVA

An example will help to explain the basic purpose of the One-Way
ANOVA. You want to explore tobacco usage among teens in three towns
that have different demographic profiles (urban, suburban, and rural). You
think that there might be differences among the three towns, and such in-
formation would be helpful for public health messaging and town planning.

1. State your null hypothesis.

There is no difference in tobacco usage among teens in three towns.
2. Identify your independent variable and your dependent variable.
“Town of residency” is your independent variable with three catego-
ries: Town A, Town B, and Town C. Using nominal scaling, you assign
a numeric label so that your ANOVA can be calculated. You use 1, 2,
and 3 as your numeric values, which represent the three towns. Your
dependent variable is number of cigarettes smoked for a one-month
period by each teen. You have also calculated the mean number of
cigarettes for each group of teens by town. The number appears differ-
ent, but you cannot really tell until an ANOVA procedure is applied
to the data.
3. Set up your data for data entry into a database.
The data for your One-Way ANOVA might be recorded in a database
that looks like table 9.1. There are three sets of teens with a total of
39 teens altogether (N). The groups are not exactly the same size, but
close enough not to seriously violate assumptions of equal group size.
The Town Code and the Number of cigarettes smoked by each teen are
the only pieces of data that are entered into your statistical formula or
statistical package.
Table 9.1. One Way ANOVA Data Set: Town of Residency and
Cigarettes Smoked for a Month
Teen Town Town Code Cigarettes Smoked
1. Traci Town A 1 123
2. Karen Town A 1 119
3. Betsy Town A 1 120
4. Maria Town A 1 103
5. Jose Town A 1 168
6. Jesus Town A 1 190
7. Angel Town A 1 120
8. Margaret Town A 1 130
9. Bonita Town A 1 103
10. Ernest Town A 1 104
11. Miguel Town A 1 189
12. Sherri Town A 1 100
Town 1 Cigarettes Smoked
(n = 12) M = 131
13. Kathy Town B 2 120
14. Giovanna Town B 2 190
15. Ginger Town B 2 189
16. Kim Town B 2 145
17. Ernesto Town B 2 171
18. Lucia Town B 2 189
19. Anna Town B 2 145
20. Eric Town B 2 177
21. Juan Town B 2 189
22. Pedro Town B 2 167
23. Angela Town B 2 156
24. Abdul Town B 2 189
Town 2 Cigarettes Smoked
(n = 12) M = 169
25. Pepe Town C 3 120
26. Maureen Town C 3 134
27. Marci Town C 3 155
28. Peter Town C 3 134
29. Paulo Town C 3 144
30. Annie Town C 3 133
31. Rebecca Town C 3 142
32. Alisha Town C 3 123
33. Steve Town C 3 110
34. Irving Town C 3 102
35. Micah Town C 3 130
36. Nina Town C 3 190
37. Dana Town C 3 100
38. Guillermo Town C 3 103
39.   Francesca Town C 3 120
Town 3 Cigarettes Smoked
(n = 15) M = 129
78  •  Chapter Nine

4. Execute the ANOVA statistical procedure to obtain an F ratio, the calcu-

lated value, and a probability level.
After you have entered your data from your database, you can execute
the ANOVA. The F ratio is the statistic in ANOVA that indicates if
you can reject the null. Parallel to the t-test procedure, the F ratio is
called your calculated value. It is usually a three- to four-digit number
with two decimal places. For our data set the calculated F ratio is 8.94.
5. Examine the p-value on the computer output that is associated with the
calculated F ratio.
Is it higher or lower than the significance level you established a priori?
The traditional significance levels are .05 or .01. You have selected a
significance level of p < .05. The actual p-value is .00.
6. Accept or reject the null hypothesis.
Your calculated value was statistically significant at the probability
level you set, p≤ .05. You reject the null hypothesis. There is a difference
in tobacco usage among teens in the three towns.
7. Conduct follow-up tests (called post hoc tests) to determine where the dif-
ferences lay.

Post Hoc Tests

When you have statistical differences with t-tests, there are only two mean
scores. You can look to see which is higher and which is lower and then draw
your conclusions and inferences. If you have a statistically significant result
with ANOVA, your work is not complete. The only answer you have ob-
tained at this point is that there are statistically significant differences among
your groups: in this example the teens from the three towns. But which
towns are different on the variable of tobacco usage? Several possibilities for
significant differences might exist:

• Between teens in Town A and Town B (A and B)

• Between teens in Town A and Town C (A and C)
• Between teens in Town B and Town C (B and C)

With ANOVA there are more than two mean scores. You cannot eyeball
the means to see which is higher and then make deductions. You have to
conduct an additional statistical procedure. A follow-up test, called a post
hoc procedure or a multiple comparison test, is designed just for this purpose.
It shows exactly where the significant differences lay after (post) a significant
 ANOVA Procedures   •  79

F ratio is obtained in ANOVA. It pinpoints which mean scores are signifi-

cantly different from each other.
There are many options of post hoc procedures to choose from. Each is
named after the data scientist who developed it.

• Fisher’s LSD
• Duncan’s new multiple range test
• Newman-Keuls
• Tukey’s HSD
• Scheffé test

Some are more conservative than others are, making it harder to reject the
null. A liberal procedure will find a significant difference between two mean
scores that are relatively close together. A conservative procedure will indi-
cate that two mean scores are significantly different only when the means are
far apart. The Fisher LSD is the most liberal and the Scheffé test, the most
conservative of the post hoc tests.
You may wonder why t-tests are not used to compare each of the com-
binations above. That would be logical except for the fact that conducting
multiple t-tests increases the chances of getting spurious results with error.
As the number of t-tests increase, the probability of getting a statistically
significant difference by chance alone also increases. This means you are ap-
proaching Type I error territory. Post hoc or multiple comparison tests adjust
the level of significance to reduce the influence of error.
For the Town example, you conducted a Scheffé multiple comparison test.
It showed that the level of cigarette-smoking in Town B was significantly
different from that in both Town A and Town C on number of cigarettes
smoked per month. However, the levels for Town A and Town C were not
different from each other. Textbox 9.1 shows how you might report your
findings in table format. There are other ways to report ANOVA findings
that can be found in professional journals in different formats.

Degrees of Freedom with ANOVA

When you set up a table for One-Way ANOVA results, you must report
the degrees of freedom associated with the ANOVA procedure. The degrees
of freedom for ANOVA procedures are a bit more complicated and can
cause some confusion. There are two numbers that compose your degrees of
freedom for ANOVA: the Between Groups degrees of freedom (df) and the
Within Groups degrees of freedom (df).
80  •  Chapter Nine

Textbox 9.1. One-Way ANOVA Table: Town of Residency and Cigarettes

Smoked per Month
Town of Residency M (SD) F ratio df p
Town A 130.75 (32.92) 8.94 2,36 .00*
Town B 168.92 (23.03)
Town C 129.33 (23.39)
Total 141.95 (31.64)
*p ≤ .05
(Scheffé post hoc: Town B is significantly different from both Town A and Town C. There is a sig-
nificantly higher number of cigarettes smoked by teens in that town compared to the other two
towns.)

• The Between Groups df is equal to the number of groups minus one. In

this case you had three towns. Your Between Groups df is 2 or (three
groups – 1).
• The Within Groups df is equal to the number in each group minus one,
and then the sum for each group. In your case the Within Groups df is
36. It is calculated like this:
Group 1 [12 – 1] = 11 +
Group 2 [12 – 1] = 11 +
Group 3 [15 – 1] = 14+ = 36

For the ANOVA example, the degrees of freedom are these two numbers:
2 and 36. You will see it often reported like this (df = 2, 36) with a comma
separating the between from the within.

Other Versions of ANOVA

There are other versions of ANOVA procedures besides One-Way Analysis
of Variance. These become more complex because of the added variables
or factors in the research designs. There are multiple F ratios, interaction
effects, and other aspects that make the following procedures more sophis-
ticated both to execute and to interpret. Here are some basics on each, but
you may need help in executing and interpreting these research designs and
corresponding statistics.
 ANOVA Procedures   •  81

Factorial ANOVAs
Two-Way ANOVA
ANOVA designs, called factorial ANOVAs, compare more than one in-
dependent variable. Let’s use the same example that we had in One-Way
ANOVA, but add another independent variable, also referred to as a “fac-
tor.” Besides town of residency, we will add school (middle or secondary).
This design is called a Two-Way ANOVA.
Now we have Town of residency as one independent variable, and School
as a second independent variable. Instead of three groups that we had in
One-Way ANOVA, we have six groups because of the two independent
variables. Sometimes, a design such as this would be listed as a 3 × 2 design.
This means there are two independent variables: one with three categories
and one with two categories. Each of the six groups’ mean scores (M) on your
one dependent variable of “Number of Cigarettes Smoked” will be compared
by the Two-Way ANOVA (see table 9.2).

