Segmenting Markets By Group Purchasing Behavior: An

HENRY ASSAEL*
Market segmentation requires definition of consumer groups by variables that
discriminate purchasing behavior. A multivariate program is described capable
of defining homogeneous groups by a large number of variables to maximize
discrimination between purchase group meons. The program has specific advantages compared to traditional methods of grouping such as cross-clossification,
regression, and discriminant analysis. It is opplied in segmenting markets by product and brand usage based on demographic and attitudinai variables.
Segmenting Markets By Group Purchasing
Behavior: An Application Of The AID
Technique
In a recent article, Bass, Tigert, and Lonsdale contend that regression techniques are poor indicators of
the value of socioeconomic variables in segmenting
markets by behavior [2]. Past studies demonstrating
low R- values between demographics and purchase rate
suggest a large within-group or unexplained variance.
Yet regression relies on the individual as the unit of
observation. As the authors suggest, it is the variance between groups, not individuals, that is critical in market
segtiientation: grouped data should be the unit of analysis.
Given the need to examine variations in grouped
data, Bass, Tigert, and Lonsdale propose multiple crossclassification analysis. A number of contingency tables
are constructed to study the simultaneous effects of two
variables at a time on mean usage rates. Yet more
than two variables may be required to define the groups
with the widest variation in usage. As the authors recognize, in such cases "the number of cells becomes
large enough to overwhelm even very large samples."
Clearly, a muitivariate method going beyond crossclassification analysis is required to define the large
number of alternative methods of segmentation at the
point of greatest discrimination in group means.
THE AUTOMATIC INTERACTION
DETECTOR PROGRAM
The Automatic Interaction Detector (AID) program,
developed by Sonquist and Morgan [ 1 , 3 , and 4J, partially overcomes the restrictions associated with crossclassification analysis; AID is a multivariate technique
for determining what variables and categories within
them combine to produce the greatest discrimination in
group means by the dependent variable (purchase frequency, brand last purchased, advertising recall, etc.).
The program divides the sample through a series of
binary splits into mutually exclusive subgroups. The
group means account for more of the total sum of
squares of the dependent variables than the means of
any other combination of predictor variables.
AID can examine up to 35 independent variables
simultaneously. The independent variables require coded
intervals, with a maximum of 63 categories per variable.
There are no coded interval requirements for the dependent variable, which is assumed to be continuous.
Diehotomous data (purchasers of Brand X vs. nonpurchasers) are frequently used as the dependent variable since they can be transformed into a continuous
variable by treating one of the categories as a proportion
(e.g., percentage of purchasers).
In operation, the program splits the sample into two
subgroups to provide the largest reduction in the unexplained sum of squares of the dependent variable.
This is accomplished as follows: group means are determined for each classification of all independent variables, and all dichotomous groupings of each variable
* Henry Assael is Associate Professor of Marketing at the
Graduate School of Business Administration, New York University. He is indebted to the Nestle Company. Inc., and lo its
former Director of Research. Mr. C. E. Wilson, for permission
to use Ihe data in the section on attitudinai segmentation.
153
Journal of Marketing
Researe/i,
Vol. v n (May 1970), 153-8
JOURNAL OF MARKETING RESEARCH, MAY 1970
154
discrimination in group means. Thus respondents might
be divided by region into Northcentral vs. rest of U.S.,
by income into $12,500 and under vs. over $12,500,
etc.
At this point, binary splits have been formed for each
predictor variable, maximizing the between sum of
squares. The between sum of squares for each variable
is divided by the total sum of squares for the group to
be split (BSSi/TSSi). This ratio is computed for eaeh
of the independent variables. AID then selects that
variable with the highest BSSi/TSS,. The first iteration
is completed when the sample is divided into two
groups accordingly. All dichotomous splits defined by
the other predictor variables are discarded.
