Marketing Strategy and Competitive Positioning pdf ebook

partitioned into a predetermined 
number of groups and then 
reassigns observation to cluster 
whose centroid is nearest
Non-hierarchical methods
Hill-climbing methods Cases are not reassigned to a 
cluster with the nearest centroid 
but moved between clusters 
dependent on the basis of a 
statistical criterion
Source: Based on Punj and Stewart (1983).
Table 8.1 
Clustering methods

---

213
POST HOC/CLUSTER-BASED SEGMENTATION APPROACHES
judgemental. The statistics produced will offer a guide as to where amalgamation of groups 
results in two quite dissimilar groups being joined. The internal homogeneity of the group 
will suffer. This is a starting point, and in some circumstances, where segmentation is very 
clear-cut, will be the best choice.
Figure 8.2 shows an example where there are three fairly clearly defined segments on 
the basis of the two dimensions studied. In this case, ‘eyeballing’ a plot of the positions 
of each object (in segmentation studies the objects are usually individual respondents) 
shows three clusters of objects scoring similarly, but not identically, on each of the two 
dimensions.
In most situations, however, there will be several dimensions on which the clustering 
is being conducted, and several candidate solutions, possibly ranging from a three-group 
to a ten-group solution. After narrowing down through examination of the statistics, the 
analyst will then need to examine the marketing implications of each solution, basically 
addressing the question: ‘If I treat these two groups separately rather than together, what 
differences will it make to my marketing to them?’ If the answer is ‘little difference’ the 
groups should usually be amalgamated. This is the creative element of segmentation, where 
judgement is crucial!
Finally, it should also be noted that lifestyle and geodemographic databases depend on 
some form of cluster analysis to group customers who are alike. The results obtained for 
ACORN and MOSAIC, for example, are based on judgement as to how many clusters are 
needed to represent the population adequately, just as tailor-made approaches are.
Once the segments have been identified, and described across other criteria, there is a 
need to validate the segments found.
8.2.4 Validating the segments
One of the beauties and problems of cluster analysis is its ability to generate seemingly 
meaningful groups out of meaningless data. This, and the confusion of algorithms, has 
frequently led to the approach being treated with scepticism. These uncertainties make 
validation an important part of segmentation research.