Table 9.2. 3 × 2 ANOVA Design: Town of Residency by School

Town A Town B Town C
Middle Middle Middle
M M M
Town A Town B Town C
Secondary Secondary Secondary
M M M

Three-Way ANOVA
This can further become a Three-Way ANOVA by adding yet another in-
dependent variable or third factor. Let’s add “coffee usage.” There are now
three independent variables with twelve separate groups on which the de-
pendent variable (M) will be measured. The three independent variables will
have three categories (Town), two categories (School), and two categories
(coffee usage). This may be called a 3 × 2 × 2 ANOVA design (see table 9.3).
82  •  Chapter Nine

Table 9.3. 3 × 2 × 2 ANOVA Design: Town of Residency by School by Coffee Usage

Town A Town B Town C
Middle Middle Middle
Coffee Coffee Coffee
M M M
Town A Town B Town C
Secondary Secondary Secondary
Coffee Coffee Coffee
M M M
Town A Town B Town C
Middle Middle Middle
No coffee No coffee No coffee
M M M
Town A Town B Town C
Secondary Secondary Secondary
No coffee No coffee No coffee
M M M

Analysis of Covariance: ANCOVA

There is another very useful statistical procedure called Analysis of Covari-
ance or ANCOVA. One purpose of this technique is oftentimes used to make
groups equivalent before they are compared on the dependent variable.
As you may recall, one of the primary rules of the t-test, as well as
ANOVA procedures, is that the groups you are studying are alike except for
your dependent variable. You want to ensure that something else (called
extraneous variance or error) does not account for differences between or
among the groups. ANCOVA adjusts for differences so that the focus of the
analysis is the dependent variable and not error.
As an example, your graduate research study is on the impact of lead-
ership training on managers at corporate versus nonprofit settings When
both groups of managers are tested on their leadership skills with a lead-
ership questionnaire, the pretest results (before any training occurs) show
statistically significant differences in mean scores between the two groups.
ANCOVA is used to statistically equate the two groups’ pretest scores on
leadership before the training is implemented. This allows you to determine
whether a real difference occurred over time at posttesting.
ANCOVA determines if the training produced different results in leader-
ship skills, when the program is concluded, by equating the mean differences
found at initial (pre) testing. The differences at pretesting could be due to
age, college degree, and other factors that make the groups nonequivalent
before the leadership training is undertaken. The pretest results on the
 ANOVA Procedures   •  83

leadership questionnaire are made equivalent so the posttest results on the

leadership questionnaire can be reliable comparisons of training efficacy.

Repeated Measures ANOVA

There are occasions when we need to measure something on a recurrent
basis. You measure the dependent variable more than once; you repeatedly
measure it. This is where the statistical procedure gets its name, Repeated
Measures ANOVA.
For example, your research might conduct employee satisfaction measure-
ments in September, again in January, and once again in September a year
later to track improvement in job satisfaction over time. Or you measure
patient response to a pharmaceutical drug intended to reduce cholesterol
levels at multiple times. All of these occasions might lend themselves to
using a Repeated Measures ANOVA procedure to determine changes over
time and to track results. Repeated measurements are needed to do that. This
statistical procedure allows for repeated testing on the dependent variable.

Chapter Summary
One-Way ANOVA is a very useful statistic and a great partner to the t-test.
It functions quite the same but affords us the opportunity to expand the
number of groups beyond two. This is often called for in many settings where
more than two categories exist for graduate student research. The conceptual
steps to One-Way ANOVA follow the t-test with the exception of the need
to conduct post hoc tests after significant results are obtained. Many statisti-
cal packages like SPSS offer this option as part of One-Way ANOVA.
For statistical questions that are more complex (in that they add more
variables to the research design), there are statistical models to accommodate
them. Factorial ANOVA, ANCOVA, and Repeated Measures ANOVA
are sophisticated statistical procedures that have been described in the most
elementary fashion here. Reference to statistical or mathematical textbooks
will present a more detailed discussion for interested graduate students.
 CHAPTER TEN

Chi-Square Procedures

In the previous two chapters we discussed parametric statistics that answer

questions about differences between groups (t-tests) and among groups
(ANOVAs). Parametric statistics are of great value and used often. Yet one
of the most valuable and popular statistics is a nonparametric procedure
called chi-square. It is also called the test of “goodness of fit.” Its symbol is X².
Unlike the t-test and ANOVA procedures, chi-square is not as powerful
to reject the null. It does not use the mean or standard deviation for com-
putation; it does not rely on an interval or ratio scaling. Because chi-square
relies on categorical data, its value lays in the statistic’s ability to answer
questions about data that are nominal and/or ordinal. Variables are measured
very often by their categories and not exact intervals. So, chi-square allows
you to answer important questions with variables measured with nominal or
ordinal scales. This is of great value to graduate student research in many
theses and dissertations.

Steps for Conducting a Chi-Square Analysis

An example of a chi-square analysis will promote the merit of this statistical
technique. Your graduate research is examining different levels of techno-
logical proficiency in new hires at your company. During the past year the
company hired 40 (n) business majors and 40 (n) engineering majors, or 80
(N) new hires in all.

85
86  •  Chapter Ten

1. State your null hypothesis.

There is no difference between business and engineering hires on
technological proficiency. (This null hypothesis sounds similar to
the hypotheses for t-tests and ANOVAs. However, chi-square is used
because it allows us to forgo many assumptions needed for t-tests and
ANOVAs, and it relies on nominal and ordinal scaling.)
2. Identify your two categorical variables.
The first is major field of study with two categories:
• Business
• Engineering
The second is technological proficiency with four categories:
• Novice
• Apprentice
• Practitioner
• Expert
3. Set up your data for data entry into a database.
Table 10.1 is the data set that indicates the major field of study and the
technological proficiency for your 80 new hires.
You have decided to code major:
• Business (1)
• Engineering (2)
Technological proficiency has been coded:
• Novice (1)
• Apprentice (2)
• Practitioner (3)
• Expert (4)
For data entry, you will only use numeric data as in table 10.1.
4. Set up a contingency table of your expected frequencies.
(As a note, you do not have to do this because statistical programs will
do this for you in the analysis. But it is good to understand how the
chi-square analysis operates.)
A contingency table of expected frequencies is an arrangement of
your categorical data into a two-way classification scheme. One of the
classifications becomes rows (across) and the other becomes columns
(down). The boxes, formed by the intersection of rows and columns,
are called cells. The cells tell us what you can expect, given the frequen-
cies of rows and columns. It is a simple application of probabilities
based on your sample numbers and their classifications. Table 10.2
displays a contingency table of expected frequencies for your data.
Table 10.1. Data Set for Eighty New Hires by Major Field of
Study and Technological Proficiency

ID # Major Proficiency ID # Major Proficiency

1. 1 2 41. 2 2
2. 1 2 42. 2 2
3. 1 2 43. 2 2
4. 1 2 44. 2 2
5. 1 2 45. 2 2
6. 1 2 46. 2 2
7. 1 2 47. 2 2
8. 1 2 48. 2 2
9. 1 2 49. 2 2
10. 1 2 50. 2 2
11. 1 1 51. 2 4
12. 1 1 52. 2 4
13. 1 1 53. 2 4
14. 1 1 54. 2 4
15. 1 1 55. 2 4
16. 1 1 56. 2 4
17. 1 1 57. 2 4
18. 1 1 58. 2 4
19. 1 1 59. 2 4
20. 1 1 60. 2 4
21. 1 3 61. 2 4
22. 1 3 62. 2 4
23. 1 3 63. 2 4
24. 1 3 64. 2 4
25. 1 3 65. 2 4
26. 1 3 66. 2 4
27. 1 3 67. 2 4
28. 1 3 68. 2 4
29. 1 3 69. 2 4
30. 1 3 70. 2 4
31. 1 3 71. 2 1
32. 1 3 72. 2 1
33. 1 3 73. 2 1
34. 1 3 74. 2 1
35. 1 3 75. 2 1
36. 1 3 76. 2 1
37. 1 3 77. 2 1
38. 1 3 78. 2 1
39. 1 3 79. 2 1
40. 1 3 80. 2 1
88  •  Chapter Ten

