Advanced Management and Marketing – MST-21306
Marketing Models: Multivariate Statistics and Marketing Analytics, 4 th ed.
Chapters: 1, 2, 3, 4, 6, 7, 8 and 9
Chapter 1: Introduction to Marketing Models
Why Go Quant?
Business decisions are strengthened with empirical evidence. As we consider any proposed strategy or
tactic, it is helpful to have information that suggests whether or not it’s a good idea to engage in that
action. Without such information, the decision is a guess. Strategic choices are more persuasive when
they are “fact-based”, that is, based on information and evidence. The strategies are also more
believable wen communicated, and therefore it is easier to get people on board, from employees to
investors to customers.
When some strategy is implemented, managers will want to know whether doing so was worth it – what
was the return on the investment (ROI) for that set of actions. Many business and marketing questions
may be answered by statistics, numbers that characterize the state of some market or the preference of
some customer segment.
What is a model?
A model is a simplified representation of the world built to help us understand the world and make
predictions about it. In modeling, we don’t believe that the model captures all the element of the real
world – we have characterized models as simplified versions of the world. We hope to build a model that
captures the essence of what is relevant for the business and marketing questions at hand.
Models in marketing and business are analogous. Say we fit a little regression model to predict likely
numbers of purchases of our brand as a function of just two things: costumers’ stated preferences and
their past brand purchases. We don’t really believe that customers think like a regression – as if they
determine their preference, multiply it by a beta weight, add it to their past brand purchases (similarly
weighted), and the predict for themselves which brand to purchase. We can acknowledge that our
statistical model of likely customer brand choice doesn’t fully capture all customer decision variables.
But, even with just our two independent variables, we can make prediction and acquire more information
than we had before we created the model.
The ultimate question in assessing a model is whether the model is useful. In addition, the question
about whether a model is useful is often best answered in comparison to another, competing model. Our
model with two predictors (using preference and past purchases to predict their next brand choice) will
provide better predictions than a model that uses something like a customer’s citizenship to predict brand
choice, and certainly our model predictions are better than a guess. Frequently in business, we want
models to offer predictions. If we have formulated a strong analytical understanding of the customer or
market phenomenon, and we’ve formulated I into a comprehensive yet parsimonious model, which we’ve
populated with reasonably valid data, then we should be able to forecast, with some confidence,
scenarios that are more (or less) likely to unfold. We gain confidence in our models and modeling ability
through experience over many scenarios.
It’s Hip to B2
For these sorts of reasons, management has become more science-like. We use data – empirical
evidence – to build models. We use the models to understand and forecast. Management is not alone.
Models and equations for predicting success have permeated many realms of life previously thought to
be immune. Modeling and forecasting are being used in still more industries.
Chapter 2: Segmentation and cluster analysis
2.1 Introduction
Segmentation, targeting and positioning (STP) comprise the strategic arm of marketing that precedes the
tactical 4Ps. A market must be segmented before we can choose which segment(s) to target and how to
position our market offering.
1
,The goals of a good segmentation scheme are two:
First, we wish to identify groups of customers who are similar to each other, perhaps with regard to
their demographics or psychographics, but ultimately with respect to their preferences and purchases
regarding our brand.
Second, we look for segments to be different from group to group. If one group likes all the same
brands that another group likes, for our purposes, we might as well combine the groups and have one
segment rather than two.
A cluster analysis algorithm will take the input variable we feed it, compute a measure of similarity
between the entities, and group together the entities that are most similar, keeping those that are more
different in different clusters. The term ‘entities’ is intended to be general, e.g., we can cluster brands, if
we have data on the perceptions of multiple brands, but for segmentation purposes, entities are
customers, consumers or businesses. We want to find groups of customers who are similar with respect
to the variables we believe are important to our business.
2.2 Input variables
It is important to remember that statistical methods are simply procedures. They crunch whatever data
you give them, this means that we can cluster just about anything. The fact that statistical methods are
simply procedures also comes with the responsibility to use each method appropriately. The variables
that marketers use can be indicators that are geographic (country, climate), demographic (age,
education), behavioral (online purchases, subscriptions) and attitudinal (brand awareness, price
sensitivity).
If many types of variables can be used in used in a cluster analysis, how does the marketer choose
which to include?
First, look around and see what data you already have in-house, but be sure not to limit your cluster
analysis to these variables (not all of them should go in) and there may be far more interest and
important variables that you should include that you simply have no data on, yet.
Second, we can supplement our data with free secondary data online.
Third, perhaps we couldn’t derive some information we wanted from our customer’s interactions with
us, and it is unlikely to already exist in some database we might access, or at least certainly not in the
form we’d like. So, we sent out a survey and use the response.
Fourth, we’ll do a little ‘pre-processing’ beginning with checking the simple descriptive statistics on
each variable before tossing it into the mix. The useful variables have to exhibit some amount of
variance.