Figure I illustrates the output of the program. The
proportion of respondents defining Brand X as their
usual brand is the dependent variable. Eight socioeconomic characteristics and level of product usage are
the independent variables. When the total sample is
examined, the maximum reduction in the unexplained
sum of squares is obtained by splitting the sample by
are examined. Using region as an example, the group
mean for those living in the East, Northcentral, South,
and West is first computed. The group means for all
dichotomous groupings are then examined—East vs.
rest of U.S., East and Northcentral vs. South and West,
etc. Each predictor variable is split into two non-overlapping subgroups providing the largest reduction in
unexplained variance. The split is chosen to maximize
the between sum of squares for the ith group (the group
to be split), so that
BSS, = {n,yi' -\- n-^-^) -
NiY,'
where:
n = size of split subgroup
N = size of parent group being split (ith group)
y = mean value of the predictor for the split sub_
group
Y — mean value of the predictor for the parent
group being split.
The division will take place at the point of greatest
Figure 1
SOCIOECONOMIC SEGMENTATION BY USUAL PURCHASERS OF BRAND X
$12,500
& under
NorthCentral
40.3%
Heavy
product
users
5.2%
9.1%
48.7^;,
4.0%
3.7%
Medium
to light
users
12.8%
55.8?
\
\
34.6%
Rest uf
U.S.
19.5%
1
512,500
&. under
1
30.0'i
Housewives
& executives
24.3%
43.0%
Over
$12,500
31.2%
20.0';;.
Females
1
61.9%
54
& under
25.9%
21.4'"
8.6%
24.8%
9.9%
^ ^
Other
occupations
Over
$12,500
17.5-;.
2
11.4%
7.7%
Over 54
6.5%
44.2%
Mure tli.111
high siliuol
28.6%
2
10.2' .:
Vhiles
6.7%
15.6%
High school
iir less
2
-
Stopping ruks invoked:
\ = Hamplc ,wre too .small
2 = Split elinihilitv iritcrion iiul met
^ = Split ri'dmihiiity irilfrioii nut me
SEGMENTING MARKETS BY GROUP PURCHASING BEHAVIOR
sex. The proportion of usual brand purchasers in the
total sample was 16.7%. Yet 24.3% of the females
considered Brand X their usual brand, compared to
only 6.7% of the males. Since sex could yield only
one dicholonious grouping, the division of the variable
was predetermined. Yet before selecting sex, all other
variables were dichotomized at their most significant
point and the BSS/TSS ratio computed.
Onee the population is divided into two subgroups,
the program treats each subgroup as a separate population and the same process is repeated. Within the female
subgroup, the program dichotomizes the remaining independent variables at the optimal point, computes a
BSSi/TSSi for each variable, and selects the variable
with the highest ratio of between to total sum of squares
for the next split. In Figure 1, a split by region into
Northcentral vs. rest of U.S. provided a maximum reduction of unexplained variance within the female subgroup. Females living in Northcentral states comprised
12.8'^c of the sample. Yet within this subgroup, 40.3%
were usual purchasers of the brand, compared to 24.3%
for all females. Similarly, the first split off the male
subgroup is by education. Given a 6.7% male market
share, the proportion drops to 3.8% if males did not
enter college.
After the second series of splits, four subgroups have
been formed. The program again treats each subgroup
as a separate population, and another iteration determines which predictor variable, when split, will best
reduee the unexplained sum of squares for that subgroup. The subgroup chosen for the next iteration is
that unsplit sample group which has the largest total
sum of squares {TSSi). The number of iterations in the
program are controlled by three stopping rules which
limit the number of subgroups formed:
1. A minimum sample criterion (set at A' = 25 in
the example cited).
2. A split eligibility criterion requiring that a group
must contain a minimum percentage of the total
original sum of squares if it is to be considered
eligible for splitting (TSS^ ^ .05 TSS. in the above
example). This requirement prevents groups with little
variation from being split.
3. Presuming the group is eligible for a split, a
split reducibiliiy criterion requiring that the size of
the between group sum of squares for the (th group
has to be a minimum percentage of the total original
sum of squares (B.SS, ^ .02 TSS,). This criterion is
applied when none of the predictor variables in the
group sufficiently reduces the unexplained variance.