Table 10.2. Contingency Table of Expected Frequencies

Business Engineering Row Total
Novice 10 (25%) 10 (25%) 20 (25%)
Cell Cell
Apprentice 10 (25%) 10 (25%) 20 (25%)
Cell Cell
Practitioner 10 (25%) 10 (25%) 20 (25%)
Cell Cell
Expert 10 (25%) 10 (25%) 20 (25%)
Cell Cell
Column Total 40 (50%) 40 (50%) N = 80
Cell Cell

In the case of this example, the columns (down) correspond to

Major field with two categories. You allocate two columns for each of
the two major field categories. The rows (across) correspond to the
variable of technological Proficiency and have four different categories.
You allocate four rows for each of the four levels. You have 80 new
hires altogether.
The classification for “columns” creates two categories of major
field. Since there are 80 new hires, 50% are expected to be business (n
= 40) and 50% are expected to be engineering (n = 40). The “rows” for
technological proficiency break into four categories. Since there are 80
new hires, 25% are expected to be Novice (n = 20), 25% are expected
to be Apprentice (n = 20), 25% are expected to be Practitioner (n =
20), and 25% are expected to be Expert (n = 20).
Next you have to calculate the cells. The expected frequency of
any cell in the table is found by multiplying the total n of the column
with the total n of the row to which the cell belongs. The product is
divided by the total sample size (N). To obtain a cell for Business, you
multiple the total of the column (40) times the total of the row (20)
and divide by 80, the sample size. You get 10 as your cell size. This is
your expected frequency for each cell. It is helpful to report the cell size
as well as its percentage. In table 10.2, Expected Frequencies has both.
You have just constructed a contingency table of expected frequen-
cies from the two nominal variables for your 80 new hires. Each cell
has 10 new hires in it because of the breakdown of the two classifica-
tion variables: major field and technological proficiency. What you
would expect for your new hires as outcomes is that ten engineering and
ten business majors will populate each of the four rubric proficiency clas-
sifications if there were no differences between the two major fields on
 Chi-Square Procedures   •  89

technological proficiency. The contingency table of expected frequen-

cies enables you to determine the nature of the relationships between
your two categorical variables.
5. Execute the chi-square statistical procedure to obtain a calculated value and
probability level.
The chi-square analysis produces a contingency table of actual frequen-
cies from the actual frequency data from your new hires. It looks like
table 10.3.

Table 10.3. Contingency Table of Actual Frequencies

Business Engineering Row Classification
Novice 10 (25%) 10 (25%) 20 (25%)
Apprentice 10 (25%) 10 (25%) 20 (25%)
Practitioner 20 (50%) 0 (0%) 20 (25%)
Expert 0 (0%) 20 (50%) 20 (25%)
Column Classification 40 (50%) 40 (50%) N = 80

It also produces a chi-square statistic of 40.00. This is called your

calculated value. The degrees of freedom (df) for a chi-square statistic
are calculated by taking the [number of rows – 1 × the number of
columns – 1]. In the case of your new hires, you have two rows (major
field) and four columns (technological proficiency). The degrees of
freedom equals [2 – 1 × 4 – 1] or 3.
6. Examine the p-value on the computer output that is associated with the chi-
square statistic.
Your calculated value for chi-square is 40.00 with three degrees of free-
dom. Your calculated value was statistically significant at the probabil-
ity level you set, p ≤ .05. The expected frequencies were significantly
different from what you observed to be true in your actual frequencies.
7. Accept or reject the null hypothesis.
You reject the null hypothesis. There is a difference between business
and engineering new hires on technological proficiency. The engi-
neering majors had more technological proficiency than did business
majors.

Reporting Chi-Square Results

From the study of new hires, major fields of study, and technological pro-
ficiency levels, here is a simple format for reporting chi-square results. The
degrees of freedom (3), the number of subjects (80), the chi-square statistic
90  •  Chapter Ten

(40.00), and the significance level (.00) are reported. The results can be as
simple in a narrative like this. X² (3, N = 80) = 40.00, p ≤ .05. Or you can
develop a table of results that looks like table 10.4 or table 10.5.

Table 10.4. Reporting the Chi-Square (X2) Results with Expected and Actual
Frequencies
New Hires X2
N = 80 Frequency Novice Apprentice Practitioner Expert df = 3
Business Expected 10 (25%) 10 (25%) 10 (25%) 10 (25%) 40.00*
40 = n Actual 10 (25%) 10 (25%) 20 (50%) 0 (00%)
Engineering Expected 10 (25%) 10 (25%) 10 (25%) 10 (25%)
40 = n Actual 10 (25%) 10 (25%) 0 (00%) 20 (50%)
* p ≤ .05

Table 10.5. Reporting the Chi-Square (X2) Results with Actual Frequencies Only
New Hires
N = 80 Novice Apprentice Practitioner Expert X2 df p
Business 10 (25%) 10 (25%) 20 (50%) 0 (00%) 40.00 3 .00*
(n = 40)
Engineering 10 (25%) 10 (25%) 0 (00%) 20 (50%)
(n = 40)

* p ≤ .05

Collapsing Categories to Make Cells Larger

Sometimes when chi-square analyses are used, the cell sizes are small. This
is problematic for the statistic to operate optimally. As a recommendation,
have at least five values in each cell. What should be done if this shortcom-
ing exists in the data? Collapse the smaller cells to create larger cells. It must
be done with caution, and it must make sense.
For example, the US Census Bureau’s coding contains over 30 Hispanic
or Latino subgroups. The major ones are Mexican, Cuban, and Puerto Rican.
There are a few other subgroups like Salvadoran or Nicaraguan. If there were
too few numbers in these two categories, it would make sense to combine
Salvadoran and Nicaraguan into one group to give the subgroup more values
for statistical analysis. You might call this combined group of two “Central
American Latinos,” for example.
 Chi-Square Procedures   •  91

Other Nonparametric Statistics

There are other statistical procedures for use with nominal and ordinal scaled
data. Statistics texts will go into more detail about them but here are a few
that graduate students might be interested in using:

• The Fisher’s exact test: This test is used to compare two independent
samples with respect to a dichotomous variable. This is similar to the
chi-square test just described, but the Fisher’s exact test works better
when there are few subjects in the data set. If that is a limitation, this
is the statistic that helps with a solution.
• The Mann Whitney U test: This statistic is used to compare groups by
using rankings on variables of interest. Both groups are independent
just as with the Independent samples t-tests. This statistic does not use
mean scores but rather ranking of data to determine differences.
• The Kruskal-Wallis H test: This test is an extension of the Mann Whit-
ney U test but used with three or more groups, parallel to the one-way
ANOVA procedure. It uses a similar methodology to the Mann Whit-
ney U Test with rankings.
• The Wilcoxon-matched pairs signed-ranks test: Similar to the correlated
t-test, this statistic uses data where there is repeated measurement. If
the data are ordinal or the assumptions of the t-test cannot be met, this
statistic is useful.

The sign test, the binomial test, and the median test are among other nonpara-
metric procedures that are useful to consider. Again, they are less powerful
than their parametric counterparts to reject the null hypothesis. This means
a Type 2 error has more likelihood of occurring. Certainly, they should be
considered as part of the array of strategies that graduate students can avail
themselves of, if the situation calls for alternatives.

Chapter Summary
The chi-square statistic, which relies on nominal and ordinal scaling de-
scribed in chapter 1, is used very often because graduate research has categori-
cal variables. The chi-square statistical procedure allows us to compare groups
against each other on important variables. Although it may not have the
sophistication of its parametric counterparts, its utility may exceed theirs. It
is a statistical technique that provides us with insights that would be impos-
sible to diagnose if you had to depend on statistics that required interval or
ratio scaling.
 CHAPTER ELEVEN

Correlation Procedures

Besides focusing on differences, graduate research focuses on relationships.

There are times when we wonder if two factors are linked. For example, we
may ask some of the following questions:

• Does the amount of coffee consumption in the evening relate to the

ability to fall asleep?
• Does the incidence of juvenile delinquency vary with the age of a teen?
• Does on-screen time by children impact their verbal communication
skills?
• Does exposure to behavioral therapy help with body weight reduction?
• Does job performance correlate with absenteeism?
• Is patient compliance related to hospital readmission?
• Does daily writing in student journals increase reading comprehension?
• Can leadership skill development impact the opportunities for
advancement?

