USING SPSS/PC+

TO ANALYZE RESEARCH DATA:

A Step-by-Step Manual

Fourth Edition

by

Lars E. Perner

Robert H. Smith School of Business

University of Maryland
College Park, MD 20742-1815, U.S.A.
Internet: [email protected]
http://www.rhsmith.umd.edu/~lperner

Copyright (C) 1989, 1990 by Lars E. Perner

 Acknowledgements

The writer would like to thank Pat Stewart of Pennsylvania State

University, Marie Rowland of the City of San Luis Obispo, California, and
numerous students enrolled in marketing research courses for their useful
feedback on previous editions of this guide.

i
 Contents

Acknowledgements ..................................................................................................... ii

To the Instructor ........................................................................................................... v

To the Student .............................................................................................................. ix

What Can SPSS/PC+ Do for Me?............................................................................. 1

Introduction................................................................................................................... 9

Step 1: Coding Your Data ......................................................................................... 11

Step 2: Writing An SPSS/PC+ Program .............................................................. 20

Step 3: Entering The Program And Data ............................................................. 30

Exercise 1 .......................................................................................................... 33

Step 4: Checking Your Data And Program For Errors ..................................... 35

Step 5: Using Statistical Procedures And Computations ................................ 38

Frequencies ........................................................................................................ 38
Creating New Variables: Compute.............................................................. 39
Recoding Variables ........................................................................................ 39
Crosstabulation............................................................................................... 41
Product Moment (Pearson) Correlation .................................................... 45
Multiple Linear Regression........................................................................... 47
Discriminant Analysis .................................................................................. 48
"""Count: "Counting" on how many variables a criterion is met ........ 49
Descriptives: A summarized version of Frequencies ............................ 49
Npar tests: Non-parametric statistical tests............................................ 50
Means: Providing a breakdown of population means by
subgroup ............................................................................................... 51
Oneway and ANOVA: Analysis of Variance ........................................... 52
Plot: Turning bivariate data into a scatterplot ....................................... 52

ii
 Reliability: Finding coefficient Alpha and other measures of
reliability in a scale ............................................................................ 54
t-test: Testing for differences in two population means. ..................... 55
Advanced features: Factor, Cluster, Hiloglinear, and MANOVA........ 56

Appendices:

A: Common Questions About SPSS/PC+........................................................ 58

B: Working With SPSS/PC+ Output ................................................................... 61
Frequencies ...................................................................................................... 61
Crosstabs .......................................................................................................... 65
Pearson Correlation ....................................................................................... 68
C: Working With Large Data Sets ....................................................................... 72
D: Using System Files ............................................................................................. 74
E: Importing Data From Lotus 123 ..................................................................... 77
F: Similarities Between SPSS/PC+ and Lotus 123........................................... 80
G: Similarities Between SPSS/PC+ and dBase III+......................................... 83
H: SPSS-X: The Mainframe Version ................................................................... 85
I: Incorporating SPSS/PC+ Output Into WordPerfect Reports ................... 86
J: Dealing With Printer Problems .......................................................................... 88
K: Statistical Significance...................................................................................... 89
L: Expanded Glossary............................................................................................... 95
M: Additional Sources of Information ............................................................. 110

iii
 To the Instructor

This text is intended to meet the needs of students enrolled in a one

term course which is only partially devoted to data analysis. In such a
situation, the instructor cannot afford to take out a major part of the class to
discuss the theoretical aspects of data analysis and processing. In fact, this
book resulted from the frustration I experienced in not being able to find a
textbook suitable for a one quarter marketing research methods course of
which only the last three weeks were devoted to data analysis.

My emphasis has been on getting the student started on doing his or

her own analysis as quickly as possible. Rather than discussing the concepts
and commands in a highly theoretical manner, I have presented some very
readable and illustrative examples. A discussion of the concepts and
mechanics behind these examples is intended to enable the student to
generalize these illustrations to their own projects.

I have taken several steps aimed at enhancing the student's

understanding of the analytical process. For example, I have chosen to
consistently place data in the same file as the SPSS/PC+ program and have
confined any treatment of the topic of system files to an appendix. While
this may at first glance seem to be a wasteful approach which requires the
user to run the same program over and over, I believe it helps the student
more clearly understand where the data comes from. For the same
pedagogical reasons, only a fixed data format is used. Not only does a fixed
data column arrangement promote a better appreciation of how each
variable is associated with every case; it also avoids inevitable sequence
displacement problems that occur when the free format is used.

I have deliberately omitted discussion of the menu driven command

system and the optional data entry module. While some users find these
features useful, I believe that a clearer understanding is reached by typing in
one's own commands. The menus can't supply such information as the
variable and value labels, which account for the majority of the program
anyway. What is more, once one gets into trouble with the menus, it is often

iv
very difficult to get out. As for typing in the data directly, rather than
through the data entry module, it is my experience that such practice helps
the user appreciate the fixed format of the data entry. Also, this method of
data entry is more efficient for large data sets since the "key puncher" is not
required to enter a carriage return after entering each variable.

These approaches, along with my avoidance of alphanumeric

variables, further allow for a better generalization to other statistical
packages and formats.

Students today are increasingly computer literate, and it is not

unusual for students to be familiar with such software programs as Lotus
123 and dBase. I have included appendices that illustrate similarities and
differences between these programs in order to help students who already
know the programs easily understand comparable procedures and
operations of SPSS/PC+. An appendix also indicates how data can be
imported from such sources as Lotus 123.

The other side of the coin is that many people still feel uneasy when it
comes to dealing with computers. Many official software manuals, and
even a number of secondary texts, tend to present the material in a
relatively abstract and "sterile" manner. At the loss of a slight degree of
generality, I have instead chosen to present examples that will help the user
"fill in the blanks" on his or her own programs. That is, I frequently use
"real" variable names rather than referring to some abstract notion such as
"varlist." In addition, I have used a great deal of humor and
anthropomorphism to put the "reluctant" computer user off guard. On that
topic, I see no reason why humorous examples cannot be as informative and
educational as boring ones. The purpose of examples is to show the student
how data can be analyzed, and while "real World" projects may be less
engrossing, the skills learned in a humorous situation can be generalized to a
routine, or perhaps even boring, situation.

The ease with which computers can perform a large number of

computations within a few moments creates the potential for a great deal of
abuse. It is almost certain that something "significant" will show up if one
performs enough tests. Thus, I have strongly emphasized the practice of

v
making a limited number of hypotheses before running the statistical
procedures. The program named "SIGNIF.EXE," which is included on the
distribution diskette for this manual, allows the user to compute the
probability of making at least one Type I error given n significance tests.
Having noted students' tendency to attempt a very large number of
analyses, it is strongly recommended that the instructor stress the concept of
accumulating error levels in class.

Since this is intended as an introductory textbook, it only discusses

those statistical techniques that would likely be encountered in an
introductory course. However, as students complete this book and the
exercises contained herein, they should be well prepared to consult the
official manuals issued by SPSS, Inc. These manuals also provide an
excellent discussion of the theory behind the statistical procedures involved.

Whenever one attempts to write a textbook for even a mildly diverse

group of readers, the question of how much background should be assumed
from the reader arises. In the present case, such a concern is particularly
relevant when it comes to deciding how statistical output should be
interpreted. My choice, for better or for worse, has been to leave any
extensive discussion of the statistical principles involved to the instructor
and/or any other textbooks and reference materials that may have been
made available to the student. I have only touched lightly on such topics as
statistical significance, although examples and brief discussions of the
techniques available might suggest applications appropriate for specific
statistical procedures. In order to accommodate users with highly diverse
statistical backgrounds, I have included, as an appendix, an expanded
glossary that explains many of the terms that the student may find
unfamiliar.

With SPSS being available in so many forms, one may wonder about
the wisdom about using the personal computer version as opposed to
mainframe versions such as SPSS-X available on most campuses. After all,
SPSS/PC+ may be installed only on a few computers on campus while
mainframe terminals are readily available to students, some of whom may
even be able to dial up the university mainframe by modem from home. I
think there are several reasons why the PC version is preferable. First, many

vi
students have already had experience on the IBM PC and thus need much
less introduction. Secondly, should the student wish to include part of the
SPSS output directly in a report, taking it from the "SPSS.LIS" output file is
much easier than downloading an output file from the mainframe. Finally,
those students who will end up using SPSS on the job are much more likely
to find a SPSS/PC+ than the mainframe version in industry.

In that same vein, a question arises as to whether one should use the
complete SPSS/PC+ program or SPSS/PC+ Studentware. While I feel that
using the "real thing" will be a better preparation for practical industry
applications, I don't think that students who only have the Studenware
edition available will be seriously shortchanged in an introductory course.

vii
 To the Student

Much has changed since I was first exposed to the Statistical Package
for the Social Sciences (SPSS) in the early 1980s. As an industry standard,
SPSS now exists in versions for many different computers. At the time,
however, I was confined to a mainframe version which did not seem terribly
user friendly.

I didn't find the textbook assigned to be appreciably more user

friendly than the program. After several chapters referring to such abstract
terminology as a "field," we finally got to write some programs, but the
coverage was still rather abstract. I have chosen a different, and hopefully
more readable approach, to writing about SPSS/PC+.

Many textbooks take the approach of presenting the theory fully

before providing any examples. I don't find that a good way to learn. Few
of us learned to talk by studying a dictionary. Most of us instead learned by
listening to others and then adapting the language to suit our own needs.

I have chosen a similar approach in writing this book. First, we will

look at some SPSS/PC+ programs and explore what they do. Looking at
these examples, we will discuss how you can generalize the techniques
presented to meet your own needs.

This book is an introduction to SPSS/PC+, and as such, it covers only

a fraction of the options available. If you go on to use SPSS/PC+
extensively, you will probably find the official manuals published by
SPSS/PC+ to be invaluable. Not only do these manuals provide a great
source of reference for the SPSS/PC+ procedures; they also provide an
excellent and very readable discussion of the statistical techniques available.

I have included some relatively humorous examples in this text.

Aside from my own exaggerated sense of humor (which has a tendency to
get me into trouble on many occasions), I think a witty approach may help
to (1) motivate you to keep up with the reading, and (2) help those people

viii
who feel uncomfortable with computers to feel more at ease with the
subject. You should not feel guilty about enjoying the reading, however.
While you may risk getting some dirty looks in the library if you laugh out
loud, the examples, although often far fetched, illustrate real research issues
and are just informative as boring examples. Why shouldn't enlightened
poodle breeders commission marketing research just like the manufacturers
of laundry detergent? Those people who split their investment funds
between an inventory of poodles and a controlling interest in Proctor &
Gamble will be just concerned about increasing levels of prejudice against
small dogs as about consumer trends toward buying generic household
products.

A few final cautions. The computer has today made it very easy to
perform statistical calculations that could literally have taken a person
months to perform in past decades. With this potential, however, comes an
opportunity for serious abuse. This can take two forms. First, anyone can do
complex statistical calculations in SPSS/PC+, but the output may not be at
all meaningful. In class, you may have discussed the distinction between
nominal, ordinal, interval, and ratio scales of measurement. However, the
computer doesn't know where your data comes from and will gladly comply
with your request to include nominal data in a procedure that really
requires interval, or even ratio, level data. Therefore, be sure you
understand the assumptions behind a statistical technique before running it.

The second potential for abuse results from our ability to

indiscriminately perform a great number of analyses at the same time.
Intuitively, it makes sense that if we try long enough to find something that
is "very unlikely," we will. Suppose, for example, that you decide to make a
telephone survey to find out the birthdays of the respondent and his or her
spouse. You call up a thousand married people at random from the phone-
book (i.e., you terminate the interview once you find out that someone is not
married). You then find out that three couples have their birthdays on the
same date. Any great discovery? Well, in each trial (call), the chance of
"success" is approximately 1/365=0.00273. Multiplying our one thousand
trials times that 0.273% chance, we would expect about 2.73 couples in our
sample to have the same birthday. Similarly, if you perform fifty
significance tests at an α=0.05 level of significance, you would expect 2.5

ix
tests to come out significant by chance alone.

x
 1
What Can SPSS/PC+ Do for Me?

SPSS/PC+ is an incredibly beautiful program. If you like computer

software, you might think of SPSS/PC+ as being as beautiful as programs
like WordPerfect and Lotus 123. Unless you are a real nature enthusiast,
you will probably find SPSS/PC+ even more impressive than a beautiful
river or mountain range. SPSS/PC+ will probably compare favorably with
the grandest piece of literature or greatest work of art you have ever seen.
And, depending on how humanistic you are, you may also find SPSS/PC+
more beautiful than the one you love. It's not surprising if, at first, you find
this statement difficult to believe, so let's get right into the features of
SPSS/PC+. You be the judge! (This chapter is intended to show you the
variety of statistical procedures available within SPSS/PC+, and may cover
statistical methods that you have not yet studied. The purpose is only to
show you what is available, and consequently, the chapter is not intended
for detailed study.)

At the most basic level, you might want to tabulate some data you
have collected in a survey or through other means. Later on in the book, we
will meet a dog breeder who is very interested in whether people own dogs
or not and what kind they prefer. Suppose he has asked you to do a survey.
After you have entered the data, you can ask SPSS/PC+ for a table that
indicates how many people gave each of the possible answers to a question:

DOGOWN Ownership of Dog

Valid Cum
Value Label Value Frequency Percent Percent Percent

Yes 1 82 46.9 49.1 49.1

No 2 82 46.9 49.1 98.2
Not sure 3 3 1.7 1.8 100.0
Missing 9 8 4.6 MISSING
------- ------- -------
TOTAL 175 100.0 100.0

Mode 1.000 Minimum 1.000 Maximum 3.000

For example, the SPSS/PC+ output indicates that eighty-two people

claimed to have a dog, eighty-two claimed not to have one, three people
 2
were not sure if they owned a dog or not, and nine people supplied answers
that were not interpretable or supplied no answer at all; hence, their
answers are "missing."

Although SPSS/PC+ does not

provide good graphics capabilities
except in an optional graphics module
that many institutions do not have,
you can use a spreadsheet or graphics
package like Lotus 123, Excel, or
Quattro to create a graph to depict the
responses given. It might look
something like this:1

OK, so this saved you some time

and provided an output that was somewhat neater and more organized
than what you would have obtained if you had done the calculations by
hand. Is that all we have to be excited about?

Of course not! We are just beginning. SPSS/PC+ allows us to do

more involved things as well. For example, we might be interested in
assessing the relationship between two or more variables. One popular
feature allowing us to do this is the "crosstabs" table. Let's suppose that you
have been hired by a major airline that wants to diversify into the
hospitality industry at its destination sites, offering consumers an integrated
vacation package. In order to establish the kinds of restaurants that will
appeal most to vacationers at each location, the airline would like to know if
there is any relationship between food preference and favorite vacation
destination. After consulting your marketing textbook, you decide to do a
crosstabulation. You think that people who prefer the Orient would be more
likely to prefer Chinese food; those people who prefer Europe would like

1
There is no need to cut and paste! Lotus graphics ("*.pic" files) can be
imported directly into WordPerfect 5.0 or 5.1.
 3
Italian and French food; and those people preferring the Continental U.S.
would prefer Western type food such as steaks, hamburgers, and fries. You
are not quite sure about those who prefer to visit Hawaii. SPSS/PC+ allows
you to test your hypotheses:2

Crosstabulation: VACPREF Preferred vacation destination

By FOODPREF Preferred dinner meal

Count Western Italian Chinese French Other

FOODPREF > Row Pct Row
Col Pct 1 2 3 4 5 Total
VACPREF
1 20 1 5 3 1 30
U.S. Mainland 66.7 3.3 16.7 10.0 3.3 20.0
83.3 3.0 8.5 10.3 20.0

2 1 1 11 1 2 16
Orient 6.3 6.3 68.8 6.3 12.5 10.7
4.2 3.0 18.6 3.4 40.0

3 2 27 4 22 55
Europe 3.6 49.1 7.3 40.0 36.7
8.3 81.8 6.8 75.9

4 1 4 39 3 2 49
Hawaii 2.0 8.2 79.6 6.1 4.1 32.7
4.2 12.1 66.1 10.3 40.0

Column 24 33 59 29 5 150
Total 16.0 22.0 39.3 19.3 3.3 100.0

Chi-Square D.F. Significance Min E.F. Cells with E.F.< 5

---------- ---- ------------ -------- ------------------

156.76735 12 .0000 .533 8 OF 20 ( 40.0%)

Statistic Value Significance

--------- ----- ------------

Cramer's V .59023
-------------------------------------------------------------------------------

If you are familiar with the Chi square statistic, you can see that there
is strong evidence to reject the null hypothesis that food and vacation
preference are "independent." As a matter of fact, the Cramer's V statistic
2
Normally, you should define hypotheses more specifically before
testing them. For now, we are just testing whether the two variables in
question (food preference and favorite vacation destination) are
dependent.
 4
even suggests a modest relationship.

But, you are of course not limited to non-parametric statistics, or

procedures that only use ordinal measures, in SPSS/PC+. You can also do a
Pearson correlation analysis if you have interval data or "better."3 Let's
suppose that the airline is interested in tourist travel and would like you to
conduct a study of how to best predict how much money a person spends on
vacation(s) every year. You decide to correlate amount spent on vacations
against various other variables.

Correlations: FOOD AUTO ENTERT DINING PHONE

VACATION .4304 .5754 .5728 .5620 .3498

( 263) ( 263) ( 263) ( 263) ( 263)
P= .000 P= .000 P= .000 P= .000 P= .000

Once you have done a correlation analysis, you might feel that the
proper next step is to do a multiple regression analysis to see if you can
improve your ability to predict based on the introduction of additional
variables. Unlike Lotus 123, SPSS/PC+ gives you several choices as to
which method you want to use (forward inclusion, backward deletion,
stepwise consideration, or "forced" entry). If a traditional method doesn't
suit your needs, you can introduce non-linear or log-linear models. Let's try
to "predict" a person's telephone bill from his or her expenditures on other
items and other demographic information.

