Market
Segmentation
Marketing Analytics
Professor Kamel Jedidi
Columbia University
Overview
n Segmentation
n Basics
n Bases
n Segmentation methods
n Hierarchical cluster analysis
n K-means clustering
n Latent class clustering
n Marketing application
n Summary
What is Segmentation?
n Market segmentation is the subdividing of
a market into distinct subsets of
customers.
n Segment members are different between
segments but similar within.
Levitt on Segmentation
To think segments means you have to think
about what drives customers, customer groups,
and the choices that are or might be available
to them.
—Levitt, Marketing Imagination
Why Segment?
Designing a
n Attributes of a Barbie:
n Face (light, ivory, tan, brown)
n Eye color (blue, green, brown)
n Lip color (rose, cranberry, red)
n Hair style (short, medium, long)
n Hair color (blond, red, brown, black)
One Barbie Appeals to None
Attribute Customer
1 Customer
2
Face Light Brown
Eye
Color Blue Green
Lip
Color Rose Cranberry
Hair
Style Long Short
Hair
Color Red Black
Benefits of Segmentation
To the Firm … To the Customer …
n Identification of valuable n Customized products &
customers services
n More targeted n Personalized
promotions & marketing experience
communications n Increased customer
n Higher Customer satisfaction
Lifetime Value (CLV)
Sustainable Profit Growth Customer Loyalty & Retention
Bases for Segmentation
n Bases — characteristics that tell us why
segments differ (the “why”)
n Descriptors — characteristics that help us
find and reach segments (the “who”)
n Behaviors — metrics that help us measure
segments’ behaviors and responsiveness
to marketing efforts (the “what”)
Segmentation Bases
Descriptors Bases Behaviors
“Who” “Why” “What”
Age Needs Usage
Income Preferences Loyalty
Education Decision Frequency of
Profession processes purchase
Media Habits Lifestyles Responsiveness
Gender Attitudes to marketing mix
Taxonomy at the Pump: Mobil’s Five
Types of Gasoline Buyers
Road Warriors: True Blues: Usually men Generation F3: (for fuel, Homebodies: Usually Price Shoppers:
Generally higher income and women with moderate food and fast): Upwardly housewives who Generally aren’t loyal to
middle-aged men who to high incomes who are mobile men and women - shuttle their children either a brand or a
drive 25,000 to 50,000 loyal to a brand and half under 25 years of around during the day particular station, and
miles a year…buy sometimes to a particular age-who are constantly and use whatever rarely buy the premium
premium with a credit station..frequently buy on the go…drive a lot and gasoline station is line…frequently on tight
card…purchase premium gasoline and pay snack heavily from the based in town or along budgets…efforts to woo
sandwiches and drinks in cash convenience store their route of travel. them have been the
from the convenience base of marketing
store…will sometimes strategies for years.
wash their cars at the
carwash.
16% of buyers
Source: Allanna 16% of buyers 27% of buyers 21% of buyers 20% of buyers
Sullivan, “Mobil Bets Drivers Pick Cappuccino over Low Prices,” The Wall Street Journal,
January 30, 1995, B1
Demographic Segmentation
Income
Shared values:
Urban, fashionable
Age, aspirations
Behavioral Segmentation
n Suppose you have a Physician Rx Volume
1 10
marketing budget to 2
3
20
30
target only ten (out of 4
5
50
100
20) physicians 6
7
90
75
8 40
through detailing. 9 45
10 55
11 70
12 15
n Which ten physicians 13
14
60
80
would you target?