The eventual output is a "tree diagram" developing
subgroups of the samples by combinations of variables.
By the fourth iteration in the female branch of Figure
1, the subgroup contained only 5.2% of the sample,
defined as females in Northcentral states who are heavy
users of the product category with family income
$12,500 and under. The proportion of usual purchasers
within this group was 61.9%, compared to the sample
155
average of 16.5%. Thus, given these four characteristics,
the probability of purchasing Brand X increases by
almost four times. Similarly, given the fact the respondent is a male laborer or foreman, the probability
of purchasing is less than one-fourth the sample average.
Figure I may be regarded as an optimal socioeconomic segmentation of the market by purchasing behavior. It is up to the analyst to determine at what point
the segments should be defined. Delineating the market
with 5.2% of the sample may not be operational. On
the other hand, 5.2% of the sample may represent such
a large proportion of total volume that it deserves separate attention.
Limitations of the Program
By considering a large number of variables simultaneously, defining the most significant combinations of
variables, and demonstrating the joint elfects of the
predictor variables, the AID program overcomes many
of the limitations inherent in standard cross-classification analysis. It represents a more effective method of
segmentation than regression analysis, since it determines the joint effects by analyzing variance between
grouped (rather than individual) data.
Yet several restrictions remain. First, the program
relies on dichotomous splits. A throe or four-way split
may reduce the unexplained variance more than a twoway split. This problem is partially resolved in that
the program can split the same variable at the next
iteration. The first split may be high vs. middle-low
income; the next split off the latter subgroup could
then be middle vs. low income, having the same effect
as an initial three-way split. Yet the first step is still
development of dichotomous splits for all variables,
possibly eliminating a three-way grouping that might
have otherwise been more discriminating. The program
may have selected income as the first two-way split
over education. Yet had the variables been defined in
a trichotomous manner, the sample might have been
split by education on the first iteration.
This illustrates another problem. The program selects the variable with the highest BSS/TSS by which
to split the sample. All subsequent splits are contingent
on the subgroups formed by the first split. Yet it is
possible a second variable was almost as discriminating
as the first. Had the program split by the second variable, the subsequent tree diagram might have been totally different. In Figure 1, sex was the first split with
a BSSi/TSSi ratio of .056. Occupation had a ratio of
.053, an insignificant difference. In this case, a split by
occupation would have made little difference, since the
division would have been by housewives vs. all others.
Yet other "near misses" could markedly change the
character of progressive iterations.
This problem can be partially resolved by judicious
use of the program. The program could be run again,
eliminating the variable which defines the first two subgroups, to give an alternate variable a chance to split.
156
JOURNAL OF MARKETING RESEARCH, MAY 1970
Income
Age
Subsequent subgroups can then be examined to determine if they differ markedly from the original run. If
the structures produced by sueh a sensitivity analysis
differ, then criteria of face validity will have to be applied to determine the most logical basis for segmenting
the market.
A third problem is that the supposedly independent
variables may be closely interdependent. As noted, the
potential split by occupation in the first group was
merely a reflection of the significance of sex. Sueh interdependencies may create spurious splits, defeating
the purpose of the program. One method of insuring
the independence of the variables is to run a prior
factor analysis and input pre-classified factor seores for
eaeh respondent rather than the original values. Yet
the difficulties of interpreting average factor scores for
a subgroup inhibits such a solution. Segmentation by
factor scores may not provide clear-eut implications
for media scheduling or promotional appeals.
Race
Number in household.
In addition, level of product usage was included as an
independent variable.
One important advantage of AID is that variables
such as occupation, race, and region can be treated as
nominal rather than sealed variables. Since the program
examines all dichotomous groupings of a variable's attributes, there need be no assumption of monotonicity—
that is, an assumed ascending or descending sequence
in the order of the variable's attributes—as might be
the case for age, income, frequency of purchase, etc.
