HYBRID SEGMENTATION: Producing A Segmentation That's Interesting And Actionable

Classical techniques of attitudinal segmentation, like Cluster Analysis, produce interesting groups that are not particularly useful. Targeting methods like CHAID produce groups related to an objective, and therefore useful, but they lack depth and are not interesting. This white paper presents a mix of the two methods - Hybrid Segmentation - as a method that produces segments that are both interesting and actionable.

Anyone who’s dealt with large groups of people – marketing to them, evaluating their opinions – will tell you: the world is complex. People are different in backgrounds, opinions, attitudes, and needs. It’s almost impossible to deal with them as a single, undifferentiated mass; in order to address the world (or the part of it that interests you) effectively, you need to break it down into manageable, meaningful groups. That’s called segmentation.

But many people shy away from segmentation, because it leaves them with a dilemma.

• Classical techniques of attitudinal segmentation, like Cluster Analysis, divide people into groups that are meaningful and interesting; but these groups are not associated either with an outcome measure (such as likelihood of purchasing a product), or with background data that would allow them to be identified so that specific messages can be addressed to them. As such, classical clustering produces groups that are interesting, but not useful.


• Targeting methods like CHAID produce groups that are related to an objective, and easily identifiable; but since CHAID doesn’t work very well with attitudinal data, the groups it produces tend to lack depth and complexity (or their complexity is uninterpretable). So CHAID-type approaches produces groups that are useful, but not interesting.

The real problem is to find a segmentation that is both useful and interesting. One possible avenue for solving this, which we at Renaissance Research & Consulting have developed successfully over a number of years, we call Hybrid Segmentation.

What is it? It is a combination of two classic segmentation/targeting techniques:

• CHAID “Tree Analysis”
• K-means Cluster Analysis

What does it do? It finds the demographic breaks that split a sample so that it segments as cleanly as possible on a set of needs or attitudes. It then produces detailed descriptive information on each segment, allowing it to be both identified, located, and interpreted.

Hybrid Segmentation can be used to answer questions like these:

• Do all consumers of a given product have the same needs and wants for it? Or do consumers of different backgrounds want different things? What are the product characteristics that motivate each of them to buy?


• Are different types of consumers attracted to different product lines? How many different lines is it worth bringing out to satisfy everybody? What would those lines look like?


• What demographics split the likely electorate into groups that coalesce around specific sets of issues? What are the issues that unite each of them?

How Does it Work?
We can best illustrate how hybrid segmentation works using the following example:

An attitude and usage study was conducted among a series of consumers who were in the market for an automobile in the next twelve months. Among other measures, the study asked the importance (on a five-point scale) of a series of automobile attributes:

• Luxury
• Sportiness
• Comfort
• Power
• Price
• Fuel Economy
• Reliability
• Safety
• Legroom
• Cargo Room

The study also collected a full battery of standard demographics on each respondent. Both the importance and demographic batteries were input into the segmentation algorithm (as separate domains).

The hybrid segmentation process is illustrated by the tree diagram below. It proceeds in a series of steps:

1. The set of demographics is scanned, one at a time, to determine which of them, if the sample were divided in two based on it (all possible break-points in each demographic are tested), create two segments whose automobile needs are as different as possible. (This is measured, as in k-means clustering, by minimizing within-cluster variance across all the attitudes.)
   
   In this example, the first split is made on gender. Breaking the sample by gender splits off a segment primarily concerned with safety (women), while the male cluster’s motivations are still heterogeneous.


2. The demographic scan is repeated for each of the two clusters, to find which demographic will split off the most attitudinally distinctive sub-cluster from either of the two.
   
   In this step, the best split divides the males on income: those with a household income of $75,000 or more form their own segment, for whom luxury and comfort are paramount.


3. The demographic scan is repeated for each of the three clusters. This time, men with incomes under $75K are split by age, forming two segments:
   1. Younger (under 35), relatively less affluent men, whose primary needs are sporty and cheap.
   2. Older, less affluent men, whose primary needs are economical and reliable.


4. The scan is repeated once more, splitting women based on whether they had children under age 6. This created two final clusters:
   1. Children under 6: Safety, Legroom and Cargo Space
   2. No children under 6: Safety and Power.

The great advantage of a hybrid segmentation is that any segments that are discovered are automatically “identified”, since they come with a “built-in” demographic profile. This has two important consequences, one substantive, the other methodological:

• Hybrid clusters are always actionable, in the sense that, since each one is associated with a particular demographic profile, a marketer always knows where to find them, and has a “leg-up” on how to talk to them (media plan), as well as knowing what to say to them.


• Hybrid clusters are self-identifying. If one wants to draw another sample containing these clusters, or to identify them in a customer database, a simple demographic screen will do the trick. On the other hand, finding purely attitudinal clusters in a new population depends on re-asking (at least a subset of) the attitudinal questions, or else depending on correlations with demographics that, because they were not part of the original clustering, often have a very sketchy relationship with the attitudes.

For applications in which segments are needed that are both attitudinally granular and attached to the “real world”, hybrid methodology such as our TargetVoice^sm system may be the way to get the best of both worlds.

This content was provided by Renaissance Research & Consulting, Inc. Visit their website at www.renaiss.com.