There are statistics that help to determine if relationships do exist, and if so,
what are the characteristics of those relationships? This is where correlations
are useful statistical techniques. They test the extent to which two variables
occur together, and how related they are. As statistics, correlations can be
descriptive and inferential at the same time. They can describe your data and
can also infer relationships from samples to populations.

93
94  •  Chapter Eleven

What Data Are Needed for Correlations?

You need two sets of variables (or paired observations) on the same individu-
als. In correlations, the first variable is called x. The second variable is called
y. Your data are paired observations of x and y on one subject. You correlate
x and y to see if there is a relationship. For the purpose of illustration, let’s
say that we wanted to know if physical fitness and weight were related for
corporate executives. We thought if you were physically fit, you would weigh
less. If you were in poor shape, you would weigh more. Fitness and weight
would be our two variables for our paired observations. We would collect
our data on all of our corporate executives. For each, we would have Weight
(x) in pounds and a Fitness (y) score. Then we would calculate a correlation
statistic and find the answer.
There are basically four questions that correlational procedures can ad-
dress. Each is followed by an example of variables that you might correlate
with each other:

1. If one variable [x] increases does the other variable [y] increase?
Example: Greater parent-to-child reading time increases school
readiness.
2. If one variable [x] decreases does the other variable [y] decrease?
Example: The less time patients spend out of bed, the less their
mobility.
3. If one variable [x] increases does the other variable [y] decrease?
Example: Increased sports participation by teens decreases cell
phone usage.
4. If one variable [x] decreases does the other variable [y] increase?
Example: Employee satisfaction declines when overtime require-
ments increase.

Correlation Coefficients (r)

The relationship between two variables, and the nature of that relationship,
is measured by a correlation coefficient, symbolized by the letter r. A correla-
tion coefficient is a two-digit decimal such as –.20 or +.78. The numeri-
cal values can range from minus one (–1.00) through zero (0) to plus one
(+1.00). However, –1.00 and +1.00 are perfect correlations; you rarely see
these numbers in the real world. (You can find numeric formulas for comput-
ing correlation coefficients in any statistical text.)
 Correlation Procedures   •  95

Strength and Direction of the Correlation

The correlation coefficient is a great statistic because a single number sum-
marizes the strength and the direction of the relationship between two vari-
ables. This is a lot of information.

Strength
The actual numeric value of the correlation coefficient tells us the strength
of the relationship. The nearer the number is to either +1.00 or –1.00, the
stronger the relationship is between the two variables. Correlations of –.88
and +.88 have the same strength. A correlation of –.88 is stronger than a
correlation of +.87 by only an infinitesimal amount.
A zero correlation (r = 0) indicates absolutely no correlation whatsoever.
The relationship between the amount of rainfall and an employee’s salary
would have no correlation (r = 0). Correlations of –.07 or +.02 are negli-
gible. Although there is no hard and firm interpretation of what constitutes
strength of all correlation coefficients, here are some suggestions of strength.

Table 11.1. Strength of Correlation Coefficients

Correlation Coefficient Strength of the Correlation
0 No correlation
±.01 to ±.30 Negligible to low
±.31 to ±.50 Low to moderate
±.51 to ±.70 Moderate to high
±.71 to ±.99 High to extremely high
±1.00 Perfect correlation

Correlations and Shared Variance

As a note, a correlation coefficient should not be interpreted as a percent-
age. Since it is a decimal, this can happen, but is incorrect. If the correlation
between salary and level of employee turnover is .50, you cannot conclude
that salary accounts for 50% of turnover. You must square the correla-
tion coefficient and multiply it by 100 to assess the shared variance of two
variables. This is called a coefficient of determination (noted by r²). For this
example, 25% (r²) is the shared variance between the two variables of salary
and turnover. You can conclude that 25% of turnover can be explained by
salary. This coefficient of determination (r²) is the percentage of variance
held in common by the two variables. Table 11.2 indicates the correlation
coefficients and the percentage of shared variance. As you can see, you must
have an extremely high correlation to assume that two variables are part of
each other.
96  •  Chapter Eleven

Table 11.2. Correlation Coefficients

and Coefficient of Determination
Correlation Coefficient of
Coefficient (r) Determination (r²)
.10 1%
.15 2%
.25 6%
.50 25%
.75 56%
.80 64%
.85 72%
.90 81%
.95 90%
.98 96%

Direction
Relationships between two variables can either be positive or negative.
That is what is meant by direction. Therefore, correlation coefficients can be
either positive or negative. A plus or a minus sign before the numeric value
indicates direction.
If a correlation is positive, it means that

• if one variable (x) increases, the other (y) increases, or

• if one variable decreases (x), the other (y) decreases.

Positive correlations sometimes have a plus sign before the decimal, but many
times it is implied. A positive correlation of +.67 might also look like .67.
If a correlation is negative, it means that

• if one variable (x) increases, the other (y) decreases, or

• if one variable (x) decreases, the other (y) increases.

There is always a minus sign before the correlation to indicate a negative

correlation. These are also called inverse correlations.

Steps for Conducting a Correlation Statistical Technique

The following example will demonstrate the qualities of asking questions
with correlational statistics. Let’s say that you have 20 preteens in your com-
munity who are at risk for dropping out of school. You believe that those who
participate in after-school clubs and extracurricular activities have better
attendance in school.
 Correlation Procedures   •  97

1. State your null hypothesis.

There is no relationship between school attendance and participation
in extracurricular activities.
2. Identify your two variables, x and y.
Your variable “x” is the total number of days in attendance, as tabu-
lated at the end of the school year. The variable “y” is the total number
of after-school and extracurricular activities the child participated in
during the school year.
3. Set up your data for data entry into a database.
Table 11.3 displays the data set for the preteens and the two variables:
attendance (x) and extracurricular activities (y).
4. Execute the correlation statistical procedure to obtain a calculated value, the
correlation coefficient (r).

Table 11.3. Data Set for Attendance (X) and Extracurricular

Activities (Y)
Attendance Out of Number of
Preteens 180 Possible Days Extracurricular Activities
1. Emmy 140 1
2. Auggie 170 7
3. Lowen 150 2
4. Elise 135 2
5. Mila 160 5
6. Luna 175 8
7. Mackie 160 6
8. Catrina 155 3
9. Jose 165 8
10. Juan 135 2
11. Wes 180 8
12. Rosalyn 165 7
13. Felicity 145 3
14. Mark 140 3
15. Miguel 155 3
16. Craig 145 1
17. Oliver 175 8
18. Alex 180 7
19. Rebecca 170 7
20. Seth 175 9

There are different correlation techniques that will be discussed shortly

and are available in statistical software programs. For these data, you have
selected a Pearson product moment correlation coefficient. The correlation
coefficient calculated is +.92 (r). From the plus sign (+) and the size of the
98  •  Chapter Eleven

correlation coefficient (.92), you can see that there is a strong, positive cor-
relation between attendance and extracurricular activities for preteens.

1. Examine the computer output to compare the p value that is associated with
the correlation coefficient.
The degrees of freedom for a Pearson product moment correlation
coefficient is (the number of pairs – 2). In this case there were twenty
pairs of data with the two variables so the degrees of freedom was 20
minus 2, or 18. As calculated, your actual p value for the data was .00
and your level of significance was p ≤. 05.
2. Accept or reject the null hypothesis.
You reject the null hypothesis. There is a significant relationship
between after-school/extracurricular activities and school attendance.
The direction and strength tell you that as one variable increases, so
does the other. Specifically, participation in extracurricular activities
boosts school attendance for preteens at risk for dropping out. Fur-
thermore, the shared variance (r²) is 84%—the correlation coefficient
squared. This means that 84% of school attendance can be explained
by the number of extracurricular activities a preteen in this community
is involved in.

Sample Size and Significance of Correlation Coefficients

The likelihood of obtaining a statistically significant correlation coefficient
is based on sample size. If you have a large sample, you often have a statisti-
cally significant correlation coefficient. For example, you can have ninety-
two pairs of data and at the .05 level of statistical significance (df = 90) and
a correlation of r = .22 would be statistically significant. However, this is a
low correlation when strength is considered. The strength of the correlation
coefficient is much more important than its statistical significance. Gradu-
ate students need to recognize this very important fact. Those who are less
informed would use statistical significance to make their case, and it would
be deceiving due to sample size alone. Look at the strength (size) of the cor-
relation coefficient and pay less attention to the p value.

