··-------·-·-~·

Chapter
Analysis of Variance and Covariance

Overview In Chapter 15, we examined tests of differences between two means or two medians. In this
(\nalysis of variance is a straightforward war_~ chapter, we discuss procedures for examining differences between more than two means or
medians. These procedures are called analysis of variance and analysis of covariance.
Although these procedures have traditionally been used for analyzing experimental data,
look at 9ifferences among more th?n tw~ they are also used for analyzing survey or observational data.
We describe the analysis of variance and covariance procedures and discuss their rel a.
tionship to other techniques. Then we describe one-way analysis of variance, the simplest of
_groups of responses measured on interval these procedures, followed by n-way analysis of variance and analysis of covariance. Special
attention is given to issues in interpretation of results as they relate to interactions, relative
importance of factors, and multiple comparisons. Some specialized topics, such as repeated
or ratio scales. ~~~~~~~~~~~
measures analysis of variance, nonmetric analysis of variance, and multivariate analysis of
variance, are briefly discussed.
Finally, we discuss the use of software in analysis of variance and covariance. Help for
Terry Grapentine, Grapentine Company, Inc. running the SPSS and SAS Learning Edition programs used in this chapter is provided in four
ways: (1) detailed step-by-step instructions are given later in the chapter, (2) you can down-·
load (from the Web site for this book) computerized demonstration movies illustrating these,
step-by-step instructions, (3) you can download screen captures with notes illustrating these
step-by-step instructions, and (4) you can refer to the Study Guide and Technology Manual,
Objectives [ After reading this chapter, the student should be able to:] a supplement that accompanies this book.

1. Discuss the scope of the analysis of variance (AN OVA) technique and its
relationship to the ttest and regression.
2. Describe one-way analysis of variance, including decomposition of the total
variation, measurement of effects, significance testing, and interpretation
of results.
3. Describe n-way analysis of variance and the testing of the significance of the
overall effect, the interaction effect, and the main effect of each factor.
4. Describe analysis of covar'1ance and show how it accounts for the ·influence
of uncontrolled independent variables
5. Explain key factors pertaining to the interpretation of results with emphasis
on interactions, relative ·importance of factors, and multiple comparisons.
6. Discuss specialized ANOVA techniques applicable to marketing such as
repeated measures ANOVA, nonmetric analysis of variance, and multivariate
analysis of variance (MAN OVA).
Real Research Analysis of Tourism Destinations
A marketing research survey conducted hy .EgeBank in Istanbul, Turkey, frn.:uscd on the importance of U.S.
tour operators' and travel agents' perceptions of sclc-cte<l Mediterranean tourist dcstina!10J18 (Egypt, Greece,
Italy, an<l Turkey). This study was conducted with the help of the Department of Tourism and Convention
Administration at the University of Nevada-Las Vegas (www.unlv.edu).
Operators/travel agents were mailed surveys based on the locations of tours, broken down as
follows: Egypt (53), Greece (130), Italy (150), and Turkey (65). The survey consisted of 4ucstions 011
affective and perceptual/cognitive evaluations of the four destinations. The four affechve 4.ucstions
were asked on a 7-point semantic differential scale, whereas the 14 rcrceptual/cognitive evaluations
were measured on a 5-point Liken scale ( I ;:::: offers very little, 2 = offers somewhat little, 3 ;:::: offors
neither little nor much, 4 = offers somewhat mud1, and 5 offers very much). The difference~ iri !he
evaluations of t!1c four locations were examined using one~way analysis of variance (A NOVA), as sc.cn
in the following table.
The ANOVA table shows that "unpleasant-pleasant" and "distressing-relaxing" affective factors
have significant differences among the four deslinations. For instance, Cireecc and Italy were pcrccivcU
w~ being significantly more relaxing than Egypt. As for the perceptual factors, eight of the I 4 factors
were significant. Turkey was perceived as a significantly hetkr value for the money than Greece
and Italy. Turkey's main strength appears lO be "good value," and the country's 1ourism agencies should

529
528
530 PART Ill • DATA COLLECTION, PREPARATION, ANALYSIS, AND REPORTING CHAPTER 16 • ANALYSIS OF VARIANCE AND COVARIANCE 531

Image Variations of Destinations Promoted to Tour Operators and Travel Agencies Real Research Electronic Shopping Risks
Turkey Egypt Greece Italy
= 34) Analysis of variance was u.se<l to test differences in preferences for electronic shopping for produc!s w 1ih
Image Items (n = 36) (n = 29) (n = 37) (n Significance
different economic and social risks. ln a 2 x 2 design, economic risk and social risk were varied at two
Affective (Scale 1-7) levels each (high, low). Preference for electronic shopping served as the dependent variable. The result:,;
6.50 indicated a significant interaction of social risk with econom1c risk. Electronic shopping was not per-
Unpleasant-pleasanl 6.14 5.62 643 0.047"
ceived favorably for high-social-risk products, rcgar<llc.'>s of the level of economic product risk, but it was
Sleepy-arousing 6.24 5.61 6.14 6.56 0.053
preferred for !ow-economic-risk products over high-economic-risk products when the level of social risk
Di.'itressing-rclaxing 5.60 4.86 6.05 6.09 0.003" was low.
Gloomy-exciting 6.20 5.83 6.32 6.71 O.U61 Despite the results of this study, the number of online shoppers continues to grow. As of 2008, more
than 875 million consumers had slwpped online, representing more th~m 80 percent of the world's onlrne
Perceptual (Scale I-SJ
population. The increase in shoppers can be attributed to hargain~secking consumers, convenience of using
Ooo<l value for money 4.62 432 389 3.27 0.000" the Internet, and, surprisingly, an added sense of safety associated with purchasing on!ine. Improved Web
Beautiful scenery sites, streamlined order taking and delivery, an<l assurance:,; of more secure payment systems have increased
and natural attractions 4.50 4.04 4.53 4.70 0.011" !he flow of new shoppers to the Internet while decreasing the traditional nsk associated with on line transac-
Good cli1nate 4.29 4.00 441 4.35 0.133 tion purclrnscs. 2 •
Interesting cultural attractions 4.76 4.79 4.67 4.79 0.781
Suitable acco1nmodation.<; 4.17 4.28 435 4.62 0.125 The tourist destination example presenteJ a situation with four categories. The t test was nut
analysis of variante
Appealing local food (cuisine) 4.44 3.57 4.19 4.85 0.000" appropriate for examining the overall difference in category means, so analysis of variance was
(ANO VA)
3.91 3.18 4.27 H,5 used instead. The electronic shopping stuJy involved a compari.'-.on of means when there
Great beaches and water sports 0.001" A '>tatist1cal technique for
4.(19 examining the d1fference5 were two factors (independent variables), each uf which was varied at two levels. In this
(Juahty of infrdstructure 3.49 2.97 368 ().0()()"

Personal safety 3.83 3.28 4.19 4.15 0.000"

Interesting hi~torical attractiom,
Unpolluted and
unspoiled environ111c1H
Goo<l nightlik·, and
enterL'.1inme11t
Standard hygiene and cleanliness
Interesting and friendly peopk
asig111f1cJnt at 0.05 level
promote this in their marketing strategies. On 1he other hand, Turkey needs to improve the perception of
its infrastnH..:ture, cleanliness, and entertainment to encourage more tour operators and travel agcm.:ies in
the United States lt) offer travel packctges to Turkey. ln 2007, 2J million tourists visited Turkey and
tourism revenues amounted to U.S.$! 8 hillion. 1 •
Analysis of vanancc
tedrni4ues can help identify
affective and perceptual
factors that differentiate
:iltemalive tou1ist
dest1rwtions.
among means for two or example, t te.'lts were not appropriate, because the effect or
each fador was not independent of
more populations the effect of the other factor (in other words, interactions were significant). Analysis of variance
4.71 4.86 4.81 4 82 0.650
provided meaningful concluslons in these studies. The relationship of analysi~ of variance to the
factors
and other techniques is considered in the next section.
t test
:l.54 334 .143 :l.59 0.784 Categorical independent
variables. The independent
v.:iridbles must be all
344 3.15 4.116 4.27 0 000"
categonca! (nonmetrn) to Relationship Among Techniques
3.29 2.79 .l.7o 4.29 IJ.OO(Y' use ANOVA Analysi.<.; of variance and analysis of covariance are used for examining the differences in the
4.34 4.24 4.35 4.32 0.956
mean values of the dependent variahle associated with the effect of the controlled independent
treatment
variables, after taking into account the intlucncc of the uncontrolled independent variables.
In ANOVA, a particular
wmb1nation of factrn level:. Essentially, analysis of variance CANOVA) is used as a test of means for two or rnore ropula-
or categories tions. The nu!l hypothesis, typically, is that all mean.'i arc equal. For example, suppDsc the.
researcher were interested in examining whether heavy, medium, light, and nonusers of cereals
one-way analysis of differed in their preference for Total cereal, measured on a 9-point Likcrt scale. The null hypoth-
variance esis that the four groups were not different in preference for Total could he tcsteJ using analysis
An .~NOVA technique in
of variance.
which d1f're 1s only one
In its silllplcst form, analysis nf variance must have a dependent variahlc (prefercrn.:e for
factor
Total cereal) that is metric (measured using an interval or ratio scale). There rnus1 also be one O!
n-way analysis of more independent variublcs (product use: heavy, medium, light, and nonuser:-.). The independent
variance variables must be all categorical (nonmetric). Categorical independent variahJes arc also called
An ANOVA model where factors. A particular combination of factor level.s, or categories, is called a treatment. One-way
two or more factors are analysis of variance involves only one categorical variahle, or a single factor. The differences in
involved. preference of heavy, medium, light, and nonusers would he examined by l>nc-way A NOVA. fn
analysi5 of covariance one-way analysis of variance, a treatment is the s..ime as a factor level (medium users con.".ititute a
(AN COVA) treatment). ff lwo or more faL.:tors are involved, the analysis is termed n-way analysis of
An c1dvanced analysis of variance. If, in addition to product use, the researcher also wanted to examine the preference for
variance procedure in whkh Total cereal of customers who are loyal and those who are not, an n-way analysis of variance
the effects of one or more would be conducted.
metric-scaled extraneous If the sci of independent variables consists of both catcgoricul and metric variahlcs, the tech-
removed from
variable:. are nique is called analysis of covariance (ANCOVA). For example, analysis of covariance woulu
the dependent v,iriable be required if the researcher wanted to examine the preference of product use groups anJ loyally
before conducting the groups, taking into account the respondent~' attitudes toward nutrition and the importance they
ANOVA
attached to hrcakfast as a meal. The latter two variables would be measured on 9-poinl Likert
covariate scales. In this case, the categorical independent variables /product use and brand loyalty) are still
A metric independent referred to as factors, wher~as the metric-independent variables (attitude toward nutrition and
v<1r'1c!ble used in ANCOVA. importance attached to breakfast) arc refctTed to as covariates.
532 PART Ill • DATA COLLECTION, PREPARATION, ANALYSIS, AND REPORTING
CHAPTER 16 • ANALYSIS OF VARIANCE AND COVARIANCE 533
FIGURE 16.1
lffl~rmnmm:nrm Statistics Associated with One-Way Analysis of Variance
Relationship
Between t Test, 2
eta (,r). The slrength of the effects of X (independent variable or factor) on Y ( <le pendent
Analysis of
Variance, Analysis
l variable) is measured by eta 2 (1/). The value of (,7 2) varies between O and I.

of Covariance, and one 1ndepen<lent One.;, More Independent ' F statistic. The null hypothesis that the category means are equal in the population is
Regression , Variable Variables tested by an F statistic based on the ratio of mean square related to X and mean square
relaied to c1rnr.
1
l
Binary>'·
i~·
;, , <:;ategorical:
!
Categ?rical
!
;Interval
Mean square. The mean square is the sum of squares divided by the appropriate degrees of
freedom.
SSb,twe,,.· Also denoted as SSx, this is the variation in Y related to the variation in the
,,! '
Faqtori~l , '
,and Interval
+
rTest
Afi,ilyil~ot , Analysis of
i t
Regression
mean.-; of the categories of X. This represents V[lriation between the categories of X, or the
portion of the sum of squares in Y related to X

Variance Covariance , ,"','Swithin' Also referred to aaS SSerror' this is the variatiun in Y due to the variation within
each of the categories of X. This variation is not accounted for by X.
_j
r----~l S5> The total variation in Y is S,\.

,, ,·,<','.·;.,":/:,.·,<·; :-.·"·i '·'More Than

On~. Factor · One·Factor,,,,. Conducting One-Way Analysis of Variance
( l The procedure for conducting one-way analysis of variance is descriheJ i11 Figure 16.2.
One'Way An~Jysi;; N,Wa{ Analysis Jt involves identifying the dependent and independent variahle.,;, decomposing the total variation,
of Variance
" ' " ,'.·»
of Variance · 'J measuring effects, testing significance, and interpreting rL:suhs. We consider these steps in detail
and illustrate them with some ,applications.

Identify the Dependent and Independent Variables

The relationship of a11alysis of variance to t tests and other techniques, such as regression
The c.JcpL:nJent variable is denoted by Y and the independent variable hy X. X is a categorical
(sec Charter 17), is shown in Figure 16. 1. All of these techniques involve a metric dependent
variahl~ having c categories. There are n observations on Y for each categmy of X, as shown in
variable. ANOVA and ANCOVA can include more than one independent variable (prnduct use,
Table 16.J. As can be seen, the sample sit.e in each category of Xis n, and the total sa1nple site
hrand loyally, attitude, and importance). Furthermore, at least one of the independent variables
N = n x c. Although the sample sizes in the categories of X (the group sizes) arc assunled to be
must be categorical, and the categorical variables may have more than two categories (in our equal for the sake of simplicity, this is not a requirement.
cxamrle, product use has four categories). At test, on the other hand, involves a .~ingle, binary
independent variable ..For example, the difference in the preferences of loyal and nonloyal Decompose the Total Variation
respondents cL>u!d he tested by conducting at test. Regression analysis, like ANOVA and
In examining the differences among means, one-way analysis of variance involves the
ANCOVA, can also involve more than one independent variable. However, all the independent decomposition of decomposition of the total variation observed in the dependent variahlc. This variation is
variables arc generaJly interval scaled, although binary or categorical variables can be accommo- the total variation measured hy the sums uf squares corrected for the mean (S'S). Analysis nf variance is so
Jated using dummy variables. For example, the relationship between preference for 'lhtal cemal, In one"way ANOVA,
named because it examines the variability or variation in the sample (dependent variable)
attitude towarJ nutrition, and importance attached to breakfast could be examined via regression separation of HlQ v,mution
and, based on the variability, determines whether there is reason to helieve lhat the population
analysis, with preference for Total serving as the dependent variable and attitude and importance observed in the dependent
means differ.
as independent variahlcs. varic1b!e into the vanat1on
tlI" cJue to the 1ndepender1t The total variation in Y, denoted by S'SY, can he dccomposeJ into two components:
VJriJbles plu'.. the variation
S,~~ = sshc{\W't'n + ,')'5','W/(fi/11
One-Way Analysis of Variance due to error.

Marketing researchers are often interested in examining the differences in the mean values of
the dependent variable for several categories of a single independent variable or factor. For FIGURE 16.2
example: Identify the .dependent and independent variables,
Conducting One-Way
• Do the various segments c.Jiffer in terms of their volume of product consu1nption?
• Do the brand evaluations of groups exposed to different commercials vary?
ANOVA l
·Decompose the total variation.
• Do retailers, wholesalers, and agents differ in their attitudes toward the firm's distribution
policies'/
o How do consumers' intentions to huy the brand vary with different price kvel.s?
l
Measure the effects.
a What is the effect of consumers' familianty with the store (measured as high, medium, and
low) on preference for the store? l
Test the significance.
The answers to these and similar questions can be determined by conducting one-way analysis
of variance. Before describing the procedure, we define the important statistics associated with
one-way analysis of variance.3
l
' Interpret the results.
534 PART 111 • DATA COLLECTION, PREPARATION, ANALYSIS, AND REPORTING
CHAPTER 16 • ANALYSIS OF VARIANCE AND COVARIANCE 535

~,:111 .. ,, Yvalucs may be obtained by examining the variation between the means. (This process is the reverse
Decomposition of the Total Variation: One-Way ANOVA of determining the variation in the means, given the population variances.) lfthe population mean is
Independent Variable

Categories
X
Total
Sample
----1 the same in all the groups, then the variation in the sample means and the sizes or the sample groups
can be used to estimate the variance of Y The reasonableness of this estimate of the Y variance
depends on whether the null l1ypolhesis is true. If the null hypothesis is true and tl1e population
means are equal, the variance esti1nate based on between-group variation is correct. On tht-
X1 X2 x, x,. other hand, if the groups have different means in the population, the variance estimate based on
Y, Y, Y, Y, Y, between-group variation will be too large. Thus, by comparing the Y variance estimates based on
Within- Y, Y, Y, Y, Y, between-group and within-group variation, we can test the null hypothesis. Deeompositinn or the
Category Variation Tola I total vaiiation in this manner also enables us to measure the effects of X on Y.
variation
= sswithi11
= ssY Measure the Effects
Y11 Y11 Yn Yn YN The effects of X on Y arc measured by SS> Because SS, is related to the vmiation in the means or the
Category Mean Y, Y2 Y1 f, Y categories of X, the relative magnitude of S.''J'x increases as the differences among the means of Yin
the categories of X increase. Tl1e relative magnitude of SS, also increases as the variations in Y within
Hctwccn-Catcgory Variation
the categories of X decrease. The strength of the ctlects of X on Y are mcasu red as rollows:
= ,",' S belwcn1
2_~ ss, (SSy - SS,.,,m)
"IJ - SSy SSy
where the subscripts between and within refer to the categories or
X. SS 0 e,wt'-en is the variation in
2
Y related to the variation in the means of the categories of X. It represents variation between the The value of "1) varies between O and I. It assumes a value or O when all the category
categories of X. In other words, S,)' 1H'twem is the portion ot the sum of squares in Y related to means arc eLJUal, indicating that X has no effect on Y. The value of 11 2 will be I when lhcre 1:-.
the independent variahle or factor X. Por this reason, S.Sbnween is also denoted as SS.1: S5,'witliin no variability within each category of X but there is some variability between categories. Tilus,
2
is the variation in Y related to the variation withiu each category of X. ,)'c'·i'wirfun is not accounted "1) is a measure of the variati'o11 in Y that is explained by the independent variable X. Not only
fur hy X. Therefore it is referrcJ to as S.)'error· The total variatitlll in Y may be decomposed as: can we measure the effects of X on Y, but we can also test for their significance.

