Quantitative methods in marketing 201
second, to classify respondents into
typologies.
The latent class model treats the manifest
categorical variables as imperfect indicators of
underlying traits, which are themselves inher-
ently unobservable. The latent class model
treats the observable (manifest) categorical
variables as imperfect indicators of underlying
traits, which are themselves inherently unob-
servable (latent). This technique is appropriate
for the analysis of data with discrete
components.
Essentially, LA attempts to ‘explain’ the
observed association between the manifest
variables by introducing one (or more) other
variables. Thus, the basic motivation behind
latent analysis is the belief that the observed
association between two or more manifest
categorical variables is due to the mixing of
heterogeneous groups. In this sense, latent
analysis can be viewed as a data-unmixing
procedure. This assumption of conditional
independence is directly analogous to the
assumption in the factor-analytic model.
The main advantage of latent analysis is
that it could be used for investigating causal
systems involving latent variables. A very
flexible computer program for maximum like-
lihood latent structure analysis, called MLLSA,
is available to marketing researchers. Latent
class models have great potential and no doubt
will be used more frequently in marketing
investigations in the future.
One of the major limitations related to LA
concerns the estimation problem, which pre-
viously made this class of models largely
inaccessible to most marketing researchers.
This problem was later solved by formulating
latent class models in the same way as in the
general framework of log-linear models. Latent
structure analysis models have been used in
segmentation research, consumer behaviour
analysis, advertising research and market struc-
ture analysis.
One of the best papers in this field is by
Dillon and Mulani (1989). A number of latent
structure models have been developed
(DeSarbo, 1993) for problems associated with
traditional customer response modelling (for
example, for more regression, conjoint analysis,
structural equation models, multidimensional
scaling, limited dependent variables, etc.). Such
latent structure models simultaneously esti-
mate market segment membership and respec-
tive model coefficients by market segment, to
optimize a common objective function.Cluster analysis
Cluster analysis is a generic label applied to a
set of techniques in order to identify ‘similar’
entities from the characteristics possessed by
these entities. The clusters should have high
homogeneity within clusters and high hetero-
geneity between clusters, and geometrically the
points within a cluster should be close together,
while different clusters should be far apart.
Cluster analysis is, in a sense, similar to
factor analysis and to multidimensional scaling
in that all three are used for reduced-space
analysis. That is, all three methods could
facilitate the presentation of the output data in
a graphical two-dimensional format that is
easier to comprehend and analyse. Cluster
analysis is primarily used for segmentation and
for decisions on marketing strategies towards
different segments and markets (Saunders,
1994), or in situations which involve grouping
products, brands, consumers, cities, distribu-
tors, etc. The main limitations of this technique
are that there are no defensible procedures for
testing the statistical significance of the emerg-
ing clusters and often various clustering meth-
ods yield differing results. There are several
types of clustering procedures. In Figure 9.2 a
hierarchical clustering of variables associated
with a marketing strategy for hotels is pre-
sented (Meidan, 1983). The diagram presents
the (level of) aggregation (or clusters) of vari-
ables that are product (i.e. hotels) character-
istics of a strategy, as defined by the suppliers
(i.e. hotel managers). Cluster 2 strategy
includes variables that indicate that hotels
adopting this strategy are more ‘aggressive’