Segmentation and
Profiling
using
SPSS for Windows
Kate Grayson
Why Segmentation?
• Used by e.g. retail and consumer product companies
• Trying to learn about and describe their customers' buying
habits, gender, age, income level, etc.
• These companies tailor their marketing and product
development strategies to each consumer group to increase
sales and build brand loyalty.
• A valuable approach in Market Research, and SPSS offers
some useful tools to facilitate this commercial process
Segmentation in SPSS
• Most of the techniques for segmentation and profiling are
exploratory
• There is no right or wrong answer, and the results are
open to interpretation
• Trying to make sense of the data or find patterns
• Iterative techniques
• If it does not make business sense then it is not a good
model!
Segmentation in SPSS
Techniques include:
• Factor Analysis / Principal Components Analysis
• Hierarchical Clustering
• K-Means Cluster
• Non-Linear Principal Components Analysis
(PRINCALS/CATPCA)
• The new Two-Step Cluster
Which Technique to Use?
Cluster Categories
Analysis
Factor Analysis
Exploratory
Confirmatory
Discriminant
Analysis
AnswerTree
Which Test to use?
• Factor Analysis - to find patterns within variables
• Categories - use if data doesn’t fit assumptions for Factor
Analysis
• Cluster Analysis - to find patterns between individuals
• Two-Step Cluster – To use with both categorical and continuous
variables
• Discriminant Analysis - to look for differences between groups,
try to predict target variable
• AnswerTree - combinations of data, to predict target
Multivariate Analysis
• These techniques are inter-related,
but don’t have to use all of them
• Can use a combination of these
techniques to segment the data
Main Considerations
• Looking for patterns or trying to make
predictions?
• Levels of Measurement of the data (categorical or
continuous)
• Sample size
• Missing values
• Does data fulfil assumptions for test?
Before you start…….
….. Check your data!
Handling Missing Data
• Check before analysis for any patterns within missing
data
• Check before analysis that missing values are defined
as missing - otherwise may compromise the model
• Be aware that most segmentation techniques ignore
any cases with missing values - so may have less
usable data than you think!
Variable and Value
Labels….
• It is worth checking the labels on your file
• SPSS may truncate long variable and value labels in
the output, making it difficult to interpret the
output
• Make sure all the useful information is at the
beginning of the variable and value labels - so even
if they are truncated, the output is still easy to read
Data Coding
• Check the direction of the coding scheme, and maybe
consider re-coding the data if the codes are counter-
intuitive
• e.g. if have a rating scale that ranges from high to low,
rather than low to high…
• ... it can be difficult to interpret output and factor
scores etc. once the data has been through several
transformations
Sample Data
• Data = usage of underarm Brand usually use: frequency w ithin sample
deodorants for men
Frequency Percent
• Three brands tested: Valid Rambo AP Spray 1013 39.5
– ‘Rambo’: the current Rambo AP Roll-on 624 24.3
Brad AP Spray 441 17.2
market leader Brad AP Roll-on 140 5.5
– ‘Brad’ : second most Clint AP Spray 278 10.8
Clint AP Roll-on 71 2.8
popular
Total 2567 100.0
– ‘Clint’ : recently launched
product
Profiling the Customers..
‘Clint’ isn’t selling as well as was hoped, so the
research aims to find out:
• Who is buying ‘Clint’?
• What sort of characteristics do they share?
• Who is buying the other deodorants tested?
• How might the marketing campaign be changed to
ensure that the correct market is targeted?
Data Collected
• Ratings of a range of lifestyle attribute questions, e.g. ‘I
tend to own the most up-to-date products’, ‘My family is
most important thing in my life’, ‘I prefer to dress and
entertain casually’ etc. (34 of these)
• Demographics: age, type of work, exercise etc.
• Brand of D/O usually use
• How see yourself in relation to others, e.g. ‘What makes
you distinctive from your friends’
Segmentation – the steps
1. Run Principal Components Analysis on ‘attribute rating’
questions, to see if any underlying dimension in the variables
2. Check using Discriminant Analysis to see if these dimensions
help predict brand used
3. Run Cluster Analysis to see if can find similarities between cases
4. Decide if other variables need to be included, e.g. categorical
demographics
5. Run Two-Step Cluster using all variables
Factor Analysis
Factor Analysis: what is it?