S,)'y = S.\'A + SScuor Test the Significance

where In one-way analysis of variance, the interest lies in testing the null hypothesis that the category
N means are equal in the pupulation. 4 In olher worJs,
ss, - ~(Y, - Y) 2
/--1
Ho: /L1 ~ /L2 = µ., = = µ.,.
ss, ~
"
L"(lj - Y) 2
1=1 Under the null hypothesis, SSx and ,)'Serror come from the same source of variation. In .such a case,
t' II
the estimate 0f the population ,·ariance of Y can he based on either between-category variation or
SSmm ~ L L(Y, 1 - Y/ within-category variation. In other words, the cstim<.1tc of the population variance of};
f 11--1
Y,. = individual observation
Yj ,ec: mean for category j <;
2
= -----
ssx
' y (c - I)
Y =- mean over the whole sample, or graru..1 mean
0

Y11 = ith observation in the )th category

= mean ·"4 uare due to X
= MS,
The logic of decomposing the total variation in Y, SS into SShetwern and SS'within in order to
1 or
examine differences in group means can be intuitively understood. Recall from Chapter 15 that if
the variation of the variable in the population was known or estimated, one could estimate how
,z ,)'Serro,
much the sample mean should v<.1ry because of random variation alone. In unalysis of v,Lriance, ,ly ~ (N =-;_:)
there ,u-e several different groups (e.g., heavy, medium, light, and nonusers). Ir lhe null hypothesis
is true and all the groups have the same mean in the population, one can estimate how much the 1nean square due to error
sample means should vary because uf sampling (random) variations alone. If the observed variation = MSnror
in the sample means is more than what would be expected by sampling variation, it is reasonahle to
conclude that this extra variability is related to differences in group means in the population. The null hypothesis may he tested by the F statistic hased on the ratio between these two
In analysis of variance, we estimate two measures of variation: within groups (,)'5,'witlli,) and estimates:
between groups (,5'S'herwee,). Within-group variation is a measure of how much the observations, Y
values, within a group va1y. This is used to estimate the vatiance within a group in the population. SSxl(c - I) MS,
F = --------~ = - ------
1t is assumed that all the groups huvc the same variation in the population. However, because it is not SSnror /(N - c) MSerrm
known tlrnt all the groups have the same mean, we cannot calculate the variance of all the observa-
tions together. The variance for each of the groups must be calculated individually, and these are
This statistic follows the F distribution, with (c - I) and (N - c) degrees of fret'clom (elf).
combined into an ''average" or "overall" variance. Likewise, another estimate of the variance of the
A table of the F distribution is given as Table 5 in the Slatistical Appendix al the end of lhc book

- -- "·----"
536 PART 111 • DATA COLLECTION, PREPARATION, ANALYSIS, AND REPORTING CHAPTER 16 • ANALYSIS OF VARIANCE AND COVARIANCE 537

As mentioned in Chapter 15, the F distribution is a probability distribution of the ratios of sample
variances. It is characterized by degrees of freedom for the numerator and degrees of freedom for
the denorninator. 5
Interpret the Results
If the null hypothesis of equal category means is not rejected, then the independent variable does
not have a significant effect on the dependent variable. On the other hand, if the null hypothesis
is rejected, then the effect of the independent variable is significant. In other words, the mean
value of the dependent variable will be different for different categories of the independent
variable. A comparison of the category mean values will indicate the nature of the effect of the
independent variab1e. Other salient issues in the interpretation of results, such as examination of
differences among specific means, are discussed later.
analysis of variance is performed on a much larger sample such as that in the Dell running case
and other cases with real data that are presented in this hook. These data were generated by an
experiment in which a major department store chain wanted to examine the effect of the level of
in-store promotion and a storewide coupon on sales. In-store promotion was varied at three
levels: high ( l), medium (2), and low (3). Couponing was manipulated at lwo levels. Either a $20
storewide coupon was distributed to potential shoppers (denoted by l) or it was not (denoted by
2 in Table l 6.2). In-store promotion and couponing were crossed, resulting in a 3 x 2 design wilh
six cells. Thirty stores were randomly selected, and five stores were randomly assigned to each
treatment condition, as shown in Table 16.2. The experiment was run for two months. Sales in
each store were measured, normalized to account for extraneous factors (store size, traffic, etc.),
and converted to a 1-to-10 scale. In addition, a qualitative assessment was made of the relative
affluence of the clientele of each store, again using a I-to-IO scale. In these scales, higher
numbers denote higher sales or more affluent clientele.

Illustrative Data
Illustrative Applications of One-Way Analysis of Variance
We illustrate the concepts discussed in this chapter using the Jata presented in Table 16.2.
For illustrative purposes, we consider only a small number of observations. ln actual practice, We illustrate one-way ANOVA first with an example .showing calculations done by hand and
then using computer analysis. Suppose that only (lne fal'.lor, namely in-store promotion, wa~
manipulated, that is, let us ignore couponing for the purpose of this illustration. The department
11n1:1.-.. , ... store is attempting to determine the effect of in-store promotion (X) on sale.s (l'). For the purpose
Coupon Level, In-Store Promotion, Sales, and Clientele Rating
of illustrating hand calculations, the data of Table 16.2 are transl,,rrned in 1,tblc 16.3 to show the
Store Coupon In-Store Clientele store ( Y1) for each level of pro1J1otion.
Number Level Promotion Sales Rating The null hypothe.~is is that the category means arc equal:
10
Ho: /Lr = J-L2 ~ J-L1
10
SPSS Data File To test the null hypothesis. the various sums of squares arc computed as follows:
JO 8

.sas.
SAS Data File
8 4
6
8
4
ss,, = (Ill - 6067) 2 + (9 - 6.067) 2 + (JO - 6.067)2 + (8 - 6067)2
+ (8 - 6.067) 2 + (9 - 6.067) 2 + (7 - 6.067)2-+ (7 - 6.067)2
+ (8 - 6.067) 2 + (8 - 6.067) 2 + (7 - 6.067) 2 + (9 - 6.067) 2
+ (9 - 6.067)2
+ (6 - 6.067)2
+ (6 - 6.067)2

10

ti) 9 Effect of In-Store Promotion on Sales

II 8
Level of In-Store Promotion
12
Store No. High Medium Low
r:l 6
4 10 Normalized Sales
14
15 I 4 Ill 5
16 2 10 7
17 9 JO 6
18 4 8 4
t9 5
20 2 6 4 2
21 1
22
8 2
23 10 9
24 4 10 4 2

2:i 4 9 Coh11nn Total.s 83 62 37

2/, 2 83 62 37
Catego,y means:~'
27 6 10 10 10
28 10 = 8.3 - 6.2
29 l
CJ1a11d mean, Y (83 -t 62 -t ,~7 ) ~ 6067
10 2 2 30
 CHAPTER 16 • ANALYSIS OF VA.RIANCE AND COVARIANCE 539
538 PART Ill • DATA COLLECTION, PREPARATION, ANALYSIS, AND REPORTING

2 2 From Table S in the Statistical Appendix, we see that for 2 and 27 degrees of freedom, the
+ (4 - 6.067)2 + (5 - 6.067)2 + (5 - 6 067) 2 + (6 - 6.067) + (4 - 6 067) critical value of F is 3.35 for a = 0.05. Because the calculated value of F is greater than
2 2
+ (5 - 6.067) 2 + (7 - 6.067)2 + (6 - 6.0(,7) 2 + (4 - 6.067) + (5 - 6 067) the critical value, we reject the null hypothesis. We conclude thal the population means for the three
2
+ (2 - 6.067)2 + (3 - 6.067) 2 + (2 - 6.067) 2 + (I - 6.067)2 + (2 -· 6 067)
levels of in-store prnmotion are indeed different. The relative magnitudes of the means for the
2 three categories indicate that a high level of in-store promotion leads to significantly higher sales.
(3.933) 2 + (2.933) 2 + (3 933) 2 + (1.933) 2 + (2.933)
We now illustrate the analysis-of-variance procedure using a computer program.
+ (l.933) 2 + (2.933) 2 + (0.933)2 + (0.933)2 + (-0.067) 2 The resulls of conducting the same analysis by computer arc presented in Table 16.4. The value
+ (1.933)2 + (1.933) 2 + (0.933)2 + (2933) 2 + (-0.067) 2
of SSx denoted by between groups is I 06.067 with 2 df; that of SS,.,.,,,, denoted by within groups
+ (-2.067) 2 + (-1.067) 2 + (-1.067) 2 + (-0067)2 + (-2.067) 2
is 79.80 with 27 df. Therefore, MS, = 106.067/2 = 53.il33, and MS,.,,,,,= 79.80/27 = 2.956.
+ c--1.067) 2 + (0.933) 2 + c-0067)2 + c-2067) 2 + c-1.061J 2 The v;tlue of F = 53.033/2.95(, = 17.944 with 2 and 27 degrees of freedom, result',ng in a proh-
+ (-4.067) 2 + (-:1.067)2 + (-4 067) 2 + (-5.067) 2 + (-4.067)2
ahility of 0.000. Because the associated probability is less than the significance level of 0.05,
185 867 the null hypothesis of equal population means is rc.1ected. Alternatively, it can be seen from
Table 5 in the Statistical Appendix that the critical value of F for 2 and 27 degrees of freedom
ss, 10\8.3 - 6.0(,7) 2 + 10(6.2 - 6067) 2 + 10(37 - 6.067) 2 is 3.35. Because the calculated value of F (17.944) is larger than the critical value, the null
10(2.233) 2 ·f JO(O 133) 2 + 10(-2.367)2 hypothesis is rejected. As can he seen from Table 16.4, the sample means, with values of 8.3,
106.067 6.2, and 3.7, are quite different. Stores with a high level of in-store promotion have the highest
2 average sales (8.3) and stores with a low level of in-store promotion have the lowest average
(JO - 8.3) 2 + (9 - 8.3) 2 + (JO - 8.3) 2 + (8 - 8.3) 2 + (9 - 83)
ssenor + (8 8 :l)2 + (9 - 8.3)2 + (7 - 8 3)2 + (7 - 8.3) 2 + (6 - 8.])2
sales (3.7). Stores with a me(J'1um \eve! of in-store promotion have an intermediate level of
2 average sales (6.2). These findings sccn1 plausible. Instead of 30 stores, if this were a large and
+ (8 - 6.2)2 + (8 - 6.2) 2 + (7 - 6.2) 2 + (9 - 6.2) 2 + (6 - 6.2)
2 representative sample, the implications would be that rnauagement seeking to increase s;.iles
+ (4 - 6.2) 2 + (5 - 6.2) 2 + (5 - 6 2) 2 + (6 - 6.2) 2 + (4 - 6.2)
should emphasize in-store proniotion.
+ (5 - 3 7) 2 I (7 3.7)2 I (6 - 3.7) 2 + (4 - 3.7)2 + (5 - 3.7)2
The pnxedure for conducting one-way analysis of variance and the illustrative application
2
+ (2 - 3 7) 2 + (3 - 3.7) 2 + (2 - 3.7) 2 + (I - 3 7) 2 + (2 - l.7)
help us understand the assumptions involved.
2
( 1.7) 2 + (0.7) 2 + ( 1.7) 2 + (-0.3)2 + (0 7)
+ (--0.3) 2 + (0.7) 2 + (-l.3) 2 + (-13)2 + (-2.3) 2
ACTIVE RESEARCH
+ (l.8) 2 + (1.8)2 + (0.8)2 + (2.8) 2 + (-0.2)2 ·-~-,
+ ( 2.2)2 + (-1.2) 2 I- (-l.2) 2 t- (--1) 2) 2 f (-2.2) 2 Experts, Novices, and Nonusers of Home Computers: Are Their Psychographics j
+ ( 1.3)2 + (:UJ 2 + (2.3)2 + (0.3) 2 + (l.l)2 ~~~fl I
+ ( J.7)2 + (-0 7) 2 + (-1.7)2 + (-27)2 + (-1 7) 2 Visit www.hp.com and conduct an Internet search using a sean.:h ~ngine and your library's online data- 1
79.80 base to obtain information on computer usage in U.S. households. l
AR the marketing director for Hewlett-Packard, how would you segment the home computer market?
It can be verified that As a marketing research analy~t working for HP, how would you determine whether the three home
computer usage segments (experts, novices, and nonusers) differ in terms of each of IO psyd1ogr.1phic
SS=SS+SS
y x errot characteristics, each mea1rnred on a7-point .scale?

as fol]ows:

185.867 = I 06 067 + 79.80

•ffl:ff=1n,
The strength of the effects of X on Y are measured as follows: One-Way ANOVA: Effect of In-Store Promotion on Store Sales
Source of Sum of Mean F F
1)
2 -
-
ssx
-~- Variation Sauares df ""
\nt___ _
l::lrP Ratio Prob.
SSy ::::.:..=---- ------
lie.tween groups
106067 106067 53.033 17.'144 tl.000
(In-store pn)ll\oUon)
185 867 SPSS Output File
Within groups
= 0.571

§.sas.
(Error) 79.800 27 2.956
TOTAL 185 867 2<J b409
In other words, 57.1 percent of the variation in sales (Y) is accounted for by in-store promotion
(X), indicating a modest effect. The null hypothesis may now be tested. Cell Means
Level of
SAS Output File In-store Promotion Count Mean
SS,l(c - I) MS,
r = . .~ - - - · = ~·~
High (I) 10 8.300
,)'Serror/(N - c) MSerror
Medium (2) 10 6.200
F = 106 067/(3 - I) Low(J) JO 3.700
79.81)()/(30 - 3) TOfAL 30 6.067

= 17.944

I
 CHAPTER 16 • ANALYSIS OF VARIANCE AND COVARIANCE 541
540 PART Ill • DATA COLLECTION, PREPARATION, ANALYSIS, AND REPORTING

• Do educational levels (less than high school, high school graduate, some college, and college
Assumptions in Analysis of Variance graduate) and age (less than 35, 35--55, more than 55) affect consumption of a brand'?
The salient assumptions in analysis of variance can be summarized as follows. • What is the effect of consumers' familiarity with a department store (high, medium,
and low) and store image (positive, neutral, aml negativc) on preference for the slore?
I. Ordinarily, the categories of the independent variable are assumed to be lixed. Inferences
are made only to the specific categories considered. This is referred to as thefixed-ejjects
model. Other models are also available. In the random-effects model, the categories or In determining such effects, n-way analysis of variance can be used. A major advantage of this
treatments are considered to be random samples from a universe of treatments. Inferences technique is that it enables the researcher to examine interactions between the factors.
are made to other categ,orics not examined in the analysis. A mixed-ejjects model results interaction Interactions occur wl1en the effects of one factor on the dependent variable depend on the level
6 When assessing the
if some treatments are considered fixed and others random. (category) of the other factors. The procedure for conducting n-way analysis of variance is
2. The error term is normally distributed, with a zero mean and a constant variance. The error relationship between similar to that for one-way analysis of variance. The statistics associated with n-way analysis of
is not related to any of the categories uf X. Modest departures from these assumptions do two variables, an variance are also defined similarly. Consider the simple case of two factors, X 1 and X2 , having
1nteract1on occurs if the
not seriously affect the validity of the analysis. Furthermore, the data can be transformed categories cl and c2. The total variation in this case is partitioned as follows:
effect of X1 depends on
to satisfy the assumption of normality or equal variances. the level of X2, and
3. The error terms are uncorrelated. [f the error terms are correlated (i.e., the observations arc vice versa SS,0101 = SS due to X 1 + SS Jue to X2 + SS due to interaction of X 1 and X 2 + SSwirhin
not independent), the F ratio can be seriously distorted.
In many data analysis situations, these assumptions are reasonably met. Analysis of variance or
is therefore a common procedure, as illustrated by the following example.
SSy = SSx 1 + S5~ 2 + SSx 1x2 + S.S'erro,

A larger effect of X 1 will be reflected in a greater mean difference in the levels of Xt and a
larger SS,,. The same is true for the effect of X 2 . The larger the interaction between X1 and X 2 , the
Real Research Viewing Ethical Perceptions from Different Lenses larger SSx,, 2 will be. On the other hand, if X 1 and X2 arc independent, the value of SS,,, 2 will be
close to zcrn. 9
A survey was conducted to examine differences in perceptions ofethica\ issues. The data were obtained from multiple,/ The strength of the joint eltect of two factors, called the overall effect, or multiple is 1i2,
31 manage.rs, 21 faculty, 97 undergraduate hu~mcss st1Hknts, and 48 MBA students. /\s pmt of the survey, The strength of the joint measured as follows:
respondents were required tl) rate five ethical ite.ms on a ~calc of I = strongly agree and ) ~ slrong!y effect of two (or more)
Ji,..,agrce with ) representing a neutral rcspon::.t~. The mean:-; for each group are shown. 011e-w<1y aruilysis of factors, or the overall
(SS,, t- SSx, + S.,~ , )
variance was conducted to examine the signillcancc of differences between groups for each survey ltcm, and effect. muliiplc 1/ -·------ .. -·-
SSy
1 2

the F and p values ohtained an.: u \so shown.

--------- significance of The significance of the overall effect may be tested by an F test, as follows:
Graduate Undergraduate F p the overall effect
Item
Students Value Value A test that some differences
No. Survey Item Managers Faculty Students
---------- (SS,, + S.1·,_ + ssx,., )/df,,
exist between some of the F = , '
3.7 3.8 3.8 4.0 0.94 0.42 treatment groups. ,)'Se,rmldt~I
Students caught cheating shouli.l
receive an F.
4.1 3.4 3.8 3.5 2.2 0.09 ~~~~·-~kl/dl~!
Plagiarism should be reported
1.7 2.7 2.8 IX.3 ().00 SSerru 1 fdf,/
Stude.nt grades shou\J be raised 1.6
lo gel employer p<1y for course. M.'-,'.x[,X2,.t1X2
34 35 3.2 11.0 0.00 ---~-··--
Use of school printers for personal 4.5
MSe1mr
p1 inhng shnuld he stopped.
1.7 1.8 2.4 2.8 114 0.00
Course Wtlrk should be simplified
where
to accorn111odate weaker stu<lcnts.
dfn = c.lt::grl':cs of freedom for the numerator
(c 1 -- 1) + (c 2 - I)+ (c 1 - l)(c 2 - I)
The fouhngs in<licMing significant differences on th1ec ol the five ethics item!-. point to 1h<.: need for mrne ('{2 -1
comm.unication among the four groups so as to helter align perceptions of cthlcal issues in management dfd = degrees of freedom for the de11011unator
education. 7 • = N- C{'2
MS = mean square

significance of the If the 0vernll effect is significant, the next step is to examine the significance of the interaction
N-Way Analysis of Variance interaction effect
8 effect. Under the null hypothesis of no interaction, the appropriate F test is:
ln marketing research, one is often concerned with the effect of more than one factor simultaneously. the s1gnif1cance
A test of
For example: of the 1nteract1on between ss,"1.1.~)<lt~
two or more independent F = ~~<.,~;mr/df;t
• How do the consumers' intentions to buy a brand vary wlth different levels of price variables.
and different levels of Uistributinn? M,:!J:2
• How do advertising levels (high, medium, and low) interact with price levels (high,
M,)'error
medium, and low) to influence a hrand's sales?
 CHAPTER 16 • ANALYSIS OF VARIANCE AND COVARIANCE 543
542 PART 111 • DATA COLLECTION, PREPARATION, ANALYSIS, AND REPORTING

II
where
df,, ~ (c 1 - l)(c 2 - I) Two-Way Analysis of Variance
Jfd = N~ C1C2 Sum of Mean Sig.
Source of Variation Squares df Square F ofF w2
If the interaction effect is found to be significant, then the effect of X 1 depends on the level
of X , and vice versa. Because the effect of one factor is not uniform, b,1t varies with the level of Main Effects
2
the other factor, it is not generally meaningful to test the significance of the main effects. In-stnr~ proniotion 106.067 2 53 ()}3 54.862 0.000 0.557
SPSS Output File
However, it is meaningful to test the significance of each main effect of each factor if the interac- Coupon 53.:m I 53.:133 55. 172 0.000 0.280

signifirance of
the main effect
A test of the s1gnif1canre
tion effect is not significant. 10
The significance of the main effect of each factor may be tested as follows for X1:
SSx/df;,
F = ·-·-----
§.sas
SAS Output File
Combined
Two-way interaction
Model
Residual (Error)
159.400
3.267
162.667
23.200 24
3
2
5
53.133
1.633
32.53.1
0.967
54.966
1.690
33.655
()000
0.206
0.000

al the main effect for SSerrm I df~t

TCHAL 185.867 29 6.409
each 1r1d1v1dua! factor = MSx,_
Cell Means
M,\\,rror I n~store Promotion Coupon Count Mean
where lligh Yes 5 9.200
df=c-1 High No 5 7.400
" I
Medium Yes 5 7.600
dfd = N -- C1C2
Medium No 5 4 800
The foregoing analysis assumes that the design was orthogonal, or balanced (the number llf cases [,ow Yes .\ 5.400
in each cell was the same). If the cell size varies, the analysj.s becomes more complex.
Low No 5 2.01/0