3
Actually, a correlation analysis is in practice applied many times
even when only ordinal level data is available. This is not sanctioned as
correct by most statisticians, but the this approach can sometimes still
yield meaningful results. When one or both of the variables depart
seriously from the assumption of interval properties, the true relationship
between the variables may be greatly underestimated. On the other hand,
a correlation will rarely provide "false positives" or suggest a relationship
that does not exist.
 5

Variable(s) Entered on Step Number

3.. COMPUTER ownership of computer

Multiple R .67310
R Square .45306
Adjusted R Square .44673
Standard Error 84.09635

Analysis of Variance
DF Sum of Squares Mean Square
Regression 3 1517312.98023 505770.99341
Residual 259 1831698.85247 7072.19634

F= 71.51541 Signif F = .0000

Equation Number 1 Dependent Variable.. PHONE total household phone bill

------------------ Variables in the Equation ------------------

Variable B SE B Beta T Sig T

DINING .09375 .01064 .58737 8.811 .0000

AUTO .22249 .10231 .14281 2.175 .0306
COMPUTER -36.34246 16.83655 -.10385 -2.159 .0318
(Constant) 408.56124 31.01232 13.174 .0000

------------- Variables not in the Equation -------------

Variable Beta In Partial Min Toler T Sig T

FOOD -.05499 -.07431 .47474 -1.197 .2324

ENTERT .02691 .01453 .13203 .233 .8156
HOME -.02361 -.02499 .36798 -.401 .6884
VCR -.02899 -.03802 .45710 -.611 .5417

End Block Number 1 PIN = .050 Limits reached.

-----------------------------------------------------------------

Not all statistical analyses have to be this involved. Maybe by now

you are getting tired of doing research for someone else. Suppose that, for
some reason, you have a good cause to believe that there is a killing to be
made in retailing tall people's clothing. Not having a preference for either
men's or women's clothing, but realizing that you need to specialize to
compete, you toss a coin and end up in the women's apparel business.
Naturally, you will want to find a geographic location where there are a lot
of tall women. Since you can't afford to research the heights of women all
over the nation, you decide to focus on Texas in the belief that Texas cities
will be the most likely spots for success. You now proceed to collect data on
the heights of adult women (in inches) in Dallas and Houston. One way to
 6
test for such differences would be to employ a t-test (two-tailed since you
have no preconceptions), but a more direct way might be to "break down"
information by city. Using the SPSS/PC+ procedure Means you get the
following result:
Summaries of HEIGHT Height of subject (inches)
By levels of CITY City of residence

Variable Value Label Mean Std Dev Cases

For Entire Population 66.6960 1.7183 200

CITY 1 Dallas 68.1830 .9038 100

CITY 2 Houston 65.2090 .8069 100

Total Cases = 230

Missing Cases = 30 OR 13.0 PCT.

Now you can optionally calculate a confidence interval for the means
of heights of the women on in each. Tentatively, it looks like Dallas might be
the best bet. (An added benefit is that by choosing this location, you will be
closer to the Oil Barons' Club).

Of course, there is always the possibility that you decided to split the
cost of doing the survey with a classmate who believes that there is more
money to be made in tall men's clothing. In that case, of course, you would
want to find out about the heights of the men in the different cities as well.
However, you would want to keep track of the heights of the men and
women separately, both because the city that has the tallest men might not
have the tallest women and because the great between-sex height
differences would greatly inflate the estimate of within-gender variability.
"Means" allows you to produce this table:
 7
Summaries of HEIGHT Height of subject
By levels of CITY City of residence
SEX Sex of respondent

Variable Value Label Mean Std Dev Cases

For Entire Population 68.9820 2.9810 200

CITY 1 Dallas 70.6840 2.6261 100

SEX 1 Male 73.1160 .8728 50
SEX 2 Female 68.2520 1.0492 50

CITY 2 Houston 67.2800 2.2614 100

SEX 1 Male 69.3700 .8929 50
SEX 2 Female 65.1900 .7875 50

Total Cases = 230

Missing Cases = 30 OR 13.0 PCT.

Now that you have been working with SPSS/PC+ for quite some time,
you are really getting good at marketing research, and you feel that you can
handle almost anything--even the unexpected. One day, you receive a
distressed phone call from Rudolph the Redneck Reindeer. Rudolph is
hysterical because his employer, an elderly man who likes to wear a red suit
during the winter months, has warned his long time sleigh puller that he
may have to lay him off because people are beginning to demand greater
sophistication from reindeer. You agree to do a survey for Rudolph to find
out how important that aspect really is to consumers. However, having
studied marketing research for some time, you realize that one question or
"item" will not give you a result that is reliable enough to give you an answer
that is dependable. You therefore decide to create a scale of "Appreciation of
Sophistication in Reindeer," where subjects will be asked to indicate their
level of agreement or disagreement with Likert type questions on a scale
ranging from "strongly agree" (1) "to neither agree nor disagree" (4) "to
strongly disagree" (7). Now you want to test whether the average score on
the questions will be reliable enough to be meaningful. You and Rudolph
hope that people will score as low as possible on that scale, suggesting that
his employer's concern is unwarranted. After you "reverse score" item #4
(which is worded in the opposite direction of the other questions), you are
ready to generate the following estimate of internal consistency:
 8

R E L I A B I L I T Y A N A L Y S I S - S C A L E (S O P H I S T)

1. SOPHIS1 Good reindeer are educated

2. SOPHIS2 Reindeer should have good manners
3. SOPHIS3 A reindeer should have a good cultural background 4. SOPHIS4 Redneck reindeer are OK
(Reverse scored)
5. SOPHIS5 Reindeer should use good grammar
6. SOPHIS6 Reindeer should be graceful
7. SOPHIS7 A reindeer's style is more important than the color of his nose

RELIABILITY COEFFICIENTS

N OF CASES = 100.0 N OF ITEMS = 7

ALPHA = .9218

Assessing internal consistency is an advanced topic, and you may not

appreciate this capability yet, but the time will come! Please be patient until
the end of the quarter.

SPSS/PC+ allows you to do many other beautiful things such as

discriminant analysis, one-way analysis of variance, multivariate analysis of
variance, factor analysis, cluster analysis, and various non-parametric
statistics. However, by now you ought to have seen enough to make an
informed judgment.

What's your verdict? I think you will agree that SPSS/PC+ is as

beautiful as WordPerfect and various other software programs. At least it is
more beautiful than a bouquet of roses or a beautiful waterfall. Now, how
does it compare to that special person?
 9
Introduction

Today, the data analysis involved in a marketing research project of

any real size is almost universally done by computer. Most statistical
procedures involve a great number of arithmetic calculations which are not
really "difficult" to do by hand, but require a tremendous amount of
repetitious work. Not only is this work boring and time consuming, but it
also provides a great deal of opportunities for little mistakes which can
seriously distort your actual results.

Today, many statistical software packages are available to help the

researcher avoid the repetitious work involved in number crunching. Not
only does such software allow us to save time, but the programs will also
allow one to do analyses which simply wouldn't have been practical to
perform in past years. While a researcher normally can't afford to literally
spend four weeks doing the calculations involved in a regression analysis of
two hundred subjects with, say, ten independent variables, the computer
will do this analysis for him or her in seconds--that is, once the data set has
been entered, or typed, into the computer.

Although the micro computer version of the Statistical Package for

the Social Sciences (SPSS/PC+) is one of the leading statistical software
packages on the market, it is by no means the only useful one available. The
group of other powerful statistical packages includes the mainframe version
(SPSS-X and SPSS v. 9.0) and such programs as the Statistical Analysis
System (SAS--available in both micro and mainframe versions) and the
Biomedical Statistical Package (BMDP). Generally, you will find that the
skills you learn while using SPSS can easily be transferred to these other
software with minor only modifications. Other programs, such as Minitab,
are slightly easier to learn but not nearly as powerful and flexible.

SPSS has many features, of which you will probably only be using a
few. It is important not to lose sight of the forest for the trees (or, in more
modern terms, not to lose sight of the computer for the chips). This manual
contains descriptions of many simple procedures to run, with more
information being available in various manuals put out by SPSS, Inc. and
 10
third party sources. (If you are unsure about particular statistical
procedures, these manuals also function very well as statistical texts since
they contain very good, real life illustrations of the statistical techniques
discussed). The attempt of this book is not to teach you all the details of
SPSS/PC+, but simply to allow you to adapt sample programs to your needs.

Please don't be intimidated by the reference to "programming" within

SPSS/PC+. All this involves is putting together a few instructions for the
computer telling it information about your data and how you would like to
analyze it. As a reasonably "user friendly" program, SPSS/PC+ accepts
relatively "English-like" commands that make a great deal of sense even
those people who don't spend most of their lives reading computer books
and magazines. Questions that SPSS/PC+ would like to have answered
relate to issues such as:

• How many variables are in the data?

• What do the different values mean?
• What happens if a person didn't answer a question?

Considering the amount of time SPSS/PC+ saves us in doing the

calculations, I think it is fair enough that SPSS/PC+ expects to get that kind
of information from us as a sort of "retainer." (In any event, it is not worth
the effort to try to bargain with SPSS/PC+ since we need it more than it
needs us).

As we go through the writing of a program for a hypothetical

questionnaire, you will be able to modify the statements of information for
that program to write one that fits your data.
 11
Step 1: Coding Your Data

Assuming that you have already collected your data, the statistical
analysis generally involves five stages:

1. Data coding, variable naming, and classification;

2. Statistical program writing;
3. Data Entry;
4. Error checking; and
5. Data analysis

We will go through each stage separately. In this chapter, we will

discuss data coding.

In Figure 1-1, you will find a sample questionnaire, commissioned by

a poodle breeder concerned about possible increasing prejudice against
small dogs, for which we will code and write an SPSS/PC+ program. Please
understand that this is not supposed to be an example of a good
questionnaire. As a matter of fact, if you plan to use one like it in your
marketing class, you should probably be prepared for a relatively low grade.
The purpose of the questionnaire, instead, is to demonstrate how to code a
number of different questions. Later in this book, we will get to the touching
story of the poodle breeder who has hired you to do a study of, among other
things, prejudice against poodles.
 12

DOG PREFERENCE QUESTIONNAIRE

We are interested in your opinion about dogs and dog care. Please take a few minutes to
respond to the questions below.

1. Do you own a dog? ___Yes ___No ___Not sure

(If no, please skip to question 6)

2. How many dogs do you own? _______

3. What is the breed of your favorite dog? _____

4. Please rank the following foods in the order you prefer to feed them to your dog:

___ generic dry dog food ___ Lucky Dog food

___ generic canned dog food
___ Kit 'n' Caboodle ___ Mighty Dog
5. How much do you spend on feeding your dog per week? $_______

6. Please write next to each of the questions below the number from the following
scale which most closely matches your level of agreement or disagreement:

1 2 3 4 5 6 7

Strongly Strongly
agree disagree

___ 1. Poodles are fragile

___ 2. Poodles are stupid
___ 3. Poodles are self-centered
___ 4. Poodles are cute
___ 5. Poodles are over-priced

7. Age:____ Sex (Please circle): Male Female

Annual household income: $_____________

1Figure 1: The Poodle Breeder's Questionnaire

 13

Whenever you plan to enter the contents of several questionnaires

into the computer, it is always a good habit to number each questionnaire.
Let's do that in the top right corner.

The need to number the questionnaires introduces an important data

coding concept. Since we will ultimately be entering the data as numbers on
one or more lines of text, we will want to determine the maximum number
of digits that a variable may take up. Suppose we administered three
hundred questionnaires and numbered them, starting with one. In order to
be able to express all these ID numbers, we would need a minimum of three
digits for the ID field (since the numbers between 100 and 300 each take up
three digits. For a questionnaire administered to less than one hundred
people, we would only need two digits for the ID.) If we assume that we
have three hundred surveys, we will eventually enter the ID number this
way:

001
002
003
..
-------> more data here <--------
..
300

It is, of course, acceptable to allocate more digits than needed to a

variable.

Now we get to actually code the questions. This includes giving the
question (or variable) a one word name, assigning a number to each possible
answer on the questionnaire4, determining how many digits are needed for
the question, determining in which columns the data will be put and,

4
SPSS/PC+ actually allows the use of alphanumeric characters, that is,
letters of the alphabet, as data. However, the use of alphanumeric data
will often cause problems which are difficult to solve and it may be a
poor practice since many other statistical programs will not allow such
data.
 14
optionally, assigning the question and answers "labels," i.e. short phrases
describing their meaning. Although this process may sound overwhelming,
it is quite simple once we go through it.

The name of the variable can simply refer to the number of the
question (e.g. "Q1," "ITEM18"), or it can be descriptive of the meaning of the
variable (e.g. "AGE," "INCOME.") The rules for naming SPSS/PC+ variables
are very similar to those for naming DOS files, i.e

• A variable must begin with a letter of the alphabet (unlike DOS, a

number is not acceptable) and may be up to eight characters long
• there may be no spaces in the name (thus, we say "ITEM1" and not
"ITEM 1").
• A variable name may not be the same as certain "reserved" words, i.e.
words that SPSS/PC+ uses for internal and programming purposes.
These reserved words are very few and far between, but you cannot
name your variables "AND," "OR," or "IF" as well as a few other
obscure names.
• SPSS/PC+ does not distinguish between upper and lower case letters.
"QUEST1," "Quest1," and "quest1" all refer to the same variable and
can be used interchangeably.

• Unlike DOS file names, a variable name cannot have an extension.

The following variable names are valid:

Q1
QUEST1
QUEST1A
QUESTA1

However, the following names are not acceptable:

1st Starts with a number

Quest 1 Contains a space
Marketshare Consists of more than eight characters
1st question Violates all of the above
 15

Let's look at question number 1. Our first task is to name it. Either
you can call it something like Q1, to keep it simple, or you can name it
something more descriptive like "DOGOWN." When you write your own
program, you have a choice; for now, we will call it DOGOWN. Next, notice
that there are three possible answers. (Your client insisted on including the
"Not sure" option since the questionnaire would be administered in the
neighborhood of a major university, making it likely that a number of absent
minded professors would be asked to respond.) We now have to assign a
number to each. Let's assign a "1" to "Yes," a "2" to "No," and a "3" to "Not
sure." Now, are these response categories enough?

Not really. Two things could happen. First, the respondent might
accidentally overlook or refuse to answer the question (a common situation
when you ask about such emotionally charged and private topics as income
or extra marital activities). The next several questions illustrate the
situation that occurs when not everybody is supposed to answer a question,
and we will discuss how to code such instances when we talk about those
questions. For this question, missing data can only arise when a person
either omits the question or provides one that is not useful. This could
happen if someone wrote a sarcastic comment instead of answering or
simply overlooked the question. Whenever a person fails to answer a
question that he or she should have answered, we will assign a numeric value.
For this question we will code it as a "9." When the question is not
applicable, just leave the space blank and the computer will assign the
response as a "system missing value." (Notice that using "9" as the missing
value for this question will result in all the numbers in between four and
eight, inclusive, not being used).

The next question, which we will call DOGCOUNT, asks how many
dogs the respondent owns. We will assume that the respondents are
reasonably normal and do not own more than 98 dogs; hence, we will
reserve only two digits for that variable. Note that here we may encounter
the kind of missing data that occurs when people legitimately omit
questions since the questions between sections two and five should only be
answered by those people who own dogs. When people "legitimately" skip
questions, we will put blank spaces in the columns designated for the
 16
variable. For those people who indicated that they own at least one dog by
answering "Yes" to the first question but failed to answer this question, we
will put in the missing value of "99."

The third question, which we will call FAVORITE, is somewhat more

complicated. This is what is called an "open ended" question; in other
words, the subject is asked to write an answer and is not given a list of
"acceptable" answers from which to choose. Therefore, we have to try to
match each written answer with some kind of code that is general enough to
be meaningful. Since more than nine breeds are likely to be mentioned, we
will reserve two digits for this question. Notice that in order to occupy two
columns, the numbers zero through nine must either be preceded by a blank
or a zero. Thus, "01" might be "Poodle;" 02, "Fox Terrier;" 03, "Yorkshire
Terrier;" and so forth. Other options such as "10" as "German Shepherd" are
possible. (Several other breeds are listed in the program that we will get to).
Again, a correctly omitted answer will result in two blank spaces and other
cases of missing data will result in a code of "99."

Note that open ended questions have a great potential for missing
data. Suppose that someone misunderstood the question and thought he or
she was asked about his or her favorite pet. If he or she answered "Polar
bear," you would most likely classify the answer as missing. You should be
prepared for certain other kinds of "missing answers." Perhaps a respondent
unsympathetic to the objectives of our research might scribble in something
like "I hate all dogs!"

To keep organized, we might consider giving very small dogs low

numbers, say, below 29; medium sized dogs intermediate numbers between
30 and 59, and big dogs the bigger numbers between 60 and 98. Depending
on our research objective, we might also like to "cluster" the numbers on the
basis of other variables such as price, friendliness/viciousness, durability,
guarding ability, or lifespan. Or, if we cared about none of those issues, we
could simply list the dogs in alphabetical order.

The next questions require very little discussion since subjects will be
responding directly with numbers. Thus, for section 4, all we have to do is to
assign one digit to each variable (i.e. each dog food that we ask the
 17
respondent to rate) and reserve "9" as a missing value. Skipping slightly
ahead, the same holds for the Likert scale questions of Section 6.

In section 5, we will assume that no one spends more than $99.98 per
week on dog costs, and we will thus assign five digits and a missing value of
"99.99." (SPSS/PC+ allows you to designate a variable as a dollar amount
rather than as plain number. However, not all programs have that
capability, so let's not get into that.) Also notice that we have are reserving
space for the period. We could arrange to use four digits instead, but why be
so stingy with the space?

For age, we will assign two digits and a missing value of "99." For sex,
"male" will be assigned "1" and "female" "2."

For the question of annual household income, we will assume that the
figure does not exceed 999,9985 for any respondent since we are located in a
university community. We will thus reserve six digits since SPSS/PC+
would not appreciate the comma (unlike the period allowed in the question
on weekly spending). Incidentally, anyone earning over a million dollars a
year would, in statistical jargon, probably be considered an "outlier"--a sort
of maverick who would probably be excluded from our analysis anyway.

The table below summarizes the lengths, missing value indicators,

and column positions of each variable. A "case," or collection of answers
from one respondent (who owns a dog), could thus look like this:
001 1 02 02 52314 03.50 66727 29 2 026000

while a case from a person who does not own a dog, and thus was not asked
to answer some of the questions, could look like this:

002 2 22252 31 1 032150

5
Remember, we need $999,999 (i.e. 999999) for missing data.
 18

Variable Missing Length Columns Label

Value
________________________________________________________________________________
ID None 3 1- 3 Identification number
OWNDOG 9 1 5 Ownership of dog
DOGCOUNT 99 2 7- 8 Number of dogs owned
FAVORITE 99 2 10-11 Breed of favorite dog
FOOD1 9 1 13 Ranking of generic dry
FOOD2 9 1 14 Ranking of generic canned
FOOD3 9 1 15 Ranking of Mighty Dog
FOOD4 9 1 16 Ranking of Lucky Dog
FOOD5 9 1 17 Ranking of Kit 'n'Caboodle
SPEND 99.99 5 19-23 Weekly expense of dog food
LIKERT1 9 1 25 Poodles are fragile
LIKERT2 9 1 26 Poodles are stupid
LIKERT3 9 1 27 Poodles are self-centered
LIKERT4 9 1 28 Poodles are cute
LIKERT5 9 1 29 Poodles are over-priced
AGE 99 2 31-32 Age
SEX 9 1 34 Sex
INCOME 999999 6 37-42 Annual household income

A look at the column assignments and the two example cases

will show that we have put spaces between some variables and not
between others. SPSS/PC+ leaves it entirely up to us whether we
want to leave such spaces or not. One of the first commands of the
program simply tells it the column positions of each variable. Thus,
we can to put spaces where they improve our ability to type in and read the
data.

