Consumer Segmentation with Bayesian Statistics

Market Segmentation with Bayesian Statistics
Esteban Ribero
The following report describes the process and results of a market segmentation analysis for
‘App Happy’ (fake name to mask the real company), a new entrant into the Consumer Entertainment
App Industry. A data set with 1800 observations from a survey conducted by the company was provided.
The survey included 40 attitudinal statements about technology, apps, shopping, leadership, and being
connected to the culture and the moment. The survey also contained a series of demographic and app
usage questions. It is believed that that sample is representative of the population of interest, although
no additional background was provided regarding the sampling methodology and data collection
method. App Happy is particularly interested in understanding the market from an attitudinal
perspective so it can better develop products and services aligned with those attitudes, as well as focus
on the most desired segments.
Attitudinal data is messy as it often is collected using a set of statement that consumers agree or
disagree with using a Likert scale. Consumers tend to skew towards one end or the other on the scale
and the ratings in one statement are often correlated with that of many other statements making it hard
to disentangle distinct dimensions underlying in the data. This is the best we can do since attitudes
cannot be directly observed but it is believed that they manifest in the agreement/disagreement with
those statements.
Exploratory Analysis
The raw data in the App Happy data set shows some diversity (dispersion) in consumer’s
attitudes but not a clearly discernable pattern can be observed to the naked eye. Figure 1 shows the
dispersion of the data across the 2 Principal Components of the variance. Notice that there is a high
concentration of points in the middle with more dispersion at the extremes of the two components. An
initial Exploratory Factor Analysis identified 3 to 4 main underlaying dimensions in the data but the
correlations among the attitudinal statements was too high (and in a single direction -positive) that is
was not useful as a technique for extracting features.
Manually creating dimensions by averaging the scores in similar statements and then running a
cluster analysis using different number of clusters using K-means algorithm was useful but there was still
not a clear and distinct pattern or solution emerging. Scree plots and silhouette scores, techniques to
identify the optimal number of clusters and quality of the groupings, indicated that the solution was
somewhere between 3 to 6 groups. However, it was not until we binarized each of 40 statements into
top agreement (bottom 2 boxes: Agree OR Agree Strongly) vs other agreement (Agree Somewhat,
Disagree Somewhat, Disagree, and Disagree Strongly) that a clear pattern and solution emerged. Figure
2 shows the resulting dispersion across the two first Principal Components. It is not unusual in marketing
research projects to binarize the data in this way after a solution is found. It provides a clearer
distinction between those that agree with some strength and those that do not, facilitating the
interpretation of results. In this case we did the binarization even before identifying a satisfactory
solution as part of the feature engineering process.

Figure 1. Figure 2.
With a recognizable pattern it was then easier to identify the appropriate solution. Although
binary data can be used as numeric variables it is more appropriate to use a clustering algorithm suited
for categorical data such as poLCA. Although a 3-clusters and a 5-clusters solution were explored, it was
clear from figure 2 that a 4-cluster solution existed and poLCA was able accurately classify the 1800
observations into those 4 groupings with precision.
Market Segmentation
The four-segment solution provides a useful segmentation from a marketing perspective. There
are two clear axes that divide the segments into four quadrants. Axis I is a combination of two sub-
dimensions that are highly and positively correlated between themselves. One dimension contains
attitudes towards technology (app/tech usage, amount of information out there, fear of being left
behind). We have called this dimension Tech influence/anxiety since it contains both statements
indicating enthusiasm about technology as well statements about feeling overwhelmed with the amount
of information and technological development and its influence in daily life. People scoring high in this
dimension are highly adept at using technology and are motivated to keep up with its developments.
This would be a highly desirable trait among a potential consumer of App Happy products. The other
sub-dimension included in this axis is Shopping enthusiasm. This dimension contains positive attitudes
towards shopping, impulse purchases, shopping for what is hot and trendy as well as appeal for luxury
and designer brands. People scoring high on this dimension would also be ideal for App Happy given
their tendency to be up to date with what is on trend, and the willingness and enjoyment to shop for
things that reflect who they are.
Axis II also contain 2 highly correlated subdimensions: Leadership and Being in-the-know.
Leadership is the clearest of all. It includes statements about being in control, being the first among
family and friends to try new things, being an opinion leader, optimistic and creative, and wanting to
stand out and take risks. Being in-the-know is more abstract including a diversity of statements that
reflect and active lifestyle and the desire to stay in touch with friends, music and TV shows. Both of
these dimensions are highly desirable for App Happy since people scoring high on Leadership will tend
to be early adopters, willing to take risks and try new things, as well as influencers of the larger market;
and people scoring high on Being in-the-know will enjoy products that keep them up to date with
culture and entertainment, areas of interest for App Happy.