Factor Level Means

Illustrative Application of N-Way Analysis of Variance Promotion Coupon Count Mean
Returning to the data of 1\thlc 16.2, let us now examine the effect of the level of in-store promo- High 10 8.100
tion and couponing on store sales. The results of running a 3 x 2 ANOVA on the computer are Medium 10 6.200
prescntccl in Table 16.5. For the main effect or level of promotion, the sum or squares SSxp' Low 10 3.700
degrees of freedom, and mean square MS).p are the same as earlier determined in Table 16.4. Yes 15 7.400
The sum of squares for couponing SSx< = 53.333 with I df, resulting in an identical value for the No 15 4.7.13
mean square MSxc· The eornhinc<l main effect is determin~d by adding the sum of squares due to
Grand Mean JO 6.067
the two main effects (SS,p + SSxc = 106.06 7 + 53.333 = 159.400) as well as adding the degrees
of freedom (2 + I ..:: 3). For the promotion and coupon interaction effect, the sum or s4uares
SSxpxc = 3.267 with (3 -· I) (2 - I) = 2 degrees of freedom, resulting in MS,pxc = 3.267/2 =
1.633. For the overall (model) effect, the sum of squares is the addition of the sum of squares for
promotion main effect, coupon main effect, and interacLion effect~ 106.067 + 53.333 + 3.267 The lest .')ti_1ti:-.t1c for the significance of the main effect or coupon is
= 162.667 with 2 + 1 + 2 = 5 degrees of freedom, resulting in a mean square nf 162.667/5 = F = (53.333/0.967)
32.533. Note, however, the error statistics are now different than in Table 16.4. This is clue to the
55.172
fact that we now have two factors instead of one. SS"'"' = 23.2 with (30 -- 3 x 2) or 24 degrees
of freedom resulting in MS"""'= 2.1.2/24 = 0.967. with I and 24 degrees of freedom, which is significant at the 0.05 kvel. Thus, a higher level of
The test statistic for the significance of the overall effect is promotion results in higher sales. The distribution of a storcwidc coupon results in higher ~ales.
The effect of each is independent of the other. If this were. a largl' an<l representative sample, tile
F = (32.'i33/0967)
implications are that management can increase sales by increasing in-store promotion and the
33.655 w;e of cnupons, independently of the other.
with 5 and 24 degrees of freedom, which is significant at the 0.05 level.
The test statistic for the significance of the interaction effect 1s

F ( I 633/0.967)
Real Research Country Affiliation Affects TV Reception
1.690
with 2 an<l 24 degrees of freedom, whlch is not significant at the 0.05 level. A study exJmincd the impact of country affiliatton on the cre<libilJty of product-Jttnbute cL1ims for TV~
Because the interaction effect is not significant, the significance of the main effects can be The dependent variahles were the following product-attribute claims: guod sound, reliability, crisr-clear
evaluated. The test statistic for the significance of the main effect of promotion is picture, and stylish design. The independent variables that wt1e manipulated consisted of price, country
affiliation, and store distribution. A 2 x 2 x 2 between-subject~ design was used Two levels of pnce,
F = (53.033/0 967) $949.95 (low) and$ l,249.()5 (high), two levels of country affiliation, Korea and the l Tnitcd .States, and two
54.862 levels of store dic;trihution, Best Buy and without Best Buy, were .<,pec1fied.

with 2 and 24 degrees or freedom, which is significant at lhe 0.05 level.

 CHAPTER 16 • ANALYSIS Of VARIANCE AND COVARIANCE 545
544 PART Ill • DATA COLLECTION, PREPARATION, ANALYSIS, AND REPORTING

Data were collected from two suburban malls in a large U.S. dty. Thirty respondents were randomly ACTIVE RESEARCH
assigned to each of the eight treatment cells for a total of 240 subjects. Table I presents the results for
manipulations that ha<l significant effects on each of the dependent variables. The Effect of Price and Quality on Preferences for Jeans
Visit www.levis.com and search the lntemet using a search engine as well as your library's online database
to find information on consumer preferences for jeans.
Table 1 Analyses for Significant Manipulations Levi's would like to conduct marketing research to increase its share of the jeans market Past studies
Univariate F df p suggest that the two most important factors determining lhe preferences fnr jeans are price (high,
Effect Dependent Variable
medium, and low) and qualily (high, medium, and low). What design would you adopt and what analysis
Good sound 7.57 l,232 0.006 would you conduct to detennine the effects of these factors on preference for jeans?
Country x price
Reliability 6.57 1,232 0.011 As Levi's marketing chief, what infonnation would you need to formulate strategies aimed at increasing
Country x price
Crisp-clear picture 6.17 1,232 0.014 market share?
Country x distribution
Reliability 6.57 1,232 0.01 I
Country x distribution
Stylish design 10.31 1,232 0.002
Country x distribution

Analysis of Covariance
When examining the differences in the mean values of the dependent variable related lo the effect
The directions of counrry-by-distribution interaction effects for the three depenc.lcnt variables are
of lhe controlled independent variables, it is often necessary to take into account the influence of
shown in Table 2. Whereas the credibility ratings for tlw crisp-clear picture, reliability, and stylish
unc<mtrolled independent variables. For example:
design claims arc improved by distributing the Korean-made TV set through Best Buy, rather than some
uthcr distributor, the same is not true or a U.S.-made set. Similarly, the directions of country-by-price * In determining how consumers' intentions to buy a brand vary wilh different levels of
interaction effects for the two dependent variables are shown in Table 3. At $1,249.95, the credibility price, attitude toward the brand may have to be taken into consideration.
ratings for the "good sound" and "reliabili!y" claims are higher for the U.S.-made TV set than for its
• In determining how different groups exposed to different commercials evaluate a brand,
Korean counterpart, hllt there is littk difference related to country affiliation wl1cn the prodm:t is priced
it may he ncccss,1ry to control for prior knowledge.
at $949.95.
• ln determining how different price levels will affect a household's cereal consumption,
it may be essential to take houschoJd size into account.

Table 2 Country-by-Distribution Interaction Means

In such cases, analysis of covariance should be w;;ed. Analysis ot covariance includes at least one
Country x Distribution Crisp-Clear Picture Reliability Stylish Design categorical inLicpe::tH..lent variahle an<l at least one interval or metric independent vaiiable. The cate-
--- gorical indepenUent variable is called afactvr, whereas the metric independent variahle is c..:alled a
Korea cov(1riate. The most common use of the t.:ovariatc is to re1110ve extraneous variation from the depen-
3.67 3.42 3.K2
Best Buy dent variable, because the effects of the factors are of major concern. The variation in the dependent
Without Bc~t Buy 3. lS 2.88 :us
variable due to the covariates is removed by an adjustment of the dependent variabh:'s mean value
United States within each treatrncnt condition. An analysis of variance is then performed on the adjusted scores. 12
3.60 3.47 3.53
Best Buy The significance of the combined effect of the covariates, as well as the effect of each covariate, is
Without Best Buy 3.77 3.65 3.75
tested by using the appropriate F tests. The coefficients for the covariates provide insigh!s into the
effect that the covariates exert on the dependent variable. Analysis of covariance is most useful when
lhe covariate is linearly rclalcd to the dependent variable and is not related to the facturs. 13
Table 3 Country-by-Price Interaction Means We again use the dala of Tahle 16.2 to illustrate analysis of covariance. Suppose that we
wanted to determine the effect or in-store promotion and couponing on sales while controlling for
Country x Price Good Sound Reliabilitt the effect of clienteJe. [tis felt that the affluence of the clientdc may also have an effect on sales of
the department store. The dependent variable consists of .stme sales. As before, promotion has three
$949.95
3.75 }40 levels and couponing has two. Clientelc measured on an interval scale serves as the covariate. The
Korea
United States 3.53 345 results arc shown in Table 16.6. As can he seen, the sum uf squares attributable to the covariate is
very small (0.838) with I df resulting in an identical value fur lhe mean square. The associated F
$1,249.95
2.90 value is 0.838/0 972 = 0.862, with I and 23 degrees or frecd,>Ill, which is nnl significant at the 0.05
Korea 3.15
3.73 3.67 level. Thus, the conclusion is that the alfluence of the dicntdc Jot."s not have an effect on the .sales
States
of the department store. H the effect of the covariate is significant, the sign of the raw coefficient
can be useLI to interpret the direclion oft he effecl on the depcndcllt variable.
This ~tudy demonstrates thal credibility of attribute claims, for products traditionally exported to the
United Slates by a <:ompany in a newly industrialized country, can he significantly improved if the same
company distributes the product through a well-known U.S. retailer aml considers making manufacturing Issues in Interpretation
investments in the United States. Specifically, three product attribute claims (crisp-clear picture, reliabil- lrnportant issues involved in the interpretation of ANOVA results i11clncle interactions, rebtivc
ity, and stylish design) are perceived as more credible when the TVs are made in Korea if they are also importance of factors, and multiple con1parisons.
distributed through a well-known U.S. retailer. Also, the "good sound" and "reliability" claims for TVs are
perceived to be more credible for a U.S-made set sold at a higher price, possibly offsetting the potential Interactions
disadvantage of higher manutacturing costs in the United States. Thus, Thonison, the manufacturer of
The Jiffcrent interactions that can arise when conducting ANOVA on two or more factors are
RCA p1oducls (www.rca.com), may be better off manufacturing its TV sets in the United St<1tes and
shown in Figure !6.3. One outcome is that ANOVA nrny indicate that there arc no intcractiuns
selling them at c:1 higher price. 11 •
(the interaction effects are not found tu be significant). The other possibility 1s that the interaction
-----~-·------------·
Chapter
Correlation is a simple but powerful way to look
at the linear relationship between two
metric variables. Multiple regression
extends this concept, enabling the
researcher to examine the relationship
between one variable and several others.
Jim McGee, Mission Research Specialist, Global Mapping International
Objectives [ After reading this chapter, the student should be able to:]
1. Discuss the concepts of product moment correlation, partial correlation,
and part correlation and show how they provide a foundation for regression
analysis.
2. Explain the nature and methods of bivariate regression analysis and describe
the general model, estimation of parameters, standardized regression
coefficient, significance testing, prediction accuracy, residual analysis, and
model cross-validation.
3. Explain the nature and methods of multiple regression analysis and the
meaning of partial regression coefficients.
4. Describe specialized techniques used in multiple regression analysis,
particularly stepwise regression, regression with dummy variables, and
analysis of variance and covariance with regression.
S. Discuss nonmetric correlation and measures such as Spearman's rho and
Kendall's tau.
560
Correlation and Regression
Overview Chapter 16 examined the relationship among the !test, analysis of variance and covariance, and
regression. This chapter describes regression analysis, which is widely used for explaining varia-
tion in market share, sales, brand preference, and other marketing results in terms of marketing
management variables such as advertising, price, distribution, and product quality However,
before discussing regression, we describe the concepts of product moment correlation and
partial correlation coefficient, which lay the conceptual Foundation for regression analysis.
In introducing regression analysis, we discuss the simple bivariate case first. We
describe estimation, standardization of the regression coefficients, testing and examination
of the strength and significance of association between variables, prediction accuracy, and
the assumptions underlying the regression model. Next, we discuss the multiple regression
model, emphasizing the interpretation of parameters, strength of association, significance
tests, and examination of residuals.
Then we cover topics of special interest in regression analysis, such as stepwise regres-
sion, multicollinec:irity, relative importance of predictor variables, and cross-validation. We
describe regression with dummy variables and the use of this procedure to conduct analysis
of variance and covariance.
Finally, we discuss the use of software in correlation and regression analysis. Help for
running the SPSS and SAS Learning Edition programs used in this chapter is provided in four
ways: (1) detailed step·by·step instructions are given later in the chapter, (2) you can down-
load (from the Web site for this book) computerized demonstration movies illustrating these
step-by-step instructions, (3) you can download screen captures with notes illustrating these
step-by-step instructions, and (4) you can refer to the Study Guide and Technology Manual,
a supplement that accompanies this book.
Real Research Regression Rings the Right Bell for Avon
Avon Products, Inc (www.avon.com), was having significant problem1> with the sales staff. The compa11y's
business, {kpcndent on sales representatives, was facing a shortage of sales reps without much hope of getting
new ones. Regression models were developed to reveal the prn,siblc variables thal were fuclmg this situation.
The models revealed that the most 1-lignificant variable was the level of the appointmenl foe that reps pay for
materials and second was the employee benefits, With data to back up its <Klions, t]1e company lowered the fee.
The company also hired .semor m:mager Michele Schneider to improve the way Avon infonneJ new hires of
their employee hcncfits program. Sdmei<lcr revamped Avon's benefits program information packet, which
yielded an infunnatjve and easy way to navigate "Guide to Your Personal Benefits." These changes re.suited in
an improvement in the recruitment and retention of sales reps. A:f. of 2009, Avon was the world'~ largest direct
.seller of beauty anU related products, selling products in more than l (X) countries. 1 •
Real Research Retailing Revolution
Many retailing experts suggest that electronic shopping will he the next revolution m retailing. Whereas many
traditional retailers experienced sluggish, single-digit sales growth in the 2000s, on! 1ne sales records were off lhe
chaits. Although e-tailing continues to make up a very small p01tion of overall retail salc:s (less than 5 per<.·cnt in
2009), the trend looks very promising for the future. A research pniject invcstiguting this trend looked for corre-
lates of consumers' preferences for ele.ctronic shopping scrvic~s. The ex.planation of consumers' preferences
was sough! in psychographic, demographic, and communication variables suggested in the literature.
561
562 PART Ill • DATA COLLECTION, PREPARATIO~, ANALYSIS, AND REPORTING CHAPTER 17 • CORRELATION AND REGRESSION 563

Good products, well-trained Division of the numerator and denominator by n - l gives

sales reps, and sophisticated
regression models have
opened the doors for Avon,
enabling it to penetrate the
±
i=l
(X; - X)(Y; -
n - 1
Y)
cosmetics market and
= ) ,;(X; - xJ2 " (Y,-- Y) 2
hccorne the world's largest
direct seller of beauty 2'.~--~
i=I
2'.~--
n - 1 i=l n - l
products.
COV,y
.'i'x Sy

covariance
A sy<;ternatic relationship
between two variables in
whtch a change 1n one
implies a corresponding
change 1n the otl1e1
(COV,,).
Multiple regression was used to analyL.e the data. The ove.rall multipk regression mo<ld was significant
at the 0.05 level. Univariate t tests indicated that the following variables ir1 the model were significant at
the 0.0.'i level or hctter: price unentation, sex, age, occupation, ethnicity, and education. None of the three
communication vanables (mass media, word of mouth, and publicity) was significantly rdated tn consumer
preference, the dependent varwhle.
The results suggest that electronic shopping is preferred by white females who arc older, bette1 educated,
working in supervisory or higher level nccupations, and price-oriented shoppers. Information of this type is
valuabk in targeting marketing efforts to electronic shoppers. 2 •
These examples illustrate some of the uses of regression analysis in determining which independent
variable~ explain a significant variation in the dependent variable of interest, the structure and form
of the relationship, the strength of the relationship, and predicted values of the dependent variable.
Fundamental to regression analysis is an understanding of the product moment correlation.
Product Moment Correlation
In these equations, X and Ydenote the sample means, and sx and sy the standard deviations.
COV,y, the covariance between X and Y, measures the extent lo which X and Y are related. The
covariance may be either positive or negative. Division hy ,\'xSy achieves standardization, so that r
varies between -1.0 and J .0. Thus, correlation is a special case of covariance, and is obtained
when the dala are standardized. Note that the correlation coefficient is an absolute number and is
not expressed in any unit of measurement. The correlation coefficient he.tween two variables will
be the same regardless of their underlying units of measurement.
As an example 1 suppose a rese<m.:her wants to explain attitudes toward a responllent's city of
residence in terms of duration of residence in the city. The attitude is measured on an 11-point scale
(I = do not like the city, l l = very muc:h like the city), mid the duration of residenc:e is measured in
terms of the number of years the respondent has lived in the city. In addition, importance attached to
the weather is also measured on an I I -point scale ( I = not important, I l = very important). In a
pretest of 12 respondents, the data shown in ll1ble 17. l arc obtained. For illustrative purposes, we
consider only a small number of observations so that we can show the calculations by hand. In actuaJ
practice, correlation and regression analyses are perfonned on a much larger sample such as that in
the Dell running case and other cases with real data that are presented in this book.
The C01Tclatio11 coefficienl may he calculated as follows:
+ 12 + 12 + 4 + 12 + 6 + 8 + 2 +- 18 + 9 -+- - -17- - -+
X= - - - - - ~ - - - - · - - -----------~--
12
= 9.333
- (6 + 9 + 8 + 3 + 10 + 4 + 5 +- 2 + l I + 9 + 1() i 2)
y = ------- · - - - - - - - ~ - ~ - - - - - - · ~ - - - · - - -
12
= 6.583

In marketing research, we are often interested in summarizing the strength nf association

II
hctween two metric variables, as in the following situations: .f!l:I~=-
Explaining Attitude Toward the City of Residence
• How strongly arc sale~ related to advertising expenditures?
• 1s there an association between market share and size of lhe sales force? Respondent Attitude Toward Duration of Importance Attached
• Are consumers' perceptions of quality related to their perceptions of prices? No. the City Residence to Weather
6 10
product moment In situations like these, the product moment correlation, r, is the most widely useJ statistic,
summarizing the strength of association between two metric (interval or ratio scaleJ) variahles, 12 II
correlation (r) SPSS Output File
A stdtistk summarizing the say X and Y. ft is an index t1se<l to Jetenninc whether a linear, or straight-line, relationship exists 12
strength of assooat1on between X and Y. It indicates the degree to which the variation in one variable, X, is related to the 4 3 4
between two rnetnc
variJbles.
variation jn another variable, Y. Because it was originally proposed by Karl Pearson, it is also
known as the Pearson correlation coefficient. lt is also refened to as simple correlation, bivariate
correlation, or merely the correlation coefficient. From a sample of n ohservations, X and Y, the
product moment correlation, r, can be cakulateJ as:
§.sas.
SAS Output File
Ill
4
12
6
8
2
II

11 18 H
"
2'.(X; - X)(Y, Y) 10 9 9 111
i=!
r= T " II 10 17

'y H\X; - X) 2;(Y, - y)2 12 2 2

----··--·-.J
564 PART Ill • DATA COLLECTION, PREPARATION, ANALYSIS, AND REPORTING
CflAPTER 17 • CORRELATION AND REGRESSION 565
n
L (X; - X)(Y; - Y) = (10 - 9.33)(6 - 6.58) + (12 - 9.33)(9 - 6.58) Hence, r2 measures the proportion of variation in one variable that is explained by the other.
1=]
Both rand r2 are symmetric measures of association. In other words, the correlation of X with Y
+ ( 12 - 9.33)(8 - 6.58) + (4 - 9.33)(3 - 6.58) is the same as the correlation of Y with X. It does not matter which variable is considered to he
+ (12 - 9.33)(10 - 6.58) + (6 - 9.33)(4 - 6.58) the dependent variable and which the independent. The product moment coefficient measures the
+ (8 - 9.33)(5 - 6.58) + (2 - 9.33)(2 - 6.58) strength of the linear relationship and is not designed to measure nonlinear relationships. Thus,
r = 0 merely indicates that there is no linear relationship between X and Y It does not mean that
+ (18 + (9 - 9.33)(9 - 6.58)
- 9.33)( 11 - 6.58)
X and Y are unrelated. There could well be a nonlinear relationship between them, which would
+ (17 - 9.33)(10 - 6.58) + (2 - 9.33)(2 - 6.58) not be captured by r (see Figure 17.1 ).
= -U.3886 + 6.4614 + 3.7914 + 19.0814 When it is computed for a population rather than a sample, the product moment correlation is
+ 9.1314 + 8.5914 + 2.1014 + 33.57]4 denoted by p, the Greek letter rho. The coeflicient r is an estimator of p. Note that the calculation
of r assumes that X and Y are metric variables whose distributions have the same shape. If lbese
+ 38.3214 - 0.7986 + 26.2314 + 33.5714 as.sumplions are not met, r is deflated and underestimates p. In marketing resemch, data obtained
179.6668 by using rating scales with a small nmnber of categories may not he strictly interval. This tends to
n
deflater, resulting in an underestimation of p. 3
L
i=l
(X; - X) 2 = ( JO - 9.33) 2 + (12 - 9.33) 2 + (12 - 9.33) 2 + (4 - 933) 2 The statistical significance of the relationship between two variables measured by using t
+ (12 - 9.33)2 + (6 - 9.33) 2 + (8 - 9.33) 2 + (2 - 9.33) 2 can be conveniently tested. The hypotheses are:
2
+ (]8 + (9 - 9.33)2 + (]7 - 9.33)2 + (2
- 9.33) - 9.33) 2 Ho: p = 0
= 0.4489 + 7.1289 + 7.1289 + 28.4089 H1 : r, ct- 0
+ 7.1289 + 11.0889 + 1.7689 + 53.7289
+ 75.1689 + 0 1089 + 58.8289 + 53.7289 The test statistic is:
304.6668
n
I = r[_ri_=._2 ]1/2
I -
L(Y, y)2 = (6 - 6.58)
2
+ (9 - 6.58)2 + (8 - 6.58) 2 + (3 - 6.58) 2
,.2
1-=I
+ ( IO - 6.58) 2 + (4 - 6.58)2 + (5 - 6.58)2 + (2 - 6.58) 2 which has at distribution with n- 2 degrees of frcedom. 4 For the crnrelation coefficient calcl11ated
2 based on the data given in Table 17. I,
+ (11 - 6 58) + (9 - 6 58)2 + (10 - 6.58) 2 + (2 - 6.58)2
= 0.3364 + 5.8564 + 2.0164 + 12.8164 112