In final preparation for the clustering, we double-check the variables we propose. Just as we want to
include variables that actually convey distinguishing information, we also don’t want to include too
many variables that convey the same information. The cluster analyses doesn’t care how many
variables of any type go in; the issue is that redundant variable implicitly get weighted more. Before
contemplating the cluster analysis, most marketing analytics run a factor analysis to look for those
redundancies.
2.3 Measures of Similarity
Once we have determined the variables that will form the basis of the segmentation, the first thing the
computer does is to compute some measure of similarity among all the customers. A natural measure is
a correlation coefficient. Correlations range from +1 (two customers have identical patterns) to -1 (two
customers have exact opposite patterns). Very roughly speaking, a cluster analysis begins by looking for
the customers who are highly correlated. The model puts those customers together into a group to reflect
their similarity. In the next iterations, the clustering algorithm brings in customers whose data were a little
less similar, and so on.
Correlations are often fine, and they’re very frequently used, in part because many people understand
them. Yet while correlations reflect relative patterns, they don’t reflect mean differences. For some
2
, purposes, the patterns are the most important information; for other purposes, volume differences
between customers matter, such as segmenting by “low” vs. “high” frequency or monetary value.
When volume matters, we don’t want to subtract out the means (as is done in the correlation coefficient).
Instead, we’d compute a (Euclidean) distance between each pair of customers. In practice, you’ll see
correlations used the most, then distances. That proportion probably ought to be switched, but it’s a
tough battle to fight, given that people are more likely to understand correlations than distances. If a
matching coefficient is desired, it’s better to use J than SMC.
2.4 Clustering algorithms
Once we have some measure of similarities (e.g. r is higher for more similar pairs) or differences (e.g., d
is higher for more different pairs), we are ready to use cluster analysis. There are many clustering
algorithms. One class of algorithms produce “hierarchical” clusters, which means that once two
customers are put into the same segment, they are always together (put customer A and B together,
they can never be broken into separate clusters again, but C can join). Other algorithms are not
constrained by this quality (if we ask for 5 segments, then customers A and B might be in the same
segment, but if we ask for 6 or 4 segments, they might be in separate segments). Hierarchical clustering
models are computationally faster.
Hierarchical clustering models
Hierarchical clustering techniques may also be further categorized by whether they are agglomerative or
divisive. In agglomerative techniques, every customer starts in his or her own segment, and with each
iteration, the model puts together customers who are similar, either by forming a new cluster with two
similar customers, or by adding a customer to an already existing cluster because he or she seems to be
like the customers in that segment. The formation of clusters, or blending customers into clusters, or
clusters into other clusters, continues until the very end, when all customers are in the same segment.
Which way you go doesn’t really matter. Also, note that the extremes in either approach, everyone in the
same cluster or everyone in a sperate cluster, are referred to as ‘trivial’ solutions, and as the term
implies, those results are uninteresting and not helpful. We wouldn’t be conducting a segmentation study
if we really believed everyone was the same (one segment), and even if “one-to-one” marketing were
ideal, the truth is, companies need some scale of economy and cannot operate efficiently if every
customer is in a separate segment.
Single-link clustering is the name of the model or algorithm that puts a customer into a segment if he
or she is similar enough to at least one member in the existing cluster. Single-link gets its name
because only a single link from within the whole cluster needs to be similar or close to the customer
potentially being added.
Complete-link clustering uses the opposite criterion. The customer is completely linked to (or similar
to, or close to) all the members in the group. Another way to think about this criterion is that we begin
by looking into each cluster for the customer who is farthest from, or most different from another
customer.
Average-link clustering looks not at the minimum distance (or maximum similarity) like single-link, nor
at the maximum distance (or minimal similarity) like complete-link, but it looks at averages.
Ward’s method is a hierarchical clustering technique that proceeds very differently from single-,
complete-, and average-link. It operationalizes the intuition that if segments or clusters are indeed
groups of similar customers, then the variability within a group should be smaller than the variability
across the groups. The technique is often called “Ward’s methods of minimizing variance” implying its
objective of seeking to assign customers to segments so as to minimize the variance within clusters.
How do we know when minimum variance is achieved? We’ll measure the extent to which this goal is
met by using an R2 index like in regression. In a standard regression, R 2 is a measure of fit that tells us
the amount of the total variance that is explained by the regression model in proportion to the total
variance. An equivalent way of saying that maximum variance is explained is to say that error variability
is minimized. In Ward’s method, the index is defined as: R 2 = (SStotal – SSerror) / SStotal. The total sum of
squares measures how close every customer is to the overall means on all the variables used. Ward’s
method begins with each of the N sample units (e.g., each costumer) in his or her own cluster. We begin
witn N different clusters, each of size 1. At this point, we calculate beginning SS total and given that the
number of clusters, r, is the same as the sample size N, then SS error would be equal to SStotal, meaning
that, at the moment we annot do worse, whe have maximum error, when we want minimum error. When
3
