Exceptional results.
What Direct Marketers Need To Know About Segmentation
Part 4 of a Multi-Part Series
By David Shepard
These four articles first appeared in Direct Magazine
In our last article we discussed the issues relating to the choice of variables that go into a Cluster Analysis. The key points were: (1) choose only those variables you want to segment around, i.e., don’t include variables that you think are irrelevant to your marketing strategy (2) standardize the variables so that scale (the size of the variables) does not become an unintended issue, $25,000 is not the same as $25m (3) consider giving more weight to some variables than others depending on your marketing objectives, and (4) use Principal Components Analysis to: (a) reduce the number of variables that will go into the solution, and (b) eliminate multi-co-linearity, the undesirable, from a modeling perspective, condition that arises because so many behavior and demographic variables are correlated with each other.
Now go run a Cluster Analysis on the data.
Not so fast.
More decisions have to be made and there’s no one right answer. Cluster Analysis is not like regression…put in the data and press the regular or logistic regression button and out comes the equation…same data in …some equation out. Unfortunately, there’s more than one way to run a cluster analysis, and the method or the process you choose will determine the outcome. Different choices…different outcomes.
Without, hopefully, presenting more information than any sane marketing person would want to know, the bottom line summary is as follows.
There are two major types of Cluster Analyses: Hierarchical and non-hierarchical.
The hierarchal methods are the easier of the two to visualize so we’ll start there. All customers start in either their own cluster and are then combined into a smaller and smaller numbers of clusters, or even easier to conceptualize, all customers start off in the same cluster and the process starts by dividing the customers into two groups, the group with the most internal variation, the least homogeneous, gets split in two and now there are three groups, and so on, and the process continues until it can no longer find a “statistical;” justification to continue. To make matters even worse there are a number of different hierarchical methods…each providing a somewhat different solution.
At this point, the marketer would ask to see profiles of the customers in each cluster. Looking at the first two cluster solution, it will be apparent that the members of each cluster will look different from each other. The same will probably be true for a three or a four cluster solution…the differences will be easy to spot…from a marketing perspective…you’ll be able to see how different strategies, different creative, different copy could be written for each cluster. After four or five clusters, it gets progressively harder to justify the additional splits…again from a marketing perspective. In most situations the clustering algorithm will produce more “statistically” different clusters, than the marketer will be able to handle. When the marketer gets to the point where he/she can not conceptualize how their approach to the last segmentation would differ from the prior segmentation…then that’s the time to stop.
To be sure that one is stopping at the right number of segments, the marketer should ask the analyst to use more than one hierarchical clustering algorithm and see if the number of desired clusters turns out to be the same in each case. Once agreement is reached with regard to the number of segments, a common practice is to run a non-hierarchal segmentation using a method that comes under the heading of a k-means solution. In a k-means solution, one must specify the number of desired clusters and the process produces a segmentation that contains that many clusters or segments. For technical reasons a k-means solution produces a “better” segmentation, once the number of segments is agreed upon. Naturally, after the k-means solution is run, profiles of the segments should be examined to be sure that the profiles more or less match the profiles produced by the hierarchical solution.
Ok, the clustering is complete…now comes the hard part. The clustering will have been done on a sample of customers. But, to be truly useful to a direct marketing person the solution has to be applied to the entire database. Packaged goods people are generally satisfied to know the segments that make up their customer base; they then profile the segments in terms of media habits and allocate their advertising budgets accordingly. Not so for direct marketers…we want to put our customers in nice homogeneous segments, profile the segments and tailor our communications, offer and products to each segment…and we don’t want to be very wrong…if we can help it…because if we’re very wrong, if the segments are not really homogenous, if the variance around the means is very wide…sure the average age is 55 but the range is between 40 and 70…then we would have done just as well using a common strategy for all segments.
To place all customers into the right segments requires another exercise, another model usually referred to as an assignment model…actually the technique that’s used most often is Discriminant Analysis and the idea is to produce a scoring equation that will allow all customers to be assigned to the cluster that comes closest to matching their personal characteristics.
If the clustering solution was based on demographic and behavior data, then the chances of building a discriminant model that will assign individuals to clusters should work fairly well. If, on the other hand, the clustering was based primarily on attitudinal values, then the probability of predicting segment membership based on available demographic and behavior data is significantly reduced. There’s no reason to assume that two persons with the same demographics and even the same behavior profile, will have the same opinion about your products and services and those offered by your competition.
Which gets us back to part one of this four part series, the key lessons to be remembered are: (1) segmentation is much trickier than prediction using relatively straight forward regression techniques… so go carefully, if not slowly and (2) make sure that your segments are really homogeneous, before you act as if they are, and when assignment is an issue, check to make sure that assignment model places customers into segments that closely resemble the segments produced by the initial analysis or market research.