Reporting Correlation Results

Reporting a correlation coefficient is very simple. Just use r and the equal
sign. For example, the previous example would be reported like this:

r (20) = +.92, p = .00

 Correlation Procedures   •  99

The number in parentheses is the number of subjects, 20. The correlation

statistic follows and then the p value for statistical significance. For our data
set, we would report in the narrative that there was a strong correlation
between extracurricular activities and school attendance. You may want to
show your actual data set. A scattergram for illustration would highlight the
impact that one variable had upon the other. What is a scattergram?

Scattergram, Also Called a Scatterplot

As we learned with frequency distributions, graphing techniques can illus-
trate a statistical message quickly and clearly. Similar to a frequency polygon,
the scattergram is a graphic representation of your correlation data. (Many
software programs can generate these for you.) You start with the vertical
(ordinate) and horizontal (abscissa) axes on a graph. For your data set, you
place the “number of extracurricular activities” on the vertical axis starting
with 1 and ending with 10, a little above the highest value in your data
set. On the horizontal axis is the total number of days in attendance in the
school year using intervals of 10 days. Then, you plot variables x and y for
each preteen on your graph. Place a dot on the graph where the two variables
intersect. Our scattergram looks like figure 11.1.

Figure 11.1. Scattergram for Attendance and Extracurriculars. Source: Author

developed
100  •  Chapter Eleven

Scattergrams are very useful because they tell us a lot about our two
variables and their relationship. Here are the pieces of information that scat-
tergrams supply.

1. Scattergrams can tell the strength of the relationship between the two variables.
• If the dots cluster close to an imaginary line, there is a very strong
relationship or correlation between the two variables.
• When the dots are scattered in an ellipse or cigar-shape, there is a
moderate correlation.
• When the dots are scattered randomly around the graph, there is a
low or negligible correlation.
The scattergram in figure 11.1 shows the dots clustering close to an
imaginary line or the line penciled in on the scatter diagram. This
illustrates a very strong relationship between the two variables. The
tighter the clustering of dots, the stronger the relationship. In the case
of our data set, the number of extracurricular activities might be con-
sidered a predictor of school attendance. The dots are clustered into a
fairly tight line.
2. Scattergrams can tell the direction of the relationship between the two variables.
• If the slope of the line falls from left to right, there is a negative
correlation.
• If the slope of the line rises from left to right, there is a positive
correlation.
The latter is the case with the data in our scattergram. The slope of
the line rises from left to right signifying a positive correlation. The
slope of the line indicates that as extracurricular activities increase, so
does school attendance. Conversely, a positive relationship also means
that as after-school and extracurricular activities decrease, so does
attendance. Just looking at the slope tells us a great deal about the
relationship between these two variables.
3. Scattergrams can also show outlier scores.
Values completely out of range with all the others should be consid-
ered for elimination in the calculation of a correlation coefficient.
They might contaminate the data set and give misleading results if a
correlation coefficient is calculated. In our example, there are no out-
lier scores. All of our x/y intersections fall close to the line. But look at
figure 11.2, which has some outliers in it.
 Correlation Procedures   •  101

Figure 11.2. Scattergram for Attendance and Extracurriculars with Outliers. Source:
Author developed

Practical Application of Correlations

There are many uses for correlations that you can avail yourselves of.

1. The most useful purpose is to see if two variables are correlated. This
is what chapter 11 has focused on.
2. Another important function of the correlation procedure is with re-
spect to tests or instruments we measure with. The reliability of a test is
reported as a correlation coefficient. It measures whether the test con-
sistently measures what it purports to measure. You can be confident of
that when you examine the correlation coefficient, reported in the test
manual as the “reliability coefficient.” If it is r = .80 or r = .75, we can
be certain that it is a pretty reliable measurement tool. Medical instru-
ments have high reliability coefficients; personality tests or vocational
interest assessments would have lower reliability coefficients.
3. A third function is performed by correlational statistics. If you are in-
terested in determining whether two raters or two observers are seeing
the same thing when they collect information, you are looking for inter-
rater reliability. For example, if two of your managers were evaluating
the customer service for the same group of bank tellers, you would want
their observations (recorded by a checklist or rubric) to be very similar.
By correlating their data, you would create an inter-rater reliability
coefficient. Hopefully, the correlational numeric value would be strong.
102  •  Chapter Eleven

Types of Correlation Techniques

• Pearson product moment correlation is a parametric statistic and requires
interval or ratio data on both the x and the y variables to calculate
it. This is the premier correlation statistic. As a parametric statistic,
the assumptions of homogeneity of variance and normal distribution
apply—just as they do with t-tests and ANOVAs. Pearson product mo-
ment correlation is called Pearson r.
• Point biserial procedure is a nonparametric statistic and uses a dichoto-
mous, nominal-scaled variable and an interval/ratio scaled variable to
calculate it. An example might be correlating tardiness (late/on time)
and frequency of behavioral problems at work.
• Spearman’s Rho is a nonparametric statistic that uses ordinal-scaled data
with ordinal-scaled data for computation. It correlates two sets of ranks
to determine their degree of equivalence. An example might be the
relationship between customer satisfaction levels and their agreement
with restaurant policy on wearing beach attire in the dining area.
• Rank biserial is a nonparametric statistic and uses a dichotomous,
nominal-scaled variable and an ordinal-scaled variable to calculate
it. An example is language spoken in the home (Spanish/English)
and satisfaction with communication from the hospital emergency
department.
• Phi coefficient is a nonparametric statistic and correlates two dichoto-
mous, nominally scaled variables. An example might be the relation-
ship between college type (public/ private) and STEM major field of
study (yes/no).

Chapter Summary
There is one last and very important point to make when correlations are
discussed. No attempt should be made to say that one variable (x) causes the
other variable (y). This is untrue. A correlation only suggests that a relation-
ship exists between the variables. It does not mean that one variable causes
the other. The use of correlational statistics is a great asset. It suggests that
relationships exist. To use them to pinpoint causality is a misuse of a highly
valuable statistical technique.
 CHAPTER TWELVE

Regression Procedures

This chapter is a brief and basic discussion of multiple regression. It is largely

a conceptual presentation of a complex statistic. Textbooks will provide
more depth for graduate students interested in this statistical procedure. As
a note, there are entire graduate level courses called Multiple Regression.
Correlation and linear regression are sometimes thought of as one and the
same. There are similarities but they are different. Correlation attempts to
determine if a relationship exists between two variables. Regression has two
distinctive purposes:

• Explanation: A variable, sometimes more than one, is used to explain an-

other variable of interest. Nurses’ satisfaction with the profession might
be explained by salary, hours, hospital resources, patient mix, supplies,
or building maintenance.
• Prediction: A variable, sometimes more than one, is used to predict
another variable of interest. Student tardiness might be predicted by
access to transportation, parental support, attitude toward school, or
food security.

Another difference is the labeling of the variables. The variables that ex-
plain or predict are the independent variables. The variable that is explained
or predicted is the dependent variable. In the previous explanation example, the
independent variables are salary, hours, hospital resources, patient mix, sup-
plies, and building maintenance. The dependent variable is nurse satisfaction.

103
104  •  Chapter Twelve

In the previous prediction example, the independent variables are access to

transportation, parental support, attitude toward school, and food security;
and the dependent variable is tardiness.
The several points you should consider before conducting a linear regres-
sion procedure are:

• Independent variables should have a moderate to strong relationship

(correlation) with the dependent variable. If they are going to be used
to predict or explain the dependent variable, it is common sense that
they should have a strong relationship with it.
• Independent variables should not have a moderate to strong relation-
ship (correlation) with each other. If the predictors are highly corre-
lated with each other, they will overlap or explain the same variance
in the dependent variable. They will be redundant and not useful for
either explanation or prediction.
• Variables should be quantitative, continuous variables. Ordinal, inter-
val, or ratio scaling are advised. (If categorical or nominal scaled vari-
ables are used as predictors, they need to be dummy coded in a special
manner. It would be advisable to get statistical advice on how to do this
because it is not a matter of just entering nominal codes.)
• Variances of the dependent variable should be similar across all levels
of the independent variable. This is called homoscedasticity in regres-
sion. (This is similar to the assumption in ANOVA and t-tests called
homogeneity of variance.)

