Nat Evans, Pitney Bowes Business Insight
It is a standard market research practice to use psychographic segmentation as a primary tool for discerning a company’s target customer. This “customer profile” creation is a primary means by which customer behavior is bucketed into distinct groups that reflect differing customer characteristics, shopping behavior and loyalty to a retail, restaurant, or consumer package brand.
Over time, the Strategy & Analytics statistical modeling team at PBBI has given a lot of thought to the idea that customer profiles, and their use in sales forecast models, may be enhanced by weighting the customer source survey data by distance. It makes sense. The farther away target (or non-) customers are, the more pronounced the profile scores may be.
For instance, a typical customer profile for a high-end department store, with a specific high income, suburban customer segment or “cluster” (Corporate Clout, say, from Acxiom’s PersonicX lifestage segmentation system) may have a score of 200, meaning that people within the segment spend two times what an average customer spends for the concept. A low-income segment (Single City Stress, for example) may have an index score of 40, or the segment spends 40 cents for every dollar that an average customer spends. Weighting the profiles by distance, however, may yield a more intensified result. You may expect the high-income cluster to go from an index score of 200 for customers only within 0 to 3 miles, to 225 for those same customers beyond 6 miles from any given store. Perhaps the Single City Stress cluster would go from 40 to 25, meaning that the farther away the cluster is from the store, the less they are willing to patronize and spend at the department store relative to other customer segments at the same distance.
In theory, it seems to be a reasonably insightful approach. In practice, the S&A modeling team created just such an analysis, and found that among several clients’ customer databases, distance has no significant bearing on the relative spend of psychographic segments at the same distance. The following box plot will give a sense for one sample profile’s distribution:
As shown, 50% of all scores’ distributions fall between approximately 50 and 110. A couple of outlying clusters find themselves floating outside the distributions (the “1”, “2”, or “3” above the whisker for each plot), but statistically, no significant difference was found between scores at different distance increments. The distributions’ medians were roughly the same; variance was the same. This pattern was consistent among several customer files we tested, and the application of several distance weighting methods yielded no statistically significant enhancement whatsoever. Goes against the hypothesis. To be sure, there exists a multitude of ways to carve up customer data, and this analysis is by no means definitive, but from a macro level, it seems the proof is in the data. This analysis also does not mean that distance in its myriad forms (straight-line, drive time, drive distance) has no influence. Obviously, it does. Distance decay portrayed on a sales per capita, or other relevant, basis is very important indeed, and is an extremely well documented predictor of consumer behavior. It’s just a matter of proper application, and what any one retailer’s customer data is truly telling an analytics researcher.