For each case, a variable will be in the same column position(s).

This can be a great help to ensure that we are on target when we type
in the data. Consider, for example, the following two cases:

003 1 02 02 52314 03.50 66727 29 2 026000

004 1 01 03 32512 05.50 67615 21 2 018500

Not only do we know that each case should end in the same
column; we also know that many of the blank spaces should be in the
same places. In this case, we have reserved blank spaces between
most of the variables, but none within the variables in the Likert scale
and rank-order sections. Putting in spaces there would make the data
seem more confusing.
 19

You should try to be as consistent as possible when coding

different questions within the same questionnaire. If for one question
you use the code "1" for yes and "2" for no, you should try to keep that
practice throughout the questionnaire.

When several people enter data, be sure that everyone

understands and agrees upon a coding system. There are numerous
ways the same variable can be coded. The variable SEX above, for
example, could have the codes "1" for male and "2" for female (as we
did); "1" for female and "2" for male; and "0" for female and "1" for
male. None is inherently better or more correct than the other
(although the first one is the most commonly used approach), but if
several individuals use different codes in the same file, a big editing
job could result.

Sometimes, the best way to code a variable does not coincide

with the coding that may be "suggested" by the questionnaire.
Consider this example:

How many persons over the age of 55 currently live in your

household?
__ a. None
__ b. One
__ c. Two
__ d. Three
__ e. Four or more

In this case, it might be tempting to start off at the beginning

and number the choices from 1 through 5. This, however, would be a
bad approach, since the code "2" would now correspond to the answer
"one person." Instead, you would be better off starting with zero and
advancing. That way, the code will correspond to the number of
 20
people actually in the household.
 21
Step 2: Writing An SPSS/PC+ Program

Now we are ready to pursue the bottom line of this text, that is,
the writing of the SPSS/PC+ program. The next chapter will discuss
how to enter the data and commands into SPSS/PC+; here, we will
just talk about what to enter. Here is a program for the questionnaire
we have been discussing. Please don't be intimidated if it looks
overwhelming at first. We will go through it line by line.
TITLE "Dog Preference Study".
DATA LIST /id 1 -3 owndog 5 dogcount 7 -8
favorite 10-11 food1 to food5 13-17
spend 19-23 (2) likert1 to likert5 25-29 age 31-32 sex 34
income 37-42.
VARIABLE LABELS
owndog "Ownership of dog"
dogcount "Number of dogs owned"
favorite "Breed of favorite dog"
food1 "Rating of generic dry dog food"
food2 "Rating of generic canned dog food"
food3 "Rating of Mighty Dog"
food4 "Rating of Lucky Dog"
food5 "Rating of Kit 'n' Caboodle"
spend "Weekly spending on dog food"
likert1 "Poodles are fragile"
likert2 "Poodles are stupid"
likert3 "Poodles are self-centered"
likert4 "Poodles are cute"
likert5 "Poodles are over -priced"
income "Annual household income".
VALUE LABELS
owndog 1 "Yes" 2 "No" 3 "Not sure"/
favorite 1 "Poodle" 2 "Fox Terrier" 3 "Yorkshire Terrier"
4 "Daschund"
10 "German Shepherd" 11 "Collie" 12 "Saint Bernard"
13 "Pit Bull" 14 "Malamute" 15 "Afghan" 16 "Cocker Spaniel"
17 "Dobermand" 18 "Golden Retriever" 19 "Rotweiler"/
food1 to food5 1 "Generic dry dog food" 2 "Generic dry cat food"
3 "Generic canned dog food" 4 "Kit 'n' Caboodle"
5 "Mighty Dog"/
sex 1 "Male" 2 "Female"/
likert1 to likert5 1 "Strongly agree" 7 "Strongly disagree".
MISSING VALUE owndog likert1 to likert5 (9)/
dogcount favorite age (99)/spend (99.99)/income 999999.
BEGIN DATA.
001 1 02 02 52314 03.50 66727 29 2 026000
002 2 22252 31 1 032150
003 2 45445 26 2 135000
004 1 01 03 32512 05.50 67615 21 2 018500

---> MORE DATA HERE <-----

999 2 77627 45 2 053000

END DATA.
FREQUENCIES VARIABLES=owndog to income/STATISTICS=all.
 22
Let's take a look at what the lines in this program look like. The
program consists of several commands which are intended to instruct
the computer about the data and what to do with it.

First, note that some lines are indented while others are not. In
general, indented lines are continuations of commands were that
started immediately at the left margin on some line above it.
SPSS/PC+ really doesn't care if you indent or not, but it will make
your program more readable. As you can see, each command
eventually ends with a period, which tells the computer to take in the
next line as a new command. If you leave out the period, the
computer will not understand your commands and will give you an
error message. Fortunately, such errors are easy to detect and correct,
so if you leave out a few periods, it only means that you will have to
do some editing after you first try to run the program. Even
experienced SPSS/PC+ users often have problems in their first
attempts at any program, but the more experience you get with the
program, the easier it gets to correct the problems.

Now, let's start from the beginning.

TITLE "Dog Preference Study".

This first command is actually optional. Eventually, SPSS/PC+

will give pages of output consisting of statistical computations, and
putting in the "title" command will put some heading that you choose
on top of every page. In terms of syntax, the word "title" is followed
by the title you wish to assign in quotes. Notice that the period,
contrary to what your junior high school English teacher taught you,
goes outside the quotes.

On The Use of Capitals

Unlike an English teacher, SPSS/PC+ doesn't

care if you use upper or lower case letters. Thus, you
can mix and match as you find it convenient.
 23

DATA LIST /id 1-3 owndog 5 dogcount 7-8

favorite 10-11 food1 to food5 13-17
spend 19-23 (2) likert1 to likert5 25-29 age 31-32 sex 34

income 37-42 .

This command is the most crucial part of the program. It may

also be the one that can cause you the most frustration.

The "data list" command tells the computer about the positions
of the variables on the data lines. If you have made a table detailing
this information, you already have all the information needed.
Otherwise, you will have to do some arithmetic now to calculate the
beginning and ending columns of each variable.

After typing in the command "data list," we will type in a slash

to indicate that we are starting the definition of a line of data. The
command will not be recognized if you leave this slash out. After
that, we will list each variable followed by the column(s) it represents.
Thus, we state that ID covers the columns one through three:
DATA LIST /id 1-3...

The variable OWNDOG only takes up one column, so no dash is

needed to indicate a range.

For a long questionnaire, we don't quite have to type in the

information for each variable individually. Notice that we are able to
consolidate the information for the Likert scale section into one
declaration:
likert1 to likert5 25-29

Since each variable takes up the same number of columns (in

this case one), SPSS/PC+ will recognize that the above declaration is
identical to the following more elaborate statement:
likert1 25 likert2 26 likert3 27 likert4 28 likert5 29
 24
This statement only works because each variable takes up the same
number of columns. If some of the listed variables took up two columns
and some just one, SPSS/PC+ would not have enough information to
determine how much each one got.

We have alluded to the fact that SPSS/PC+ knows that

item1 to item6

is the same as

item1 item2 item3 item4 item5 item6 .

However, SPSS/PC+ does not allow you to create five variables

in the variable list by saying

item1a to item1e .

You would thus have to type in:

item1a item1b item1c item1d item1e .

 25

Technical Note

Once you have defined a range of variables

such as

item1a item1b item1c item1d item1e

you can refer to the variables as

item1a to item1e

in subsequent procedures since SPSS/PC+ now

knows the sequence in which they occur in the
program.

This is no different from the fact that, in

ensuing procedures, you will, using the example of
our questionnaire, be able to refer to

owndog to likert5

which would include the variables OWNDOG,

DOGCOUNT, FOOD1 to FOOD5, SPEND, and
LIKERT1 to LIKERT5.
26
 27

Handling Long Questionnaires

SPSS/PC+ only allows each line of text to be

eighty characters long. That means that when all
information from one questionnaire can't fit into eighty
columns, you have to use more than one line for each
case.

To tell SPSS/PC+ that you are continuing a

case on a new line, put a slash (/) before the first
variable to go on the following line. If you have the ID1
number and ITEM1 to ITEM50 on the first line and ID2
and ITEM51 to ITEM100 on the second line, your data
list might look like this:

data list /id1 1-3 item1 to item50 5-54

/id2 1-3 item51 to item100 5-54.

If you have a very long questionnaire requiring

more than three lines of text, see Appendix C for some
strategies.
 28
VARIABLE LABELS
owndog "Ownership of dog"
dogcount "Number of dogs owned"
favorite "Breed of favorite dog"
food1 "Rating of generic dry dog food"
food2 "Rating of generic canned dog food"
food3 "Rating of Mighty Dog"
food4 "Rating of Lucky Dog"
food5 "Rating of Kit 'n' Caboodle"
spend "Weekly spending on dog food"

The "variable labels" command is rather long, so only the first

few lines appear above. Here, we are assigning a "label," or
descriptive phrase, to each variable or question. Note that the
variable name is first listed and its "label" is then put in quotes. Each
variable label may be up to forty characters long. The main purpose
of the variable label is to remind you or the reader of your report what
the question is about. The computer really doesn't care what label
you give a variable but just feeds it back to you on certain statistical
procedures. Be sure to put in the period at the end of the last label!
(Note that there is no period at the end of each individual label).
VALUE LABELS
owndog 1 "Yes" 2 "No" 9 "Not sure"/
favorite 1 "Poodle" 2 "Fox Terrier" 3 "Yorkshire Terrier"
4 " Daschund"
10 "German Shepherd" 11 "Collie" 12 "Saint Bernard"
13 "Pit Bull" 14 "Malamute" 15 "Afghan" 16 "Cocker Spaniel"
17 " Dobermand" 18 "Golden Retriever" 19 " Rotweiler"/
food1 to food5 1 "Generic dry dog food" 2 "Generic dry cat food"
3 "Generic canned dog food" 4 "Kit 'n' Caboodle"
5 "Mighty Dog"/
sex 1 "Male" 2 "Female"/
likert1 to likert5 1 "Strongly agree" 7 "Strongly disagree".

The value labels command assigns names or labels to the

values, or "answers," associated with each variable. First, the variable
name is stated. Then each value is listed, followed by its description
in quotes. Note that there is a slash at the end of each question. Also
note that if a set of labels applies to more than one question, they can
be listed simultaneously. For example, if "1" means yes and "2" means
no for Q1 to Q17 and Q28, you could say
VALUE LABELS q1 to q17 q28 1 'yes' 2 'no'.

Note that some questions don't have value labels. It is not

meaningful, for example, to assign labels to the values associated with
 29
AGE and INCOME since these are self-explanatory. This is the case
for most interval and ratio scaled variables; the quantity expressed
usually carries its own meaning. (In some cases, you will probably
want to express the unit of measurement, such as pounds, inches,
years, or dollars, in the variable label).

Both the "variable labels" and "value labels" commands are

optional and are available solely for your convenience. Putting them
in will, however, tend to improve the readability of your output.

MISSING VALUE owndog likert1 to likert5 (9)/

dogcount favorite age (99)/spend (99.99)/income 999999.

We discussed the meaning of missing values in the chapter on

coding. Note that SPSS/PC+ will automatically interpret blanks as
missing values; thus you only put in those missing values you have
defined as referring to a non-legitimate missing answer. From the
above, you can see that the syntax is the variable name(s) followed by
the missing value in parentheses, separating each range by a slash.
BEGIN DATA.
END DATA.

These two commands, both followed by a period, tell the

computer that the data will now begin and stop, respectively. In
between, you can then type in the data as it is defined in the data list.
 30

Sending Output
To the Printer
By default, SPSS/PC+ displays the statistical output on the
screen. If, instead, you would like to send the output to the printer,
simply put the following two lines in your program:

set more off.

set printer on.

Whatever SPSS/PC+ produces after these commands are

encountered in the program will be sent to the printer. Thus, if you
don't want to send your whole program to the printer, you can put the
commands immediately before the statistical procedures.

The "set more off" command frees you from having to press
<RETURN> or <SPACE> at the end of each screenful of data. This
means that the output may "pass you by" before you have a chance to
read it. If this happens, you may either want to leave out this
command or wait to read the output until you have it printed on paper.

Some printers, particularly laser printers and other printers that

user single sheets instead of continuous "tractor" fed paper, cause
sometimes cause particular problems. See Appendix M for more
details.
 31

When typing in the data, be sure to check that you are "on
target" with respect to the columns. Generally, all the lines should be
equally long. Also, be sure to check that, when you have typed in a
complete entry, you are one column farther out than the last position
listed in the data list.
FREQUENCIES VARIABLES=all/STATISTICS=all.

The "frequencies" command is an example of a statistical

procedure or command--SPSS/PC+'s raison d'être. Once we have
looked through the data for errors, we will go over other statistical
commands, which normally go here in the program, after any
recoding and computations, which we will also discuss.
 32
Step 3: Entering The Program And Data

There are several ways you can enter the program and data
into SPSS/PC+. Since SPSS/PC+ uses an ASCII, ("plain text") file to
handle the input data, you can either use REVIEW, the editor supplied
with SPSS/PC+, or a word processor such as WordStar, WordPerfect,6
or Microsoft Word. When using a word processor, be sure to set the
margins so that the lines can be long enough. The default margins for
most word processors will normally allow only about sixty-five
characters on a line.7 Be sure to save the file as an ASCII file, i.e. not
in word processing format.

LOTUS 123 provides a nice way of entering data for small

questionnaires. See Appendix E for details.

To use REVIEW, start SPSS/PC+. How you will do this may

depend on the setting of the computer you will be using. In some
cases, there will be a "menu" on the computer, and all you have to do
is to enter the number or letter that corresponds to SPSS/PC+. On
other computers, you may have to start from the DOS prompt. If
SPSS/PC+ is found on the "\SPSS" directory of the hard drive, you
might type in the underlined part of the following:
C:\>CD \SPSS
C:\SPSS>SPSSPC

If neither of these approaches work, you will have to find

"local" instructions for what do to. Fortunately, most of the rest of this
manual's approaches will be more universally applicable.

6
Use the text-in/text-out feature (<CONTROL> <F5>) to create an
ASCII file.
7Also notice that some word processors set margins in terms of length
rather than characters. In the newer versions of WordPerfect or Microsoft
Word, margins are by default set in terms of inches. In such programs,
you may wish to switch to a smaller font instead of adjusting the margins.
 33
Once you are in SPSS/PC+, a logo will first flash and you will
then be presented with a menu. Press <ALT>-M, then <F3>
<RETURN>. Now specify the name you want to give the file8 that
will contain your program and data and press <RETURN>. For
example, if you were entering the questionnaire about dogs, you
might call it "A:POODLE.SPS." (It is traditional to use the file name
extension ".SPS," but if you like to be different, this convention is not
required).

You are now in REVIEW, the editor associated with SPSS/PC+.

To move around in the text you create, use the arrows to move up
and down and in the right and left directions. When you are ready to
save, move to the top of the file (press <CONTROL>-<HOME>), press
<F9>, and press <RETURN> twice.

If your file already exists, you will probably be brought into the
document at the bottom. To go to the top, press <CONTROL>-
<HOME>. (As you might expect, <CONTROL>-<END> will bring you
to the bottom of the document). You can use the arrows on the
keyboard to move one space or line at a time. The <INS> key will
toggle between insert and write-over (that is, whether the computer
will type new text on top of existing text or move the old text over to
make room for the new). (The default is "on"; however, you may want
to turn it off if you are editing data and you want to overwrite some
incorrect contents).

One tricky situation involves the insertion of a new line on the

very top of the file. (Suppose you want to insert "SET PRINTER ON"
above the DATA LIST statement.) To do this, go to the top and
inserting a blank space at the beginning of the line. You can now
press <RETURN> and the blank line will be inserted at the top of the
file.

For some reason, REVIEW will occasionally save only what is

8If your floppy disk is in Drive A:, the filename should start with "A:,"
e.g. "A:poodle.sps" in our case.
 34
below the cursor. Therefore, you should always be sure to go to the top of
the file before saving it. The complete sequence to save, including this
first step, is:
<CONTROL>-HOME>
<F9> <RETURN> filename <RETURN>

where "filename" represents a new filename and specification you

may optionally give the file. If you want to keep the "suggested"
filename, just press <RETURN>.

Remember that SPSS/PC+ does not automatically save your

data. Therefore, it is not a bad idea to save every fifteen to twenty
minutes to guard yourself against a power surge or other interruption
which might destroy your data. Also, it is a good practice to have at
least two diskettes with the one being used as a backup.

You may not have time to type in all of your program and your
entire set of data in one sitting. You can leave REVIEW at any time
and resume at a later point.
 35

Essential REVIEW
Commands
SAVE FILE
<CONTROL>-<HOME>
<F9> <RETURN> <RETURN>
GO TO THE TOP OF THE FILE
<CONTROL>-<HOME>
MOVE UP OR DOWN ONE LINE OR MOVE LEFT OR RIGHT
Use the cursor keys
INSERT A LINE ABOVE THE CURRENT LINE
<SPACE> <RETURN>
EXIT FROM SPSS/PC+
(Be sure to save first). <F10> E EXIT. <RETURN>

Exercise

You are now ready to enter data into SPSS/PC+ and analyze it.
Your first exercise is simply to type in the following program and run
it. From the menu, call up SPSS/PC+, then press <ALT>-M <F3>
<RETURN>, followed by the filename "A:exerc1.sps" (make sure you
have a formatted floppy disk in Drive A:), and press <RETURN>
again.

Now type in the following program, substituting only your

name in the "TITLE" line. Make sure that you get the punctuation
right--SPSS/PC+ is even pickier than your junior high school English
teacher!
set printer on.
 36
set screen off.
title "YOUR NAME".
data list /id 1 -3 age 5 -6 educ 8-9 sex 11 income 13-15
vacation 17-20.
variable labels
age 'age of respondent'
educ 'years of formal education completed'
sex 'sex of respondent'
income 'annual household income in hundreds of dollars'
vacation 'amount spent by household on vacations last year'.
value labels
sex 1 'male' 2 'female'.
missing value
age,educ(99)/sex(9)/income(999)/vacation(9999).
begin data.
001 52 13 1 232 3270
2 50 13 1 235 4279
3 57 12 1 274 3643
4 30 10 1 217 3504
5 74 17 1 379 4247
6 63 15 1 353 3715
7 53 15 2 277 4059
8 57 16 1 294 4154
9 35 10 2 266 2908
10 38 11 2 234 3602
11 29 14 2 243 3492
12 24 11 1 185 3391
13 61 15 2 316 3637
14 47 12 2 225 3422
15 49 13 1 259 4289
end data.
frequencies variables=age to vacation/statistics=all.