The tree diagram in Figure 1 presents a series of
profiles by usual purehasers of Brand X. The 5.2% of
the sample represented by females living in the Northcentral region who are heavy users of the product with
income under $12,500, has a 61.9% probability that
Brand X is their usual brand compared to a sample
average of 16.5%. Figure 2 demonstrates that this
group accounts for 20% of all usual purehasers of
Brand X. The least likely user is effectively defined by
only two characteristics: sex and education. The 15.6%
of the sample who were males who did not complete
high school had a 3.8% probability of claiming Brand
X as their usual brand. This group accounts for only
3.5% of Brand X usual purchasers. An individual in
the extreme high probability group had 16 times as
much chance of being a usual purchaser of Brand X
as an individual in the extreme low probability group.
Similarly, intermediate segments can be identified with
purchaser probabilities ranging from 6.5% to 31.2%.
SOCIOECONOMIC SEGMENTATION
Figures I and 2 illustrate the results of the program
in a socioeconomic segmentation of a consumer packaged goods market. Four hundred users of the product
category were interviewed. The dependent variable was
the proportion of respondents defining Brand X as their
usual brand. Eight socioeconomic variables were utilized
as predictor variables:
Sex
Occupation
Region
Education
Figure 2
MARKET CONCENTRATION OF USUAL PURCHASERS OF BRAND X
Tliis %
28.6%
24.7%
2S.9%
5.2%
15.6%
of sample
aceinints for
Females
Females \
in
\
North- \
Central \
wlio are \
heavy users \
with income \
under
\
S 12,500
\
in
rest of
U.S.
who are
housewives
\
\\
A,
\
\
Males
with a
betler iliaii
liigli soliuol
educa tion
\
otlier females
\
\
\
\
\
\
or
\
\
\
\
\
\
\
\
executives
and under 55
\
Males
\
with
\ high school
\
or less
\
\
\
\
Tliis % of
usual purchasers
38.4%
20.0%
of Brand X
\
\
26.6%
\
11.5',;
\3.5';
157
SEGMENTING MARKETS BY GROUP PURCHASING BEHAVIOR
As a representation of existing market potential,
Figures I and 2 may be a valuable guide to media
selcclion. Bass., Tigert, und Lonsdale suggest that an
important criterion in media selection is the proportion
of the vehicle's audience in high or low market potential
segments. If vehicles can be selected based on the Interaction of sex, region, income, and education, then the
cost per unit of market potential delivered may be minimized. This is particularly valid when one-twentieth of
the sample accounts for one-fifth of the market.
The AID program is most valuable in demonstrating
the interaction between variables. Region was a diseriminator only within the female group. Income and
level of u.sage were important only within the femaleNorthcentral subsegnient. Moreover, education had an
opposite effect on usage level for males compared to
females in the Northcentral states.
The female orientation of the brand was well known
prior to the analysis. Yet the sharp interaction between
sex, region, rate of use, and income was not suspected,
nor could it have been uncovered without a multivariate
program for data reduction.
ATTITUDINAL SEGMENTATION BY AID
In some cases, perceptual rather than socioeconomic
segmentation of the market may be desirable as a first
step. In one application, the objective was to define
differences in attitudes for a commonly used beverage
by usage rate. Thirteen attitudinai criteria were developed. A national sample of 2,000 housewives was then
Figure 3
ATTITUDINAL SEGMENTATION BY USAGE LEVEL
y
Restores
energy
9-10
Yearround
drink
8-10
47.5%
5.3%
70.5%
ii.r;
Easy to
prepare
10
<
U.7-A
35.3%
31.7%
E;isy to
prepare
0-9
Yearround
drink
0-7
1
3
•
37.1'/.
21.r;
Low in
calories
8-10
Easy lo
prepare
9-10
68.3%
/
22.2'r
29.i:;,
i.r:,.
38.0?;
Restores
energy
0-8
Low in
calories
0-7
2
10.2%
Economical
0-7
/
17.S';
30.3:;
\
Stopping rules invoked:
_!_= Sanifiic size loo small
2_= Spin cligibiiiiy mterion not
met
/ met
_3_= Spin rcdiitihilily criterirm not
met
wi met
24.0'.;.
