Developing an Effective Segmentation Scheme

In our last article (“Customer Segmentation: Problems, Solutions,” DIRECT, March 1), we defined the objective of a survey-developed attitudinal segmentation scheme that would separate some hypothetical chain-store customers into four clusters based on a set of major dimensions: price, service, product depth, product quality, overall shopping experience and brand image. And we asked how we'd go about doing it.

Table 1

ACTUAL	PROJECTED 1	PROJECTED 2	PROJECTED 3	PROJECTED 4	TOTAL
1	24%	4%	2%	0%	30%
2	3	25	2	1	30
3	1	2	20	2	25
4	0	1	1	13	15

Table 2

ACTUAL	PROJECTED 1	PROJECTED 2	PROJECTED 3	PROJECTED 4	TOTAL
1	15%	10%	3%	2%	30%
2	6	18	5	1	30
3	5	4	12	4	25
4	3	3	2	7	15

If we follow the traditional path we would probably hire a market research company to develop an in-depth questionnaire that would be presented to a random sample of a few thousand of our customers in their homes or through direct mail — maybe even over the phone.

Results of the survey would be processed, and for sure the research firm would identify four or five major segments since segmentation projects based on survey data rarely fail to produce meaningful results.

Let's call these segments:

1. Bargain hunters (30% of the sample). They only buy “on deal,” never pay full price and aren't terribly concerned with quality.

2. Quality seekers (30% of the sample). They'll pay full price if they think the quality is there, and they'll spend a long time looking for the right combination of price and quality.

3. Convenience shoppers (25% of the sample). They're pressed for time, buy only what they came looking for, aren't price conscious — and they hate to shop.

4. Old-timers (15% of the sample). They've been shopping with you for years and can't remember why.

Now, let's assume that we buy into these results — that is, we believe our customer file probably does break down into these four segments, and in proportions fairly close to those shown here.

How do we go about placing all our customers into the right segment? Remember, the segmentation was based on just a sample of our customers, and it's not practical or economically feasible to send the same questionnaire to all of them.

The answer has to be predictive modeling. We need to be able to anticipate which segment a customer is most likely to belong to. But how likely are we to succeed? To answer this we have to consider the data we have about both those who answered the survey and those who need to be scored and placed into segments.

Both groups have behavior/transaction data and/or demographics in common. Thus the question now becomes: Are the attitudes reflected in our segmentation likely to be correlated with the available behavior and demographic data?

(Survey respondents make up our modeling data set. The variable we will try to predict is segment membership, and the predictor variables will be all the demographic and behavior data we have about the respondents. If the model works it will be applied to all the other customers in the database, for whom the same data is known.)

My guess is that in this case the answer may be yes — not perfectly, of course, but sufficiently correlated so that we can come reasonably close to the results we would have gotten if the entire database of customers was surveyed.

Let's look at what we have to work with.

Segment 1: Bargain hunters should be identifiable on the basis of the prices they paid for the items purchased — assuming the database could yield this information.

Segment 2: Quality seekers should also be identifiable by examining the products purchased and the prices paid. Here, however, a lot of additional analysis and coding probably will be required if “quality” is not easily identified from data maintained on the database.

Segment 3: Convenience shoppers might be identified by gender (from the description noted earlier, I'd guess they were men) and by their shopping patterns — they probably buy relatively few items on the same day.

Segment 4: Old-timers may display a relatively constant purchase behavior over a fairly long period of time, and they may be chronologically older, as well as being on the file longer.

So, in this hypothetical case, we have a segmentation scheme that seems to make sense. For one thing, it conforms to our understanding of the business. For another, it should be reasonably predictable through modeling because the attitudes discovered would seem to be correlated with the information on our database.

This might not always be the case. It may be that the segmentation, though correct, may yield segments that are not correlated with available behavior or demographic data. For example, in financial services, an important attitudinal distinction exists between investors who like to do their own research and make their own decisions, and those who want someone else to do the hard work. But this distinction may not be correlated to any known demographic or any behavior information kept on the database.

In these situations it may be appropriate to first build a segmentation model based on demographic and behavior data and then research the segments. And again, within this framework there are numerous options.

One could segment first using just demographics and then once these segments have been determined, use behavior data to divide the demographic segments into “high value” and “low value” customers. Then, based on these two dimensional segments, one could use research to understand the differences in behavior; or demographic data could be combined with behavior data to create a similar number of segments and then those segments could be researched.

It's fair to ask, then, if a segmentation based on survey data may or may not be projectable to the entire universe, should it be done at all?

One answer is to do it and see what happens. Of course, that's not a very good answer.

Another solution is to do some limited research and a quick and dirty segmentation before rolling out the full survey and test to see if the segmentation appears to be projectable enough to be used.

Table 1 (see page 49) shows the results of what would have been a very successful prediction, with 82% of the names modeled falling into the correct segment.

In Table 2 (also on page 49) the correct classification is only 52%. It's directionally correct but not nearly as accurate. It's also the kind of result that's likely to get a direct marketer into trouble.

The trouble comes about like this: Using the results in Table 2 we're right in our placement about 50% of the time…much better than the chance rate of 26%. (The chance correct rate is about 26% ={ (30/100)_ + (30/100)_ +(25/100)_ +(15/100)_}

But that means were also wrong about 50% of the time. Now, if our creative team develops packages for each segment that are very targeted — packages that are perfectly appropriate for members of the segment when the prediction is correct, but not appropriate for misplaced members — they could possibly reduce response among these incorrectly placed customers by more than they increase response among correctly assigned customers.

When we test these targeted packages against a non-targeted control (one that appeals somewhat to all segments and turns off none of them), it's not unusual to see the general control beating the targeted packages.

Does this mean we shouldn't be doing segmentation, attitudinal or otherwise? Of course not! But it does mean that we have to proceed carefully.

How? We'll answer this question in the May 1 issue of DIRECT and begin our in-depth look at the tools of segmentation: factor and cluster analysis.

This is the second in a series of articles on customer segmentation.