You are now ready to run the program. Press <CONTROL>-

<HOME> to go to the top, <F9> <RETURN> <RETURN> to save, and
<F10> <RETURN> to run. If you have made a syntactic mistake,
SPSS/PC+ may point it out to you and you will have to fix it and
rerun.

Analysis

As is evident, this survey contains demographic information

about surveyed individuals as well as information on how much they
spent on vacations.

1. Using the FREQUENCIES output you will receive, find the

mean, median, mode, and standard deviation for each variable. Do
those seem to be representative of the population at large?
 37
Step 4: Checking Your DataAnd Program For Errors

So you thought that you were finally done with the program
and data entry? Well, not quite yet! Your data is probably good
enough to be published in a tabloid magazine as it is, but there is one
additional step that a conscientious researcher must take.

When you type in a large amount of data, there is a significant

chance that you might make a typographical error. Not all errors can
be caught and some errors won't make that much difference, but some
are relatively easy to catch and should be eliminated.
Now that you have all your data entered, make sure that there
is nothing below the "end data" line. (In our example, you would
delete the "frequencies" line).

Now press <CONTROL>-<HOME> to go to the top of the file,

press <F9> followed by <RETURN> <RETURN> to save, and <F10> to
run. Unless you have turned the printer option on, you will get one
screen at a time. When a complete screen has been displayed, you will
hear an obnoxious beep and you will be prompted with the message
"MORE." Press <RETURN> to see the next screen.

The program may point out some errors in your program. Such
errors are often caused by (1) omission of a period, quote, or slash, (2)
the misspelling of a command or variable name, or (3) other
"typographical" error. The computer will beep and stop after each
screen of information has been displayed. Note down any errors and
press <RETURN> to continue.

Once the program has come to an end, press <F3> <RETURN>,

followed by your filename and <RETURN>. You are now back in
editing mode and you can now fix any problems you have been able to
identify. Continue running the program this way until all errors have
been fixed.

Note that taking care of one error may fix several other
complaints that SPSS/PC+ had in the previous run. If, for example,
 38
you leave out a period or slash, SPSS/PC+ may encounter several
subsequent "errors"--expressions that are not allowed in the given
context. In other words, if you left out some punctuation, SPSS/PC+
may expect something that is not forthcoming and will continue to
complain.

Also note the way SPSS/PC+ chooses to describe your error

may not be very informative. The reference to "an unrecognized
expression," for example, can mean almost anything. Instead, focus
on where error occurs. Should there have been a period immediately
before? Did you misspell a command?

When you are satisfied that the errors have been removed from
the program, add the line
frequencies variables=all.

to the bottom of the file, press <CONTROL>-<HOME> to go to the

top, <F9> <RETURN> <RETURN> to save, and then <F10>
<RETURN> to run the program. After going through the data
definition part of the program, the computer will display the
frequency counts of each variable. You should now look for
"illegitimate" values for each variable. Let's take a look at the below
example:

DOGOWN Ownership of Dog

Valid Cum
Value Label Value Frequency Percent Percent Percent

Yes 1 82 46.9 49.1 49.1

No 2 81 46.3 48.5 97.6
Not sure 3 3 1.7 1.8 99.4
4 1 .6 .6 100.0
Missing 9 8 4.6 MISSING
------- ------- -------
TOTAL 175 100.0 100.0

Valid Cases 167 Missing Cases 8

In this case, it is quite evident that an error has been made since
there is no such legitimate value as "4" for this question. That is, you
either own a dog, don't own a dog, don't know if you own a dog, or
 39
refuse to answer the question. Therefore, the code "4" cannot
represent a acceptable answer. We now know that something went
wrong, and we will want to track down the error. Also note that we
really would have no way of detecting if the value of "3" had been
entered one time too many (at the expense of some other code) since
that would not show up as an illegitimate value. Be sure to note down
all unacceptable values. (In this case, there is only this one
"objectionable" value). Note that the period indicates a "system
missing value" (or a blank) and that "9" is our defined missing value to
be used when the given answer is not usable).

When you have found all the illegitimate values in the

program, run the program again, this time putting in the following
two lines at the bottom:
PROCESS IF (dogown EQ 4).
FREQUENCIES VARIABLES=id.

In the above example, you would modify the part in

parentheses to meet the condition relevant to your case. On the left
side of the "EQ" put the name of the variable that gave you an
offending value and on the right side, put the value in question. Now
the computer will select only the case that has given you the problem.
The next line will give you the case number of the problem variable.
When you run the program, you will identify the offending case and
you can make appropriate corrections.

Once you have weeded out the incorrect values, you may want
to run the frequencies check again to see if you got them all or if new
ones have come about as a result of editing.
 40
Step 5: Using Statistical ProceduresAnd Computations

Statistical commands normally go at the bottom of the file,

after any computation and recoding commands.

Frequencies

You have already been exposed to the Frequencies command,

which provides a frequency count answers to each variable specified.
The Frequencies command can give you more information, however.
By saying
FREQUENCIES VARIABLES=all/STATISTICS=all.

you will get the mean, standard deviation, median, mode, and various
other statistics associated with the distribution.
DOGS Number of dogs owned or leased

Valid Cum
Value Label Value Frequency Percent Percent Percent

.00 42 63.6 66.7 66.7

1.00 13 19.7 20.6 87.3
2.00 4 6.1 6.3 93.7
3.00 1 1.5 1.6 95.2
4.00 1 1.5 1.6 96.8
6.00 1 1.5 1.6 98.4
41.00 1 1.5 1.6 100.0
99.00 3 4.5 MISSING
------- ------- -------
TOTAL 66 100.0 100.0

Mean 1.190 Std Err .656 Median .000

Mode .000 Std Dev 5.208 Variance 27.124
Kurtosis 57.522 S E Kurt .595 Skewness 7.448
S E Skew .302 Range 41.000 Minimum .000
Maximum 41.000 Sum 75.000

Valid Cases 63 Missing Cases 3

For example, this table indicates that 42 of the people own or

lease no dogs, 13 people have one, four people have two, and three,
four, six, and forty-one dogs are possessed each by one person. We
have missing data for three people, and the mean number of dogs
owned is 1.19 (although the median is 0.00). It looks as though our
 41
average has been brought up quite a bit by the person who owns
forty-one.

Creating New Variables: Compute

Computations can be very useful. Suppose you have collected

data on how much people spend during Christmas for presents
(GIFTS), comestibles (FOOD), travel (TRAVEL), and additional
Christmas related expenses (OTHER). If the variable names are the
ones given in the parentheses, you can calculate total Christmas
expenses (EXPENSES) by
COMPUTE expenses=gifts+food+travel+other.

If for some reason you wanted to find the average of those

figures, you would say
COMPUTE avgexp=(gifts+food+travel+other)/4.

Recoding Variables

Recoding can sometimes be useful when doing crosstabs and

other nominal statistics where you would like to "collapse" the data to
make it more interpretable. In our questionnaire, we might want to
collapse the AGE variable:
RECODE age (0 THRU 25=1) (26 THRU 40=2)
(41 THRU 65=3) (66 THRU 98=4).

As is evident from the example above, you first state the name
of the variable. Each value range is then specified in parentheses,
followed by an equal sign, and the desired recoded value.

Details on specific statistical procedures are found in the

SPSS/PC+ manuals, however, the syntax for a few procedures is
contained below.
 42

Reverse Scoring
When using Likert scales and other measures of opinion or attitude, it is
sometimes desirable to word questions in the opposite direction of what one is
looking for. There are two reasons for this approach. First, wording the question
one way may be clearer or more natural than wording it the other way. Secondly, it
may be desirable to reverse the polarity of the question to prevent respondents from
simply checking the same answer for each question.

For example, when we attempted to measure prejudice against poodles in

our questionnaire, we included the item "Poodles are cute." Since strong
agreement with this question (Likert4) could be expected to signify lack of
prejudice, we should "reverse score" it before we add it to the others to compute
the scale of prejudice against poodles.

In reverse scoring, we will essentially turn the scale upside down. That is,
we will convert the highest value to the lowest, the lowest value to the highest, and
so forth. In this example, since we have a seven point scale, the command would
look like this:
RECODE likert4 (7=1) (6=2) (5=3) (3=5) (2=6) (1=7).

(Notice that "(4=4)" is superfluous. That is, if the person neither agrees
nor disagrees, that fact is not going to change with the polarity of the question).
 43
Crosstabulation

A crosstabulation allows us to explore the relationship between

two variables by tabulating one against the other. Consider this
example:

Crosstabulation: HEIGHT Height of customer

By ANIMAL Species of preferred stuffed animal

Count Giraffe Elephant Bear Cow Pig Dog

ANIMAL > Row
1 2 3 4 5 6 Total
HEIGHT
1 6 10 14 5 21 4 60
Short 33.3

2 4 10 16 5 19 6 60
Medium 33.3

3 10 6 17 5 15 7 60
Tall 33.3

Column 20 26 47 15 55 17 180
Total 11.1 14.4 26.1 8.3 30.6 9.4 100.0

Chi-Square D.F. Significance Min E.F. Cells with E.F.< 5

---------- ---- ------------ -------- ------------------

6.17035 10 .8008 5.000 None With ANIMAL

As you will note, the values of one variable are listed

horizontally, and the values of the other are listed vertically. In the
cells, you see the number of subjects falling into the "intersection" of
the two. The table shows, for example, that fifteen people are both
tall and prefer pigs. You will notice that some statistics are also
provided. (We will soon get to why anyone would care about the
relationship between these two variables).

The general syntax for crosstabs table is:

CROSSTABS TABLES=firstvar by secondv/STATISTICS=all.

where "firstvar" and "secondv" are the two variables you want to
tabulate against each other. Notice that the optional statistics take up
 44
a lot of room. If you only want Chi square ( χ2), you can reduce the
output by substituting "STATISTICS=1" for "STATISTICS=all."

Or, you could select the statistics available from this list:

Statistics Available in
Crosstabs
1 Chi square
2 Phi or Cramer's V, depending on the
number of variables
3 Contingency coefficient
4 Lambda
5 Uncertainty coefficient
6 Kendall's Tau-b
7 Kendall's Tau-c
9 Somers' d
10 Eta
11 Pearson's r

Ordinarily, you probably wouldn't want to look at anything

more than statistics numbers one and two. However, calculating the
others won't take the computer very long at all, so there is little
penalty in saying "STATISTICS=all."

If you want more detail, you can get row and column
percentages, i.e. the percentage of the row and column that each cell
contributes, by putting in the "/OPTIONS=3,4" parameter. Thus, if
you were tabulating "firstvar" with "secondv" and wanted these
features as well as Chi square and Cramer's V, the command would
be:
CROSSTABS TABLES=firstvar BY secondvar
/STATISTICS=1,2
/OPTIONS=3,4.
 45

Note that the period goes at the very end of all the
subcommands. You would not place a period after "firstvar BY
secondvar" if you included additional subcommands as we did in this
case.

If you want to create multiple tables with one command, you

can specify a variable list on each side of the "BY" part of the
command. For example, you could say:
CROSSTABS TABLES=var1 to var5 by var6 to var10/
STATISTICS=1.

Notice, however, that this would create 5x5=25 tables! It is

easy to write a statement, without realizing it, that would create
hundreds of tables. This would most likely provide you with severe
information overload and would make any sort of meaningful
interpretation impossible. For example, if you have thirty questions
and you want to see how "everything relates to everything," you
might think about saying "tables=var1 to var30 by var1 to var30."
However, this would create 420 non-redundant tables! You would be
sitting by the printer for a long time and would probably not have
time to interpret all of them.
To find out in advance how many tables you would get by
trying a list of variables against the same list, use the formula

T = (n2-n)/2

where T is the resultant number of tables and n is the number of items

in the list.

To find out how many tables would result from running two
different lists against each other, multiply the number of variables in
each list by each other. For example,

CROSSTABS TABLES=var1 to var10 by var15 to var20

would result in (10*6)=60 tables.

 46
Let's suppose that you have been hired to do a marketing study
for a manufacturer of stuffed animals. The manufacturer wants to
start a poster media campaign to promote his stuffed animals to the
public. A media consultant (of questionable reputation) that he has
hired believes that advertising will be most effective if it is placed at
eye level. In order to enable the manufacturer to target customers of
different height, the manufacturer has asked you to find out whether
there is a relationship between a person's height and his or her
preferred stuffed animal species. You collect data and run the
following crosstabulation:
Crosstabulation: HEIGHT Height of customer
By ANIMAL Species of preferred stuffed animal

Count Giraffe Elephant Bear Cow Pig Dog

ANIMAL > Row
1 2 3 4 5 6 Total
HEIGHT
1 6 10 14 5 21 4 60
Short 33.3

2 4 10 16 5 19 6 60
Medium 33.3

3 10 6 17 5 15 7 60
Tall 33.3

Column 20 26 47 15 55 17 180
Total 11.1 14.4 26.1 8.3 30.6 9.4 100.0

Chi-Square D.F. Significance Min E.F. Cells with E.F.< 5

---------- ---- ------------ -------- ------------------

6.17035 10 .8008 5.000 None With ANIMAL

Statistic Symmetric Dependent Dependent

--------- --------- ------------- -------------

Lambda .04898 .08333 .01600

Uncertainty Coefficient .01260 .01585 .01045
Somers' D -.02035 -.01877 -.02222
Eta .15250 .05750

Statistic Value Significance

--------- ----- ------------

Cramer's V .13092
Contingency Coefficient .18205
Kendall's Tau B -.02042 .3730
Kendall's Tau C -.02222 .3730
Pearson's R -.03093 .3401
Gamma -.02806

Number of Missing Observations = 0

 47

Looking at the intersection of the two variables in the table, we

can see six people are both short and prefer giraffes, 16 medium sized
people prefer bears, etc. That is a large amount of information not
easily interpretable without any kind of statistical summary.

Is there a relationship between the two variables? It is difficult

to say just from looking at the table. However, the Chi square ( χ2)
statistic will test the null hypothesis that the two variables are
"independent," i.e. that knowing information about the one does not
tell us anything about the likelihood of the other. Normally, we
require the significance level to be less than 0.05. Since we did not
come anywhere near that this time around, we conclude that there is
not enough evidence to support the height hypothesis, and we
recommend to the manufacturer that he find another method of
segmenting his advertising.

Product Moment (Pearson) Correlation

The general syntax for Pearson Correlation in SPSS/PC+ is:

CORRELATION varlist WITH varlist

/OPTIONS=options .

In practical terms, if we wanted to generate the below matrix of

the (boring) variables X60 to X62 correlated against X63 to x65, we
would write:
CORRELATION x60 to x62 WITH x63 to x65/OPTIONS=5.
 48
Correlations: X63 X64 X65

X60 .0267 -.0900 .0448 (<---coefficient)

( 72) ( 72) ( 72) (<---no. of cases)
P= .412 P= .226 P= .354 (<---significance)9

X61 .0350 .1456 -.0413

( 72) ( 72) ( 72)
P= .385 P= .111 P= .365

X62 .1223 -.2251 .1382

( 72) ( 72) ( 72)
P= .153 P= .029 P= .123

(Coefficient / (Cases) / 1 -tailed Significance)

" . " is printed if a coefficient cannot be computed

The "/OPTIONS=5" statement provides for the printing of the

number of cases and the one-tailed significance levels associated with
the respective correlation coefficient.

If you are not sure whether to expect a positive or negative

relationship between the two variables, you may want a two-tailed
probability associated with the correlation coefficient. You would
then add option 3, making the above statement:
CORRELATION x60 to x62 WITH x63 to x65
/OPTIONS=3,5.

If you want to create a square matrix of variables correlated

against each other, you can simply say
CORRELATION varlist / OPTIONS=options.

Thus,
CORRELATION x60 to x65/OPTIONS=5.

would create a matrix of x60 to x65 against each other. Notice that
close to half of the correlations would be redundant (i.e. X62 against

9
Note that, unless you change the options, this significance level is
one-tailed.
 49
X63 is the same as X63 against X62) and the coefficients of the
correlations on the diagonal would be all ones since we are correlating
each variable with itself. (That is, a variable perfectly "predicts"
itself).

Multiple Linear Regression

There are three main methods of doing linear regression

analysis within SPSS/PC+10: forward inclusion, backward
elimination, and stepwise. To run an analysis, you must make a
decision as to which one to use. Theoretical and philosophical
considerations will enter your decision as to which method to select,
but if you wanted to keep things simple, you would most likely use
forward inclusion.

A regression analysis contains exactly one dependent variable

and one or more independent variables. The syntax for the
command is:
REGRESSION VARIABLES=varlist/DEPENDENT=depvar
/METHOD=method.

Thus, to do a forward inclusion regression analysis trying to

"predict" X38 from X3,X4,X5,X6, and X7, you would write:
REGRESSION VARIABLES=x38,x3 to x7/DEPENDENT=x38
/METHOD=forward.

If you wanted to do a stepwise or backward elimination

procedure, you would substitute the words "STEPWISE" and
"BACKWARD," respectively, for the term "FORWARD."

Please note that, when doing a regression analysis, you are not
guaranteed any result. If no predictor variables are significant at the
10
Some more esoteric varieties (such as hierchical regression) are also
available but will not be discussed.
 50
first step, the process will simply tell you that the "PIN limit" of .05 (or
whatever level you may have specified if you chose to override the
default) has been reached and terminate. This is a frequent outcome,
reflecting an empirical reality, and does not indicate that an error has
been made.

Also beware that the computer can calculate a meaningless

regression equation if the underlying assumptions of the model are
not met.

Discriminant Analysis

Discriminant analysis is a relatively complex topic. In order to

perform this procedure appropriately, you will need a solid
background in statistics. It is expected that you will have obtained
this background from a statistical text or other appropriate source
prior to reading this manual.

The procedure DSCRIMINANT (note the lack of "i" in the

spelling) is found in the Advanced Module of SPSS/PC+. This means
that you may not have access to this feature in industry even if your
company has the Base version of SPSS/PC+.

The syntax for the procedure is:

DSCRIMINANT GROUPS= vargroup (val1,val2)/
VARIABLES=predictor variables
/STATISTICS=statistics.