• Looks for relationships between continuous variables
(based on correlations), in this case ‘attribute rating’
questions
• Derives underlying constructs or dimensions in the data
• Tries to reduce a large number of variables to a small
number of factors which explain most of the variance in
the data
• If can’t interpret the resulting solution then no good!
Run
Principal Components Analysis
on 34 rated attributes
Factor Analysis Results
The best solution produced 9 factors, interpreted below:
• F1: High computer use
• F2: Rules, need to conform
• F3: Party animal
• F4: Family man
• F5: Likes new products, experiments
• F6: Likes pampering, pays more for trusted brands
• F7: Cautious, follower rather than leader for new products
• F8: Relaxed, casual
• F9: Home loving
Do these factors help?
Run Discriminant Analysis to see if can predict D/O used
Combined Groups Plot
4
Brand usually use
Rambo AP Spray
Rambo AP Roll-on
Brad AP Spray
Brad AP Roll-on
2 Clint AP Spray
Clint AP Roll-on
Group Centroid
Function 2
Rambo AP Spray
0
Brad AP Roll-on Rambo AP Roll-on
Clint AP Spray
Clint AP Roll-on
Brad AP Spray
-2
-4
-7.5 -5.0 -2.5 0.0 2.5 5.0 7.5
Function 1
Factor Analysis Results
• The factors are good at predicting ‘Rambo’ usage, but not
at differentiating between ‘Brad’ and ‘Clint’
• So try instead investigating relationships between cases –
using Cluster Analysis
• Options for clustering are:
• Hierarchical Cluster
• K-Means Cluster
• Two-Step Cluster
Hierarchical Cluster
• This is often thought of as the ‘proper cluster’
method
• Looking for natural groupings within the data
• Bases groupings upon the similarity or
dissimilarity between cases, rather than variables
• Very iterative technique – time consuming!
Clustering Data - Diagram
= data point:
one case
Decisions before Cluster:
•Which variables to use?
•Which distance measures between cases to use?
•Which criteria for creating clusters to choose?
NB
The quality of the analysis will always depend upon the variables
used
Cluster Analysis will always find a solution!
It is not possible to assess in the analysis itself how appropriate a
variable is
Stages of Hierarchical Cluster:
Select variables for analysis (carefully!)
Build and assess model
Save cluster membership
If required, create cluster matrix for K-Means
NB
Because based on cases, need to make sure data is measured
on same scale - if not, data should be standardized
Run
Hierarchical Cluster Analysis
on Saved Factor Variables
Decision with D/O Data
• I can’t get a very good (i.e. useful to the
business) model from Hierarchical Cluster
analysis
• Also, I want to be able to include both
categorical and continuous variables in the same
model
• So I decide to use Two-Step Cluster instead
Two-Step Cluster
Two-Step Cluster
• The TwoStep Cluster Analysis procedure is an exploratory tool
designed to reveal natural groupings (or clusters) within a data
set that would otherwise not be apparent.
• The algorithm employed by this procedure has several features
that differentiate it from traditional clustering techniques:
– The ability to create clusters based on both categorical and
continuous variables.
– Automatic selection of the number of clusters.
– The ability to analyze large data files efficiently.
TwoStep Cluster
• Uses scalable cluster analysis algorithm
• This algorithm can handle both continuous and categorical variables
or attributes and requires only one data pass in the procedure
• The first step of the procedure pre-clusters the records into many
small sub-clusters
• Then it clusters the sub-clusters created in the pre-cluster step into
the desired number of clusters
• If the desired number of clusters is unknown, TwoStep Cluster
analysis automatically finds the proper number of clusters
Two-Step Cluster
This is unlike other clustering methods in SPSS - if the
desired number of clusters is unknown, TwoStep Cluster
analysis automatically finds the proper number of clusters
Or you can pre-specify the number of clusters required -
flexibility
Run
Two-Step Cluster Analysis
on Saved Factor Variables
and Categorical Variables
Link to more information
• More useful information about Two-Step Cluster
can be found at the following websites:
• [Link]
S/Algorith.120/twostep_cluster.pdf
– NB This was the handout for the talk, with algorithm etc.