+- 11.6%4 + 6.6564 + 2.4964 + 20. 9764 1 ~ 0.9361[- ~ :::_2_]

· I - (0.9361 )2
+ 19.5364 + 5.8564 + 11.6964 + 20.9764 = 8.414 .
= 120.9168
and the degrees of freedom = 12 - 2 = 10. From the t distribution table (Table 4 in the Statistical
Thus, Appendix), the critical value oft for a two-tailed test ancl n = 0.05 is 2.228. Hence, the null
hypothesis of no relationship between X and Y is rcjecle.J. This, along with the positive sign of r,
179.6668 indicates that attitude toward tllc city is positively related to the duration of rcsiden{.'e in the city.
r = ~--:::-~--_ - -----. - -
\/(304.6668)(120.9 ](18) Moreover, the high value of r indicates that this rek1tionship is strong. If this were a large and
= 0.9361 repre&cntalive sample, the implication woulJ he that managers, city officials, and politicians
wishing to reach people with a favorable attilude toward the city should target long-lime resi-
In this example, r = 0. 9361, a value close to 1.0. This means that respondents' duration of dents of that city.
residence in the city is strongly associated with their attitude toward the city. Furthermore, the ln conducting multivariate data anaJy,.·~is, il is often useful to examine the simple correlation
positive sign of r implies a positive relationship; the longer the duration of residence, the more between each pair of variables. These results are presented in the form of a correlation matrix,
favorable the attitude and vice versu. which indicates the coefficient of correlation between each pair of variables. Usually, only the
Because r indicates the degree to which variation in one variable is related to variation in lower triangular portion of the matrix is considered. The diagonal elements all equal I .00,
another, it can also be expressed jn terms of the decomposition of the total variation (see Chapter 16 ). because a variable correlates perfectly with itself. The upper triangular portion of the 1natrix is a
In otber words, mirror irn.:igc of the lower triangular portion, because r is a symmetric rne.1sure of association.
The form of a correlation matrix for five variables, V 1 through V5, is as follows.
Explained variation
r 2 = -~---------~
Total variation
FIGURE 17.1
ssx
ss,,
A Nonlinear
Relationship for
y~~-\
4 • I
•
Total variation - Error variation
---------------------
Which r = 0 3 \ I
~ .\_!;
Total variation
SSy SSermr I I
ss, -3 -2 -1 0 I 2 3
X

-,

566 PART 111 • DATA COLLECTION, PREPARATION, ANALYSIS, AND REPORTING
p,;Hti(d correldlion
<:oefficient
A lllf'dSlHP, of the
association between two
v,:mables after controlling
or ddJUSt1ng for the effects
of orie or rnore add1t1onal
varic1bles
CHAPTER 17 • CORRELATION AND REGRESSION 567
Partial con-elations have an order associated with them. The order indicates how many variables
v, v2 V3 V4 Vs are being acljusted or controlled. The simple correlation coefficient, r, has a zero-order, as it does
v, not control for any additional variables when measuring the association between two variables.
Vz 0.5 The coefficient r,y., is a first-order partial correlation coefficient, as it controls for the effect of one
v, 0.3 0.4 additional variable, Z. A second-order partial cmrelation coefficient controls for the effects of two
variables, a third-order for the effects of three variables, and so on. The higher-order pa,tial correla-
v. 0.1 0.3 0.6
tions are calculated similarly. The (n + !)th-order partial coetficient may be calculated hy replacing
V5 0.2 0.5 0.3 0.7
the simple correlation coeffkients on the right side of the preceding equation with the n th"order
pa1tial coefficients.
Partial correlations can be helpful for detecting spu,ious relationships (see Chapter 15). The
Although a matrix of simple correlations provides insights into pairwise associations,
relationship between X and Y is spurious if it is solely due lo the fact that Xis associated with Z,
sometimes researchers want to examine the association between two variables after controlling
which is indeed the true predictor of Y. In this case, the correlation between X and Y disappears
for one or more other variables. In the latter case, partial correlation should be estimated.
when the effect of Z is controlled. Consider a case in which consumption of a cereal hrand (CJ JS
positively associated with income([), with rci = 0.28. Because this brand was popularly priced,
income was not expected to be a significant factor. Therefore, the rese.archer .suspected that this
Partial Correlation
relationship was spurious. The sample results also indicated that income is positively associated
Whereas the product mmnent or sitnple correlalion 1s a measure of association describing the linear
with househoh.l size (H), rhi = 0.48, and that household size is assnciated with cereal consumption,
association between two variables, a partial correlation coefficient mea.sures the association
r,·h = 0.56. These figures seem to indicate that the real predictor of cereal consumption is not
between lwo variahles after controlling for or adjusting for the effects of one or more additional
income but household size. To test this assertion, the first-order partial correlation hetwccn cereal
variables. This statistic is used to answer the following questions: consumption and income is calculated, controlling for the effect of household size. The reader can
• How strongly are sales related to advertising expenditures when the effect of price is verify that this partial correlation, 'n.h' is 0.02, and the initial correlation hetween cereal consump-
controllc<l? tion and income varnshes when the household size is controlled. Therefore, the corrclatiun hetwecn
• Is there an association between market share and size of the sales force after adjusting for income and cereal consumption is spurious. The special case when a partial correlation is larger
the effect of sales promotion? than its respective zero-order con-elation involves a suppressor effect (see Chapter 15). 5
• Are consumers' perceptions of quality related to their perceptions of prices when the effect part correlation Another correlation coefficient of interest is the part correlation coefficient. This coefficient
of brand image is controlled? coefficient represents the correlation between Y and X when the linear effects of the other i ndependcnt
A rnea:iure of the variables have heen removed from X but not from Y The part correlation coefficient, ry(x. z)' is calcu-
As in these situations, suppose one wanted to calculate the association between X and Y correlation between Y and lated as follows:
after cont rolling for a third variable, Z. Conceptually, one would first remove the effect of X when the linear effects of
the other independE nt r xy ~ r xz r y,:.
Z from X. To do this, one would predict the values of X based on a knowledge of Z by using the 1
product moment correlation between X and Z, rxz· The pn:dicted value of Xis then subtracted vanables have been ry(tz) = \/J---~~1
removed from X but not
from the actual value of X to construct an adJustcd value of X. [n a similar manner, the values
from Y The part correlation between attitude toward the city and the duration of residence, when the lin-
of Y are adjusted to remove the effects of Z The product moment correlation between the
adjusted values of X and the adjusted values of Y is the partial correlation coefficient between ear effeds of the importance attached to weather have been removed from the duration resi- or
X and Y. after controlling for the effect of Z, and is denoted by rxy.t Statistically, because the dence, can be calculated as:
simple correlation between two variables completely describes the linear relationship between
0.9361 - (0.5495)(0.7334)
them, the partjal coJTelation coefficient can be calculated hy a knowledge of the simple corre-
latiom, alone, without using inJividual observations.
ry(x,.s,) - Vl-=- (O.'i495)'
rxy - (rx,)(ry,) 0.63806
r.x.vz = --·---- --------
VI- r;,Vl - r;,
To continue our example, suppose the researcher wanted to calculate the association between Real Research Selling Ads to Home Shoppers
attitude toward the city, Y, and duration of residence, Xi' after controlling for a third variable,
importance attached to weather, X 2 . These data are presented in Table 17. I. Tlie simple correla- Advertisements play a very im[XJ1tant rnlc in fonning attitudes/preferences for brands. Often adve1tisern use
tions bet ween the variables are: celebrity spokespersons as a credible source to influence consumers' attitude,<, and purchase intentions.
Another type of source credihi!ity is corporate credibility, which can also influence consumer reactions to
ry,, = D.93~1 ryx, = 0.7334 rx,x, = 0.5495 advertisements and shape brand attitudes. [n general, it has heen found that for low-involvement products,
attitude toward the adve11isemcnt mediates brand cognition (beliefs about the brand) and attitude toward the
The required partial correlation is calculated as follows: brand. What would happen to the effect of this mediating variable when produds are purchased through a
0.9361 - (0.5495)(0.7334) home shopping network? Home Shopping Budape~t in Hungary conducted research to assess tile impact or
advertisements toward purchase. A surw.y was conducted where scvernl measures we1e taken, such a'> attitude
ryx,.x, - VI - (0.5495) 2 Vl - (0.7334)2 toward the product, attitude toward lhc hrand, attitude toward the ad characlcristics., brand cognitions, and St)
= 0.9386 on. It was hyrothcsized that in a home shopping network, advertisements largely <lctenuined attitude toward
the brand. In order to find the degree of association of uttitudc. toward the ad with both altitude towarci
As can be seen, controlling for the effect of importance attached to weather has little effect on the the brand and brand cognition, a partial correlation coefficient could he computed. Tbe p,utlal corrclatioll
associalion between attitude toward the city and duration of residence. Thus, regardless of the would be calculated between attitude toward the brand and brand cognition .iftcr controlling for the
importance they attach to weather, those who have stayed in a city longer have more favorable effects of attitude toward the ad on the two variables. If attitude toward the ad is ~ignificantly high, then the
attitudes toward the city a11d vice versa.
partial correlation coefficient should be significantly less than the product moment con'Clation between hrand
I
568 PART Ill • DATA COLLECTION, PREPARA:'10N, ANALYSIS, AND REPORTING
nonmetl'ic. correlation
A correlation measure for
two nonmetric variables
that relies on rankings to
compute the correlation
regression analysis
A stat1st1C.al procedure for
analyzing dS'.>Ociat1ve
rcldt1onsh1ps between a
metric dependent vcir1cJble
and one or more
1ndepend('nt variables
bivariate regression
A procedure for deriving a
rnathematKal relationship,
1n the form of an equation,
between a single metric
dependent va1·1able and a
smgle metric indepenrlerit
variable
cognition and attitude toward the brand. Research was conducted that supported this hypothesis. Then, Saatchi
& Saatchi (www.saatchi.com) designed the ads aired on Home Shopping Budapest to generate positive
6
attitude toward the advertising, and this turned out to be a major competitive weapon for the network.
The partial correlation coefficient is generally viewed as more important than the part correlation
coefficient because it can be used to determine spurious and suppressor effects. The product
moment correlation, partial correlation, and the part correlation coefficients all assume that the
data are interval or ratio scaled. If the data do not meet these requirements, the researcher should
consider the use of nonmetric correlation.
Nonmetric Correlation
At times, the researcher may have to compute the correlation coefficient between two variables
that are nonmetric. It may be recalled that nonmetric variables do not have interval or ratio scale
properties and do not assume a normal distribution. If the nonmetric variables are ordinal and
numeric, Spearman's rho, P.,·, and Kendall's tau, T, are two measures of nonmetric correlation
that can be used to examine the correlation between them. Both these measures use rankings
rather than the absolute values of the variables and the basic concepts underlying them are quite
similar. Both vary from-I .0 to I .0 (see Chapter 15).
In the absence of ties, Spearman's Ps yields a closer approximation to the Pear.son product
moment correlation coefficient, p, than Kendall's T. In these cases, the absolute magnitude of T
tends to be smaller than Pearson's p. On the other hand, when the data contain a large number of
tied ranks, Kendall's T seems more appropriate. As a rnle of thumb, Kendall's Tis to be preferred
when a large nurnher of cases fall into a relatively small number of categories (thereby leading to a
large number of ties). Conversely, the use of Spearman's p, is preferable when we have a relatively
larger number of categories (thereby having fewer ties). 7
The product moment as well as the pm1ial and part correlation coefficients provide a conceptual
foundation for bivariate as well as multiple regression analysis.
Regression Analysis
Regression analysis is a powerful and flexible procedure for analyzing associative relationships
between a metric dependent variable and one or more independent variables. It can he used in the
following ways:
1. Determine whether the independent variables explain a signiticant variation in the dependent
v:uiable: wl1ether a relationship exists
2. Determine how much of the variation in the dependent variable can be ex plaineU by the
independent variables: strength of the relationship
3. Determine the structure or form of the relationship: the mathematical equation relating
the independent and dependent variables
4. Predict the values of the dependent variable
5. Control for other independent variables when evaluating tht'. contributions of a specific
variable or set of variables
Although the independent variables may explain the variation in the dependent variable, this
does not necessarily imply causation. The use of the terms dependent or criterion variahle.s, and
independent or predictor variables, in regression analysis arises from the mathematical relationship
between the variables. These terms do not imply that the criterion variable is dependent on the
independent variables in a causal sense. Regression analysis is concerned. with the nature and
degree of association between variables and does not imply or assume any causality.
Bivariate Regression
Bivariate regression is a procedure for deriving a mathematical relationship, in the form uf an
e4uatiun, between a single metric dependent or criterion vmiable and a singJe metric independent
or predictor variable. The analysis is similar fo many ways to c.letermining the simple correlation
between two variables. However, because an equation has to be derived, one variable must be
CHAPTER 17 • CORRELATION AND REGRESSION 569
identified as the dependent and the other as the independent variable. The examples given earlier
in the context of simple correlation can be translated into the regression context.
•
• Can variation in sales be explained in terms of variation in advettising expenditures? What
is the structure and form of this relationship, and can it be rnodele<l mathemalically by an
equation describing a straight line?
• Can the variation in market share be accounted for by the size of the sales force?
• Are consumers' perceptions of quality determined by their pcrce.plions of price?
Before discussing the procedure for conducting bivariate regression, we define some important
statistics.
Statistics Associated with Bivariate Regression Analysis
The following statistics and statistical terms are associated with bivariate regression analysis.
Bivariate regression model. The basic regression equation is Y, = /3n + {3 1 X, + e,, where
Y = dependent or criterion variable, X = independent or predictor variable, {3 0 = intercept
of the line, [3 1 = slope of the line, and e; is the error term associated with the ith observation.
Coefficient of determination. The strength uf association is measured hy the coefficient
of determination, r2. It varies between O and I and signifies the proportion of the total
variation in Y that is accountec.l for by the variation in X.
Estima!ed or predicted value. The estimated or prcdicled value of Y, is 9, = a + hx,
where Y; is the predicted value of Y,, and a and bare estimators of /Jo and {3 1, respectively.
Regression coefficient. The estimated parameler bis usually referred to as the nonstan-
dardized regression coefficient.
Scattergram. A scatter diagram, or scattergram, is a plot of the values of !wo variables for
all thi: cases or observations.
Standard error of estimate. This statistic, SE/c'. is the standard deviation of the actual Y
values from the predicted 9 values.
Standard error. The standard deviation of h, SEb, -is called the standard error.
Standardized regression coefficient. Also termed the beta coejji<'icnt or beta weight, this
is the slope obtained by the regression of Yon X when the data are standardized.
Sum of squared errors. The distances of all the points from the regression line are squared and
added together to arrive at the sum of .squared errors, which is a measure of total en-or, ~f}
t statistic. A t statistic with n - 2 degrees of freedom can be used to test the null hypothe~is
that no linear relalionship exists between X and Y, or
Ho: /3 1 = 0, where t = -t-.
.SEi,
Conducting Bivariate Regression Analysis
The steps involved in conducting biv2riate regression analysis are described in Figure 17 .2.
Suppose the researcher wants to explain attitudes toward the city of residence m tenns of the
duration of residence (see Tabk 17.1). In deriving such relationships, it is often useful to first
examine a scatter diagram.
Plot the Scatter Diagram
A scatter diagram, or scattergram, is a plot of the values of two variahle1-i for all the cases or
observations. It is customary to plot the dependent variable on the vertical axis and the indepen-
dent variable on the horizontal axis. A scatter diagram is useful for determining the form of the
relationship between the variables. A plot can alert the researcher tu patterns in the data, or to
possihlc problems. Any unusual combinations of the two variables can he easily identified.
 ,-,----------------------------------------- ----·-- ·--·"-

Chapter
Discriminant and Logit Analysis

Overview
Often you have measured different groups of This chapter discusses the techniques of discriminant analysis and logit analysis, We begin
by examining the relationship of discriminant and logit analysis to analysis of variance
(Chapter 16) and regression analysis (Chapter 17), We present a model and describe the
metric variables. general procedure for conducting discriminant analysis, with emphasis on formulation,
estimation, determination of significance, interpretation, and validation of the results. The
procedure is illustrated with an example of two-group discriminant analysis, followed by an
Discriminant analysis is a useful way to example of multiple (three-group) discriminant analysis. The stepwise discriminant analysis
procedure is also covered. When the dependent variable is binary, the logit model can also
be used instead of two-group discriminant analysis. We explain the log it model and discuss
its relative merits versus discriminant and regression analysis.
answer the questions ... _are the groups Finally, we discuss the use of software in discriminant and logit analysis. Help for running
the SPSS and SAS Learning ,Edition programs used in this chapter is provided in four ways:
(1) detailed step-by-step instructions are given later in the chapter, (2) you can download
different? On what variables are they (from the Web site for this book) computerized demonstration movies illustrating these
step-by-step instructions, (3) you can download screen captures with notes illustrating these
step-by-step instructions, and (4) you can refer to the Study Guide and Technology Manual,
most different? . Can I predict which a supplement that accompanies this book.

---------~g_ro_u~p_a~p~e_r_s_cm belongs to using these Real Research Rebate Redeemers

variables? A study of 294 consumers wa.s undertaken to determine the correlates of rebate proneness, or the
chan-icteristics of cotu,urners who respond favorably to rehalc prorn()lJOns. The predictor variables were
four factors related to houscholJ shopping altitudes and behaviors, and selected demographic characteristics
Jomie Baker Prewitt, Senior Vice President/Director.
Multiple discrimimmt
Decision Sciences, Burke, Inc. analysis can help uJentify
the factors that diffen;nliatc
frequent users, light users,
and nonu/'.ers of rebate:,,

Objectives [ After reading this chapter, the student should be able to: J
1. Describe the concept of discriminant analysis, its objectives, and its applications
-~~~G-'6
~
in marketing research.
2. Outline the procedures for conducting discriminant analysis, including the
~\:.e~~si
.• ,,. .., .......,, c;ct'1 .
formulation of the problem, estimation of the discriminant function coefficients,
determination of significance, interpretation, and validation.
3. Discuss multiple discriminant analysis and the distinction between two-group
and multiple discriminant analysis.
4. Explain stepwise discriminant analysis and describe the Mahalanobis
procedure.
5. Describe the binary logit model and its advantages over discriminant
and regression analysis.