In the GROUPS specification, you indicate the values for each

of the two11 relevant groups within the appropriate grouping
variable. The variable specification parameter allows you to specify

11
Technically, it is possible to specify more than two groups. If, for
example, you were specified "GROUPS= CLEVEL(2,3,4)," you could
involve sophomores, juniors, and seniors in your analysis.
 51
which variables you wish to attempt to use as predictors. Thus, if you
were trying to predict purchase of your product (purchase=1) vs. non-
purchase of your product (i.e. no purchase or a purchase of your
competitor's product) (purchase=2), using income, sex, age,
education, and various other demographic variables (DEMO1 to
DEMO15), your procedure might look like this:
DSCRIMINANT GROUPS=purchase(1,2)/
VARIABLES=income,sex,age,educ,demo1 to
demo10/STATISTICS=1,2.

Since the specification of "STATISTICS=ALL" will normally

result in a very voluminous output, it is recommended that you
confine yourself to "STATISTICS=1,2," which will give you the mean
and standard deviations for each potential predictor variable broken
down by the two groups.

For more information, see the SPSS/PC+ Advanced Statistics

Manual.

Count: "Counting" on how many variables a criterion is met

Sometimes, you may be interested in "counting" the number of

variables on which subjects meet a certain criterion. In less abstract
language, you might be interested in knowing how many of a list of
products have been bought by a particular consumer. For example, if
we wanted to know how many major brands of soft drinks a person
had bought and we had each of the major brands coded as an
indicator variable (i.e. 1=purchase, 0=non-purchase), we might make
a count as follows:
COUNT sdrink=drink1 TO drink10(1).

In the above example, we are interested in those variables

among "drink1" to "drink10" to which people gave the answer 1
(purchase).
 52

Descriptives: A summarized version of Frequencies.

Descriptives (known as Condescriptive in the mainframe version

and in certain other packages) provides a summary of the statistics
which appear in the frequencies table, such as the mean, median, and
mode, without the explicit listing of the frequency of each response.
This may be useful under certain circumstances, particularly when
you are using interval data with many different values. For example,
if you were to measure annual household income with a precision of
$100, you might expect to get a large number of discrete answers if
from a survey of several hundred people. Thus, you might be better
off with a set of summary statistics including the mean, median,
mode, and standard deviation, at least for inclusion in your report.
(Note: You may still want to do a complete frequencies printout to
look for outliers).

Using the income example, the following command might

create the below table:
DESCRIPTIVES VARIABLES=income
/STATISTICS=1,5,13,10,11.

Number of Valid Observations (Listwise) = 49.00

Variable Mean Std Dev Minimum Maximum N Label

INCOME 270.43 46.10 209 356 49 annual household inc

It goes without saying, but is worth reiterating, that

"descriptives" would not generate a meaningful output for variables
in an nominal type scale.

For those technically interested, Descriptives can also be used to

create z-scores, i.e. standardized scores measuring relative deviation
from the mean rather than absolute magnitude. (This is useful when
you want to run a regression analysis on data that has an odd
distribution).
 53

Npar tests: Non-parametric statistical tests.

SPSS/PC+ allows you to perform various non-parametric tests

such as Wilcoxon, Cochran, Friedman, Mann-Whitney, and various
others. See the SPSS/PC+ Base Manual for details.

Means: Providing a breakdown of population means by subgroup.

Again, the title sounds a little esoteric, but the basic concept is
quite simple. Sometimes, a single population mean may be deceptive
because it is heavily influenced by some underlying variable.
Suppose, for example, that we know that the average height of a
group is 69 inches (5'9"). We might find it useful to break down this
figure by men and women, realizing that a lot of the variability is due
to between group (sex) differences:

Summaries of HEIGHT
By levels of SEX

Variable Value Label Mean Std Dev Cases

For Entire Population 58.9350 5.0615 200

SEX 1 male 62.3200 3.5587 100

SEX 2 female 55.5500 3.9603 100

Total Cases = 200

But we can do better than this. Suppose that we also know that
half of the people are basketball players (bball=1) and half are not
(bball=0). We can now do a breakdown at two levels. Notice that sex
is still the outermost criterion for distinction since this is considered
the stronger source of influence. (That is, we expect the average male
non-basketball player to be taller than the average female player).
 54
Summaries of HEIGHT
By levels of SEX
BBALL basketball player or not

Variable Value Label Mean Std Dev Cases

For Entire Population 58.9350 5.0615 200

SEX 1 male 62.3200 3.5587 100

BBALL 0 no 60.0800 2.6942 50
BBALL 1 yes 64.5600 2.8440 50

SEX 2 female 55.5500 3.9603 100

BBALL 0 no 53.1600 3.0463 50
BBALL 1 yes 57.9400 3.2789 50

Total Cases = 200

ONEWAY and ANOVA: Analysis of Variance

SPSS/PC+ allows for one way as well as more complex

methods of analysis of variance. For ONEWAY the syntax is:
ONEWAY VARIABLES=dependents BY independent(min,max).

where dependents refers to the list of dependent variables to be tested

and independent refers to the criterion variable on which subjects are
to be distinguished. "Min" and "max" refer to the minimum and
maximum values of the independent variable on which subjects are to
be distinguished. Thus, for example, if we have the following
occupational groups:

1=construction
2=other blue collar
3=professional
4=other white collar
5=other

and were are interested in testing for income differences between the
first four groups, our statement would look as follows:
ONEWAY VARIABLES=income BY occup(1,4).
 55
The ANOVA program is somewhat more complex and should
be attempted only by individuals well versed in statistics. Details are
in the SPSS/PC+ Base Manual.

Plot: Turning bivariate data into a scatterplot.

When you do a correlation analysis, you may come up with a

result like r=.33. From this, you may be able to deduce that the one
variable accounts for something like (.33)0.5=11.1% of the variability
of the other. However, perhaps a plot of the two variables against
each other might be more illustrative and help in your interpretation.
To create a plot, use the PLOT command. The syntax for a plot
suitable for the above situation is:
PLOT FORMAT=REGRESSION/PLOT=var1 with var2.

where "var1" and "var2" are the names of the two variables you want
to plot.

For example, this plot displays the relationship between

expenditures on chocolate and spending at service stations among a
group of hypothetical consumers. The correlation is approximately
.33.n
 56
PLOT OF CHOCOL WITH AUTO

W 9
e 1
e
k 1
l
y 6

c R
h 1 21 11
o 1 1 2124 2123 12
c 3 1 1 1 2313111 1
o 1 23 211332 21 1
l R 1 112 22 1 1
a 1 1
t
e 0 1 1

30 50 70 90
20 40 60 80

Monthly spending at service stations

(The numbers indicate how many people fall approximately at

that point in the intersection of the two variables).

Plot allows a number of options which are described in the

manual.

Reliability: Finding coefficient Alpha and other measures of

reliability in a scale

If you ask a question or about a concept in many several

different ways or versions, you may be interested in attempting to
infer the reliability of the measure from the repeated measurements.
Of particular interest may be coefficient alpha. Suppose you have
collected a survey containing a scale consisting of variables
PREJUD16 to PREJUD25, all measuring prejudice against poodles. To
find coefficient alpha, the command would be:
RELIABILITY VARIABLES=prejud16 TO prejud25
/SCALE=prejud16 TO prejud25
/MODEL=ALPHA.

(It is not a mistake that the variable specification, "prejud16 TO

prejud25," appears identically twice in the command).
 57

Other models, such as split half and parallel forms, are

described in the Advanced Statistics manual.

t-test: Testing for differences in two population means.

There are two kinds of t-test commonly used. The one is used to
test for differences between groups, as if we were going to test for
differences in height between males (sex=1) and females (sex=2). The
syntax for this kind of t-test is
T-TEST GROUPS=sex(1,2)/VARIABLES=height.

Another example: Is there a difference in the number of soft

drinks consumed (sdrinks) by males and females?
T-TEST GROUPS=sex(1,2)/VARIABLES=sdrinks.

Independent samples of SEX

Group 1: SEX EQ 1 Group 2: SEX EQ 2

t-test for: SDRINKS beverages consumed prior to campaign

Number Standard Standard

of Cases Mean Deviation Error

Group 1 100 28.2300 10.744 1.074

Group 2 100 30.4100 10.181 1.018

Pooled Variance Estimate Separate Variance Estimate

F 2-Tail t Degrees of 2 -Tail t Degrees of 2-Tail

Value Prob. Value Freedom Prob. Value Freedom Prob.

1.11 .594 -1.47 198 .142 -1.47 197.43 .142

The other kind is used to test for significant differences in

repeated measures or same-subject variables. Suppose we are
interested in finding out if people changed their consumption
patterns of soft drinks after an advertising campaign. If we collected
data on consumption both prior to (predrnk) and after (postdrnk) the
 58
campaign, the syntax might look as follows:
t-test pairs=predrnk with postdrnk.

Paired samples t-test: PREDRNK beverages consumed prior to campaign

POSTDRNK beverages consumed after campaign

Variable Number Standard Standard

of Cases Mean Deviation Error

PREDRNK 200 29.3200 10.497 .742

POSTDRNK 200 30.7800 9.887 .699

(Difference) Standard Standard 2 -Tail t Degrees of 2-Tail

Mean Deviation Error Corr. Prob. Value Freedom Prob.

-1.4600 13.797 .976 .085 .233 -1.50 199 .136

Since the significance level associated with this test depends on

whether the test is one or two tailed, it is particularly critical that you
make hypotheses prior to running this test. For example, it might be
reasonable to hypothesize that soft drink consumption would increase
after the campaign, thus avoiding the need to allow for the left tail
(that is, the probability that sales could also decline as a result of the
campaign). In any event, we must be careful to notice that even if
there is a significant change, we cannot positively say that it was
caused by the advertising campaign. We hope, however, that this is
the best explanation available.

Advanced features: Factor, Cluster, Hiloglinear, and MANOVA.

The SPSS/PC+ Advanced Statistics Manual contains

information about these more complex procedures. In general, these
should be attempted only by individuals very knowledgeable about
the particular procedures since quite meaningless results can be
generated and misinterpreted by an inexperienced user.
 59

Using Only Selected Cases

For some analyses, you may want to include only selected
subjects in your calculations. For example, if you want to figure out
the average age of only those respondents to your survey who are dog
owners, you might enter the following command:
PROCESS IF (dogown EQ 1).
DESCRIPTIVES VARIABLES=age.

In the parentheses, we indicate a logical condition. In this case,

the logical condition is that the value on the dog ownership variable is
1.

Other logical "operators" are:

GT Greater than
GE Greater than or equal
LT Less than
LE Less than or equal
NE Not equal

Thus, if we wanted a Frequencies table of the number of dogs

owned by people whose age is less than 35, we would say
PROCESS IF (age LT 35).
FREQUENCIES VARIABLES=dogown.
 60
Appendix A
Common Questions About
SPSS/PC+

Q.: I use SPSS/PC+ on a computer that is connected to a dot matrix

printer. Is there any way I can print the output from another
computer that has a better printer attached?

A.: Yes. SPSS/PC+ output is sent to the file C:\SPSS\SPSS.LIS.

You can import this file into a word-processor and then move
the file over to the other computer (See Appendix I for
information on how to import SPSS/PC+ output into
WordPerfect). Caution: Beware that SPSS/PC+ output is often
voluminous--printing all of it could be time consuming.

Q.: Sometimes, the arrow keys don't work while I am in SPSS/PC+.

Did I do anything wrong?

A.: No. There is a bug that will often "invalidate" the arrow keys
after an SPSS/PC+ program has been run. The best way to get
around this problem is to save your data and reboot the system.

Q.: I have some LOTUS data I would like to analyze with

SPSS/PC+. Do I have to retype it all?

A.: No. Either print it to an ASCII file (using PrintFile) or use the
IMPORT command in SPSS/PC+.

Q.: Can I do nice graphics in SPSS/PC+?

A.: Although a graphics module is available, it has not been

purchased by all institutions. Therefore, you are better off
 61
creating graphs with a program such as LOTUS or Harvard
Graphics. (For example, by typing in the information from a
frequencies run, you could create a nice pie-chart).

Q.: I am really sold on SPSS/PC+ and I would love to have it at

home. How do I go about getting it?

A.: Since SPSS/PC+ is a commercial product, it is copyrighted, and

you will have to buy your own copy. The Base version lists for
about $800 and contains the basic modules for such procedures
as frequencies, crosstabs, t-test, ANOVA, regression,
descriptives, and means. The Advanced supplement costs
about the same and contains the modules for reliability, factor
analysis, and other advanced functions. You will need eight to
ten megabytes on your hard drive. You can, however, buy a
more limited student version for about $40.

Q.: Sometimes when I press <ALT>-M <F3>, the computer beeps

and says that the selection is not available in this mode. What
went wrong?

A.: The <ALT>-M is a so called "toggle" switch which changes

between the menu and "type-in" modes. Apparently, you were
already in type-in mode before you started the command.
Simply repeat the procedure (<ALT>-M <F3>), and you should
be able to continue.

Q.: I have noticed that some people use the menu system in
SPSS/PC+ to write programs. Why isn't it covered in this text?

A.: It's true that the menus free you from typing in certain
commands in their entirety, but in the author's view, they
create more problems than it solves.
 62
Q.: What about the Data Entry module? Doesn't that save time?

A.: Some people find it easier to use this module, but since it has to
be purchased as a separate product, it may not be available to
you in industry, and it is therefore not a good idea to become
dependent on it. Also, if you are entering a long questionnaire,
it is inefficient and frustrating to have to press <RETURN>
after entering each variable.

Q.: I just have a short survey and all I want is some averages and
perhaps some standard deviations. Do I really have to use
SPSS/PC+ for this?

A.: No. Lotus 123 will allow you to compute the mean, standard
deviation, maximum, and minimum for a data range. In
addition, you can do a regression analysis (DataRegression).
 63
Appendix B

Working With SPSS/PC+

Output

Everything should be made as simple

as possible, but no simpler than that.
--Albert Einstein

SPSS/PC+ allows you to do very complex analyses, including

some whose interpretation would be challenging even for individuals
holding advanced degrees in statistics and related fields. Therefore,
we cannot cover the majority of analyses available, but in this section,
we will attempt to look at the output from some of the procedures
typically used in a beginning research methods class. Specifically, we
will look at how to prepare output from the frequencies, crosstabs,
and Pearson correlation procedures for analysis and interpretation.

Frequencies

Recall that the fundamental purpose of statistics is to make

inferences about some underlying population.12 Unless the sample
size employed is very large or unless the variability of the data is very
low, statistics based on a sample will rarely ever be exactly equal to
the population parameters. In some cases, however, statistics can be
good estimates of the underlying population parameters, and we can
use statistical principles to evaluate the likely accuracy of our
estimate.

One approach to the assessment of the accuracy of a statistical

estimate is the use of a confidence interval. A confidence interval
refers to the range within which a given proportion of statistics will
occur in repeated sampling of the same size from the respective
12
Figures based upon information about the entire population is
known as a parameter. By definition, there are no confidence intervals
associated with parameters.
 64
population. In less abstract terminology, if obtain, from sample data,
a 95% confidence interval of $34,598 to $38,710 for the mean
household income in a neighborhood in which we are considering
opening a convenience store, we can be ninety-five percent certain
that the true population mean is within this range. There is, however,
a 2.5% (5%/2) chance that the income is higher and a corresponding
chance that they are lower.
 65

Warning!

Missing data can severely distort your statistics since

SPSS/PC+ can only base its calculations on "valid" cases--that is,
the cases that have a meaningful value for the variable in question.
This is problem is particularly evident when we calculate the mean
of a variable.

When you encounter a variable that has many cases missing,

examine whether the missing cases may be systematically different
from the valid cases. For example, perhaps those people who did
not respond to a question of how many library books they checked
out last month don't care that much about libraries and checked
out fewer books than those who were conscientious enough to
answer the question.

Also, what is your population of interest? If you are

interested to find out how much the average person spends on dog
food, then those people who do not own dogs should not be
counted as missing--instead, we would count those that appear to
have relatively normal eating habits as spending $0.00. If we
counted only those people who have dogs, we would greatly
overestimate the market.

Depending on the situation involved, you might be interested in

one of two different kinds of confidence intervals, i.e. those for either
 66
a mean or a proportion.

Confidence interval for a mean. We illustrated the idea that a

sample mean can be used to generate a confidence interval for a
population mean above. Now, let's look at some SPSS/PC+ output 13
to see how generate the data:
EDUC years of formal education completed

Valid Cum
Value Label Value Frequency Percent Percent Percent

7 1 .4 .4 .4
8 1 .4 .4 .8
9 15 5.4 5.7 6.5
10 15 5.4 5.7 12.2
11 16 5.7 6.1 18.3
12 27 9.7 10.3 28.5
13 39 14.0 14.8 43.3
14 36 12.9 13.7 57.0
15 36 12.9 13.7 70.7
16 30 10.8 11.4 82.1
17 18 6.5 6.8 89.0
18 20 7.2 7.6 96.6
19 5 1.8 1.9 98.5
20 2 .7 .8 99.2
21 2 .7 .8 100.0
99 16 5.7 MISSING
------- ------- -------
TOTAL 279 100.0 100.0

Mean 13.970 Std Err .167 Median 14.000

Mode 13.000 Std Dev 2.710 Variance 7.343
Kurtosis -.408 S E Kurt .299 Skewness -.023
S E Skew .150 Range 14.000 Minimum 7.000
Maximum 21.000 Sum 3674.000

In the above survey, the respondents claimed to have a mean

level of education of 13.97 years (i.e. being 0.03 years short of
completing two years of college). The sample standard deviation is
2.71 years. Using the Lotus 123 spreadsheet MEANTINT.WK1 on the
distribution disk, get the result14

13
For this example, we will use information from a Frequencies printout,
but you could also get the needed information from the procedure Means .
14
For formulas, see an introductory statistics or research methods text.
 67

Variable Sample Sample Sample Lower Upper

Mean Std size (n) limit limit
----------------------------------------------------------------------
Education 13.97 2.71 263 13.64 14.30

From the above, we can conclude with 95% certainty that the
"true" population mean is somewhere between 13.64 and 14.30 years.
If we wanted to be more certain than that, the confidence interval
would be wider.

Confidence intervals for a proportion. As you know, an

important step in business strategy formation is competitor analysis.
Tallon Termite & Pest Control, a California company that uses liquid
nitrogen to kill termites with "coldness instead of chemicals," is
concerned about aardvarks, the only other natural enemy that
termites have. The aardvark is commonly known as the ant-eater, but
that's a misnomer since, according to unverified company
intelligence, two out of three aardvarks prefer termites over ants.
Although aardvarks reportedly chew with their mouths open, which
is not a pretty sight, they are could nevertheless represent a serious
competitive threat.