• Also useful:
• [Link]
• [Link]
Some of the output produced
by the Two-Step Cluster
Analysis is reproduced in the
next few slides
Brand usually use by Cluster
‘Clint’ spray seems to be associated with Cluster 6, with the
roll-on version being associated with Clusters 4 and 2
branduse Brand usually use
Percent
Cluster
1 2 3 4 5 6 Combined
Rambo AP Spray .0% 18.1% 52.3% .0% 29.6% .0% 100.0%
Rambo AP Roll-on 70.3% 29.7% .0% .0% .0% .0% 100.0%
Brad AP Spray .0% .0% .0% 100.0% .0% .0% 100.0%
Brad AP Roll-on .0% 3.6% .0% 96.4% .0% .0% 100.0%
Clint AP Spray .0% .4% .0% .0% .0% 99.6% 100.0%
Clint AP Roll-on .0% 14.3% .0% 85.7% .0% .0% 100.0%
Employment Status by Cluster
Cluster 2 (‘Clint’ roll-on) is largely made up of part-time, retired
and not working respondents, Cluster 4 also has a high number of
retired respondents, while Cluster 6 ‘Clint’ spray) also has a high
percentage of part-time and unemployed.
employ Employment Status
Percent
Cluster
1 2 3 4 5 6 Combined
Full time
24.5% 2.3% 29.7% 13.8% 16.8% 12.9% 100.0%
employment
Part-time
11.9% 61.9% 4.8% 2.4% .0% 19.0% 100.0%
employment
Not employed .0% 79.3% .0% 9.8% .0% 10.9% 100.0%
Student .0% 91.9% .0% 1.0% .0% 7.1% 100.0%
Retired .0% 61.1% .0% 33.3% 5.6% .0% 100.0%
Age Group by Cluster
Cluster 2 (‘Clint’ roll-on) is largely made up of the younger and
older age groups, Cluster 4 also has a high percentage of
older respondents. Cluster 6 is more from 25 years upwards
agerseu Age of respondent
Percent
Cluster
1 2 3 4 5 6 Combined
Under 18 .0% 96.8% .0% 3.2% .0% .0% 100.0%
18-24 .0% 57.2% .0% 6.9% 27.6% 8.3% 100.0%
25-34 18.3% 11.1% 47.4% 10.5% .0% 12.7% 100.0%
35-44 23.0% 3.8% 44.6% 14.2% .0% 14.4% 100.0%
45-54 29.7% 5.5% .0% 13.1% 39.5% 12.2% 100.0%
55-64 15.5% 38.7% .0% 15.2% 19.1% 11.6% 100.0%
65 or over .0% 68.8% .0% 18.8% .0% 12.5% 100.0%
TwoStep Cluster Number = 4
Cluster 4 F1: High computer use
(‘Clint’ roll-on) F8: Relaxed, casual
has below F2: Rules, need to conform
average F9: Home loving
Variable
computer use F4: Family man
and need to F7: Cautious, follower rather than
leader for new products
conform, F5: Likes new products,
experiments
above F6: Likes pampering, pays more for
trusted brands
average F3: Party animal
on
‘Home Loving’ -30 -20 -10 0
& ‘Family
Man’
TwoStep Cluster Number = 6
Cluster 6 F1: High computer use
(‘Clint’ spray) F2: Rules, need to conform
has above F8: Relaxed, casual
average F7: Cautious, follower rather than
leader for new products
Variable
scores on F4: Family man
‘Relaxed, Casual’ F6: Likes pampering, pays more for
trusted brands
but not much F9: Home loving
else – this is F5: Likes new products,
experiments
Mr Laid Back! F3: Party animal
-40 -20 0 20 40
Summary of Findings
• Profiling of this data suggests that ‘Clint’ is not targeting
the expected market
• ‘Clint’ is often not seen as sufficiently different from
‘Brad’, it has no perceived USP
• ‘Clint’ is being used by a high percentage of older, retired,
and part-time or not employed consumers, which may be a
result of the aggressive product launch campaign with free
samples, discounted prices etc.
• ‘Clint’ marketing needs some more work!
Summary of Segmenting and
Profiling this data using SPSS
• Principal Components Analysis helped investigate
relationships between the rated attribute variables
• Hierarchical Cluster was used to try and find similarities
between cases, using the factors derived from PCA
• Two-Step Cluster was then used to enable clustering of
both continuous and categorical variables in the same
model
• Useful conclusions were drawn about the market
positioning of ‘Clint’ deodorant