Table 1, shows the list of attitudinal statements grouped by the four dimension described above
as well as the level of agreement for each of the four segments. The level of agreement is calculated as
an index by dividing the percentage of people in each segment that agree with the statement by the
percentage of agreement among all respondents and multiplying this by 100. The index then represents
the likelihood of a consumer in each segment to agree more (or less) with the statement than the
average. For instance, an index of 100 means that a consumer in segment x has the same probability to
agree with the statement than the average. An index of 160 means that a consumer in segment x is 60%
more likely to agree with the statement than the average, and an index of 75 means that a consumer in
segment x is 25% less likely to agree with the statement than the average. The overall level of
agreement with each statement is provided for reference. A high or low index for a statement with a
high level of agreement is even more indicative of an important difference between the groups.
Table 1. Attitudinal indexes for each statement and dimension by segment.

Notice the names of the segments in Figure 1. This is deliberate because the segments have a
logical appeal to App Happy. Although a more descriptive name can be provided, we wanted to make
the business implications take the lead. Figure 3 show the segments plotted in the quadrants created by
the crossing of the two main Axes and their sub-dimensions.
The NOW segment
As can be seen in Figure 1 and Table 1, the first segment “NOW”, it scores high on all dimensions
making it the ideal segment to start marketing to with a new product designed for technology
enthusiasts that enjoy shopping, being in-the-know, and are leaders. They make up about 19% of the
market. The difficulty marketing to them, relative to the other segments, is low and the potential
rewards can be high. Take a look at their demographic and app behaviors: Their estimated median age is
28.4 years old, slightly skewing towards females, more than half are single but a high number of them
(47%) have children 12 years old or younger. They have a decent median income and skew towards
iPhones (iOS) vs other operating systems. They have the second highest number of apps in their mobile
devices (21.4) and are the ones showing higher than average interest for a variety of apps related to
entertainment, shopping and news. Their most distinct statements are “I cannot get enough Apps”, ” I
am influenced by what is hot and what is not”, “I am very active and always on the go”, and “I love
showing off my new Apps to other”.
Figure 3. Segmentation Scheme for App Happy
The THEN segment
The THEN segment (18% of the market), shares most of the NOW enthusiasm for technology
and shopping -although to a lesser degree- but lacks the Leadership attitudes and Desire to being in-the-
know. It is the youngest off all the segments, more likely to be single, less likely to have kids under 12,
high multicultural index, lower income, and a lower average number of apps downloaded on their
mobiles (18.6). However, consumers in this segment show a variety of interests for apps about Tv