Let's suppose that you have been hired by Tallon Termite & Pest
Control, Inc., to assess the extent to which aardvarks constitute
serious competition nitrogen program. You are directed to find out
snack preferences of the rivals by sampling one hundred aardvarks
and asking if they prefer ants or termites. The results are as follow:

PREF Snack preference

Valid Cum
Value Label Value Frequency Percent Percent Percent

Ants 1 26 26.0 26.0 26.0

Termites 2 74 74.0 74.0 100.0
------- ------- -------
TOTAL 100 100.0 100.0
 68

Using the Lotus template PROPINT.WK1 on the distribution

disk, we get the following results, where p is the proportion of
aardvarks preferring termites and q is 1-p15, we obtain

Variable p q Sample Lower Upper

size (n) limit limit
----------------------------------------------------------------------
Snack preference 0.74 0.26 100 0.65 0.83

Since we can conclude aardvarks are potentially a serious

threat (we are 95% sure that between 65% and 83% of aardvarks
prefer termites over ants), your next task is to find out what
percentage of home owners in the company's trade area considers
aardvarks' table manners offensive. Because the methodology is the
same, the task is left to you. (If funding is available, you might
explore whether attitudes are influenced by whether a home owner
has children of an age where they are particularly susceptible to
influence).

Crosstabs

The crosstabulation feature is one of the more popular

programs in SPSS/PC+ since it allows the user to assess a relationship
between two variables when one or both of them are nominally
scaled. Note that in this example, the assignment of numbers to
various categories of animals is nominal or arbitrary while the
scaling for height is ordinal:

15
q is not always the equivalent of the proportion preferring ants. In
some cases, there may be several ways in which something can be "not p"
without qualifying as q. For example, suppose that the aardvarks were
also given the choice "other." In that case, q would be equal to the
proportion preferring ants plus the proportion preferring other snacks.
 69
Crosstabulation: HEIGHT Height of customer
By ANIMAL Species of preferred stuffed animal

Count Giraffe Elephant Bear Cow Pig Dog

ANIMAL > Row
1 2 3 4 5 6 Total
HEIGHT
1 6 10 14 5 21 4 60
Short 33.3

2 4 10 16 5 19 6 60
Medium 33.3

3 10 6 17 5 15 7 60
Tall 33.3

Column 20 26 47 15 55 17 180
Total 11.1 14.4 26.1 8.3 30.6 9.4 100.0

Chi-Square D.F. Significance Min E.F. Cells with E.F.< 5

---------- ---- ------------ -------- ------------------

6.17035 10 .8008 5.000 None With ANIMAL

Statistic Symmetric Dependent Dependent

--------- --------- ------------- -------------

Lambda .04898 .08333 .01600

Uncertainty Coefficient .01260 .01585 .01045
Somers' D -.02035 -.01877 -.02222
Eta .15250 .05750

Statistic Value Significance

--------- ----- ------------

Cramer's V .13092
Contingency Coefficient .18205
Kendall's Tau B -.02042 .3730
Kendall's Tau C -.02222 .3730
Pearson's R -.03093 .3401
Gamma -.02806

Number of Missing Observations = 0

Crosstabs provides a number of statistical tests. One of the
more well known of these is the Chi square ( χ2) test of independence,
which assesses the probability that one could obtain a relationship in
the sample as strong as the one obtained if the sample were taken
from a population in which no relationship actually existed. The
formula compares the ratio of the actual count in each cell to that
expected under no relationship, given the distributions of each
variable:
 70

where k is the number of "cells" (i.e. different combinations of the two

variables), and Oi and Ei are the observed and "expected" number of
cases in the respective cell. SPSS/PC+ calculates the expected
number of cases in each cell by multiplying the proportion of cases
meeting each condition and then multiplying by the total number of
cases. For example, if we have one hundred and eighty cases (n=180)
and sixty people are tall (nT=60) and forty-seven people prefer bears
(nB), then, if there is no relationship between the two variables, we
would expect a total of
(60/180) * (47/180) * 180 = 15.67

people to fall into the category of being both tall and preferring bears.
This figure compares to the actual, or observed, count of seventeen
people falling into that cell.

To each Chi square value corresponds a figure for degrees of

freedom and a significance level. In our case, the significance level is
0.8008, giving us insufficient evidence to reject the null hypothesis
that the two variables are independent. (Even if the Chi square value
had been significant, we would not be able to conclude anything more
than that some kind of relationship probably exists between the
variables).

On the negative side, Chi square has some serious drawbacks:

1. It is not a powerful test, and we can only count on confirming a

 71
relationship if it is strong. If you have interval data or better
for both variables, you would be better off using correlation or
some other more powerful technique. Also, you are unlikely to
pick up a significant relationship with a small sample size.

2. It is not likely to yield significant results when there are many

"levels," or possible values, associated with each variable.
Variables having many levels should probably collapsed (e.g.
on an age scale, you might collapse age 35-50 to one value).

3. When a large number of cells are empty, the test may not be
valid.

By specifying "/STATISTICS=ALL" when running Crosstabs,

you get several other nominal level statistics. More details are in the
SPSS/PC+ Base Manual.

Pearson Correlation

The Pearson correlation coefficient is used to assess the extent

to which two variables can be used to "predict" each other. It is
theoretically assumed that the two variables are both at least at the
interval level; however, in practice, the Pearson correlation coefficient
is often used when one or both of the variables in question are only
ordinal. When assumptions are violated in this manner, any true
relationship between the two variables may be underestimated or
entirely overlooked. Using a variable that does not achieve interval
or ratio scaling could also lead to an overestimate of the true
relationship, but that is less likely.

In practice, many researchers include Likert scale items, which

are ordinally scaled variables, in correlation analysis.
 72

WARNING!
Pearson correlation coefficients are completely
meaningless when one or more of the variables is
nominal (categorical). If we survey people for their
preferences for soft-drinks and "1" corresponds to
Coke, "2" corresponds to Pepsi, and "3" corresponds to
"Sprite," the fact that Sprite drinkers may be special
and "reaching for more" doesn't make them exactly
three times as much of anything as Coke drinkers. (To
use categorical data in correlation or regression
analysis, you can compute indicator variables, which
have metric properties, and use those instead).

Also note that correlation is intended to explore any linear

relationship between two variables. That is, it is assumed that the
change in one variable as a result of the change in the other will be
constant regardless of the magnitude of either variables. For example,
if we correlated grocery spending against income, it would be
assumed that a $1,000 increase in come would result in the same
increase in grocery spending whether the original income were
$18,000 or $249,500. If this assumption is not viable, you may want to
transform one or both of the variables. This topic is beyond the scope
of this book, but is generally covered in methodology texts.

Assuming you specified "Options=5" when you ran the analysis,

your output will consist of three parts. Let's take a look at a fictitious
example between income and vacation spending:

Correlations: VACATION

INCOME .5636
( 263)
P= .000
 73
The first figure is the sample correlation coefficient. This
ranges between -1.0, when there is a perfect negative correlation
between the two variables, and 1.0, when there is a perfect positive
correlation between the variables.

On the second line, the number of cases on which the

correlation is based appears in parentheses. Sometimes, this may be a
lower figure than your total number of subjects because there may be
missing data among some of your subjects16. Some people, for
example, might have felt that one or more of these questions was too
personal to answer.

Finally, on the last line is the significance level associated with

the correlation coefficient. As we noted above, the figure of .5636 is
the sample correlation coefficient and may not be the actual
population coefficient.17 The significance level is the probability that the
correlation coefficient is zero, not the probability that it differs from the
figure shown (which it probably will). It cannot be concluded from
the above information (without making further calculations) that the
population correlation coefficient is greater than, say, 0.35. However,
the sample coefficient is our best available estimate.

Now, what is the meaning of a correlation coefficient? In

general, the rule says that the square of the value of the coefficient
represents the proportion of the variability of the one variable
"explained" by the other. For example, in the above example,
2
.5636 = .3176,

suggesting that income "explains" about 31.76% of the variation in

vacation spending between people.

16
Note that a case will be missing in correlation if the only one of the
two variables is missing.
17
Because we are taking a finite sample from a population, we expect
our sample results differ somewhat from those of the actual population.
 74

Correlation coefficients are not very intuitive. Sometimes, a

visual representation is more informative. Using the Plot command,
we get the following illustration:
PLOT OF INCOME WITH VACATION

a 400 1
n 1 1
n 1 1 11 12 1
u 11 3211 1 31 1 1 R
a 1 3515221412 1 1
l 300 2131115341221213 11
2 22 23422 5434111
h 2 1 1 2 15216352 36 42 1
o 1 2112521314132 1 1
u 1 1 1 121 13 22 1 11
s 200 1 111 121 1 1
e R 1 11 12 1 2 522 2 1
h 21 1 1 1
o
l
d 100 1

2000 2800 3600 4400

2400 3200 4000 4800

amount spent by household on vacations l

(The numbers in the box represent the count of cases falling at

that coordinate. In a perfect correlation, all the numbers would have
been on a diagonal; with a complete lack of correlation, the numbers
would have been scattered around the box without pattern).
 75
Appendix C

Working With
Large Data Sets

In Chapter 2, we discussed how a case can take up more than

one line. We noted that in the "data list" command, we mark the
beginning of a new line by a slash. Thus, if we have ID1 and questions
1 through 50 on line 1 and ID2 and questions 51 through 100 on line 2,
our data list might look like this:
data list /id1 1-3 q1 to q50 5-54
/id2 1-3 q51 to q100 5-54.

Two cases might look like this:

001 80053594603668932848906761501037537601345684443947962
001 23073741805274028467403200920734931437934283227406462
002 37523313809651941755581798497615730965614784546728377
002 19023558532353565640285040717661639182211365682678986

Now suppose that instead of just two lines per case, we have a
data set that requires five lines per subject. Two cases, using that
system, might look like this:
001 51881042413879258528815244790246113226620946180660912
001 17360960447695181568034889642259507514245342577273341
001 90272540808655925656834500842489437576499768779375834
001 85877256735693738165767968166614884079224430157401661
001 93770572368521277158218564000020348310063127470916776
002 28192629327531949569666071329351801708701532387853052
002 46719233470503286424781463574656343764730263589509917
002 56568263273220578001682471949627692765870597258624634
002 92326111451257030599355546291915734015614726564885112
002 85184875444974819662113452795996745459258090787424135

That format is not very readable. If you accidentally leave out

one line from a case or put it in twice, the computer would warn you
that you that the number of lines is not evenly divisible by the number
of lines per case, but it might be difficult, or at least laborious, to track
down where the problem occurred. Therefore, we might want to take
two preventive steps:

1. First, assign a variable to each line to identify its number. Thus,

 76
you might have a data list beginning like this:
data list /id1 1-3 line1 4 q1 to q50 6-55
id2 1-3 line2 4 q51 to q100 6-55

and going on like that until five lines have been created.
For every case, then "LINE1" would be equal to "1,"
"LINE2," to 2, etc.

2. At the end of the five lines, create a data line that only contains
the one "variable" called "BOGUS." This will create a blank line
that makes it more obvious where each case ends. Note that
you must assign the columns that "BOGUS" could take up even
though the variable will be "system missing" for all cases. We
will just say that it occupies column 1 of line 6. The complete
data list from the above might look like this:
data list /id1 1-3 line1 4 q1 to q50 6-55
/id2 1-3 line2 4 q51 to q100 6-55
/id3 1-3 line3 4 q101 to q150 6-55
/id4 1-3 line4 4 q151 to q200 6-55
/id5 1-3 line5 4 q201 to q250 6-55
/bogus 1.

Two cases from this data set might look like this:
0011 51881042413879258528815244790246113226620946180660912
0012 17360960447695181568034889642259507514245342577273341
0013 90272540808655925656834500842489437576499768779375834
0014 85877256735693738165767968166614884079224430157401661
0015 93770572368521277158218564000020348310063127470916776

0021 28192629327531949569666071329351801708701532387853052
0022 46719233470503286424781463574656343764730263589509917
0023 56568263273220578001682471949627692765870597258624634
0024 92326111451257030599355546291915734015614726564885112
0025 85184875444974819662113452795996745459258090787424135

If your data set becomes so big that it takes a long time to run,
you might consider making a "system file," which will increase the
execution speed considerably. See Appendix D for more information.
 77
Appendix D

Using System Files

So far, we have had to run our program each time we wanted

to execute a series of one or more statistical procedures. This is
generally the easiest way to work when your data set is small.
However, as you noticed, running the program and having the
computer "read" the data each time takes a great deal of time. The
bigger the data set, the more time it takes to run the program.

Creating a "system file," which stores the data in a way that

enables the computer to access it more quickly, may help you make
the reading of the data go faster. There are a few costs associated
with this, however:

• It takes time to create the system file

• The system file takes up a great deal of space on the hard disk.
If the computer is found in a college computer lab, there may
not be space enough on the hard disk for everyone to store his
or her data as a system file, and you may not be able to fit all
your data on a floppy disk.
• Although variable and value labels will show up in your output
as usual, you no longer have much of the information you used
to create your program in your file. If you forget a variable
name, you may either have to refer to a printout or "get" the file
to find this information.

To create a system file, first type in the SPSS/PC+ program and

the data as you normally would. When finished, you may want to run
a Frequencies procedure to find out if there are any errors in the data.
You can also do any data modifications (such as Recode and Compute
statements) now. When you are satisfied with the shape of your data,
add a "save" statement to the very end of your program:
SAVE /OUTFILE="filename .SYS".
 78
where "filename" refers to the name you want to give the system file
you are creating. If your data file is named "SURVEY.SPS," for
example, you might want to name the system file "SURVEY.SYS." (On
a floppy disk in Drive A:, this would be "A:SURVEY.SYS." Note,
however, that there might not be enough space on a floppy disk to
accommodate a system file containing a large data set.)

Now you are ready to access the system file for further analysis.
To do that, create a new file (say, "FASTSURV.SPS"). Put in a "get"
statement as the very first line in this file:
GET /FILE="filename .SYS" .

where "filename" refers to the name of your system file.

After the "get" statement, you can add any statistical

procedures and run the program as normal. Thus, in our example, a
run involving a system file might involve the following:
GET /FILE="fastsurv.sps".
CORRELATION age WITH income.
CROSSTABS TABLES=dogpref BY sex.
 79

Advanced Tip

Your data file might contain some redundant and

unnecessary information that you don't want to include
in your system file. For example, if you have five lines
of data, you might have five lines of data (e.g. ID1, ID2,
ID3, ID4, and ID5) and five line identification numbers
(e.g. LINE1, LINE2, LINE3, LINE4, and LINE5). To
exclude these variables from your system file, use the
"drop" subcommand of the Save command. If we want
to get rid of the above variables and our file and we
want to name our system file "SURVEY.SYS," the
command would look like this:
SAVE /OUTFILE="survey.SYS"/DROP=id1,
id2,id3,id4,id5,line1,line2,
line3,line4,line5.

WARNING: In this case, "ID1 to ID5" would not

be equivalent to "ID1,ID2,ID3,ID4,ID5" since there are
a lot of variables that come in between the
identification numbers. Thus, if you said "/DROP=ID1
to ID5," you would be excluding all but the last line of
text!
 80
Appendix E

Importing Data From

Lotus 123

Sometimes, you may want to analyze some data you already

have typed into LOTUS 123 in SPSS/PC+. There are at least two
situations in which this could happen. First, suppose started an
assignment in LOTUS 123. Later, you realize that you wish to do
some statistical analyses not available in LOTUS. The second
possibility is that you may want to do some of your data input on a
computer that has LOUTS, but not SPSS/PC+, installed.

The following will illustrate how you can import sets of data
that consist of at most one line per case into SPSS/PC+. (If you have
more than one line of data, you will most likely have to use an
SPSS/PC+ system file. See Appendix D and the section on Import in
the SPSS/PC+ Base Manual.) Since you already have the data
entered into LOTUS, it is assumed that you are reasonably familiar
with the program.
 81

Warning!

Lotus is not very strict about whether you put your cases
across rows or columns. That is, when you put the first variable
from a case in a cell, you have a choice between putting the next
variables to the right of this variable or below. For example, if you
put VAR1 in A1, you can either put variables 2 through 5 in A2..A5
or B1..E1.

SPSS/PC+ is a little more choosy. In SPSS/PC+, variables

associated with the same case must occur in rows. This means that
if you put the first variable in A1, the next one should go in A2.
Then variable 1 of the next case might then go in B1.

First, get into LOTUS 123 and bring up the file that contains the
data you wish to use in SPSS/PC+. Make sure that your data is
aligned properly in the columns by issuing the "/Range Format Fixed"
sequence of commands.

Now make sure that the data does not take up more than eighty
columns or spaces. If it does, you might try to reduce the length by
first reducing excess columns assigned to any one variable. For
example, while LOTUS assigns nine digits to each column by default,
you don't need that many for a variable like age. (For reasons beyond
the scope of this book, you need one more column than the number of
digits of the greatest number.) To reduce the number of columns
reserved for age to three, get into the column containing the age and
issue the directive "/Worksheet Column Set width" and then specify
3. If your data still does not fit, you might think about only taking
some of the variables into SPSS/PC+. You can do that by moving only
those variables you want into another area of the spreadsheet.
 82

To take the data to SPSS/PC+, we will now "print" the variables

to an ASCII (i.e. plain text) file. You may be familiar with the "/Print
Printer" command. Now instead say "/Print File." The computer will
prompt you for the name of the output file. Let's call it
"A:LOTUSDAT.SPS." (Make sure you have a formatted disk in drive
A:.) Press "Range" and specify the range you want to take into
SPSS/PC+. Now specify "Options" in the print menu and choose
"Margins Left" and specify "0." Finally, press "Align Go." When the
computer has finished "printing" to the disk, that is, when the menu
comes back, choose "Quit," then "/Quit" to exit LOTUS 123.

Now get into SPSS/PC+ and get into the file

"A:LOTUSDAT.SPS" (i.e. by <ALT>-M <F3>).
 83

Warning!

Even when you "print" to a file, LOTUS still

thinks in terms of a "sheet" of paper. Since a "sheet" in
LOTUS cannot be set to be longer than 100 lines (with
the default being 66 lines), you will have some empty
lines when each "page" ends in the file. Once you get
into SPSS/PC+, be sure to delete these empty lines.

Once you are in SPSS/PC+, it is time to put in the SPSS/PC+

commands. You can get the information for the data list command by
cursoring over the beginning and end of each variable and looking at
the position indicator in the bottom right part of the screen. Then
insert the needed statements such as Data list, Begin data, End data,
and any Variable labels, Value labels, and Missing value definitions that
you would like.