Bivariate or Simple Regression

The most basic type of regression is called bivariate or simple regression. As
with simple correlation, there are only two variables that give the regression
model its name. In this regression procedure the independent variable is used
to explain or predict the dependent variable.
As an example, a high school math teacher felt that midterm grades could
explain attitudes toward math. She has the data on her 2022 freshmen stu-
dents that contain two variables: their actual fall midterm grades and their
attitude toward math scores. She conducts a bivariate multiple regression
procedure. The regression produces a statistic called a Multiple R (R). With
just two variables, the bivariate regression Multiple R looks just like the
Pearson correlation coefficient (r). The difference is that with regression,
the purpose is explanation and not just description. If a scattergram were
constructed, dots intersecting the two variables (attitude toward math and
 Regression Procedures   •  105

midterm exam grade) would form an ellipse, approximating a straight line.

The independent variable would show its effectiveness in explaining the
dependent variable. This line is called the regression line or line of best fit.

The Coefficient of Determination (R2)

In order to tell if the regression model is strong, the coefficient of determination
(R2) is examined, just as it is with the correlation coefficient discussed in the
last chapter. The R2 measures the proportion of dependent variable that is
explained by the independent variable. If the R2 was .04, only 4% of the vari-
ance in the dependent variable (attitude toward math) was explained by the
independent variable, midterm grade (not much). If the R2 was .84, 84% of
the variance in the dependent variable (attitude toward math) was explained
by the midterm grade (a great deal).

The Regression Equation

The bivariate regression procedure is very useful because it provides elements
for building an equation. This equation can be used in predicting the depen-
dent variable with subsequent sets of sample data. Regression procedures have
been used by college admissions officers to predict who should be admitted
to their institutions. It is used in weather forecasting to predict hurricane
seasons. It is used for many events for which data are known to predict the
unknown future.
Continuing with the high school math teacher example, the bivariate
regression output with the 2022 freshmen student data produced a coefficient
of determination R2 of .75, very strong. It looks like the regression 2022 data
could be useful in predicting attitudes of her 2023 class of freshmen. She
decides to use the regression procedure statistics from last year’s freshmen
class (2022) to predict attitudes of this year’s freshmen class (2023). Here is how
she does this.
On the bivariate multiple regression output from 2022 there are elements
that provide for building a regression equation. The equation looks like this:

Ŷ = a + b times X

The variable Ŷ is the unknown dependent variable being predicted: the at-
titude toward math. The variable X is the known independent variable, the
actual midterm grade for the 2023 students. The bivariate regression statisti-
cal output that the teacher used with the 2022 class data provides the num-
ber for the a (the constant) and b (the regression coefficient or slope). (As
106  •  Chapter Twelve

graduate students you need to know where to find a and b on the computer
output like SPSS.)
For each student in her 2023 class, the math teacher uses the equation by
putting in the value for a (the constant), the value for b, and the freshman’s
actual 2023 midterm grade (X) to predict their attitude toward math (Ŷ).
The equation is calculated like this: you add the value for a with b times the
midterm grade (X) and you get a predicted score on attitude toward math
(Ŷ). The equation is below.

Ŷ (attitude toward math) = a + b × X (midterm grade)

The teacher can now identify those freshmen whose attitudes are predicted
to be negative, and then provide support as they begin the second half of
their freshmen year of high school. The bivariate regression equation devel-
oped from the 2022 class of freshmen will help to identify those 2023 fresh-
men whose attitude toward math may be low.

Multiple Regression: Correlating More than Two Variables

Many research settings are multivariate environments. This means that more
than one factor usually accounts for what is happening. So how do graduate
students address this fact? Multiple regression procedures help with this. As
with any statistical software program, a click of the mouse produces statistical
output, but it is not always correctly executed. Multiple regression may be
an area where more statistical expertise is required so that your findings are
reliable and valid.
Like bivariate regression, multiple regression has one dependent variable
(Ŷ). Also, there are the same reasons for its use, explanation, and prediction.
Another similarity is that a regression equation is produced. Lastly, both are
inferential statistics, which look to the independent variables to account for
the variance in the dependent variable. The key difference is that there is
more than one independent variable (X).
The multiple regression equation with one dependent variable (Ŷ) and
two independent variables looks like this:

Ŷ = a + b1 × (X1) + b2 × (X2)

(X1) and (X2) are the two independent variables. There is always a constant
(a). There is a regression coefficient or slope (b) for each independent vari-
able (X).
 Regression Procedures   •  107

There could be a third independent variable and the equation would look
similar. There is always one constant (a). There is a regression coefficient or
slope (b) for each independent variable (X).

Ŷ = a + b1 × (X1) + b2 × (X2) + b3 × (X3)

The introduction of multiple independent variables is intended to assess the

extent to which each can explain or predict the dependent variable. Previous
assumptions are worth repeating:

• Independent variables must have a strong relationship with the depen-

dent variable.
• Independent variables should have a low correlation with each other.

Some Commonly Used Forms of Multiple Regression

Here are three forms of multiple regression procedures that graduate students
might use:

• Simultaneous multiple regression enters the independent variables all at

once into the procedure to see if they together can explain the vari-
ance in the dependent variable. This is like a stew where all meats and
vegetables are thrown into the pot together at once. There is no order
to the entry of independent variables.
• Stepwise multiple regression uses a stepwise entry of the independent
variables based on their incremental, nonoverlapping ability to explain
the dependent variable. This procedure selects out the most efficacious
independent variable first that explains the most variance in the de-
pendent variable. Then, it selects the next independent variable that
explains any variance that is left. And on with the next independent
variable until the incremental variance explained by predictors is
exhausted.
• Hierarchical multiple regression has a controlled entry of the independent
variables based on a predetermined plan. For this procedure there is a
deliberate method designed by whomever is executing the regression
procedure. For example, when exploring what factors might predict col-
lege selectivity, average household income of a town might be entered
first, as a controlling variable. Here it is acknowledged that household
income of the town likely explains much variance in college selectivity.
However, variables such as high school rank, number of AP credits, or
size of school might also be considered important factors for their effect
108  •  Chapter Twelve

on college selectivity. Conducting the regression procedure this way

allows the variance from income to be extracted first or controlled, but
then allows the other variables the chance to explain or predict the
college selectivity.

Steps for Conducting Multiple Regression

The following example will demonstrate the process of using a multiple
regression procedure. A department manager wants to explore what influ-
ences organizational climate among the 77 employees in the department.
The perception is that organizational climate is explained by two factors: job
satisfaction (measured by a total score on a survey) and the years employed.
The department manager has employee scores on an instrument measuring
organizational climate as well.

1. State the null hypotheses. Organizational climate (Ŷ) is not explained by

a combination of job satisfaction (X1) and years employed (X2).
2. Identify the multiple regression variables. The dependent variable is or-
ganizational climate (Ŷ) and the two independent variables are job
satisfaction (X1) and years employed (X2).
3. Set up the data for data entry into a statistical software database. There
are data on 77 employees with numeric values on organizational cli-
mate (Ŷ), job satisfaction (X1), and years employed (X2) in the database.
4. Execute descriptive and correlational statistics to examine the data before
running the multiple regression procedure.

These statistics are provided in statistical packages like SPSS as part of the
regression procedure. Table 12.1 displays the descriptive statistics on the
three variables under study. The means (M) and standard deviation (SD) are
produced for each variable X1, X2, and Ŷ.

Table 12.1. Descriptive Statistics for the Multiple

Regression
Variables N M SD
Organizational climate (Ŷ) 77 79.94 14.25
Job satisfaction (X1) 77 78.79 13.27
Years employed (X2) 77 17.01 8.83

Table 12.2 produces the correlations among the three variables. This is an
important piece of information. It indicates that the independent variables
 Regression Procedures   •  109

are correlated with the dependent variable to differing degrees, and the in-
dependent variables have a low correlation with each other. That is what is
desired when a multiple regression procedure is used.