Easy lo
prepare
0-8
21.6%
i:,>;
S.7';
Good
taste
8-10
5.7%
15.4',;
15.9';
Good
tasle
0-7
4..V'
23.8';^
Economical
8-10
12.4';
2
JOURNAL OF MARKETING RESEARCH, MAY 1970
158
asked to rate the degree to which the beverage satisfied
each criterion on a ten point scale, with ten being the
ideal. The thirteen criteria were:
1.
2.
3.
4.
5.
6.
7.
8.
9.
10.
11.
12.
13.
Good tasting
Good for serving guests
Good at mealtimes
Refreshing
Easy to prepare
Provides lift and pickup
Good for the family
Low in calories
Economical
Good as year-round drink
Restores energy
Thirst-quenching
Pure ingredients.
Figure 3 presents the delineation of segments by these
criteria. The dependent variable is the number of units
consumed by the family in the last week. "Restores
energy" was the criterion that most sharply differentiated
usage rates. Those rating the drink close to the ideal
on this criterion drank twice as much of the beverage.
Once again, the interactions demonstrated were the
most important output. The 5.3% of the sample that
rated the beverage close to the ideal on both "yearround drink" and drink that "restores energy," yet less
than ideal on "ease of preparation," consumed 70.5
units compared to the sample average of 23.3. Not
surprisingly, the 8.7% of the sample rating the beverage
low on energy, "ease of preparation," and taste consumed only 4.3 units. As before, the usage rate between
the most and least concentrated segments was 16 to one.
Examination of the detailed output also revealed:
1. "Year-round drink" is an important discriminator only within the subgroup that rates the beverage
high on energy.
2. "Ease of preparation" increases usage only
among those who rate the beverage intermediate to
low on energy.
3. "Good tasting" is an important discriminator
primarily among those rating the beverage low on
"ease of preparation."
In segmenting usage rates by attitudes. Figure 3 provides important implications for sales appeals. Sharply
differentiated segments require different appeals to reinforce or change perceptions of the product. As an
example, an important subsegment—29% consuming
approximately the same amount as the sample average
—rates the product low on "restores energy," yet close
to the ideal on "ease of preparation" and calories. The
inclusion of a saliency scale in the questionnaire to determine the importance of each attitudinai criterion permitted a judgment as to whether the weak criterion had
to be strengthened or the strong criteria reinforced for
maximum impact. Since a relatively small percentage
of respondents in this subgroup considered "energy" an
important criterion, it was obvious this was not a point
that had to be delivered despite the low attitudinai
rating.
SUMMARY AND CONCLUSION
Both regression and cross-classification analysis are
inadequate means of segmenting markets. Correlation
coefficients may not reflect significant differences between grouped data. Cross-classification analysis may
demonstrate such differences, yet is extremely limited in
going beyond the two variable case imd demonstrating
interactions among a large number of potential market
discriminators.
Although having certain serious defects, the AID program is a significant improvement over standard crossclassification analysis. It is capable of analyzing and
reducing large combinations of data to produce the
greatest discrimination between subgroups. The program was effective in segmenting product and brand
usage by demographic and attitudinai variables, producing strategic implications for media selection and the
delivery of copy appeals to concentrated market segments.
REFERENCES
1. Henry Assael, John H. Kofron and Walter Burgi, "Advertising Performance as a Function of Print Ad Characteristics," Journal of Advertising Research, 7 (June 1967), 20-6.
2. Frank M. Bass, Douglas J. Tigert and Ronald T. Lonsdale,
"Marketing Segmentation; Group Versus Individual Behavior," Jouriiai of Marketing Research, 5 (August I96H),
264-70.
3. J. N. Morgan and J. A. Sonquist, "Problems in the Analysis
of Survey Data and a Proposal," Journal of tiie American
Slatisticai A.'i.sociation, 58 (June 1963), 415-35.
4. —-——, The Determination of Interaction Effects, Monograph No. 35, Ann Arbor: Survey Research Center, Institute
for Social Research. University of Michigan 1964.