15 120
16 35
17 65
18 75
19 85
20 5
Behavioral Segmentation
Physician Rx Volume
15 120
5 100
6 90
19 85
14 80
7 75
18 75
11 70
17 65
13 60
10 55
4 50
9 45
8 40
16 35
3 30
2 20
12 15
1 10
20 5
Behavioral Segmentation
Brand Share by Physician
n Suppose you have Physician
15
Rx Volume Our Brand B
120 0% 0%
C D
0% 100%
data on physician 5
6
100
90
55% 0% 45%
93% 0% 7%
0%
0%
19 85 16% 69% 14% 1%
loyalty to different 14
7
80
75
0% 0% 100%
22% 5% 73%
0%
0%
drugs. 18
11
75
70
52% 48%
21% 13%
0%
8% 58%
0%
17 65 1% 99% 0% 0%
13 60 40% 60% 0% 0%
10 55 18% 0% 82% 0%
How would you
4 50 96% 0% 0% 3%
n 9 45 0% 0% 1% 99%
8 40 35% 0% 7% 58%
improve your 16
3
35
30
56% 17% 20%
0% 1% 0% 99%
7%
targeting? 2
12
20
15
51% 0%
0% 0%
0% 49%
0% 100%
1 10 90% 0% 7% 0%
20 5 48% 52% 0% 0%
Behavioral Segmentation
Brand Share by Physician
Physician Rx Volume Our Brand B C D Segment
15 120 0% 0% 0% 100% Lost
5 100 55% 0% 45% 0% Competitive
6 90 93% 0% 7% 0% Loyal
19 85 16% 69% 14% 1% Switchable
14 80 0% 0% 100% 0% Lost
7 75 25% 5% 70% 0% Switchable
18 75 52% 48% 0% 0% Competitive
11 70 21% 13% 8% 58% Switchable
17 65 1% 99% 0% 0% Lost
13 60 40% 60% 0% 0% Switchable
10 55 10% 0% 90% 0% Lost
4 50 96% 0% 0% 4% Loyal
9 45 0% 0% 1% 99% Lost
8 40 35% 0% 7% 58% Switchable
16 35 56% 17% 20% 7% Competitive
3 30 0% 1% 0% 99% Lost
2 20 51% 0% 0% 49% Competitive
12 15 0% 0% 0% 100% Lost
1 10 90% 0% 7% 3% Loyal
20 5 48% 52% 0% 0% Competitive
Segmenting the Pregnancy
Test-Kit Market
n Quidel wants to
enter the B2C
pregnancy test-kit
market.
n How would you
segment this
market?
Segmenting the Pregnancy
Test-Kit Market
Segments
“Hopeful’’ ‘‘Fearful’’
Brand Conceive RapidVue
Name
Package Pink box, Brick lettering
Smiling Baby Mauve background
Shelf Near ovulation
testing kit Near Condoms
Location
Price $9.99 $6.99
Cluster Analysis
n A class of techniques used to
classify objects into groups
n Objects within a group should be as
similar as possible
n Objects belonging to different groups
should be as dissimilar as possible
Example
ID Income in $K Education in Yrs
1 20 18
2 20 16
3 30 20
4 25 15
5 25 12
6 30 22
Data Plot
Euclidean Distance
n D12 = (20 − 20)& +(18 − 16)& =2
ID Income in $K Education in Yrs
1 20 18
2 20 16
3 30 20
4 25 15
5 25 12
6 30 22
The Distance Matrix D
ID 1 2 3 4 5 6
1 0 2.0 10.2 5.8 7.8 10.8
2 0 10.8 5.1 6.4 11.6
3 0 7.1 9.4 2.0
4 0 3.0 8.6
5 0 11.2
6 0
Hierarchical Clustering Algorithm
n STEP 1: Select Min {Dij} and join i and j at that distance
D12 = 2.0 à join subjects 1 and 2 in one group (cluster)
D36 = 2.0 à join 3 and 6 in another group
ID 1 2 3 4 5 6
1 0 2.0 10.2 5.8 7.8 10.8
2 0 10.8 5.1 6.4 11.6
3 0 7.1 9.4 2.0
4 0 3.0 8.6
5 0 11.2
6 0
Dendogram Construction
8
Joining Distance
1 2 4 5 3 6 ID
Algorithm—Continued
n STEP 2: Update D using minimum distance rule
D[ij]k = Min [Dik, Djk]
n Example:
D[12]3 = Min [10.2, 10.8] = 10.2
D[12]6 = Min [10.8, 11.6] = 10.8
D[12],[3,6] = Min [10.2, 10.8] = 10.2
ID 1 2 3 4 5 6
1 0 2.0 10.2 5.8 7.8 10.8
2 0 10.8 5.1 6.4 11.6
3 0 7.1 9.4 2.0
4 0 3.0 8.6
5 0 11.2
6 0
Linkage Rules
n Minimum (single) linkage
n Average linkage
n Maximum (complete) linkage
n Ward (based on minimizing the within-
cluster variability of distances)
Updated Distance Matrix
[1,2] [3,6] 4 5
[1,2] 0 10.2 5.1 6.4
[3,6] 0 7.1 9.4
4 0 3.0
5 0
Algorithm—Continued
n STEP 3: Pick min distance Dij and join i and j.