Table 12.2. Correlation Matrix of Variables in the Multiple Regression

Organizational Job
Climate (Ŷ) Satisfaction (X1) Years Employed (X2)
r r r
Organizational climate (Ŷ) 1.00 .88 .40
Job satisfaction (X1) .88 1.00 .37
Years employed (X2) .40 .37 1.00

5. Execute a simultaneous multiple regression procedure where the two inde-

pendent variables of job satisfaction (X1) and years employed (X2) will
attempt to explain organizational climate (Ŷ).
6. Examine the p value that is produced on the summary table for multiple
regression.
Statistical packages like SPSS provide many pieces of statistical infor-
mation in the summary table like that in table 12.3. Look carefully for
these: Multiple R, R2, F ratio, degrees of freedom, and p value.
Table 12.3. Partial Summary Table for Multiple Regression
R R2 F df1 df2 p
.88 .78 127.30 2 74 .00*

*p ≤ .05

For this example, a Multiple R of .88 was obtained. This is a very

strong correlation, but more importantly the R2 or coefficient of deter-
mination was .78. This indicates that 78% of the variation in organi-
zational climate was explained by job satisfaction and years employed.
The F ratio was 127.30 with 2 and 74 degrees of freedom. The p value
was .00. The statistical significance of the regression procedure is re-
ported like this: R = .88, F (2,74) = 127.30, p ≤ .05.
7. Accept or reject the null hypothesis. Organizational climate (Ŷ) is ex-
plained by a combination of both job satisfaction (X1) and years
employed (X2). The null is rejected. However, were each of the two
independent variables similarly successful in explaining the variance?
One way to assess this is by using a t-test, which the multiple regres-
sion procedure provides as part of the SPSS statistical output. The first
t-test value for job satisfaction (X1) is 14.21 with a probability value of
110  •  Chapter Twelve

.00. It is lower than the significance level of p ≤ .05 and is a statistically

significant independent variable. However, the second T-test value for
years employed (X2) is 1.50 with a probability value of .14, not statisti-
cally significant. This independent variable did not add much to the
explanation of organizational climate, the dependent variable (Ŷ). It
was largely explained by job satisfaction (X1).
8. Set up scattergrams to display the findings from the multiple regression. The
line of best fit in figure 12.1 shows that there is a strong relationship
between job satisfaction (X1) and organizational climate (Ŷ). The
dots that intersect the two variables create a linear picture, forming a
straight line. The slope rising from left to right tells us of a positive re-
lationship. The scattergram in figure 12.2 for the years employed (X2)
show little relationship to organizational climate (Ŷ).
If the manager wanted to use these data to monitor organizational
climate in the department over time (such as year to year), the regres-
sion equation could be generated from data in table 12.4.

Ŷ = a + b1 × (X1) + b2 × (X2)

(Ŷ) = 6.16 (a) + .91 (b1) × (job satisfaction X1) + .14 (b2) × (years empl. X2)

Figure 12.1. Scattergram: Organizational Climate by Job Satisfaction. Source: Author

developed
 Regression Procedures   •  111

Figure 12.2. Scattergram: Organizational Climate by Years Employed. Source: Author

developed

Table 12.4. Constant, Regression Coefficients, and t-Tests

B t value p
Constant (a) 6.16
Job satisfaction (X1) .91 14.21 .00*
Years employed (X2) .14 1.50 .14
*p < .01

Chapter Summary
The occasion to undertake a regression procedure may be commonplace
due to the nature of correlated variables in research settings. Given that
fact, graduate students should consider using this statistic if the situation
warrants. Some expert assistance may be worth the investment, but the in-
sights produced are invaluable. The concepts behind this statistic are fairly
straightforward. However, some understanding of this statistic is necessary
because a click of the mouse will produce output, but it may be misinforma-
tion if done incorrectly.
 CHAPTER THIRTEEN

Practical Tips for Graduate Students

Starting the Research Process

Select a general research topic before you dive into the library or the internet and
drown in the resources. Go through this brief evaluation before you make a
commitment to an area of inquiry. This exercise will reduce your stress and
contribute to a topic that you will enjoy and will be acceptable to your degree
granting institution. There are several ways that you can determine if your
topic is a good one. Here are some strategies that you can use to assess the
value of your research topic:

• Consider your own interest in the topic. Are you genuinely interested
in it and willing to spend days, weeks, months, and maybe even years
focused on it? Are you willing to do the work to become an expert on
this topic? Is it one you would really enjoy exploring?
• Evaluate the importance of the topic. Is it something that will be of
benefit to the field you are part of? Will your work on this topic add to
the body of knowledge? Is there some practical value that other profes-
sionals in your field can use or apply that makes your topic of actionable
value?
• Is your topic relevant and timely? Is it something that others in your
field have written about in the recent past? Is it an area of discussion
right now, or is it passé? Is there a new treatment that you would like to
compare to another one or one that is older? Is there a gap in the body
of knowledge that you would like to fill? Are there contradictions in

113
114  •  Chapter Thirteen

results published in the literature that you would like to address? Does it
build upon someone else’s work but has a different twist? Is it an original
idea or has it already been done with conclusive results?
• Is the time required to complete the research on this topic realistic? Is
the topic something you can research within a timeframe or is the com-
mitment of time too great for you to undertake as a graduate student?
• What is the level of difficulty? Is your topic one that has sensitivity,
geography, or complexity that will make the process of investigating it
filled with extra stress? Does controversy surround it? Do you have to
travel far and wide to access the data sources? Is it so complex that you
will drown in the process of doing the research?
• What are the out-of-pocket expenses associated with your topic? Ana-
lyze the costs of pursuing it. Is the topic one that fits within a student’s
limited budget or will the cost exceed the pocketbook?

Use multiple library and online resources to establish keywords to get the ball
rolling; add keywords as you move along. “Research the research” in the litera-
ture review process. Here are a range of resources that can help you search
the literature, thoroughly and efficiently. Remember to add keywords as you
go along with the purpose of “researching the research.”

• Books (primary and secondary sources)

• Journals (professional association, trade journals, popular magazines)
• Newspapers (The New York Times, The Wall Street Journal, The Wash-
ington Post)
• Videotapes (your own, documentaries, films)
• Indices: ask a Reference Department librarian for help to find the right
ones (Reader’s Guide, Education Index, Index Medicus, ERIC, and others)
• Federal, state, governmental reports (Census Data, Public Health, De-
partment of Labor, etc.)
• Interviews
• Doctoral dissertations or graduate theses in your own college or univer-
sity library
• Dissertation Abstracts—a librarian can help you find this source
• Websites and blogs

Take your time in the library and online when reviewing the literature.
Avoid costly and embarrassing mistakes because you were rushing. It is in-
credibly easy to make literature review errors when you are first starting out
on this task. Listed below are some of the most common ones that graduate
 Practical Tips for Graduate Students   •  115

students make during their review of the literature. You will see that avoiding
them usually requires just a little slowing down, and thoughtful planning on
your part. Literature review is oftentimes done in a hurry instead of system-
atically. Sources are overlooked or skipped because

• you rely on secondary sources and forget the primary sources;

• you read journal article results sections only and not the methodology
sections, which give hints to producing good dissertations or theses;
• your topic is too broad; the literature becomes unwieldy and massive;
• you make copies with pages missing;
• you neglect to generate new key words as you pursue the literature
review; or
• you forget to use the reference section of journal articles to generate
lists of more pertinent articles.

Use index cards to take notes and to guide your actual writing. It seems
antiquated but you can shuffle the cards. They are physical and can be moved
in a different order for easy writing of narratives.

Selecting the Research Design

Research design and methodology are plans that promote systematic man-
agement of statistical data collection. Design and methodology dictate what
you need to answer your research questions. There are many types of research
designs for graduate research. Some are mixed methods, which means that
a combination of methods is used—possibly quantitative and qualitative ap-
proaches. Here are a few of the basic designs.