The program will now run just like any other program.
 84
Appendix F

Similarities Between SPSS/PC+

and Lotus 123

There are some similarities between some SPSS/PC+ functions

and those in Lotus 123 that you may already know. Let's suppose you
want to add up a person's total utility expenditures from his or her
phone, water, natural gas, and power bills. In Lotus we might have
the data entered as follows:

A B C D E F

1 Phone Water Gas Power

2 ---------------------------------------------------------------
3 33.99 11.44 23.44 29.95
4 144.54 12.34 53.60 44.45
5 67.41 9.50 22.34 34.35
6 34.45 10.51 34.45 29.45

To add up the first person's total utility expenditures, you

might type in the formula
@SUM(A3..D3)

in cell E3. Alternatively, you could use the expression

+a3+b3+c3+d3 .

After you completed the calculation for the first case, you
would have to copy the formula from E3 to E4 through E6:
/cE3 <CR> E4..E6 <CR>.

Now look at how similar the procedure is in SPSS/PC+!

Assuming that we already have created the variables PHONE,
WATER, GAS, and POWER, we can use the latter approach from lotus
 85
and say:
COMPUTE total=phone+water+gas+power.

(Unfortunately, there is no function comparable to @SUM in

SPSS/PC+). For practical purposes, SPSS/PC+ will automatically do
the "cell copying" for us! (Remember, for statistical purposes we
generally work with the entire data set without specific interest in
individual cases).

SPSS/PC+ provides other functions as well. These include both

the traditional ones such as the absolute value, e.g.
COMPUTE differnc=abs(score1-score2).

and more esoteric ones such as logarithmic and trigonometric

functions18, which are also available in Lotus (See p. C-24 of the
SPSS/PC+ v. 2.0 Base Manual for more information).

You can also use arithmetic expressions in calculations. While

in Lotus you might multiply by putting the expression
+A12*A13

in cell A14, you might say

COMPUTE interest=prncpal*rate.

in SPSS/PC+.

18
For those interested in econometric and forecasting methodology, it
is also possible to use lagged data. Details are found in the manual.
 86

Important Note!

In Lotus, you are used to the automatic

updating which takes place when you change the
values in a cell on which a formula is based. In
SPSS/PC+, such updating will generally not take
place. Consider this example of a command
sequence in SPSS/PC+:
COMPUTE discret=income-expenses.
COMPUTE income=income*(1-taxrate).

SPSS/PC+ will not adjust discretionary

income for the effect of taxes in this example! To
accomplish this task, the second COMPUTE
statement would have to precede the first. In other
words, the effect of the change in the income
calculation is not "retroactive" as it would have been
in Lotus if you had gone back to change the formula
in the cell in which (pre-tax) income was computed.
It will, however, be in effect any time after the change
has been performed.
 87
Appendix G

Similarities Between SPSS/PC+

and dBase III+

Readers familiar with dBase III and similar database programs

may notice certain conceptual similarities in SPSS/PC+. Although
understanding how database software functions is not necessary to
understanding how SPSS/PC+ works, it may help you understand
some concepts more easily.

First of all, both SPSS/PC+ and dBase clearly divide the data
into individual cases or "records." In dBase, the data associated with
each entity is kept within a single record. Thus, if we have a database
of student grades, a dBase record might look like this:

Record 1

ID 5551515555
MIDTERM 84
PAPER 95
FINAL 88

In SPSS/PC+, each case is assigned a specific number of lines.

Thus, our SPSS/PC+ case might look like this:

5551515555 084095098

Each record or case is further subdivided. In dBase, we call

each variable a "field." For example, the field "MIDTERM" contains
the student's midterm score of 84. In SPSS/PC+, the variables are
 88
assigned specific column positions. If we look immediately after the
blank space following the ID number (probably the student's social
security number), we can see the entry of "084," signifying the score of
84. Notice that the separation of the variables is completely up to us;
we defined a blank space between the ID number and the scores, but
none between the scores themselves.

dBase allows us to compute several statistical functions. Thus,

we could use the "average()" and "std()" functions to get much the
same information we could get with "descriptives" command in
SPSS/PC+. However, dBase does not give us a direct way to do
anything as complicated as crosstabs, correlation, regression, or even
frequencies.

Sometimes, we are only concerned with part of our sample.

Thus, in dBase, if we wanted to find the average on the final for those
people who scored below 70 on the midterm, we could say
AVERAGE(final) FOR midterm<70

while in SPSS/PC+, we might issue the commands

PROCESS IF (midterm LT 70).
DESCRIPTIVES VARIABLES=final.
 89
Appendix H

SPSS-X: The Mainframe Version

As mentioned in the introduction, a mainframe version, known

as SPSS-X, is also available on many university campuses. There are
certain circumstances under which it may be desirable to use the
mainframe. Some features are not readily available on the PC or may
take too long to execute. This will normally not be a problem except
you wish to undertake very advanced analyses.

Another advantage of using a mainframe computer is that

many university computer centers have a "high speed" printer
connected to their mainframes. These printers can often print out a
hundred pages in less than one minute. Although you normally have
to wait for the printouts to be distributed to an "out" box at some
central location, you at least save the time you would have to sit in
front of a slower "ordinary" printer.

SPSS-X commands are very similar the SPSS/PC+ commands,

the main exception being that there are normally no periods after the
end of each command. You may also have to include some "file
handle" commands under certain circumstances and, of course, you
will have to deal with some computer commands very different from
those of the IBM PC.

Extensive documentation on the use of SPSS-X is available from

SPSS, Inc. You may also have to get some handouts on your
mainframe's implementation of SPSS-X.
 90
Appendix I

Incorporating SPSS/PC+ Output Into

WordPerfect Reports

Sometimes, you might like to include the results of a statistical

analysis in a report. While it is possible to manually type the results
from your computerized analysis into a word processing document, it
can be cumbersome and the complexity involved can lead to a
number of mistakes. Fortunately, there is an easier way.

SPSS/PC+ creates an output file to which everything you see on

the screen or on printer output is sent. This file, called
"\SPSS\SPSS.LIS," is overwritten every time you restart SPSS/PC+, so
you will have to "rescue" the material immediately if you need it.19

Although this manual will illustrate the use of WordPerfect 5.0

or 5.1 to "grab" important output, most word processors will provide
the same capability.20

Immediately after you exit SPSS/PC+, select WordPerfect from

the menu. If you have already created the document in which you
would like to include SPSS/PC+ output, call it up now. Regardless of
whether you take that step or not, now switch to document #2
(<SHIFT>-<F3>). Use the "text-in" function to retrieve the file
"\SPSS\SPSS.LIS":
<CONTROL>-<F5> 1 2 C:\SPSS\SPSS.LIS <RETURN>

Find the part of the output you would like to include in your

19
Alternatively, can use the "SET RESULTS" command to direct your
output to a permanent file (see manual).
20
SPSS.LIS is an ASCII file and should be edited and read in
accordingly.
 91
document and use the block function (<ALT>-<F4>) to mark the
appropriate text. To copy the text, use <CONTROL>-<F4> to copy
and press <SHIFT>-<F3> to move back to document #1. Use the
cursors to move where you want the text inserted, then press
<RETURN>.

If you want more text from the output, you can repeat the
process.

NOTE: In some cases, the lines of SPSS/PC+ tables are more

than 65 columns long. If you encounter such a table, you will want to
be sure to use a font smaller than the default Courier, 10 pitch. Use
the base font option (<CONTROL> <F8> 4) to select. (The SPSS/PC+
output included in this text was printed in Letter Gothic, 16.67 pitch).

Also, beware that you should never use a proportionally spaced

font for tables. (Helvetica and Times Roman are two of the most
common proportionally spaced fonts). While text printed in a
proportionally spaced font looks very beautiful, this scheme tends to
distort tables that depend on proper alignment of variables.
 92
Appendix J
Dealing With Printer Problems

Unless you specify otherwise, SPSS/PC+ assumes that you are

printing to printer that uses continuous sheets of paper, each of which
is sixty-six lines long. This is fine when you print to dot matrix
printers whose paper is "tractor fed."

However, you may run into some problems when you try to
print to a laser printer or other kind printer that prints on stacks of
single sheets of paper. This happens because, although there is room
for sixty-six lines of text on a laser printer, not all of the space on the
paper is available for the printer's use. Thus, the pages may become
disalligned; that is, the page breaks may not occur at the right places.

To deal with this problem, enter the following series of

commands where you would otherwise place the "set printer on"
command:
SET SCREEN ON.
SET MORE OFF.
SET LENGTH=59.
SET PRINTER ON.
SET EJECT ON.

Be sure to enter the commands in the suggested order.

NOTE: When SPSS/PC+ is done printing, you may need to

press the <FORM FEED> key on the printer to get the last page out.
 93
Appendix K

Statistical Significance

When studying a subject such as statistics, it is easy to lose sight

of the forrest for the trees (or, in more contemporary terms, to lose
sight of the computer for the chips). In order to get the "big picture" in
statistics, it is important that you remember the fact that statistics is
all about making inferences about an underlying population from a
limited sample. If you do a study of grocery shoppers, you really don't
care in particular about the people you are interviewing. A sample
made up of 100 people is in and of itself irrelevant (it is a very small
part of the total market). You are really interested in generalizing the
responses of the sample to the population as a whole. This is not to
say that the selection of your sample is not important. In order to
make sure that the sample is representative of the population, the
researcher should use a sampling scheme that ensures proper
selection. Depending on the circumstances, one may want to use
either a random or systemematic sample or some variation of these.

As you know from introductory statistics, a sample mean will

rarely ever equal the "true" population mean. You may recall the
Central Limit Theorem, which suggests that the larger the sample
size, the closer its mean will be to the actual population mean. In
general, the principle that a larger sample size will allow you to make
more precise statistical inferences applies to other statistical
procedures, too.

Every profession has its peculiar jargon, and statistics is no

exception. One of the common terms in this field is that of statistical
"significance." Despite what the word suggests, statistical significance
has nothing to do with importance. As ironic as it sounds, something
can be very significant but not at all important. Consider this fact:
There is a statistically "significant" relationship between the number
of people at a Southern California beach and the number of leaves on
the surrounding eucalyptus trees. (This is, in terms of another piece of
 94
statistical jargon, a spurious correlation in the sense that both of these
phenomena are "caused" by a common factor [the season of the year]
although there is no intercausal relationship between the two).

In the broadest possible terms, statistical significance refers to

the probability that one would observe a sample statistic as extreme
as the one obtained if, in fact, there were no such relationship in the
population.

You will recall from your statistics classes the use of the null
and alternative (or "research") hypotheses (Ho and Ha). You assume
the null hypothesis, which states that there is no relationship, to be
true until you find overwhelming evidence that it is not--in that case,
you "accept" the alternative hypothesis, knowing that there is a
certain chance (e.g. 5%) that the relationship does not exist.

Consider this fictitious example: A researcher samples five

athletes and five non-athletes to see how many soft drinks they each
consumed last week. The results are as follows:
Soft drinks Standard
consumed deviation n
_______________________________________________
Athletes 19.1 7.7 5
Non-athletes 14.9 6.5 5

Intuitively, would you be confident in concluding that athletes

drink more soft drinks than non-athletes? Although you could do a t-
test to be sure, you would probably conclude without doing so that
you don't have enough evidence to conclude that there is a difference.
(This would particularly be the case if you had you had not made a
hypothesis in advance that athletes drink more soft drinks than non-
athletes). The fact is that the variance within each group is so high
compared to the difference in the two sample means suggests that no
confident conclusion is possible. Now let's suppose, however, that we
sampled ten thousand athletes and ten thousand non-athletes and got
the following results:
 95
Soft drinks Standard
consumed deviation n
_______________________________________________
Athletes 19.1 3.1 10,000
Non-athletes 14.9 2.5 10,000

In this case, the variance is much smaller relative to the

differences in the sample means, and although you might not feel
confident intuitively concluding that there is a difference, the
evidence now seems impressive enough to attempt a formal t-test.

As previously indicated, it is possible to generalize the notion of

significance to other statistical tests. Consider the example of
correlation. Recall that the Pearson correlation coefficient ranges
from -1, a perfect negative relationship, to 1, a perfect positive
relationship. When the population coefficient r=0, there is no
relationship. Let us suppose that there is in fact no relationship
between GPA and the number of soft drinks consumed the previous
week. However, a group of students survey of fifty individuals and
find a sample correlation of r=0.05. Does this mean that there is a
small but real relationship between the two variables? No! The
correlation is so close to zero that it may just have happened that the
number came out this way. If a "competing" group of students
wishing to prove that the relationship was negative attempted
another survey, they might find r=-0.08. Another group might find
r=.01. Thus, the only thing we seem to be able to conclude is that, if
any relationship exists in the population, it is probably trivial.

This suggests that a certain amount of "error" variance takes

place. In a limited sample size, the correlation coefficient obtained
may not be the true population coefficient, even if the sample is a
perfect random sample. For this reason, you will find a significance
value below the correlation coefficient and sample size if you use the
parameters in the correlation statement described in this text; that is,
for example:
CORRELATION var1 WITH var2
/OPTIONS=5.

If you choose a 95% level of confidence, for example, you would

 96
say that there is a relationship only when the significance level is less
than 0.05. This significance level only implies that there is either a
positive or negative relationship, not how big it is. (Beware that these
are one-tailed levels of significance. This means that you would have
had to make a directional hypothesis (e.g. "Soft drink consumption is
positively associated with GPA," not "Soft drink consumption is
associated (either positively or negatively) with GPA).

When running the correlation procedure, you will tend to find

that, for a small sample size, a relationship will have to be strong to
show up as significant. Also beware that even if a relationship shows
up as significant, the true correlation coefficient may be greatly
different from what was obtained. For example, you might observe a
sample coefficient of 0.41. However, the 95% confidence interval for
the true coefficient may be something like 0.25-0.57--quite a bound!

To reiterate, it is useful, in general, to look upon significance as

the probability that a relationship as extreme as what was observed
would have happened in a sample of a given size if the null
hypothesis were true; that is, if no relationship actually existed. You
will find significance levels in virtually all statistical tests including,
but not limited to, regression, chi square, discriminant analysis, t-
tests, and correlation. Some analyses such as regression may contain
several significance tests. There will be significance levels at each
step (as each additional variable is entered into or removed from the
equation). Further, each step will have a separate significance test for
each variable, measuring the confidence we can have that the
particular variable should in fact be in the "true" regression variable.
The overall regression significance test examines whether we in fact
have an equation which significantly "explains" a portion of the
variance in the population dependent variable.

A related topic is that of simultaneous significance testing. This

is a relatively difficult subject to grasp, but one that is nevertheless
very important.

Remember that, if we choose the 0.05 level of significance, each

 97
time we run a correlation analysis, there is a five percent chance of
committing what is called a "Type I" error, i.e. concluding that there is
a significant relationship when there in fact isn't. Now, let's suppose
that we run twenty analyses and find that one is significant. Should
we conclude with confidence that a relationship exists?

Doing so may be risky. We ran twenty analyses, in each of

which we ran a 5% risk of being wrong. Multiplying twenty (the
number of analyses we ran) by 0.05 (the probability of each one being
wrong), we find that the "expected" number of "false alarms" is 1.
Thus, the relationship may not be truly significant.

We can also look at this issue another way. If we make one

significance test at α=0.05, the probability of a Type I error is 0.05. If
testing all our hypotheses requires us to make ten such tests, what is
the probability of making at least one Type I error? This probability is
not 10(0.05)=0.50, although that may seem an intuitive answer.
Instead, we must use the Bernouli Formula. When you encountered
this formula in finite mathematics, you most likely used it to calculate
the something like the probability that there would be two or more
left-handed people in a class of thirty.

The Bernouli Formula involves what in finite mathematics is

called a "combination," or the number different ways one can select a
sample of a desired size from a population of a given size.21 The
formula involves only basic arithmetic, but lots of it, so the
distribution disk includes a program called "SIGNIF.EXE" which will
calculate the probability of getting at different numbers of significant
outcomes under a given number of tests at the desired significance
level. (With the distribution disk in Drive A:, type in "A:SIGNIF.EXE"
at the DOS prompt to run the program).

21
The theory behind Bernouli trials and combinations is beyond the
scope of this text but is generally discussed in introductory finite
mathematics texts.
 98
Statisticians have never satisfactorily resolved the issues and
problems that surround simultaneous significance testing. However,
there are some ways of limiting its potential dangers:

• Be sure to make your hypotheses before running your tests and

do not run any more tests than are necessary to answer the
questions raised by your hypotheses. Whenever possible,
hypotheses should be one-tailed, i.e. predicting the direction of
the relationship.
• Confine your tests to those for which you have an adequate
sample size. If you have only thirty subjects, for example, it is
unlikely that you will find a truly significant correlational
relationship, so running that procedure is not appropriate.
• Realize that sporadic significant tests are perhaps significant by
chance. (If, for example, you obtained three significant
independent tests out of forty, you could use the Bernouli
Formula to figure out the probability of obtaining that many
"successes" by chance.

Significance is not an easy topic, and you may have to do a lot

of thinking in your project to find out how the concept relates to your
particular findings.
 99
Appendix L

Expanded Glossary

Statistics

alpha (α ): (1) the maximum probability of a Type I error allowed in a

statistical test. The level of α=0.05 is frequently used in most
tests. (2) The level of internal consistency in an alpha
reliability analysis; α is a conservative estimate of the test-
retest correlation between a scale and a hypothetical parallel
version. (See also reliability.)

alternative (research) hypothesis (Ha): The hypothesis that the

researcher is trying to support. For example, if you are trying to
prove that athletes on the average drink more soft drinks than
non-athletes, the alternative hypothesis, Ha, might be µa>µn,
where µa is the mean number of soft drinks consumed by
athletes and µn is the mean number of soft drinks consumed by
non-athletes. Note that the alternative hypothesis can either be
one tailed, where we hypothesize that one mean is greater than
the other, or two-tailed, where we merely hypothesize that the
two means are different.

analysis of variance (ANOVA): A statistical method that enables the

user to assess the contribution of one or more independent
variables to the variability of the dependent variable. For
example, a researcher may wish to test the hypothesis that the
number of soft drinks consumed by a student is influenced by
both major (1=business, 2=liberal arts, 3=engineering, 4=other)
and athletic status (1=athlete, 2=non-athlete). ANOVA is a
very flexible statistical procedure that allows the user to create
sophisticated statistical models. (See also ONEWAY).

case: One entity or unit of data on one or more variables; usually one
subject, individual, or other unit (such as a firm). For example,
 100
when you administer a questionnaire, each person is a case, but
if you have both a pre-test and a post-test, the two tests for the
same subject together constitute a case.