shows, Entertainment, Gaming and Social Networking. Given their high score on the Technology and
Shopping dimension they are still ideal prospects and may not require a lot of marketing efforts
specifically directed at them since they will follow the aspirational NOW segment and App Happy just
needs to make sure they are extending their marketing effort to include them. They won’t be early
adopter and may need reassurance from their more adventurous NOW counterparts, but once App
Happy has established itself among the NOW, THEN will likely follow.
The NEXT segment
The NEXT segment is highly desirable for a couple of reasons: They are the biggest group of all
(representing about 35% of the market), have the highest median income at $68,189, high education
level, skew toward iOS operating system, have the largest average amount of apps downloaded on their
devices at 23.4, and show a variety of interests for app about music, shopping, specific publications, and
other. They are a bit older with a median age of 33.9 and are less likely to be single although as likely as
the average to have children under 12 years old. They score high on the Being in-the-know dimension as
well as the leadership dimension. All ideal traits. What makes them less ideal than the other two
segments described above is that they score low on the Technology and Shopping dimensions and so
will require an extra effort to convinced them. They appear to be more discerning and cautious
consumers, not easily persuaded by the what is hot and less likely to make purchases out of impulse.
The extra difficulty convincing consumers in this segment may however be worth it since they have the
potential for the biggest reward among all segments. It may be advisable to start with the other
segments first without loosing sight of this bigger and more discerning segment that may be accessible
once enough experience and success have been collected from the easier NOW and THEN segments.
Convincing NEXT would be the ultimate goal.
The NOT segment
Good market segmentation provides prioritization as well as focus. That means sometime
ignoring certain groups. Consumers in the NOT segment lack all the characteristics that makes them
appealing to App Happy. They have the lowest interest in apps and Technology, lack enthusiasm for
Shopping, don’t care much about Being in-the-know, and won’t take the risk to lead and try new things.
They are the antithesis of the NOW segment scoring the lowest on all dimensions. They may share
similar demographic characteristics to the NEXT segment but lack all of the potential for a Consumer
Entertainment app. Although they represent a sizable portion of the market at 28%, It is advisable to
ignore this segment as it will be extremely difficult to inspire, and the potential reward may be low.
Focusing on the other three segments may increase the chances of success in this marketplace.
Typing Tool and Consumer Classification
The NOW THEN NEXT NOT framework described above can be a powerful tool for App Happy as is
develops and further explore the opportunity for products and services in the Consumer Entertainment
App industry. The key to its power is the easy classification of consumer into segments that have clear
business and marketing implications. Since App Happy is early in its effort to market to this industry is
expected, and recommended, to do further research to hone-in on the specifics of a product or service.
It would be key to be able to classify new research participants into each of the segments. To do this we
will develop a Typing Tool using one or more of the following approaches.

1. We could train a Machine Learning classifier such as a Random Forrest to identify a subset of the 40
attitudinal statements (no more than 10, ideally between 4 to 8) and include some demographic and
other info such as the number of apps, and app interest to make a short battery of questions with
high predictive power that can be included in the screeners that are used to recruit participants. A
simple web app or Excel macro will host the algorithm so recruiters can easily type in the answer to
these questions and get a suggested segment and the probabilities of belonging to each of the
segments just in case of borderline cases.
2. A faster and easier approach would be to use the conditional-item-response-probabilities provided
already by poLCA, the model-based algorithm used for the segmentation, to select the most
predictive statements of each class and simply use those statements and multiply the probabilities
to get the posterior probabilities of class membership once we know the answer to the selected set
of statements (if we don’t want to use them all). This will have the same effect as the Random
Forrest classifier but there will be no need for additional training. The information is already in the
model used for the segmentation. This is one of the advantages of poLCA vs other methods such as
k-means.
3. A simplified approach can be to craft a decision tree with the top statements in each dimension or
developing new statements that better reflect the sentiment of the two Axes described here, and
create a heuristic (rule of thumb) that would give immediate segment assignment to the user.
4. In the future, when more behavioral patterns and media consumption is collected about consumers
in each segment, it may be possible to forgo the attitudinal statement altogether and develop a
classifier trained purely on behavioral and other inferred data. This could include 1st
party and 3rd
party data. This will be a more involved process and will require large amounts of data.
Conclusion
Market segmentation is as much of an art as it is a science. Alternative segmentation schemes
are possible, and we must balance statistical rigor with marketing and business sense. Fortunately, we
found a solution that achieved both objectives here. The NOW THEN NEXT NOT scheme provides
actionable segments, clear distinctions across easily to understand dimensions, a well-balanced size of
the segments, and direct and clear marketing and business implications. The poLCA algorithm used here
was able to fully capture the natural structure of the data once it was binarized without need to further
reduce the dimensionality of the data or perform additional transformations. An additional benefit of
poLCA is that it already provides the conditional-item-response-probabilities by statement for each class
and so a typing tool can be easily constructed from these making the bridge between clustering method
and future classification seamless. The segmentation scheme and insights uncovered here can be used
immediately and we look forward to continuing our understanding of the opportunity for App Happy in
the Consumer Entertainment App industry.

Consumer Segmentation with Bayesian Statistics

More Related Content

What's hot (18)

Similar to Consumer Segmentation with Bayesian Statistics (20)

More from Esteban Ribero (8)

Recently uploaded (20)

Consumer Segmentation with Bayesian Statistics