D45 = 3.0 à 4 and 5 are joined
Updated Dendogram
7
Joining
Distance
1 2 4 5 3 6
ID
Algorithm—Continued
n STEP 4: Update the distance matrix as in step 2
[1,2] [3,6] [4,5]
[1,2] 0 10.2 5.1
[3,6] 0 7.1
[4,5] 0
Algorithm
n STEP 5: Select min {Dij} and join i and j
[1,2] and [4,5] are joined
Updated Dendogram
7
Joining
Distance
1 2 4 5 3 6
ID
Algorithm—Continued
n STEP 6: Update distance matrix
[1,2,4,5] [3,6]
[1,2,4,5] 0 7.1
[3,6] 0
§ STEP 7: Join [1,2,4,5] and [3,6] and Stop
Final Dendogram
7
Joining
Distance
1 2 4 5 3 6
ID
Summary
Hierarchical Cluster Analysis is:
n A numerical procedure which attempts to separate a set of
observations into groups/clusters
n Members of the same group/cluster are more similar than
members of different clusters
n Agglomerative – seeks to join objects sequentially until get
one large cluster
n Obtain a tree or “dendogram” representation
n One of the most popular technique used for market
segmentation
A Marketing Application
n Attribute importance data collected
from 72 students (24 MBAs and 49
undergrad) using the constant-sum
method
Attribute Allocation
Trendy/innovative
Styling
Reliability
Sportiness
Performance
Comfort
Total 100 Points
n Are there different benefit segments?
How many segments? How are they
different?
Excerpt from the Dataset
Data Pre-processing
n It is wise to standardize the variables,
especially if they are measured with
different scales
n To perform hierarchical cluster analysis in
R, use hclust()
n hclust() requires the distance matrix as
data input
Data Pre-processing
Euclidean distances between the first five respondents
hclust()
output
Four Clusters?
The Four-Cluster Solution
% 25 40 23 12
Three Clusters?
The Three-Cluster Solution
% 37 40 23
The Three-Cluster Solution
% 37 40 23
Performance
Comfort
Appearance
Cluster 3:
Appearance-Driven Segment
NbClust() in R Uses 26 Criteria to
Determine the Number of Clusters
Are the Segments Identifiable?
Are the Segments Meaningful?
K-Means Clustering
n K-means requires the specification of the number of
clusters in advance, say S=3.
n K-Means algorithm:
1. Start by randomly assigning each subject to a cluster, s=1,…,S
2. Compute the centroid of each cluster and the distance of each
subject to each of the clusters centroids
3. Reassign each subject to the cluster with closest centroid
4. Repeat steps 2 and 3 until no further reassignment is possible
(i.e., when the within-cluster variance is minimized)
K-Means Clustering
NbClust() in R Uses 26 Criteria to
Determine the Number of Clusters
Concordance between kmeans()
and hclust() cluster memberships
Hit Rate:
(17+27+15)/73
= 81%
Latent Class Analysis—Mclust()
n Uses a statistical model (vs. numerical algorithm) to
form clusters
n Assumes that data follow a finite mixture of normal
distributions
n Estimates a family of models and selects the best
based on the Bayesian Information Criterion (BIC)
n Outputs cluster means and cluster membership for
each subject
n Requires a large sample size
Finite Mixture of Two Univariate
Normals
Finite Mixture of Two Bivariate
Normals
Figure is generated by Mathematica.
Latent Class Analysis
Segments Interpretation
Segments Interpretation
Styling/Reliab
Sport/Comfort
Are the Segments Identifiable?
Are the Segments Meaningful?
Summary
n Market segmentation is a core concept in marketing
n Customized offers to different groups of customers
n Better targeted resources and higher CLV
n Be clear about which criteria to use for segmentation
n Use multiple methods for clustering (Hierarchical, K-
Means, Latent Class)
n Use judgment and statistical criteria to decide on number of
clusters
n Make sure that clusters are interpretable, identifiable, and
managerially meaningful