Historical Research Design: The purpose is to collect, verify, and synthesize

evidence to establish facts that defend or refute your hypothesis. It
uses primary and secondary sources and lots of qualitative data sources
such as logs, diaries, official records, reports, and other documents. The
limitation is that the sources must be both authentic and valid. Hypo-
thetical examples include the evolution of bottle feeding for infants or
the evolution of the laptop computer.
Case and Field Research Design: Also called ethnographic research, it uses
direct observation to give a complete snapshot of a case that is being
studied. It is useful when not much is known about a phenomenon.
It uses few subjects. Hypothetical examples are the investigation of
116  •  Chapter Thirteen

hospice services on a terminally ill patient or observations of Muslim

children in the American classroom.
Descriptive or Survey Research Design: It attempts to describe and explain
conditions of the present by using many subjects and questionnaires to
fully describe a phenomenon. Survey research design/survey method-
ology is one of the most popular for graduate research. These can be
executed online, by mail, by phone, or administered to intact groups.
The latter is the most cost effective and quickest way to collect your
data! You don’t need an example of a survey in this day and age, but
one might be the public’s view of American leaders.
Correlational or Prospective Research Design: It attempts to explore relation-
ships and/or to make predictions. It uses one set of subjects with two or
more variables for each to examine relationships. This research design
might show the relationship between smoking, exercise, and diet on
heart health, or maternal exercise during pregnancy and its impact on
newborns.
Causal Comparative or Ex Post Facto Research Design: This research design
attempts to explore cause and effect relationships where causes already
exist and cannot be manipulated. It uses what already exists and looks
backward to explain why. This research design might look at divorced
couples and examine causes by looking back on their relationships
while married, or surgical patient recovery and satisfaction with their
hospital stay.
Developmental or Time Series Research Design: Data are collected at certain
points in time going forward. There is an emphasis on time patterns
and longitudinal growth or change. This research might examine the
impact of exposure to STEM content in college to see if it positively
impacts students’ choice of careers, or monthly peer counseling and its
impact on juvenile delinquency.
Experimental Research Design: This design is most appropriate in controlled
settings such as laboratories. The design assumes random selection of
subjects and random assignment to groups (experimental and control).
It attempts to explore cause and effect relationships where causes can
be manipulated to produce different kinds of effects. Because of the re-
quirement of random assignment, this design can be difficult to execute
in the real-world (nonlaboratory) setting. An example of experimental
research is the impact of an experimental antidepressant on patients
with one group receiving a placebo. Another example is water with and
without fluoride and the effect on dental health.
 Practical Tips for Graduate Students   •  117

Quasi-Experimental Research Design: These typically have some of the

features of the true experiment such as control groups with a before
and after intervention. They are not done in the laboratory setting but
in the field that creates a less perfect comparison. However, they may
provide more real-world application. An example of this type of design
might be the comparison of two sets of business leaders with leader-
ship training for one group and the other with none to determine if
corporate metrics were impacted. Another example is two innovative
curricula and their impact on third graders’ reading abilities.

Writing Your Dissertation or Thesis

You have designed your study, collected data, and analyzed it; you are almost
finished at this point. Take the time to end on a high note. Here are some
tips in writing your dissertation or thesis.

Use a Standard Organizational Framework to Format the Text

When you begin to write your findings, it is imperative that you address the
format. A recommendation is to use a standard organizational framework to
format the text. The school you are enrolled in probably has a template for
you to follow. Use it. Also, find peers who have done similar dissertations or
theses at the school you are enrolled at and look at theirs for guidance. Orga-
nization is important for telling your story; for explaining what your graduate
research and statistics have uncovered.
You also need to keep your advisors connected to your findings. They
need to follow a logical sequence to understand the purpose of the study, the
methodology, and how findings were arrived at. A well-constructed narrative
will provide the reviewers with all of the necessary information including

• the theoretical rationale for the research;

• the research questions and hypotheses to be answered;
• the design and methodology;
• the subjects and sampling procedures;
• the methods and procedures of data collection;
• the instruments selected to collect data, their reliability, and validity;
• the statistical findings; and
• the implications, limitations, and follow-up research opportunities.
118  •  Chapter Thirteen

Keep the Text Concise and to the Point

Dissertations and theses do not have to be three inches thick or 500 pages
long. Use concise text that is to the point. Some graduate students bury a main
finding in so much verbiage that the reader becomes confused or misses it alto-
gether. Sometimes two points are grouped together, also adding to confusion.
As simple as it sounds, make your points one at a time. Keep the text simple,
clear, and readable without run-on sentences and extraneous descriptive lan-
guage. High-sounding language is not impressive and will only diminish the
impact of your findings. Make the paragraphs straightforward. Boil it down,
and then boil it down again. If the tables are well-conceived, a lot of narrative
is unnecessary. Do not add statistical jargon. If it is absolutely essential, make
it user friendly by defining it and its purpose. Otherwise, leave it out.

Use Tables to Present and Illustrate Your Findings

Tables can be your best friend in writing the main body—the statistical
findings. They should be used to both present and illustrate your statistical
findings. Usually, the statistical tables are developed after the statistics are
tabulated in SPSS or another software package. This provides you with the
means to have the tables to use when you tackle the narrative portion. Writ-
ing directly from the tables, generated by the study, will take the guesswork
out of organizing the findings. It will be easy to present the findings logically
in a way your advisors can follow.
Keep the tables simple. Some tables are intimidating because of the large
amounts of numbers they present. It is recommended that only one or two
variables be presented in one table. If you are reporting gender and grade
levels, that is enough in one table. Instead of putting several variables in one
table, use several tables to simplify the presentation.
Do not report numbers without percentages. In your tables that describe
categorical variables or those variables with nominal or ordinal scales, do
not list the numbers in each category without the corresponding percent-
ages. With raw numbers only, you make the reader work hard; why should
they? Again, make the data easy to grasp. Percentages convert raw data in
categories into meaningful information. The message is immediate and clear.

Consider Including Graphs If They Add to the

Comprehension of the Findings
Many graduate students fill up pages in their dissertations or theses with
graphs. There is an abundance of pie charts, histograms, scattergrams, and a
host of other pictures cluttering the statistical findings. Consider including
graphs if they add to the comprehension of your statistical findings.
 Practical Tips for Graduate Students   •  119

The prudent use of frequency polygons, histograms, bar charts, picto-

graphs, pie charts, and scattergrams is strongly urged—if they add to the
results of your work. If not, leave them out. Above all, use them if there is an
inherent message that is conveyed. Do not add “so what” graphs. If you have
a graph, it should stand alone without text narrative or numbers to tell some
story. Many times, they do not add anything except color and design. This is
not a project in graphic arts; it is a serious research study.

Have Your Dissertation or Thesis Read by an Independent Reviewer

Having your work read by an independent reviewer is an important factor
in quality control. This final piece of advice for writing your dissertation or
thesis is probably the last thing you will want to hear but it is strongly urged.
It is one of the most critical steps in producing the final document. You are
tired of this whole process. You have done a lot of work. And you want to be
finished. But, as a graduate student, you have been married to this work for
months and even years. You may be too familiar with the subject matter or
too impressed with your own writing style. As a result, sufficient detail may
be lacking in the explanation of some findings or too much abstract language
used in others.
Typographical errors leave a bad impression and take away from your
credibility. If words are misspelled, or there are grammatical errors, this can
only work against you, regardless of the quality of the research you are pre-
senting. Have a colleague, a family member, or friend critically review your
work before it goes to the next step. It may save a lot of blood, sweat, and
tears. You have climbed the highest mountain and are almost at the top.
Don’t let this last step be the spoiler.

Chapter Summary
Research is hard work, and it requires a commitment over time. One of the
greatest barriers to completing the thesis or doctorate is implementing the
research methods, including statistics. This book is intended to help you
move through the barriers that seem formidable but are not. With a little
persistence and time, you can master the research challenge and become the
success that you set out to be when you signed on to your academic journey.
In the process, you will find that conducting research can be gratifying in so
many ways. The content of this book is to help you find joy in that process,
and maybe even to continue to do research after your degree is obtained.
Good luck.
 About the Authors

Susan Rovezzi Carroll is president of Words & Numbers Research, Inc.

She received her doctorate in 1981 and served as a faculty member at the
University of Connecticut prior to founding the research and evaluation firm
in 1984. Her expertise is in research design, instrumentation, and statistical
analysis. Dr. Carroll has provided guidance and support to hundreds of gradu-
ate students in the United States and abroad. Her website for graduate stu-
dent support is www.dissertation-statistics.com. Her office email is dr.carroll
@optonline.net.

David J. Carroll is vice president of Words & Numbers Research, Inc. He

received an MSW in planning and evaluation in 1977. David has expertise
in demographic analysis, development of social indicators, and strategic
planning. He has conducted comprehensive studies on children’s well-being,
health equity, and women’s economic security. David published several pub-
lications and books. He can be reached at [email protected].

Since 1984, Words & Numbers Research, Inc. has conducted numerous evalu-
ation and research studies in the fields of education, healthcare, manufactur-
ing, banking, human services, and others. External evaluation services have
been provided to the National Science Foundation and the US Department of
Education. The firm conducts surveys, focus groups, and in-depth interviews.
The office can be reached by phone at 860-489-5639, by email at wordsnum
@optonline.net or on website at www.wordsandnumbersresearch.com.

121