600 601
 PART 111 • DATA COLLECTION, PREPARATION, ANALYSIS, AND REPORTING
602
dis.crimini::lnt analysis
A technique for analyzmg
rnarket1no research data
wheri the criterion or
Uepemlent variable 1s
catt'qoricdl ,md the
wecii(tor or 1ndt:ipendent
vciric1bles are interval
1n ndture
di~triminant functions
The li11Pcir r ornb1nat1on
of independent vanables
develnped by d1scrnninant
ancJlys1s that will best
d1scr 1rn1nalP between
the c<1tE-'gorics of the
dependent vari<1blL'
tW(Jw9roup
di:;crirninant analysis
Ll1~rnrn1na11t analysis
techniq11e where the
criterion variable ha'.>
two (c1tegor1es
multiple dis.crimin.:mt
z.uM!ysis
[)1scnrnmant an,1ly">1'>
tt:"drnique where the
rnter,011 variable involves
three or more Cdtegories
CHAPTER 18 • DISCRIMINANT AND LOGIT ANALYSIS 603
(sex, age, and income). The depenctent variable was the respondent's degree of rebate proneness, of Relationship of Discriminant and Log it Analysis to ANOVA
which Lhrec levels were identified. Respondents who reported no rebate-triggered purchases during the
pasl 12 months were classified as nonusers; those who reported one or two such purchases as light users; and Regression
and those with more than two purchases, frequent users of rebates. Multiple discriminant analysis was The relationship among discriminant analysis, analysis of variance (ANOVA), and regression
used to analyze the data. analysis is shown in Table IS.I. We explain this relationship with an example in which the
Two primary findings emerged. First, consumers' perception of the effort/value relationship was the
researcher is attempting to explain the amount of life insurance purchased in terms of age and
most effective variable in discriminating among fre4uent, light, and nonusers of rebate offers. Clearly,
income. All three procedures involve a single criterion or dependent variable and multiple
rebate-sensitive consumers ..issociate less effort with fulfilling the requirements of the rebate purchase, and
they are willing to accept a 1datively smnller refund than other customers. Second, consumers who are predictor or independent variables. However, the nature of these variables differs. In analysis or
aware of the regular prices of products, so that they recognize bargains, are more likely than others to variance and regression analysis, the dependent variable is metric or interval scaled (amount of
respond to rehate offers life insurance purchased in dollars), whereas in discriminant analysis it is categorical (amount
These findings were utilized by Dell (www.dell.com) when it offered up to $150 ca.sh rebates on its of life insurance purchased classified as high, medium, or low). The independent variables are
notebook computers during April 2009. The company felt that this would encourage the rebate-sensitive categorical in the case of analysis of variance (age and income arc each classified as high,
customers lo choose Dell notebooks. 1 • medium, or low) but metric in the case of regression and discriminant analysis (age in years and
income in dollars, i.e., both measured on a ratio scale).
The rebate proneness example examined three groups (nonusers, light users, and frequent users Two-group discriminant analysis, in which the dependent variable has only two categories,
of rehates). Significant intergrour differences were found using multiple rredictor variables. is closely related to multiple regression analysis. In this case, multiple regression, in which the
An examination of differences across groups lies at the heart of the basic concept of discrimi- dependent variable is coded as a Oor I dummy variable, results in partial tegressiou coefficients
nant analysis. that are proportional to discriminant function coefficients (see the following section on the dis-
criminant analysis model). The nature of dependent and independent variables in the binary logit
model is similar to that in two-group discriminant analysis.
Basic Concept of Discriminant Analysis
Discri1ninant analysis is a technique for analyzing data when the criterion or dependent
variable is categmical and the predictor or imlcpendent variables are metric, i.e., measured on at
Discriminant Analysis Model
least interval scales. 2 For example, the depcnJent variable may be the choice of a brand of discriminant analysis The discriminant analysis model involves linear comhinations of the following form:
personal <.:omputcr (brand A, B, or C) and the independent variabJes may be ratings of attrihutes model
or PCs on a 7-point Liker! .scale. The objectives of discriminant analysis are as follows:
The statistical model on D = b0 + b 1X 1 + /J 2X2 + h1X1 + · · · + hv(,
which d1scnrninant and!ys1s
1s based where
I. Development of discriminant functions, or linear comhinations of the predictor or
independent variables, which will best discriminate between the catcgorie.s of the criterion D = discriminant score
or d~pendent variable (groups) h's = discriminant coefficient or weight
2. Examination of whether significant differences exist among the groups, in terms of the
X's = predictor or in<lcpcnde.nt variable
predictor vuriable.s
3. Determination of which predictor variahlcs contribute to most of the intergroup The coefficients, or weights (h), are estimated S() that the groups differ as much as possible
differences on the values of the discriminant function. This occurs when the ratio of hetween-grnup sum of
4. Classification of cases to one of the groups hase<l on the values of the predictor variables squares to within-group .sum of squares for the discriminant scores is at a maximum. Any other
5. Evaluation of the accuracy or
classification linear cornhination of the predictors will result in a smaller ratio.
We give a brief geometrical cxpo:-.ition of two-group discriminant analysis. Suppose we had
Discriminant analysis techniques arc described by the number of categories possessed by two groups, GI and G2, and each member of these groups was measured on two variables X I mH.l X 2.
the criterion variable. When the criterion variahle has two categories, the technique is known as A scatter diagram of the two groups is shown in Figure 18.1, where X 1 and X2 arc the two axes.
two-group discrirninant analysis. \\'hen three or more categories are involved, the technique is Members of CH arc denoted b), I and rnc.rnbcn-; of G2 by 2. The resultant dlipscs encompas.'-i
referred to as multiple discriminant analysis. The main distinction is that, in the two-group some specified percentage of the points (members), say 93 percent in each group. A straight line
case, it is possihle to derive only one discriminant function. In multiple discriminant analysis, is drawn through the two points where the ellipses intersect amt then projected to a new axis, J>. The
more than one function rnay be computed.'
Examples of discriminant analysis abound in marketing research. This tccbnilluc can be
•t!l:Jll:llf;
used to answer questions such as: Similarities and Differences Among ANOVA, Regression,
and Discriminant/Logit Analysis
• In terms of demographic characteristics, how do customers who exhibit store loyalty differ
from those who do not? Discriminant/
• Do heavy, medium, and light users of soft drinks differ in terms of thell" cunsumption AN OVA Regression Log it Analysis
of frozen foods'' Similarities
• What psychographic characteristics help differentiate between price-sensitive and Number of dependent variable~ One One One
non-price-sensitive huyers of groccrjes? Number of independent variahles Multiple Multiple Multiple
• Do the various market segments differ in their media consumption habits?
Differences
• In terms of lifestyles, what are the differences between heavy patrons of regional
department store chains and patrons of national chains? Nature of the dependent variables Metric Metric Categorical/B 1nary
What are the distinguishing characteristics of consumers who respond to direct mail Nature of the independent variables Categorical Metric Metric
solicitations?
604 PART 111 • DATA COLLECTION, PREPARATION, ANALYSIS, AND REPORTING
FIGURE 18.1
Xz ' computed by averaging the separate covariance matrices for all the groups.
A Geometric
Interpretation
GI ''
of Two-Group
Discriminant
Analysis
'
' Large values of A (near I) indicate that group means do not seem to be differenl. Small
XJ
G2'
l)
overlap between the univariate distributions Gl' and G2', represented by the shaded area in
Figure IX.I, is smaller than would be obtained by any other line drawn through the ellipses
rcpre.senting the scatter plDts. Thus, the groups differ as much as possible on the/) axis. Several
statistics are associated with discriminant analysis.
Statistics Associated with Discriminant Analysis
The imrn11ant statistics; associated with discriminant analysis include the following.
Cunonical correlation. Canonical correlation measures the extent of association between
lhc discriminant scores and the groups. It is a measure of association between the single
<li~criminant function and the set of dummy variables that (le fine the grour memhership.
Centroid. The centroid is the mean values for the discriminant scores for a particular
group. There are as many centroids as there are groups, because there is one for each
group. The means for a group on all the functions are the group centroids.
Class;Jlcation matrix. Sometimes also called confusion or prediction matrix, the classifica-
tion matrix contains the number of correctly classified and misclassified cases. The correctly
dassifieJ cases appear on the <liagonal, because the predicted amt uctual groups arc the
same. The off-diagonal elements represent cases that have been incorrectly clas:,;ifte<l. The
sum of the diagonal elements divided by the total number of cases represents the hit ratio.
Discrimirwntfunction coefficient;·. The discrirninant function coefficients (un.standardiied)
arc the multipliers uf variahlcs, when the variables are in the original units of measurement.
Discriminant scores. The unstandardized coefficients arc multiplied by the values of lhe
variables. These products are summeJ and a<ldcd to the constant term to obtain the
discriminant scores.
Eigenvalue. For each di~criminant function, the eigenvalue is the ratio of between-group to
within-group sums of squares. Large eigenvalues imply .superior functions.
F values and their significance. These are cakulatcd frum a one-way ANOVA, with the
grouping variable serving as the categorical independent variable. Each predictor, in turn,
serves as the metric dependent variable in the ANOVA.
Group means and group standard deviations. These are computed fnr each predictor for
each group.
analysis sample
Part of the tutdl sample that
1s used for estimation of the
d1scrirnt11dnt function
validation s.rmple
That part of the total
sa1nple used to check the
results of the est1mdtion
sample
FIGURE 18.2
Conducting
Discriminant
Analysis
CHAPTER 18 • DISCRIMINANT AND LOGIT ANALYSIS 605
Pooled within-group correlation matrix. The pooled within-group correlation matrix is
Standardized discriminant function coefficients. The stanclarJized discriminant function
coefficients are the discriminant function coefficients and arc used as the multipliers when
the variables have been standardized to a mean of Oand a variance of I.
Structure correlations. Also referred to as discriminant loadings, the structure correlations
represent the simple correlations between the predictors and the discriminant function.
Total correlation matrix. If the cases are treated as if they were from a 1inglc sample and
the conelations computed, a total correlation matrix is obtained.
Wilks' A. Sometimes also called the U statistic, Wilks' A for each predictor is the ratio of the
within-group sum of squares to the total sum of squares. lts value varies belween Oand 1.
values of A (near 0) indicate that the group means seem to be different.
The assumptions in discriminant analysis are that each of the groups is a sample from a
multivariate normal population and all of the populations have the same covariance tnntrix..
The role of these assumptions and the st.rtistics just described can be better understood by
examining the proceJur~ for conducting discriminant analy.,.;is.
Conducting Discriminant Analysis
The steps involved in conducting discri1ninant analysis consist of formulation, estimation,
determination of significance, interpretation, and vallclation (8ee Figure 18.2). The:,;c steps arc
discussed and illustrated within the context of two-group discriminant analysis. Discriminant
analysis with more than two groups is discussed later in this chapter.
Formulate the Problem
The first step in discriminant analysis is to formulate the problem by identifying the objec:tivcs,
the criterion variable, and the independent variahles. The criterion VJriahle must com,ist of lwo
or more mutually exclusive and col1ectivcly exhaustive categories. When the dependent variahlc
is interval or ratio scaled, it rnu!it first be converted into categories. For example, attitude toward
the brand, measured on a 7 -point scctle, could be catcgmized as unfavorable (I, 2, 3), neutral (4 ),
or favorable (5, 6, 7). Alternatively, one could plot the distribution of lhc depcndcut variable and
form groups of equal size by determining the appropriate cutoff points for each category. The
predictor variables should be selected based nn a theoretical model or previous rcscan::h, or, in
the case of exploratory research, the experience of the researcher should guide their sele.ction.
The next step is to divide the sample into two parts. One parl of the sample, caJled the
estimation or analysis sample, is used for estimation of the discriminant function. The other
part, called the holdout or validation sample, is reserved for validating the discriminant func-
tion. When the sample is large enough, it can he split in half. One half &ervcs as the analysis
Formulate the problem.,
i
.Estimate the discrimiri.ant
function coefficients.
i
Determine the significance
of the discriminant function.
i
Interpret the results.
i
Assess the validity of
discriminant analysis.
l
606 PART Ill • DATA COLLECTION, PREPARATION, ANALYSIS, AND REPORTING CHAPTER 18 • DISCRIMINANT AND LOGIT ANALYSIS 607

sample and the other is used for validation. The role of the halves is then interchanged and the IM:mlf:
analysis is repeated. This is called double cross-validation and is similar to the procedure
Information on Resort Visits: Holdout Sample
discussed in regression analysis (Chapter 17).
Often the distribution of the number of cases in the analysis and validation samples follows Annual Attitude Importance
the distribution in the total sample. For instance, if the total sample contained 50 percent loyal Resort Family Toward Attached to Household Age of Head Amount Spent on
No. Visit Income ($000) Travel Family Vacation Size of Household Family Vacation
and 50 percent non loyal consumers, then the analysis and validation samples would each contain
50 percent loyal and 50 percent nonloyal consumers. On the other hand, if the sample contained 50.8 4 7 3 45 M(2J
25 percent loyal and 75 percent nonloyal consumers, the analysis and validation samples would 63.6 7 4 7 55 H(J)
be selected to reflect the same distribution (25 percent versus 75 percent). 54.0 6 7 4 58 M(2)
Finally, it has been suggested that the validation of the disc1iminant function should be conducted 4 45.0 5 4 3 60 M (2)
repeatedly. Each tirnc, the sample should be split into different analysis and validation pmts. The
5 68.0 6 6 6 46 H(l)
discriminant function should be estimated and the validation analysis carried out. Thus, the validation
4 6 62.l 5 6 3 56 fl OJ
SPSS Data File assessment is based on a number of trials. More rigorous methods have also been suggested.
7 35.0 4 3 4 54 J.(I)

sas
To better illustrate two-group discriminant analysis, let us look at an example. Suppose we
49.6 5 3 5 39 L(I)
want to determine the salient characteristics of families that have visited a vacation resort
during the last two years. Data were obtained from a pretest sample of 42 households. Of these, 39.4 6 5 3 44 H (3)
JO households shown in Table 18.2 were included in the analysis sample and the remaining 12 10 37 .0 2 6 5 51 L(I)
SAS Data File ti 54.5 7 3 3 37 M(2)
[2 38.2 2 2 3 49 L(l)
l71{:i.'Sf:
Information on Resort Visits: Analysis Sample

II
Annual Attitude Importance
Resort Family Toward Attached to Household Age of Head Amount Spent I shown in Tahle 18.3 were rart of the validation sample. For illustrntive purposes, we consider
No. Visit
--- -------
Income ($000) Travel Family Vacation Size nf Uru 1cohrdrl
u , , ,uuA"u,u An Family Vacation
u t only a small numher of observations. In actual practice, discriminant analysis is performed on a
much larger sample such as that i;1 the Dell running case an<l other cases with real dala that arc
50.2 43 M(2)
presented in thls book. The househo!Js that visited a resort during the last two years are coded
70.3 4 61 11 (3)
as I; those that did not, as 2 (VISIT). Both the analysis and validation samples were balanced in
62.9 6 52 H (3) terms of VISIT. As can be seen, the analysis sample cont1Jins 15 households in each category,
48.5 5 36 L(I) SPSS Data File whereas the validation sample has six in each category. Data were also ohlaincd on annual
52 7
75.0
4<>.2
57.0
3
4
4

6
55
68
62
51
11(.l)
H (3)
M(2)
M (2)
§.sas. SAS Data File
family income (INCOME), attitude toward travel (TRAVEL, measured on a lJ-point scale),
importance attached to family vacation (VACATION, n1easurcd on a 9-point :-;cale), household
size (HSIZE), and age of the head of the household (AGE).

4 57 II (3) Estimate the Discriminant Function Coefficients

64.1
6 45 H (J) Once the analysis ~ample has been identified. as in Table J 8.2, we can estimate the discrimi-
tll I 68 I
44 H (1) direct method nant function coefficients. Two hroaU approaches arc availahlc. The direct method invoJve.s
JI I 73.4 6 7
An app1oach to estimating the discriminant function so that all the predictors are included simultaneously. In
12 I 71.9 ~ 4 64 H (3)
Uiscnmmant analysis that this case, eai;,;h independent variable is included, regardless of it.i, discriminating power. This
13 I 56.2 8 6 54 M(2) involves est1mc1t1ng the method is appropriate when, based on previous research ur a theoretical model, the researcher
I 49 ..l 2 3 56 H (3) d1srnm1nant function so
14 wants the discrimination to he hased on all the prcJictors. J\n aliernative approach is the
62.0 6 2 5X II (l) that all the predictors are
15 I stepwise. method. In stepwise discrin1inant analysis, the predictor variahl12s are entered
58 L(I) included sir11ult,:H1eously
16 2 32.1 4 3 sequentially, based on their ability to discriminate among groups. This method, described in
36.2 2 55 l.(I) stepwise discriminant more detail later, is appropriate when the researcher wants to select a ,o;,ubset of the predictors
17 2
t8 2 43.2 2 57 M(2) analysis for inclusion in the discriminant function.
2 50.4 2 4 37 M(2) D1scnminant analysis The results of running two-group discriminant analysis on the data of Table 18.2 using a
J9
M(2) 111 which the predictors popular computer program are presen!ed in Table 18.4. Some intuitive feel for the results may he
20 2 44.t 6 3 42
are entered sequentially obtained by examining the group means and standard deviations. It appears that the twn groups
:J8.3 6 2 45 L(I)
21 2 based on their ability
57 M(2) are more wiJe]y separated in terms of income than other variables. Ti1erc appears to be 111ore of
22 2 55.0 2 2 to disrnrrnnate between
L(I) a separation on the importance attached to family vacation than on attitude toward travel. The
23 2 46.I 51 thf> groups
difference between the two groups on age of the head of the household is small, and the standard
2 35.0 6 4 5 64 L(I)
24 deviation of this variable is large.
37.3 2 7 4 54 l.(t)
25 2 The pooled within-groups correlation matrix indicates low correlations hetween the predictors.
41.8 3 56 M (2)
26 2 Multicollinearity is unlikely to be a proble1n. The l'iignificance of the univariate F ratios indicates
57.0 2 36 M (2) that when the predictors are consitlernd individually, only income, impmiance Df vac:1tion, and
27 2
3.l.4 6 2 50 L(I) household size significantly differentiate between those who visited a resort and those who did not.
28 2
37.5 2 48 L(l) Because there are two groups, only one discriminant function is estimated. The eigenvalue
29 2
4 t.:l 2 42 L(t) associated with this function is 1.7862 and it accounts for I 00 percent of the explained variance.
\0 2
The canonical cone1ation associated with this function is 0.8007. The square of this correlation,
 Chapter Factor Analysis

Factor analysis allows us to look at grn1_:l_,p_s_ __ Overview In analysis of variance (Chapter 16), regression (Chapter 17), and discriminant analysis
(Chapter 18), one of the variables is clearly identified as the dependent variable. We now turn
to a procedure, factor analysis, in which variables are not classified as independent or depen-
of variables that tend to be correlated dent. Instead, the whole set of interdependent relationships among variables is examined.
--- This chapter discusses the basic concept of factor analysis and gives an exposition of the;
factor model. We describe the steps in factor analysis and illustrate them in the context of
to each other and identifL1Jnderlying principal components analysis. Next, we present an application of common factor analysis.
Finally, we discuss the use of software in factor analysis. Help for running the SPSS and
SAS Learning Edition programs used in this chapter is provided in four ways: (1) detailed
dimensions that ex£lain these step-by-step instructions are given later in the chapter, (2) you can download (from the Web
site for this book) computerized demonstration movies illustrating these step-by-step
instructions, (3) you can dowr.load screen captures with notes illustrating these step-by-step
correlations. instructions, and (4) you can refer to the Study Guide and Technology Manual, a supplement
that accompanies this book.
To begin, we provide some examples to illustrate the usefulness of factor analysis.
William n Neal, Senior Partner, SDR Consulting
Real Research Factor Analysis Earns Interest at Banks