Central Limit Theorem: The theorem which holds that, regardless of

the distribution of a particular variable, the means of any large
sample from the given distribution will follow a normal
distribution. Thus, we are really talking about "means of
means," or in less formal terms, averages of averages. Knowing
this property allows us to perform certain tests without
assuming that a variable involved is normally distributed. The
distribution of the means is dependent on both the sample size
and the variability of the population. Taking this information
into consideration, we can calculate the standard error of the
mean, a concept comparable to the standard deviation of a
variable. The standard error of the mean is appreciably smaller
than the standard deviation of the variable; the formula is
Install Equa tion Editor and double -
click here to view equation.

One application of the Central Limit Theorem is the

calculation for the mean of a "large" sample (i.e. a sample that
has an n greater than 30).

Chi square ( χ 2): A statistical test that determines the probability that
two nominal22 variables, given the sample data, are
"independent," i.e. whether knowing one variable will help
"predict" the other. The formula is

22
Chi square may be used on ordinal, interval, or ratio scaled
variables, but more efficient tests are available for such variables.
 101
where k is the number of "cells" (i.e. different
combinations of the two variables), and Oi and Ei are the
observed and expected number of cases in the respective
cell. (See also crosstabulation.)

confidence interval: In context of the mean of a variable, the interval

that is estimated with a given certainty (e.g. 95%) to contain the
true mean. The Central Limit Theorem allows us to estimate
confidence interval for the mean of a variable can be calculated
by using the standard error of the mean (s/n0.5). The formula
is:

Install Equa tion Editor and double -

click here to view equation.

where s is the sample standard deviation, n is the sample size,

and z α/2 is the two-tailed z-value for the given level of
significance.

coding: The process of translating research data (such as

questionnaire responses) into consistent numerical codes for
input into the computer. For example, to the question of "Did
you vote in the last election?" we might assign "1" to "yes," "2" to
"no" and "9" for missing data.

correlation: A measure of how well two variables "predict" each

other. Correlation can either take the form of the Pearson
Product-Moment Correlation, which assumes interval data, or
Spearman Rank-Order Correlation, which assumes only ordinal
data. (See also regression.)

crosstabulation: A table showing the relationship between two variables.

For example, the below table shows a hypothetical
relationship between a person's height and stuffed
animal preference. Of the respondents, six people were
both short and preferred giraffes.
 102

Crosstabulation: HEIGHT Height of customer

By ANIMAL Species of preferred stuffed animal

Count Giraffe Elephant Bear Cow Pig Dog

ANIMAL > Row
1 2 3 4 5 6 Total
HEIGHT
1 6 10 14 5 21 4 60
Short 33.3

2 4 10 16 5 19 6 60
Medium 33.3

3 10 6 17 5 15 7 60
Tall 33.3

Column 20 26 47 15 55 17 180
Total 11.1 14.4 26.1 8.3 30.6 9.4 100.0

Chi-Square D.F. Significance Min E.F. Cells with E.F.< 5

---------- ---- ------------ -------- ------------------

6.17035 10 .8008 5.000 None With ANIMAL

The SPSS/PC+ crosstabs procedure allows you to calculate

several statistical tests and functions such as Chi square ( χ 2).
(See also Chi square).

data list: The command in SPSS/PC+ which informs the computer of

the names and positions of the variables in the data. For
example, the line
data list /id 1-3 age 5-6 sex 7.

means that for every case, the ID number is contained in

columns one through three; age in columns five through six;
and sex in column seven. The data list is a required command.

data set: The total collection of cases and variables included in a file.
For example, if you administer a questionnaire to one hundred
people and enter the answers to all the questions into the
computer, this would be your data set.

dependent variable: The variable we are trying to "predict" from one

or more independent variables. For example, if we wish to test
whether athletes on the average consume more soft drinks than
 103
non-athletes, the number of soft drinks consumed is the
dependent variable and athletic status is the independent
variable. See also independent variable.

descriptives: a procedure that calculates the mean and various

optional statistics associated with a variable.

"dummy" variable: see indicator variable.

frequencies: A procedure that indicates the number of cases

associated with each value of a variable. For example, the
below table indicates the number of people who gave each of
various answers to the question of whether they owned a dog.
Various statistics, such as the mean, median, mode, standard
deviation, and standard error of the mean, optionally are
available with this procedure.
DOGOWN Ownership of Dog

Valid Cum
Value Label Value Frequency Percent Percent Percent

Yes 1 82 46.9 49.1 49.1

No 2 82 46.9 49.1 98.2
Not sure 3 3 1.7 1.8 100.0
Missing 9 8 4.6 MISSING
------- ------- -------
TOTAL 175 100.0 100.0

Mode 1.000 Minimum 1.000 Maximum 3.000

independent variable: A variable used to "predict" a dependent

variable. See dependent variable.

indicator variable: A binary variable used to designate whether a

case falls into a category or not. Usually, "1" denotes that the
subject has the characteristic and "0" that it does not. Indicator
variables are most commonly used in correlation or regression
analyses where they can be used to represent such dichotomies
as sex (1=Female, 0=Male) or residential status (1=Urban,
0=Other).

internal level of measurement: See levels of measurement.

 104

levels of measurement: Variables can be divided into four categories

or levels of measurement.

At the "lowest" level are nominal scales, where the numbers are
arbitrary. Thus, if, for a given variable, the code "1"
corresponds to "Democrat" and "2" corresponds to "Republican,"
the scale is not meant to imply that Republicans are exactly
twice as much of anything as Democrats. In fact, we could
reverse the two codes without making any real difference.

In ordinal (or monotonic) scaling, the measure increases or

decreases with the number. Thus, if we assign a "7" to "strongly
agree", "5" to "slightly agree," and a "4" to "neither agree nor
disagree," the person who responds "7" agrees more than the
other two. We cannot say, however, that the person who
chooses "7" agrees 1.75 times (7/4) as much as the person who
is neutral.

In an interval scale, the units of increment are equal, but there is

no absolute zero. A classic example of this is a temperature
scale, where 0 oF or 0 oC are not absolute minima.

A ratio scale has an absolute zero, and increments are equal.

Using a scale of income, a person who earns $40,000 earns
exactly twice as much as the person who earns $20,000. (Note,
however, that the scale is no longer a ratio scale if it has been
collapsed [1=$0-10,000, 2=$10,001-20,000, etc.])

Likert scale: A scale which asks the respondent to rate his or her
opinion on a bi-polar scale. For example, a respondent might be
asked to rate his or her level of agreement with a statement
using a scale such as:

Strongly Strongly
agree disagree
1 2 3 4 5 6 7
 105

Other dimensions might include, but are not limited to,

"Extremely satisfied/Not at all satisfied," and "Extremely
important/Not at all important."

mean: The qoutient of sum of all the cases of a variable to the sample
size; i.e., for the sample mean,

Install Equa tion Editor and double -

click here to view equation.

This quantity is also known as the arithmetic mean or, in less

precise language, the average.

median: The middle value of a variable, i.e. the value of a variable

such that an equal number of variables are less than and
greater than or equal to this quantity. If n is even, the median
is the mean of the two middle values. For example, in the set {1,
3, 5, 7, 8}, the median is five. In the set {1, 3, 4, 5, 7, 8}, the
median is (4+5)/2=4.5. Note that even if we rearrange the
order of the set to {3, 8, 7, 1, 2, 5}, the median is still 4.5 since the
result is based on the ranked values.

missing data: The condition that arises when data is not available or
not applicable on one or more variables for one or more cases.
Missing data can take two forms.

The first form occurs when a subject fails to respond to a

question that he or she should have answered. (This omission
may be deliberate or accidental). We often designate a missing
value consisting of one more "9s" for this purpose, although the
exact value designated is arbitrary. For example, we might
designate the missing value of "99" for the variable "age" since
 106
we normally assign two digits for this variable.

The other kind of missing data occurs when a person is not

supposed to respond to a question. For example, if we have
already determined that a person does not own a dog, it does
not make sense to ask him or her to respond to a question about
his or her expenditures on dog food (except, of course, if we are
surveying people who tend to be on unusual diets). We
frequently leave the data columns assigned to the variable
blank in such cases.

mode: The value that occurs most frequently on a given variable. For
example, in the set {1, 2, 3, 3, 3, 3, 4, 4}, the mode is 3.

nominal level of measurement: See levels of measurement.

multiple r: See R2, the square of this quantity, which is a more

interpretable statistic.

non-parametric statistics: Statistical procedures that do not make

assumptions about the distribution or level of measurement of
the data. For example, Chi square ( χ 2) is a non-parametric
statistic.

null hypothesis (Ho): The hypothesis that no statistical relationship

or difference exists between two variables in the population.
For example, if we were interested in testing the hypothesis that
athletes drink more soft drinks than non-athletes, the null
hypothesis would be µa=µn. See also alternative hypothesis.

ONEWAY Analysis of Variance: A method of testing for differences

in group means on a variable. For example, we might use the
ONEWAY procedure to test the hypothesis that there is a
difference in the average number of soft drinks consumed by
students, professors, and clerical workers. While it would have
been possible to make pairwise comparisons between these
groups using t-tests, error levels would accumulate since an
 107
alpha probability of a Type I error is associated with each test.

ordinal level of measurement: see levels of measurement.

parametric statistics: Statistical procedures that assume an interval

or ratio level scale of measurement. For example, Pearson
product-moment correlation and multiple regression are
parametric procedures.

population: The complete group about which we want to make

generalizations from our statistical data. Specifying a
population is important but can be difficult. For example, if we
are interested in the soft drink consumption habits of athletes,
is our population "all professional athletes in the United States,"
"all professional, amateur, and recreational athletes in the
World," or some other group?

Technical note: In defining a population, statisticians generally

refer to the actual units of comparison rather than the subjects
with which they are associated. Thus, to a statistician, the
population in the above example would be the soft drink
consumption counts scores and not the athletes. However, in
the social sciences and in this text, the term is more loosely used
to refer to the subjects, that is, in this case, the athletes
themselves.

random sample: A sample selected in a manner such that each

element has the same probability of being picked from the
population. Random samples are typically selected using a
random number table or through random pickings by a
computer.

ratio level of measurement: See levels of measurement.

recoding: The process of systematically changing values of a variable

into others. In SPSS/PC+, recoding can take several forms.
First, we may want to "collapse" a variable. For example,
 108
supposed we asked a person to name his or her favorite kind of
food. We might have a separate code for several kinds of Asian
food (E.g. Chinese=11; Japanese=12; Korean=13; Thai=14) and
several kinds of Italian food (pizza=21; pasta=22). Part of our
recode statement might thus look like this:
recode food (11 thru 19=10) (21 thru 29=20).

Secondly, we may want to "reverse" score a variable. For

example, a scale of prejudice against poodles might include
both of these Likert scale items:
___. Poodles are stupid
___. Poodles are loving dogs.

Note that the two variables go in the opposite direction of each

other; agreeing with the first (poodle1) would imply a negative
attitude toward poodles while agreeing with the second
(poodle2) one would imply a positive attitude. Thus, we should
recode the second variable so that a high score will become a
low score and vice versa. If we asked the subject to respond
based on a seven point Likert scale, our recode command might
look like this:
recode poodle2 (1=7) (2=6) (3=5) (5=3)
(6=2) (7=1).

regression: A statistical procedure which attempts to predict a

dependent variable from one or more independent variables.
Regression requires interval or ratio data, although, in practice,
it is often used with ordinal data. In such cases, interpretation
is somewhat open to question. Regression differs from
correlation in that more than one variable can be used to
predict the dependent variable.

reliability: The extent to which a variable consistently measures

some phenomenon. Reliability can be assessed in various ways.
If we have two "parallel" or similar measures of the same
variables, such as two comparable versions of the same test, we
can correlate these. Another method is to administer the same
 109
"test" or questionnaire twice and correlate the scores against
each other. A third method is the measurement of "internal
consistency" of a scale consisting of several variables through
coefficient alpha. See also alpha.

r2: The proportion of the total variance in a dependent variable

explained by a regression equation. r2 is the square of r in a
simple correlation or multiple r in multiple regression analysis.

reverse scoring: see recoding.

sample: A subset of the population selected for analysis. Often, it is

not possible to survey or examine the entire population, making
it necessary to generalize from a sample. For example, it would
be logistically impossible, let alone cost effective, to contact all
athletes to find out how many soft drinks they consume. Thus,
we might ask only a limited number, such as 200, and try to
generalize to the entire population with some level of
confidence.

significance: The probability that some statistical phenomenon

would happen by chance in a sample of a given size if no
difference or relationship actually existed in the population.
For example, in a t-test, if p=0.03, differences as large those
observed would be expected only in three percent of samples if
no actual difference exists in the population.

simultaneous significance testing: The condition that arises when

several significance tests are made at the same time. Since a
separate alpha ( α) value, or probability of a Type I error, is
associated with each test, error levels accumulate. Thus, for
example, if we perform six independent tests at the same time,
there is a higher probability (approximately, but not exactly,
.30) that at least one test will be significant at the 0.05 level
even if no relationship or difference exists in the population.
 110
standard deviation: A measure of variability within a variable. The
standard deviation of a sample is:
Install Equa tion Editor and double -
click here to view equation.

When calculating by hand, the following formula is more

efficient to use:

Install Equa tion Editor and double -

click here to view equation.

standard error of the mean: An estimate of the variability of the

mean of a sample from a given population. The greater the
variability (i.e. standard deviation) of the variable in the
sample, the less precisely we can estimate the true mean of the
population for a given sample size. See also Central Limit
Theorem.

t-test: A statistical method used to test for differences in the means of

either (1) the same variable between two different groups or (2)
between two different variables relating to the same subject.

The first form is illustrated by the testing of our hypothesis that

athletes drink more soft drinks than non-athletes.

The second form is frequently used to test for differences

between a pre-test and post-test. That is, we first give the
subjects a test, we then make some "intervention," and then test
again to see if it appears that the intervention has had an effect.
For example, suppose we first determine how many soft drinks
 111
of our brand that each member of a random sample of people in
two different regions consume. We then conduct a promotional
campaign in the one region and none in the other region (which
will serve as our control group). (Assume we have selected
regions that are relatively similar in terms of climate and social
stratification.) After the promotional campaign is over, we
again determine how many soft drinks are consumed by people
in each sample. If there is a significant difference between
consumption before and after for those people in the region
where the campaign was conducted, we might conclude that
the campaign had an effect on consumption of our brand.
However, if there were also significant differences in the
consumption in the control group (which received no
promotional campaign), we should be alert to the possibility
that the increase in sales has actually been caused by some
other effect (such as changes in the weather or a change in
consumer preferences). In any event, we can never be
absolutely sure that increase in consumption has been caused
by our intervention, but if all other reasonable causes have been
eliminated, we can make a good case for that assumption.

type I error: The error of incorrectly rejecting the Null Hypothesis

(Ho) when it is in fact true. For example, our sample might lead
us to conclude with, say, ninety-five percent confidence, that
athletes drink more soft drinks than non-athletes when, on the
average, the two groups actually consume the same number.
The probability of committing a type I error is denoted by the
letter α. Note that a type I error is not the result of an
inappropriate application of a statistical procedure. Rather, it
occurs as a result of our selection of a test which will lead to a
false positive a certain percentage of the time. We can adjust
this probability of false positives to our liking, but doing so
involves a tradeoff against the probability of failing to accept
the alternative hypothesis when it is in fact true. See also alpha,
type I error.

type II error: The error of incorrectly of failing to reject the null

 112
hypothesis (Ho) when the alternative hypothesis (Ha) is in fact
true. For example, our statistical test might lead us to
determine that there is not enough evidence to conclude from
our sample that athletes drink more soft drinks than non-
athletes; our sample may show differences, but those
differences may not large enough to be statistically significant.
Note that a type II error is not the result of an improperly
applied statistical procedure. Rather, it is the result of the finite
sensitivity of the test.

value labels: An optional SPSS/PC+ command which allows us to

assign labels to the values (i.e., in a questionnaire, the answers
to a question) associated with a variable. For example, for the
question "Did you consume one or more soft drinks today?" we
might assign the label of "Yes" to the value of "1" and "2" to the
value of "No." Some variables, such as age in years, should not
have value labels assigned since the answers are self-
explanatory. However, value labels might be used if the
variable is collapsed, as in the following example:
value labels
age 1 '0-15' 2 '16-25' 3 '26-40' 4 '40-55'
5 'Over 55'.

variable labels: An optional SPSS/PC+ command which allows us to

assign a "label," or brief description, to a variable. The variable
label is intended to help us identify the meaning of the variable
and has no effect on the computer's computations. If one
question on a survey asks how many soft drinks a person
consumes on an average day, a good variable label might be
"Soft drinks consumed on an average day" but, if by some
accident we typed in "Expenditures on Christmas presents," the
computer would not know that anything was wrong.

variable name: A name used by the computer to identify a variable.

For example, we might assign the name AGE to the variable
describing the subject's age. A variable name has to comply
with approximately the same rules as the ones that apply for
naming DOS files; i.e. a filename must be no longer than eight
 113
characters long, must begin with a letter of the alphabet, and
the remaining letters must be either letters, numbers, or certain
other characters such as underscore (_).

Computer Terms

ASCII file: A plain text file that can be edited. SPSS/PC+ data files
are ASCII files while system files, word processing files, and
executable files are normally not ASCII files.

cursor: The small blinking character that tells you where on the
screen you are editing.

file: An electronic means of storing data, normally on a disk. A file

has a file name and the data is accessed by referring to the
filename.

system file: An SPSS/PC+ file in which the data, value labels, and
variable labels are compressed so that they can be more
quickly accessed by SPSS/PC+. System files are normally only
worth the effort if the data set is very large. For more
information, see Appendix D.
 114
Appendix

Additional Sources of Information

There are both official and unofficial SPSS/PC+ manuals on the

market. The official manual put out by SPSS, Inc., comes in five
volumes, all of which are written by Marija Norusis:

• SPSS PC/+ Base Manual: Covers basic topics such as

frequencies, crosstabs, correlation, condescriptive analysis, and
data recoding and variable computing.

• SPSS/PC+ Advanced Statistics Manual: Covers more

advanced topics such as factor analysis, discriminant analysis,
and cluster analysis.

• SPSS/PC+ Version 3.0 Update Manual: Covers additions to

SPSS/PC+ made in Version 3.0 such as the inclusion of
Spearman Rank-Order Correlation, non-linear regression, and a
new data checking program.

• SPSS PC/+ Trend: Covers forecasting and time series

programs available within SPSS/PC+.

• SPSS PC/+ Tables: Covers special table and presentation tools

available within SPSS/PC+.

While the official manuals are good for reference purposes,

some more readable (but less detailed) unofficial books are also
available:

• Marija J. Norusis, The SPSS/PC+ Guide to Data Analysis

(Chicago, SPSS, Inc.: 1986).

• Neil Frude, SPSS/PC+ Introductory Guide (New York:

 115
Springer Verlag, 1987).