Objectives [ After reading this chapter, the student should be able to:] H1)W do L:onsurners evaluate h,mks? Respondents in a survey were asked to rate the importunrc of 15 hank
attributes. A 5-point scale ranging from not important to very important was employed. These data were
1. Describe the concept of factor analysis and explain how it is different from analyzed via principal components analysis.
analysis of variance, multiple regression, and discriminant analysis. A four-factor solution resulted, with the factors heing labeled a.~ traditional ~ervices, (.;Ol1Vcnience,
vi~1hility, and competence. TradiLJonal services mclu<led interest mies on loans, reputation in the conununity,
2. Discuss the procedure for conducting factor analysis, including problem
low rates for checkrng, friendly and personal izeJ &crvice, easy-to-read monthly statements, and obtain.ability
formulation, construction of the correlation matrix, selection of an appropriate
of loans. Convenience was comprised of convenient branch location, convenient ATM locations, speed of
method, determination of the number of factors, rotation, and interpretation
service, and convenient hanking hours. The visibility factor mcluded recommendations frorn friends and
of factors.
3. Understand the distinction between principal component factor analysis Factor analysi'> helped
and common factor analysis methods. JPMorgan Chase & Co.
lo 1Jenl1fy the drn1cnsions
4. Explain the selection of surrogate variables and their application, with
consumers use to evaluate
emphasis on their use in subsequent analysis. banks and to Jevelop appro-
5. Describe the procedure for determining the fit of a factor analysis model priate nrnrketmg stiategies
c.·nablmg it to become one
using the observed and the reproduced correlations.
of the largest lJ .S. banks

634
635
636 PART Ill • DATA COLLECTION, PREPARATION, ANALYSIS, AND REPORTING CHAPTER 19 • FACTOR ANALYSIS 637

relatives, attractiveness of the physical structure, community involvement, and obtainability of loans. (Figure 19.1 ), we can select home is best place and football as independent variables, and
Competence consisted of employee competence and availability of auxiliary banking services. It was drop the other five variables to avoid problems due to multicollinearity (see Chapter 17).
concluded that consumers eva!uate<l banks using the four basic factors of traditional services, convenience,
visibility, and competence, and hanks must excel on these factors to project a good image. By emrhasizing All these uses are exploratory in nature and, therefore, factor analysis is also called exploratory
these factors, JPMorgan Chase & Co. became one of the largest U.S. hanks and hought the hanking opera- factor analysis (EFA). The technique has numerous applirntions in marketing research. For
1
tions of bankrupt rival Washington Mutual in September 2008. • example:

e It can be used in market segmentation for identifying the underlying variables on which to
Basic Concept group the customers. New car buyers might he grouped based 011 the relative emphasis
Factor analysis is a general name denoting a class of procedures primarily used for data reduc- they place on economy, convenience, performance, comfort, and luxury. This 1night result
faOor' :.rnalr,is tion and summarization. In marketing research, there may be a large numher of variables, most in five segments: economy seekers, convenience seekers, performance seekers, comfort
A class of procedures of which are correlated and which must be reduced to a manageable level. Relationships among seekers, and luxury seekers.
prinia1ily used for rlata sets of many interrelated variables are examined and represented in terms of a few underlying • In product research, factor analysis can be employed to determine the brand attributes that
reduction and factors. For example, store image may he measured by asking respondents to evaluatl! stores on int1uence consumer choice. 'foothpaste brands might be evaluated in terms of protection
summarization. a series of items on a semantic differential scale. These item evaluations may then be analyzed to against cavities, whiteness of teeth, taste, fresh breath, and price.
determine the factors underlying store image. • In advertising studies, factor analysis can he useJ to understand the media <..:onsumption
[n analysis of variance, multiple regressiun, and discriminant analysis, one variahle is habits of the target market. The users of frozen foods may be heavy viewers of cable TV,
considered as the dependL:nt or criterion variable, and the others as indepen<lent or predictor vari- see a lot or
movies, and listen to country music.
ables. However, no such distinction is made in factor analysis. Rather, factor analysis is an • In pricing studies, it can be used to identify the characteristics of price-sensitive consumers.
2 For example, these consumers might he methodical, economy minded, and home centered.
interdependence technique in that an entire set of interdependent relationships is examined.
in t1::nJepr~r1<lt!n<.<s? Factor analysis is used in the following circumstances:
tet..hniqw~
Mult1vJn,1te statistiral
1. To identify underlying dimensions, or factors. that explain the correlations among a Factor Analysis Model
set of variahles. For ex.ample, a set of lifestyle statements may be used to measure the
techniques 1n wh1d1 MathematicalJy, factor analy~is is smncwhat similar tn multiple regression ,malysis. in that each
the whole set of
psychographic profiles of consumers. These statements may then he factor analyzed
variable is expressed as a linear combination of llnclerlying factor!-;. The amount of variance a
1ntPrdepende11t to identify the underlying psychographic factors, as illustrated in the department store
variable shares with all other vmiablcs included in the analysis is referred to as communality. The
rl'lat1on'::.hiµ":> 1s example. This is atso illustrated in Figure 19.1 derived based on empirical analysis,
covariation among the variahles is described in terms of a small numher of common factors plus
exrlr111ned where the seven psychographic variables can be represented hy two factors. In this
a uniq11e faclor for each variable. These factors are not overtly observed. If the variables are stan-
figure, factor I can he interpreted as homebody versus socialite, and factor 2 can be
fattor·~ dardized, the factor model may he represented as:
interpreted as sports versus movies/plays.
An underly1nq d1n1ens1on
tllcll explain:, the
2. Tll identify a new, smalh.:r set of uncorrelated variables to replai.:<.·. the original set of cone lated X, = A;,F1 + A,2F2 + A,1F, + + A,mF'm + Vi l/1
variables in subsequent multjvariate analysis (regre~sion or discriminant analysis). For example,
<Orrl'ldtions amnnq a set where
of variable<..
the psychographic factors identified may be used as int..!epcndent variables in explaining the
diff~n.:nccs between loyal and non loyal consumers. Thus, instead of the seven correlated X1 ith standardized variable
psychographic variables (1fFigure 19.1, we caii use the two uncorrelated factors, A11 = standardi1.cd multiple regression coefficient of variable ion common factor j
i.e., homchody versus socialite, and sports versus movies/plays, in suhscqucnt analysis.
F common factor
3. To identify a smaller set of salient variables from a larger set for use in suhsequent multi-
V1 slandardi1.cd regression coeflicicnt of variahk ion uniqut' factor i
variate analysis. For example, a few of the original lifestyle statements that correlate highly
with the identified factors may be used as independent variables to explain the differences U 1 = the unique fac!or for variable i
between lhe loyal and nonloyal users. Spccilically, based on theory and empirical results m number of common fat:tors

The unique factors are uncorrelated with each other and wilh the common factors. 3 The
FIGURE 19,1 Factor 2
common factors themselves can be expressed as linear combinations of the observed vanahles.
Factors Underlying
Selected Baseball
F; = W,1X1 + W,2X2 + W,1X3 + + w,,x,
footbaJI
Psychographics where
and Lifestyles
F1 = estimate of ith factor
Evcnmg at home w, weight or factor score coefficient
number of variahles

It i.s possible to select weights or factor score coefficients so that the first factor ex.plains the
largest portion of the total variance. Then a second set of weights can be selected, so that the second
Go to a party Horne is best place
factor accounts for most of the residual variance, suhjecl to being uncunelated with the first factor.
This same principle could be applied to selecting additional weights fnr the additional factors. Thus,
the factors can he c:,;timated so that their factor scores, unlike the values of the original variahles, are
not correlated. Furthermore, the first factor an:oun1s for the highest variance in the data, the second
Plays factor the second highest, and so on. A simplified graphical illustration of factor analysis in the case
Movies of two variables is presented in Figure 19.2. Severn! statistics arc associated with factor analysis.
638 PART Ill • DATA COLLECTION, PREPARATION, ANALYSIS, AND REPORTING CHAPTER 19 • FACTOR ANALYSIS 639

r'IGURE 19.2 x, FIGURE 19.3

Fonnulate the problem.
Graphical r'r=W,1X1+W,2 X2 Conducting Factor
Illustration Analysis +
Construct the correlation matrix.
of Factor
Analysis
+
Determine the method of factor analysis.
/
+
Determine the number of factors.

L..--------+ x, +
Rotate the factors.

+
Statistics Associated with Factor Analysis Interpret the factors.

The key statistics associated with factor analysis are as follows: l

Bartlett~..,, test ofspherit.·ity. Bartlett's test of sphericity is a test statistic used to examine
tl1e hypothesis that the variables are uncorrelated in the population. [n other words, the
population correlation matrix is an identity matrix; each variable correlates perfectly with
itself' (r = I) hut has mi correlation with the other variables (r = 0).
Correlation matrix. A correlation matrix is a lower triangle matrix showing the simple cor-
relations, r, between all possible pairs of variables incluJcd in the analysis. The diagonal
elements, which are aJl 1, are usually omitted.
Communality. Communality is the amount of variance a variable shares with all the
other variahles heing cnnsidered. This is also the proportion of variance explained hy
the common factors.
Eigenvallle. The eigenvalue represents the total variance explained by each factor.
Factor loadings. Factor loadings are simple conelations between the variables and the factors.
Factor loading plot. A fador loading plot is a plot of the original variables using the factor
loadings as coordinates.
Factor matrix. A factor matrix contains the factor loadings of all the variables on all the
factors cxlraclcd.
Factor scores. Factor scores are composite scores estimated for each re~pondent on the
derived factors.
Factor scores coefficient matrix. This mntrix contains the weights, or fa<.:tnr score coeffi-
cients, used to comhine the standardized variables to obtain factor scores.
Kaiser-Meyer-Olkin (KMO) measure of sampling adequacy. The Kaiser-Meycr-Olkin
(KMO) measure of sampling adequacy is an index useLl to examine the appropriateness of
factor analysis. High values (ht'tween 0.5 and LO) indicate factor analy!'>is is appropriate.
Values helow 0.5 imply that factor analysis may not be appropriate.
Percentage of variance. This is the percentage of the total variance attributed to each factor.
Residua[f.i. Residuals me the differences between the ohscrved correlations, as given in the
input co1Telation matrix, and the reprodw...:ed correlations, as estimated from the factor matrix.
Scree plot. A scree plot is a plot of the eigenvalues against the number of factors in order
of extraction.
In the next section, we descrihe the uses of these statistics in the context of the procedure for
conducting facto1 analysis.
Conducting Factor Analysis
The steps involved in conducting factor analysis are illustrated in Figure 19.3. The first step is
to define the factor analysis problem and identify the variables to be factor analyzed. Then a
correlation matrix of these variables is constructed and a method of factor analysis selected.
l
Calculate the ·'select ihe
factor , surrogate
Scores. variables.
I l
I
Determine the model fit.
The researcher dt'ddcs on the number of factors to be extracted and the method of ro1ation.
Next, the rotated factors should be interpreted. Depending upon the objectives, the factor scores
may he. calculated, or surrogate variahles selected, to represent !he factors rn suhsequent niultr-
variatc analysis. Finally, the fit of the factor analysis model is determined. We discuss these
steps in more detail in the following scctions. 4
Formulate the Problem
Problem forrnubtion includes several tasks. Pirst, lhc objectives of factor analysis should be
idcntitieJ. The variables to be inclrn.Icct in the factor analysis should he specified based on past
research, theory, and judgment uf the researcher. It is important that the variable:,; be apprnpn-
ately measured on an interval or ratio scale. An appropnatc sample size should be used. 1\s a
rough guideline, there should be at lea~t four or five times as many observatiDns (sample size) as
there are variahlcs. 5 In many marketing n:search .situations, the :-.ample -~izc i:,, small and this ratio
is considerably lower. ln these cases, the results should be intcrpte.ted cautiously.
To illustrate factor analysis, suppose the rese~ud1cr wants to determine the underlying ht;ncfits
consumers seek from the purchase of a toothpaste. A sample of 30 respondents was interviewed
using mall-intercept interviewing. The respondents were asked to indicate their degree of agreement
with the following statements using a 7-point scale (I = strongly disagree, 7 . .:: :-.tnmgly agree):
V1: It is important to buy a toothpaste that prevents cavities.
V2 : l Jike a toothpaste that gives shiny 1ceth.
V1 : A toothpaste should .strengthen your gum.s.
V4 : I prefer a toothpaste that freshens breath.
V1 : Prevention of tooth decay is not an important benefit offered by a toothpaste..
V(); The most important consideration in buying a toothpaste is attractive teeth.
The data obtained arc given in Table 19.1. For illustrative i,urposes, we consider only a small
numhcr of ohservations. In actual practice, factor analysis is performed on a much larger sample
such as that in the Dell running case and other cases with real data that are presented in Lbis hook..
A c.:orrclation matrix wa.<.; construc.:tcd based on these ratings data.
 Cluster Analysis

Overview Like factor analysis (Chapter 19), cluster analysis exarrnnes an entire set of interdependent
relationships. Cluster analysis makes no distinction between dependent and independent
Cluster analysis helps us identify groups or variables. Rather, interdependent relationships between the whole set of variables are ex,m,-
ined. The primary objective of duster analysis is to classify objects into relatively homogeneous
groups based on the set of variables considered. Objects in a group are relatively similar in
segments that are more like each other terms of these variables and different from objects in other groups. When used in this manner,
cluster analysis is the obverse of factor analysis, in that it reduces the number of objects, not the
number of variables, by grouping them into a much smaller number of clusters.
than they are like members of other This chapter describes the basic concept of cluster analysis. The steps involved in
conducting cluster analysis are discussed and illustrated in the context of hierarchical cluster-
ing by using a popular computer program. Then an application of nonhierarchical clustering
groups or segme_11ts. is presented, followed by the TwoStep procedure and a discussion of clustering of variables.
Finally, we discuss the use of software in cluster analysis. Help for running the SPSS and SAS
Learning Edition programs used in this chapter is provided in four ways: (1) detailed step-by-step
Torn Myers, Senior Vice President, instructions are given later in the chapter, (2) you can download (from the Web site for this book)
Client Services Manager, Burke, Inc computerized demonstration movies illustrating these step-by-step instructions, (3) you can
download screen captures with notes illustrating these step-by-step instructions, and (4) you can
refer to the Study Guide and Technology Manual, a supplement that accompanies this book.
To begin, we provide some examples to illustrate the usefulness of cluster analysis.
Objectives [ After reading this chapter, the student should be able to: ]

1. Describe the basic concept and scope of cluster analysis and its importance Real Research Ice Cream Shops for "Hot" Regions
in marketing research.
H.iagen-Dan Shoppe Co. (www.liangen-da,.s.com), with more than 850 retail ice cream shups in over
2. Discuss the statistics associated with cluster analysis. 50 countries 111 2009, was interested in expanding its ~ustomer ba~c. The ohJcdive was to identify potcn·
3. Explain the procedure for conducting cluster analysis, including formulating tial consumer segments that could generate additional sales. Geodemography, a method of clustering
the problem, selecting a distance measure, selecting a clustering procedure,
deciding on the number of clusters, and interpreting and profiling clusters. Hiwgcn-Daz~ increased its
penetration hy idcnt1fy1ng
4. Describe the purpose and methods for evaluating the quality of clustering gcode1rnigraphic cl11stc1s
results and assessing reliability and validity. uffcring potential for
increased ice cre,1m sales.
5. Discuss the applications of nonhierarchical clustering and clustering of
variables.

661
660
 PART Ill • DATA COLLECTION, PREPARATION, ANALYSIS, AND REPORTING
CHAPTER 20 • CLUSTER ANALYSIS 663
662

consumers based on geographic, demographic, and lifestyle characteristics, was employed for this
purpose. Primary research was conducted to develop demographic and psychographic profiles of Hiiagen-
Dazs Shoppe users, including frequency of purchase, time of the day they came in, day of the week, and
other product use variables. The addresses and zip codes of the respondents were also obtaine<l. The
respondents were then assigned to 40 geodemographic clusters based on the clustering procedure
developed by Nielsen Claritas (www.claritas.com). For each geodemographk cluster, the profile of
Htiagen-Dazs customers was 1.:'.0mpared to the cluster profile to determine the degree of penetration. Using
this information, Htiagen-Dazs was also able to identify several potential customer groups from which to
attrnct traffic. Jn addition to expanding Haagen-Dazs' c,astomer base, product advertising was established
to target new customers accordingly. New products were introduced. As of 2009, the Hiiagen-Dazs brand
was owned by General Mills. However, in the United States and Canada, Haagen-Dazs products were
produced by Nestle under a preexisting license. 1 •
Both cluster analysis and discriminant analysis are concerned wilh classification. However,
discriminant analysis requires prior knowledge of the cluster or group membership for each
object or case included to develop the classification rule. In contrasl, in cluster analysis !here is
no a priori information about the group or cluster membership for any of the objects. Groups Dr
clusters arc suggested by !he data, not defined a priori. 4
Cluster analysis has been used in marketing for a variety of purposes, inclmling the following:'
• Segmenting the market: For example, consumers may be cl11steretl on the basis of benefits
sought from !he purchase of a product. Each cluster would cousisl of consumers who are
relatively homogeneous in terms of the benefits !hey seek. 6 This approach is called benefit
segrnentation.

The Hiiagen-Da1,s example illustrates the use of clustering to arrive at homogeneous segments
for the purpose of formulating specific marketing strategies. Real Research The Vacationing Demanders, Educationalists, and Escapists

Basic Concept
Cluster analysis is a class of techniques used to classify objects or cases into relatively homoge-
neous groups called clusters. Objects in each cluster tend to be similar to each other and dissim-
ilar to ohjects in the other clusters. Cluster analysis is also called classijh:atum unalysis, or
numerical taxonomy. 2 We will he concerned with clustering procedures that assign each object to
one and only one cluster. 3 Figure 20. l shows an ideal clustering situation, in which the clusters
are distinctly separated on two variables: quality consciousness (variable I) and price sensitivity
(variable 2). Note that each consumer falls into one cluster and there are no overlapping areas.
Figure 20.2, on the other hand, presents a clustering situation that is more likely to be encoun-
tered in praclice. In Figure 20.2, the boundaries for some of the cluslers are not clear-cut, and the
classification of some consumers is not obvious, hecause many ol them could be grouped into
one cluster or another.
FIGURE 20.1
An Ideal Clustering
Situation 0 0
In a study examining decision-making patterns nmong intern;Jtional vacationers, 260 resimndcnts provided
mforrna!ion on ~ix psychographic orientations: psychological, edw.:ational, social, relaxational, physiological,
and aesthetic. Cluster analysis w<1s used to group respondents into psychographic segments. The results
suggested that there. were three meaningful segments b;Jse<l upon these lifestyles. The J!rst se.gmcnt (:',3 percent)
consisted 1Jf individual'> who were high on nearly all lifestyle scales. This group w,is called the "dcmanders."
The second group (20 percent) was high on the educational scale and was named the ''educationalists." The
last group (26 percent) wc1s high on relaxation and !ow on sncial scales and was named the "e!icapists."
Specific marketing strategies were fonnulak<l to attract vacationers in each segment. ln order to recover from
the nfterma!h of the economic downturn in 2008-2009, Thailand made a special clfort ID reach lhe "escapists"
segment in 20 IO, because the country would appeal the most to these vacationers, given it~ many rclaxatio11
oppnr1unities rich in natural beauty. 7 •
• Understa,ulin~ buyer behaviors: Cluster analysis can be. used to iJcntify homogeneous
groups of huye.rs. Then the buying behavior of each group may be cx~nnined separately,
as in the cJepai1ment store project, where re~pondents were clustered on the basis of
self-reported importance attached to each factor of the choice criteria utilizcJ in sekcting
a department store. Cluster analysis has also he.en use<l to identify the kinds of strategies
automobile purchasers use to obtain external information.

0'. A
V
• Identifying new product opportunities: By clustering brands and products, competitive sets
1
>
within the market can be determined. Brands in the same duster compete rnorc fiercely with
each other than with brands in other clusters. A firm i.:an examine its cun~nt (1fferings com-
w pared to those of its competitors to identify potential new product opportunities.
• Selecting test markets: Ry grouping cities into homogeneous clusters, it is po,<,siblc to
Variable 2 select comparabk cities to test various marketing strategies.
• Reducing data: Cluster analysis can be used as a general data reduction tool to develop
FIGURE 20.2 clusters ur subgroup~ of <lala that are more manageable than individual observations.

A Practical
. Suhscquent multivariate analysis is conducted on the clusters rather than on the individual

·-
observations. For example, to describe differences in consumers' product usage behavior,

..•.... ........... .
Clustering the consumers may first be clustered into groups. The differences among the groups may
Situation
-~·· tl1Lm be examined using multiple discriminant analysis .

:g" .:.: .·...:•..................., Statistics Associated with Cluster Analysis

·:a
> ....... .~ :'.,
Before discussing the statistics associateJ wilh cluster analysis, it should be mentioned that mo~t
clustering methods arc relatively simple procedures that are not supported hy an extensive body of
statistical reasoning. Rather, most clustering methods are heuristics, which arc based on
• algorithms. Thus, cluster analysis contrasts sharply with analysis of variance, regrcssinn, discrim-
inant analysis, and factor analysis, which are based upon an exlensive hudy of statistical
reasoning. Although many clustering rnelhods have important statistical properties, the funda-
mental simplicity of these methods needs to be recognized. 8 The f,)llowing statistics and concepts
Variable 2 are associated wlth cluster analysis.
 PART Ill • DATA COLLECTION, PREPARATION, ANALYSIS, AND REPORTING
CHAPTER 20 • CLUSTER ANALYSIS 665
664

Agglomeration schedule, An agglomeration schedule gives information on the objects or
cases being combined at each stage of a hierarchical clustering process,
Cluster centroid, The cluster centroid is the mean values of the variables for all the cases
or objects in a particular cluster.
Cluster centers, The cluster centers are the initial starting points in non hierarchical
clustering, Clusters are built around these centers or seeds,
Cluster membership, Cluster membership indicates the cluster to which each object or
case belongs.
Dendrogram, A dendrogram, or tree graph, is a graphical device for displaying clustering
results. Vertical lines represent clusters that are joined together. The position of the line on
the scale indicates the distances at which clusters were joined. The dendrogram is read
from left to right. Figure 20.8 is a dendrogram.
Distances between cluster centers, These distances indicate how separated the individual
pairs of clusters are. Clusters that are widely separated are distinct, and therefore de,sirable.
Icicle plot, An icicle plot is a graphical display of clustering results, so called hecaLtsc it
resembles a row of icicles hanging from the eaves of a house. The columns correspond to
the objects being clustered, and the rows correspond to the number of clusters. An icicle
plot is read from bottom to top. Figure 20.7 is an icicle plot.
Similarity/distance coefficient matrix, A similarity/distance coefficient matnx is a
lower-triangle matrix containing pairwise distances between ohjects or cases.
Conducting Cluster Analysis
The steps involved in conducting cluster analysis are listed in Figure 20.3, Tile first step is to
formulate the clustering problem by defining the variables on which the clustt:ring will be
based. Then an appropriate distance mca:mre must be selected. The distance measure deter-
mines how similar or dissimilar the ohjects being clustered are. Several clustering pruccdun.:s
have been developed and the researcher should select one that is appruprlate for the prohlem at
hand. Deciding on the number of clusters requires judgment on the part of the researcher. The
euclidean distance
1 he '>quare root of the sum
of the squared d1tferences
1n values for each variable
To illustrate, we consider a clustering of consumers based on attitudes toward shopping.
Based on past research, six attitudinal variables were identified. Consumers were asked lo
express their degree of agreement with the following statements on a 7-point scale (I = disagree,
7 = agree):
V1: Shopping is fun.
V2: Shopping is bad for your budget.
V1: I combine shopping v.ith eating out,
V4 : I try to get the best buys when shopping.
V,: I don't care about shopping,
V6 : You can save a lot of money by comparing prices.
Data obtained from a pretest sample of 20 respondents arc shown in Table 20.1. A small
sample size has been used to i1lustrate the clustering proces,l.i. In actual practice, cluster analy ...,is
is performed on a much larger sample such as that in the Dell running case and other cases with
real data that are presented in this hook.
Select a Distance or Similarity Measure
Because the objective of clustering is to group similar ohjects together, some measure is needed
to assess how similar or different the objects are. The rnost common approach is to measure sim-
ilarity in terms of distance between pairs of objects. Objects with smaller dislances between
them are more similar to each Llther than arc those at larger Jista111..:es. There are several ways tn
compute the distance between two objects.9
The most commonly used measure of similarity is the euclidean Liistance or its square. The
euclidean distance is the square root of the sum of the squared differences in values for each
variah!c. Other distance measures are also available. The city-biork or Manhauan distance
between two objects is the sum of the absolute differences in values for each variable. The

II
deriveJ clusters should be interpreted in terms of the variables used to cluster them and profiled !llm:1111:9'-I•Jt
in terms of additional salient variables. Finally, the researcher must asses,<., the validity of the Attitudinal Data for Clustering
clustering process. Case No. '1 V2 V3 v. v, v6
4 7
Formulate the Problem 4
3
Perhaps the most import.ant part of formulating the clustering problem is selecting the variahles
SPSS Data File 2 0 4
on which the clustering is based. Inclusion of even one or two irrelevant variables may distort an

§.sas
4 4 6 4
otherwise useful clustering solution. Basically, the set of variables selected should describe the
similarity between objects in te-rms that are rckvant to the marketing research prohlcm. The 3 2
variables should he selected based on past research, theory, or a consideration of the hypotheses 4 6
being tested. ln exploratory research, the researcher should exen.:ise. judgment and intuition. 6
SAS Data File
3 7
2 4
Fortiii,lai/th& probie~: ·" , ''
FIGURE 20.3
Conducting Cluster
Analysis
.· · ... ·r . . ·.. ,
Select 'a distance measUJ:e,'
10
11
12
5
3
4
2
b

.,, \, ,. t,. ...... , . .

Select.a clustering procedure,:
IJ
14 4
2
6 4
5 4
4
15 4

. . ···········:i·.·······,.····
Decide on the number of clusters;.
. i
In
17 4 4
4
7
18 7 2 4
Interpret .and profile clusters,
19 6 3 2
"., .. ' ., ! ·.· . . ,, "' 20 3
Assess the val\dity .o(.duiter!ng:

··---~-~-
 Multidimensional Scaling and Conjoint Analysis

Often, relationships are easier to see if you can 6. Di'.:>cuss the bds1c concepts of OJrJJ01t1t <-111cily~;i'.;, con11c1~;t it w11l I MD~), ;ir11J
discuss its Vcnious uppli(<Jtior1.c,.
7. UcscribE~ the procf'.!dure for conduc1insJ conjoi11L irnJud1nq furmul<1L
or create a chart that ing the probk-nn, construc1in~ the stim1d1, dccidin~J of i11pt1I d,itrJ,
selectin~J a conJuint atic-dy:;i~! prou~durc, irilL~rp1el111~-J t!1l~ 1c,ldl'., rJIH1 d'.l'.~c~,:,
1

ing reliability and validity.

illustrates the relationships ... and that is 8. Define the conu-)pt of hybrid c:rnijoiril ar1<1lysis ,md uxpl<1111 how ii c;irnpli!it)'.)
the dcJli1-collt:U1on task

the goal of multidimensional scaling. Overview This chapter on data analysis p1e~,en!s lwo relcJted tec:h111q11es ror
tions and preferences: rnultidirnens1rn ldl ~;cc.1l!nq (MIJS) arid c or\)01t1t WL'
illustrate !he steps involved ir1 conducting MDS '-rnd discus'._, 1he ;i1no11q MD'.;, id(
Conjoint analysis, on the other hand, helps tor analysis, and cliscrirninc.ml an<.1lys1'.;. Thr.:n we dE;'.;cribr; con101nt ut1~ily\1:, <.111d p1c:._;c:11t d
by-stt~p procedure for it. We also provide hr 1ef cov('1 dCJr-' of liyhrid crn1Jt ,111t
Finally, we discuss Lhe use of in MDS and c.oriJoint <.mdly~,12, Hc,lp l[)t I tm111nq the• '.-;IJSS
-···· us profile which attributes contribute and SAS programs used ifl this chcrptc,r is prc,vidc-d in lnur wc1ys: (1 I ddcrrlt!d r,lt,p by c;lep
instructions are given later in the chapter, (2) you cm download (frorn lfH:-; Wc;b '.;il<] for l~w. book)
computerized dQrnonslratton movies 1llu'.1tr(.1t1r1g these step-l>y-step 1nst1ud1or1~,. C{) you c·,1n
__ most heavily to a person's choice among download screen cuptures with nott~s illusl1<1lin~:i thf~Sl·~ step by ~)lep in~;lrt1Ll1cm\ drid (4) yo11
can refer to the Study Ciwde and /(x:hriology Manuc1/, ,1 suppll:1rH~nl th,it accumpcm1c~ tlii'.;
book.
a variety of offerings made up of different
Real Research Colas Collide
-~---- ·-· --------····-·--
combinations of these attributes.
------ In a survey, respondents were aske<l to rank-order all the possible pairs of JO brands of soft drinks in tcrn1s
of their similarity. These data were analyzed via multidjmension:..i.l st;:tling and resulted in the folJowing
Kuna! Gupta, Vice President/Senior Consultant, Decision Sciences, Burke, Inc. spatial representation of soft drinks.

0.8
•
Dr. Pepper
0.6
Objectives [ After reading this chapter, the student should be able to:]
().4 ... Slice
1. Discuss the basic concept and scope of multidimensional scaling (MDS} in
marketing research and describe its various applications.
L. •
0.2
2. Describe the steps involved in multidimensional scaling of perception data, • Coke Classic 7-Up
including formulating the problem, obtaining input data, selecting an MDS . ().() • •
procedure, deciding on the number of dimensions, labeling the dimensions Pepsi
•
Diet Slice
and interpreting the configuration, and assessing reliability and validity. --0.2
... Diet Pepsi
3. Explain the multidimensional scaling of preference data and distinguish
... • •
between internal and external analysis of preferences.
4. Explain correspondence analysis and discuss its advantages and disadvantages.
-0.4
~
Tab
• •Diet Coke Diet 7-Up

-0.6 ~

5. Understand the relationship among MDS, discriminant analysis, and factor ...
analysis. -0.8 I I

688 1.5 -1.0 -0.5 0.(J 0.5 I.I) 1.5 61:19

690 PART Ill • DATA COLLECTION, PREPARATION, ANALYSIS, AND REPORTING
MOS has been used to
understand consumer' per-
ceptions of soft <l1 mks and
the competitive structure of
the soft drink market
Real Research
multidimensional
scaling (MDS)
A class ot procedures for
representing perceptions
and preferences of
respondents spatially by
means of a visual display
From other inlormat1on obtained in the questionnaire, the hnriznnlal ax.is was labeled .is "CnLt Flavor."
Tab was pcrceiwd tn he the tll()St cola tlavorcd and 7-Up the least cola flavurccl. The vertical axis wa~
labeled as "Dietness," with Tab being pcrcclvcd to be the most dietetic and Dr. Pepper the most nun<lictctic
Note that Pepsi aml Coke Classic were pcrcewed to be very .c,irnilar, as ind1catcd by the1r close11css in the
perceptual map. Clnse similarity was also perceived between 7-lJp and Slice, Diet 7-Up and Diet Slier. and
Tab, Diet Coke, and Diet Pepsi. Note tbat Dr. Pepper i<; perceived to be relatively dissimilar lo the other
brands. Such MDS maps are very useful Ill understanding the competitive structure of the soft drink market.
'Il1e ( \ica-Coia Company ha:-. utiliLed techniques such ,is MDS to understand how consumers perceive their
prodLKts as well as those of competitors and, as a result, reaped rich rewards by rnJintaining an iron grip on
thc lJ .S. carbonated soft drink market that was cstima!ed tu be about $70 billion in 2009. 1 •
What Do Customers Look For in a Computer Printer?
Printronix. (www.printrouix.com), a manuf~1cturer of computer printers in Irvine, California, recently
spnn.<-(>red a natiDnwidc conjDint analysis project using interaclive software provided hy Trade-Off
Research Services. The ohicctive of this direct-mail project was to identify the huying habits of present and
future customers a:,; well as those of purchasers of competitive product.'-..
"We a1c in a markl'.t-driven, competitive pnntct industry where the customer has many choi<.:es and
options," says Jack Andersen, vice p1esident ol domesttc marketing fur Pnntronix. "It is critical fot the
growth of this company that Wlc know why custorners huy or reject some printers over others."
Printronix mailed 1,600 Jiskcttc surveys to a pre4ualificd list of dcci~ion makers. The survey~ were
divided according to the pnce range of the printers with only slight diffcrc11ccs in the survey questions in
both groups. Prequatifying also determined whether or not deciswn makers were planning on purchasing
new equipment and when, and the willingness of the decis1011 rnakers tu participate in the survey.
RL"'.sidts received by Printronix. management will help the company better undcrstanJ the computer
printer maiketplace. The company will be ahlc to identify ils customer base, what the "hot" buttons are, an<l
what type of products customers want now and in the future. Purthcrrnore, tht.: tc~ulls tabulatl"'.d will be able
to provide insight into current and future product need~.
Particular marketing strategies can also be dcvclopc<l, for example, on how to engineer the product,
how to advertise the product, and how he:-;t to sell it. "It's critic.:1! to focus your product markeling message
tn the needs of the buyLT," says Andersen. "lf we have designed certain elements into our products and then
fail to promote them, we lose market share. That's the botlom lrnc, so il's important to know as much as
possible about the buyer."
Upon completion of the project and in characterizing the marketplace, Printronix will be able to cross-
reference various respondents (e.g, MIS managers only; companies under $10 mi\J1on; IBM-PC users only)
to help identify and define vertical market potentials. Additionally, ''what-if' analysts can be generated from
the results. For example, ii" the speed of a printer were increased, with everything else remaining constant as
CHAPTER 21 • MULTIDIMENSIONAL SCALING AND CONJOINT ANALYSIS 691
it relates to quality and perfonnance, how would t!iis new input affect the current marketplace share? Such
surveys and other sophisticated marketing have helped Printronix. to understand the market needs and
customer preferences better. The company has also developed cutting-edge proUucts including RFII) prin~.'::·~
and thermal bar~code printers to cater to the evolving neC'ds of these markcts. 2 •
The first example illustrates the derivation am! use of perceptual maps, which lie at the heart ,,r
MDS. The second example involves trade-offs respondents make when evaluating alternative,,
The conjoint analysis procedure is based on the.se tmde-offs.
Basic Concepts in Multidimensional Scaling (MDS)
Multidimensional scaling (MOS) is a class of procedures for representing perceptions and pref-
erences of respondents spatially by means of a visual display. Perceived or psychological rela-
tionships among stimuli are represented as geometric relationships among points in a nrnlt1di-
mensional space. These geometric representations are often called spatio! maps. The axes of the
spatial map are assumed to denote the psychological ha.ses or underlying dimensions respon-
dents use to form perceptions and pretercnces for stirnuli.1 MDS has been used in marketing to
identify:
1. The number and nature of din1en.sion.s consumers use to perceive different hrands in the
marketplace
2. The positioning of current brands on these dimensions
3. The positioning of consumer....,• ideal brand on these dimensi<ms
lnfonnation provided- by MOS has been used for a variety of marketing applications, including:
• lma}{e measurement. Compan> the customers' and noncustmners' perceptions of the firm
with the firm's perceplions of itself and thus identify perceptual gaps.
• Market segmentation. Posjfo.rn brands and consumers in the same space and thus identify
groups of consumers with relatively homogeneous perceptions.
• New product development. To look for gaps in the spatial map, which indicate potcntiul
oppo11unities for positioning new products. Also, to evaluate new product concepts and
existing brands on a test basis to determine how consumers perceive the new concepts.
The proportion of preferences for each new product is one indkator of its success.
• Assessing advertisin1-: effectiveness. Spatial maps can be used to dete:nnine whether
advertising has been successful in achieving the desired hrand positioning.
• Pricing analysis. Spatial maps developed with a.nt.l without pricing inf"ormation can be
compared tu determine the i1npact of pricing.
• Channel decisions. Judgments on compatibility uf brands with different retail outlets could
lead to spatial maps w.;eful for making channel decisions.
• Attitude scale construction. MDS tcc:hniques can he used to dcvelnp thi; <1ppropriate
dimensionality and configuration of the attitude space.
Statistics and Terms Associated with MDS
The important statistics and terms associated with MOS inclt1de the following:
Similarity judgments. Similarity judgments are ratings on all possible pairs of brands or
other stimuli in terms of their similarity w,ing a Likert-type scale.
Preference rankings. Preference rankings arc rank orderings of the brands or other
stimuli from the nl(JS! preferred to the least preferred. They are normally obtained
from the respondents.
Stress. Thjs js a lack-of-fit measure; higher values of strc~s indicate poorer tits.
R~square. R-square is a squared cotTelation index that indicates the proportion o~v~riance
of the optimally scaled data that can be accounted for by the MDS procedure. Thts 15 a
goodness-of-tit measure.
Spatial map. Perceived relationships among hrands or other stimuli are represe~ted as
geom~tric rclati<rnships among points in a multidimensional space called aspattal map.
"·· <'""~tet~+r,,):~~~&;,
692 PART 111 • DATA COLLECTION, PREPARATION, ANALYSIS, AND REPORTING CHAPTER 21 • MULTIDIMENSIONAL SCALING AND CONJOINT ANALYSIS 693

FIGURF 21.1 Formulate the problem. FIGURE 21.2 MDS Input Dala
Conducting Input Data for
Multidimensional 1
Obtain input data.
Multidimensional
Scaling r ----~- l
Scaling

1 PerGeptions Preferences

Select an MDS procedure. I

1 L --i
Direct Derived
Decide on the number of dimensions.
(Similarity (Attribute

l
Label the dimensions ·.
Judgments) Ratings)

and interpret the configuration. own criteria. Respondents are often required to rate all possible pairs or brands ur stimuli in terms of
·. . 1' simi !arity on a Llkcrt scale. These data are referred to as similarity judgments. For example, si1nilmity
judgments on al1 the possible pairs of toothpaste bnrnds may be obtained in the following rnanner·
Assess reliability and validity.

Very Very
Coordinates. Coordinates indicate the positioning of a brand or a stimulus in a spatial map. Dissimilar Similar
Unfolding. The representation of both brands and respondents as points in the same space Crest vs. Colgate 2
is referred to as unfolding. Aqua-Fresh vs Crest 2
Crest vs. Aim 2 h

Conducting Multidimensional Scaling

Figure 21.1 shows the steps in MOS. The researcher must formulate the MDS prohlcm care-
fully because a variety of data may be used as input into MOS. The researcher must also
Culgate vs A.qua-Fresh 2 (,
determine an appropriate form in which data should he obtained and select an MDS proce-
dure for analy ,.ing the data. An important aspect of the solution involves Lletermining the
number of dimensions for the spatial map. Also, the ax.es of the map should be labeled
and ihe derived configuration interpreted. Finally, the researcher must assess the quality The number of pairs to be evaluated is n(n - I )/2, where n is th~ number l)f ~tin,uli. ()thcr
of the results obtained. 4 We descrihe each of these steps, beginning with problem procedures are also availabk. Respondents could he asked to rank-order al! the possible pairs
formulation. derived approoches from the most similar to the least si1nilar. In another mdtwd, the respondent rank-onlcrs the
In MDS, attribute-ba':>ed brands in terms of their similarity to an anchor hrand. Each hrn.nd, in turn, serves as the a11chor.
Formulate the Problem approarhe<, to rdlect111(_J In our example, the direct approach was adoptCd. Subjects were askt'd to provide. similarity
formulating the proh\em rel.1uircs that the researcher specify the pmvosc for which the MDS µercept1on data requiring judgments for all 45, (Ill X 912), pairs of toothpaste hran<ls, using a 7-point scale. The data
result.'> would hi.: useJ and select the brands or other stimuli to be included in the analysis. The the resriondents to rc1te the obtained from one respondent are given in Table 21.J .'l
stimuli on tlw 1dt'11t1f1ed
numher of brands or stimuli sG!ecte<l and the spe<:ific brands included determine the nature of
attributes using semantic PERCEPTION DATA: DERIVED APPROACHES Derived approaches lo ml lecling perccplion data
the resulting dimensions and configurations. At a minimum, eight brands or stimuli should be
d1fferent1al or l1kert scales. me attribute-based approaches requiring the respondent,<,, to rate the brands or stimuli un the
included so as to obtain a well-defined spatial map. Including more than 2.5 brands is likely to be
cumbersome and may result in respondent fatigue.
The decision regarding whkh speciric bran<ls or stimuli to include should be made care-
fully. Suppose a researcher is interested in obtaining consumer perceptions of automohiles. Similarity Ratings of Toothpaste Brands
[f luxury autnmohiles are not included in the stimulus set, this dimension may nnt emerge in
Aqua- Plus Ultra Close-
the results. The choice of the number and specific hrands or stimuli to be included should be Fresh Crest Colgate Aim Gleem White Brite Up Pepsod_e~ensody_n..:._
base<l on the statement uf the marketing research problem, theory, and the juJgment of the
researcher.
Multit:imensional scaling will be illustrated in the context of obtaining a spatial map for IO Crest
SPSS Data File
toothpaste brands. These brands a.re Aqua-Fresh, Crest, Colgate, Aim, Glecm, Plus White, Ultra Colgate 6
Brite, Close-Up, Pepsodcnt, and Sensodyne. Given the list of brands, the next question, then, is: Aim 4
How should we obtain data on these 10 branJs? Gleern 2 J 4 5
Plus White 3 3 4 4
Obtain Input Data SAS Data File llltra Brite 2 2 2 :,
As shown in Figure 21.2, input data obtained from the respondents may be related lo perceptions Close-Up 2 2 2 2
or preferences. Perception data, which may be direct or derived, is discussed first. Pepsl)(!ent 2 2 2 2
PERCEPTION DATA: DIRECT APPROACHES In direct approaches to gathering perception data, the Sensodyne I 2 4 2
respondents are asked to judge how similar or dissimilar the various brands or stimuli are, using their
700 PART Ill • DATA COLLECTION, PREPARATION, ANALYSIS, AND REPORTING
hunt Space Configuration
of Automobile Brands and
Consumer Preferences
(Illu~trative Output)
tonespondr:nce
analysis
An MDS te( hnique tor
scaling qual1tc1t1ve dnta that
sc_dlt!'> the rows drid rnlurnr1)
uf the input contingency
tdhle in corresponding units
so thc1t each cJn be
displdyod in the Sd!lle
luw-d1rner1:.ional spdce
Sportiness
(high)
Porsche• II
• Audi
• Jaguar
• Saturn
High prestige/
expensive
I • Toyota • Honda ~
Economy
Continental Buick• • Chevrolet
• • Cadillac • Chrysler
@ eLcxus • vw
• Mercedes • Hyundai
(low)
The preference data consisted nf a simple rank order of the brands according to consumers' prefer-
L'nccs. Respondent:.,' ideal points arc also located in the same spatial representation. Each ideal poinl repre-
s<.::nt.s tl1e locus of pre-ference of a particular respondent. Thus, respondent I (dcno!t:d hy 11) prefers the
sporty cars: Porsche, Jaguar, and Audt. Respondent 2 (Uenoted by 12) on the nthcr hand, prefers luxury cars:
Continental, Mercedes, Lexus, and Cadillac..
Such analy.\i.s can be done at the indivi<lual-respondent level, enabling the re.searcher to segn1cnl the
market ,1ccording to similarities in the respondents' ideal points. Alternatively, the rcsplrn<lents can be clus-
tered based on their similarity with resr,ed to the original preference ranking and idea! points c:stahlished
for ew.:h ~egment. 1 1 •
Whereas, so far, we have consiJcrcd only quantitative data, qualit1tive data can al.so he mapped
using procedures such as con-cspondence analysis.
Correspondence Analysis
Correspondence analysis is an MOS tcchni4ue for scaling qualitative Jata in m..J.rketing
research. The input data are in the form of a contingency table indicating a qualitatlve association
between the rows and columns. Correspondence analysis scales the rows and columns in corre-
sponding units, so that each can he <lisplayetl graphically in the same low-dimensional space.
These :,;patial maps provide insights into (I) similarities and differences within the rows with
respect to a given column category; (2) similarities and differences within the column categories
with respect to a given row category; and (3) relationship among the rows and colurnns. 12
The interpretation of results in correspondence analysis is similar to that in principal cnmpo~
nents analysis (Chapter 19), given the sirnil...trity of the algorithms. Correspondence analysis results
in the grouping of categories (activities, brands, or other stimuli) found within lhe L:ontingency
table, just as principal components analysis involves the grouping of the variahles. The results are
interpreted in terms of proximities among the rows and columns of the contingency tahle.
Categories that are closer together are more similar in underlying structure.
The advantage ol' correspondence analysis, as compared to other multidimensional scaling
techniques, is that it reduces the data-collection demands imposed on the respondents, because
only binary or categorical data are obtained. The respondents are merely a~ked to chc~k which
attributes apply to each <>f several brands. The input data are the number of "yes" responses for
each brand on each attribute. The brands and the atlributes are then displayed in the same
multidimensional space. The disadvantage is that between-set (i.e., between column and row)
distances cannot be meaningfully interpreted. Currespon<lence analysis is an exploratory data
analysis technique that is not suitable for hypothesis testing. 13
MDS, including correspondence analy:-.is, is not the only procedure available for obtaining
perceptual maps. Two other techniques that we have discussed before, discriminant analysis
(Chapter 18) and factor analysis (Chapter 19), can also be used for this purpose.
conjoint analysis
A technique that attempts
to deterrrnne the relative
importance consumers
attach to 'idl1ent dttr1butes
and thP utilities they attach
to the levels of attributes
CHAPTER 21 • MULTIDIMENSIONAL SCALING AND CONJOINT ANALYSIS 701
Relationship Among MDS, Factor Analysis,
and Discriminant Analysis
If the attribute-based approach is used to obtain input data, spatial maps can also be obtained hy
using factor or discriminant analysis. In this approach, each respondent rates n brands on m
attributes. By factor analyzing the data, one could derive, for each respondent, n factor scores for
each factor, one for each brand (see Chapter 19). By plotting brand scores on the factors, a spa-
tial map could be obtained for each respondent. If an aggregate map is desired, the factor score
for each brand for each factor can be averaged across respondents. The dimensions would be
labeled by examining the factm loadings, which are estimates of the correlations between
attribute ratings and underlying factors.
The goal of discriminant analysis is to .select the linear combinations of attrihules that best
discriminale between the brands or stimuli (Chapter 18). To develop spati:il maps by means of
discriminant analysis, the dependent variable is the brand rated and the independent or predictor
variables are the attribute ratings. A .spatial map can he obtained hy plotting the discriminant
scores for the brands. The discriminant scores are the ratings on the perceptual Uimensions,
based on the attributes that best distinguish the brands. The Jimensiuns can be labeled. by t:xam~
ining the discriminant weights, or the weightings of attributes that make up a discriminant func-
tion or dimension. 14
Basic Concepts in Conjoint Analysis
Conjoint analysis attempts to determine the relative importance consumers attach to salient
attrihutes and the utilities they a*ttach to the levels of attrihutes. This information is Jerivcd from
consumers' evaluations of brands, or brand profiles composed of these attrihntcs and thcfr levels.
The respondents are presented with stin1uli that ctrnsist of combin..itions of attrihute levels. They
are asked to evaluate these stimuli in terms of their desirability. Conjoint procedures attempt to
assign values to the levels of each attrihute, so that the resulting values or utilities attached to the
,..,timuli match, a..., clo8ely as po.,;.sible, the input evaluations provided hy the rC.'-fll)lldent.'i. The
underlying assumption is that any set uf stimuli, such as products, brands, or stores, is evaluated
as a bundle of attributes_ I.I
Like rnultiJirnensional scaling, conjoint analysis relies on respondents' subjective evalua-
tions. However, in MDS, the stimuli are products or hrands. In conjoint analysis, the stimuli
are comhinatinns of attribute levels J.etcrmineJ by the researcher. The goal in MDS is tn
<level op a spatial map depicting the stimuli in a multidimensional perceptual or preference
space. Conjoint analysis, on the other hand, seeks to develop the part-worth m utilily functions
describing the utility consumers attach to the levels of each attribute. The two techniques are
complementary.
Conjoint analysis has been used in 1narketing for a variety of purposes. including:
• Determining the relative importance of attributes in the consumer choit:c proccs.~. A stan-
dard output from conjoint analysis consists of derived relative importance weights for all
the attributes used to construct the stimuli used in the evaluation task. The relative impor-
tance weights indicate which attributes are important in influencing con:..uiner choice.
• Estimating market share of brands that differ in attribute levels. The utilities <.k·nved from
conjoint analysis can he used as input into a choice simulator to dcterllline the share of
choices, and hence the market share, of di ffercnt brands.
• Determining the composition of the most preferred hrand. The hramJ features can be varied
in terms of attribute levels and the corresponding utilities determined. The brand features
that yield the highest utility indicate the composition of the most preferred brand.
• Segmenting the market ba.se<l on si1nilarity of preferences for attrihute levels. The part-
worth functions derived for the attributes may be used as a hasis for clustering respondents
to arrive at homogeneous preference segments. 16
Applications of conjoint analysis have been made in consumer gonds, industrial goods,
financial, and other services. Moreover, these applications have spanned all areas of marketing.
A survey of conjoint analysis reported applications in the areas of new pro<luct/cnnccpt identifi-
cation, competitive analysis, pricing, market segmentation, atlvcrtising, aud distrihution. 17

702 PART 111 • DATA COLLECTION, PREPARATION, ANALYSIS, AND REPORTING
Statistics and Terms Associated with Conjoint Analysis
The important statistics and terms associated with conjoint analysis include:
Part-worth functions. The part-worth functions or utility.fimctions describe the utility
consumers attach to the levels of each attribute.
Relative importance weights. The relative importance weights are estimated and indicate
which attributes are important in influencing consumer choice.
Attribute levels. The attribute levels denote the values assumed by the attributes.
Full profiles. Full profiles or complete profiles of brands are constructed in terms or all the
attributes by using the attribute levels specified by the design.
Pairwise tables. In pairwise tables, the respondents evaluate two attributes at a time until
all the required pairs of attributes have been evaluated.
Cyclical designs. Cyclical designs are designs employed to reduce the number of paireJ
comparisons.
Fractional factorial designs. Fractional factorial designs arc designs employed to reduce
the number of stimulus profiles to be evaluated in the full profile approach.
Orthogonal arrays. Orthogonal arrays are a special class or fractional designs that enable
the efficient estimation of all main effects.
Internal validity. This involves correlations of the predicted evaluations for the holdout or
va!i<lation stimuli wilh those obtained from the respondents.
Conducting Conjoint Analysis
Figure 21.8 lists the steps in conjoint analysis. Formulating the problem involves identifying the
salient atlributcs and their levels. These attrihutes and levels are used for constructing the stimuli
to be used in a conjoint evaluation task. The respondents rate or rank the stimuli using a suitable
scale and the data obtained arc analyzed. The results arc interpreted and their reliabil'ity and
validity assessed.
Formulate the Problem
In formulating the conjoint analysis problem, the researcher must identify the attributes and
attribute levels to he used in constructing the stimuli. Attrihutc levels denote the values assumed
by the attributes. From a theoretical standpoint, the attributes .selected should be salient in influ-
encing consumer preference and choice. For example, in the choice of an automobile brand,
price, gas mileage, interim space, and so forth should be included. From u managerial perspec-
tive, the attributes and their levels should be actionable. To tell a manager that consumers prefer
FIGURE 21.8
Conducting
Conjoint Analysis
CHAPTER 21 • MULTIDIMENSIONAL SCALING AND CONJOINT ANALYSIS 703
a sporty car to one Ihat is conservative looking is not helpful, unless sportiness and conservative-
ness are defined in terms of attributes over which a manager has control. The attributes can be
identified through discussions with management and industry experts, analysis of secondary
data, qualitative research, and pilot surveys. A typical conjoint analysis study involves six or
seven attributes.
Once the salient attributes have been identified, their appropriate levels should he
selected. The number of attribute levels determines the number of parameters that will he
estimated and also influences the number of stimuli that will be evaluated by the respondents.
To minimize the respondent evaluation task, and yet estimate the parameters with reasonable
accuracy, it is desirable to re,trict the number of attribute levels. The utility or part-worlh
function for the levels of an attribute n1ay be nonlinear. For example, a consumer ,nay prefer
a medium-sized car to either a small or large one. Likewise, the ulility for pnce may be non-
linear. The loss of utility in going from a low to a medium price may be mud1 Slllaller than
the loss in utility in going from a medium to a high price. In these cases, at least three levels
should be used. Some attributes, though, may naturally occur in binary form (two levels): a
car does or does not have a sunroof.
The attribute levels selected will affect the consumer evaluations. If the price of an automo-
bile brand is varied at $10,000, $12,000, and $14,000, price will be relatively unimp011ant. On
the other hand, if the price is varied at$ I 0,000, $20,000, and $30,000, it will be an important fac-
tor. Hence, the researcher should take into account the attrihute levels prevalent in the market-
place and the objectives of the study. Using attribute levels thal are beyond the range reflected in
the marketplace will decrease lhe believability of the evaluation task, but it will increa.,e the
accuracy with which the parametern are estimated. The general guideline is to select attrihute
levels so that the ranges are s()mewhat greater than that prevalent in the marketplace but not so
large as to adversely impact the believability of the evaluation task.
We illustrate the conjoint methodology by considering the prohlem of how students evaluate
sneakers. Qualitative research identified three attributes as salient: the sole, the upper, and the
price. 18 Each was defined in tcnns or three levels, as shown in Table 21.2. These attributes and
their levels were used for constructing the conjoint analysi~ stimuli. Note that to kcL'P the illus-
tration simple, we are using only a limited number of attributes, that is, only three. It has hccn
argued that pictorial stimuli should he used when consumers' marketplace choices are strongly
guided by the product·~ styling, such that the choices are heavily based on an im;pL'.ction of actual
products or pictures of products. 19
Construct the Stimuli
Two broad approaches are available for constructing conjoint analysis stimuli: the pairwise approach
and the full-profile procedure. In the pairwise approach, al.'-io called twoJuctor £'valuations, the
respondents evaluate two attributes at a ti111c until all the possible pair:-. of attributes have been evalu-
ated. This approach is illustrated in the context of the sneaker example in Pigure 21.9. For each pair,
respondents evaluate all the combinations of levels of hoth the attributes, which are presented in a
matrix. in the full~profile approach, also called multiple~factor evaluations, full or complete pro ti Jes
ll1l'f:) ........
Sneaker Attributes and Levels
Attribute Level No. Description
Sole 3 Ruhbcr
2 Polyurethane
Plastic
Upper 3 Leather
Canvas
Nylon
Price 3 $30.00
2 $60.00
$9000
 CHAPTER 21 • MULTIDIMENSIONAL SCALING AND CONJOINT ANALYSIS 705
704 PART 111 • DATA COLLECTION, PREPARATION, ANALYSIS, AND REPORTING

FIGURE 21.9
~m:,.,..,... ...,
Sneaker Profiles and Their Ratings
Pairwise Approach to Collecting Conjoint Data
Sole Attribute Levers•
Sole
Plastic Rubber Polyurethane Plastic Preference
Rub her Polyurethane

$3000
I I Profile No. Sole Upper
1
Price
I
Rating
9
Leather
2 2 7
,.
~ Canvas
"
-~ $60.00 1 3 3 5
p., 2
;:, 2 1 6
2 2
Nylon $90.00 3 I 6
I
2
Price
$30.00 $60.(XJ $90.00
dThe nttnbute levels correspond to those 1n !"able 21.2.

Leathc1

t validation purposes. Input Jata were obtained for both the estimation and validation stimuli.
o.
Q.
Canvas
;:, However, before the data could be obtained, it was necessary to decide on the form of tl1e input data.

Nylon Decide on the Form of Input Data

As in the case of MDS, conjoi11t analysis input data can be either nornnctric or metric. For
nonmctric data, the respondents arc typically required to provide rank order evaluations. Por the
pairwlse approach, respondents rank all the cells tlf each matrix in tern1s (Jf their dcsirnbil ity. For
of brands are constructed for all the attributes. Typically, each profile is described on a separate inJ.ex the full-profile approach, they rank all the stimulus profiles. Rankings involve relative evalua-
c.:u·d. This approach is illustrated in the context of the sneaker example in Tahle 21 .3. tions of the attnhute levels. Proponents of nmking data believe that such data accurately rellcct
It is not necessary to evaluate all the possible combinations, nor is it feasible in all cases. the hehavinr of consumers in the marketplace.
In the pairwise approach, it is posslhle to reduce the numh~r of paired comparisons hy using Jn the metric form, the respondents provide ratings, rather than rankings. In this case, !he
cyclical designs. Likewise, in the full-profile approach, the number of stimulus profiles can be judgments are typically made independently. Advocates of rating data hclievc that they are more
greatly reduced hy means of fractional factorial designs. A special class of fractional designs, convenient for the respondents and easier to analy7.e than rankings. In recent years, Lhe use of
called orthogonal array.I', allows for the erficient estjmation of all main effects. Orthogonal ratings has become increasingly common.
arrays permit th~ measurement of all main effects of interest on an uncorrelated basis. These In conjoint analysis, the dependent variable is usually preforence or intention to buy. Jn other
designs assume that all interactions are negligible. 20 Generally, two sets of data ;.ire obtained. words, respondents provide ratings or rankings in term.s of their preference or intentions to buy.
One, the estimation set, is used to calculate the part-worth functions for the attrihutc levels. However, the C(JOjoint mcthodolc>gy is flexible and can accomrno<late a range of other dependent
The other, the holdout set, is used to assess reliability am! validity. variables, including actual purchase or choice.
The advantage of the pairwise approach is that it is easier for the respondents to provide In evaluating sneaker profiles, respondents were rc:quired to provide preference ratings for
these judgments. However, its relative Jisa<lvantagc is that it requires more evaluations than the sneakers dcscrihcd hy the nine profiles in the estimation set. These ratings were obtained
the full-profile appro,1ch. Also, the evaluation task may be unrealistic when only two attributes u.sing a 9-point Likert scale (l = not preferred, 9 = greatly preferred). Ratings ohtai11cd from
arc being evaluated simultaneously. Studies comparing the two approaches indicate that both one respondent are shown in 111ble 21.4.
methods yield comparable utilities, yet the full-profile approach is more commonly used.
The sneaker example follows the full-profile approach. Given three attributes, defined at three Select a Conjoint Analysis Procedure
levels each, a total of 3 X 3 X 3 = 27 profiles can be constructed. To reduce the respondent evalu- conjoint analysis The basic conjoint analysis model may be represented by the following forrnuia:2 1
ation task, a fractional factorial design was employed and a set of nine profiles was constructed to model Ill k,
constitute the estimation stimuli set (see Table 21.4). Another set of nine stimuli was constructed for The mathematical model
expressrng the fundamental
U(X) = LL"ifXiJ
i=lyl
relationship between
attributes anrJ utilrty in
where
;,:
con1rnnt analysis. U(X) = overall utility of an alternative
Full-Profile Approach to Collecting Conjoint Data
l\, -=-- the part-worth contribution or utility associated with
Example of a Sneaker Product Profile the )th level (J = I, 2, . k) of the ith attribute (i = I, 2, . m)
Made of rubber k1 = number of level.\ of attribute i
Sole
m = number of attributes
Upper Made of nylon
x.: = I if the jlh level of the ith auributc is present
11
Price $30.00 = 0 otlK:rwise