
Click for a Printerfriendly Version
 Adobe PDF
Superiority of Tree Analysis Over RFM:
How It Enhances Regression
By Jim Wheaton
Principal, Wheaton Group
Original Version of an article that appeared in the
August 12, 1996 issue of "DM News"
Last month's "RFM Cells: The "Kudzu'
of Segmentation" showed why statisticsbased predictive models
— regression, neural networks, genetic algorithms, and the
like — are superior to RFM Cells. The reason is that
predictive models are more stable than RFM Cells, easier to implement,
and substantially more powerful.
If RFM is inferior, then why is it so popular? One reason
is fear of the unknown. Many of us are suspicious of statisticsbased
predictive modeling techniques because we don't understand
them. If we're having difficulty understanding exactly
what "Regression Deciles 8 through 10" are, for example,
it's going to be tough explaining to the CEO why we shouldn't
mail them!
For all of its shortcomings, RFM is understandable. Every marketer,
for example, can explain why a customer with just one previous order
— five years ago, for $5 — should not be contacted.
But even database marketers who insist on an easytounderstand,
Celldriven segmentation strategy should dump RFM. This is
because there exist statisticsbased tree analysis tools that are
far superior.
One such product is "CHAID" (ChiSquare Automatic Interaction
Detection). Another is "CART" (Classification
and Regression Trees). This is not meant to be a product plug
— I'm sure that there are other excellent tools in the
market place with which I'm not familiar. P.C. versions
of these products are available, and can be executed by marketers
without an advanced degree in statistics.
Tree analysis is similar to RFM in that each node ("group")
is defined by easytointerpret "and" statements.
Assume that we want to segment a customer file by response to a
previous mailing. To keep matters simple, we'll build
a tree model off just six variable types: Months Since Last
Order, Number of Orders, Average Order Size, Merchandise Categories,
Age, and Gender.
The modeling process for tree analysis is a series of formal statistical
procedures that divides our customer file into multiple groups —
with response rates that are significantly different from each other.
Each of these groups is defined by a subset of our six variable
types. These subsets can be very different from each other.
We'll end up with — say — 31 groups, including the
following (with their corresponding performance):
 4050 year old female jewelry buyers, with four or more purchases,
averaging $500+, and at least one purchase within the past six months
— 8% response rate.
 3035 year old male electronics buyers, with three or more purchases,
and at least one within the past twelve months — 4% response
rate.
 1825 year old male sporting goods buyers, with just one order,
38+ months ago, for under $15 — 0.5% response rate.
Clearly, these groups are no more difficult to understand than RFM
Cells. They'll be just as easy to implement. But,
they'll be much more stable than RFM Cells because they're
the product of a formal statistical process.
Having said all of this, our research group prefers regression rather
than tree analysis as its primary tool for predictive modeling.
Regression has its advantages when the goal is to rankorder a prospect
or customer file on predicted future behavior, not the least of
which is ease of implementation.
Nevertheless, we are heavy users of tree analysis as an enhancement
to regression. For one, tree analysis is a wonderful tool
for creating what are known as interaction predictor variables.
These are defined as the synergy in explanatory power that often
results when multiple variables are combined. Frequently,
such synergistic interaction variables are important components
in subsequent regression models. The following example explains
how this works:
 Assume that we're selling Florida beachfront condominiums,
and that we're building a regression model to rankorder a
prospect file by likelihood of purchase. While performing
the upfront exploratory analysis that's required for any
successful model, we notice that older individuals have a response
rate that's moderately higher than average. This makes
sense because retirees often settle in Florida.
 Then, we discover that affluent individuals also have a moderately
higher response rate. That's not surprising either.
After all, beachfront property isn't cheap.
 In a preliminary step before the actual regression, let's
assume that age and income are entered into a tree analysis —
along with many other potential predictor variables. As described
in the previous example, the tree analysis will divide our prospect
file into several groups — with response rates that are significantly
different from each other.
 Let's say that one of these groups generated by the tree analysis
is 6572 year olds with incomes exceeding $110,000. And, that
this older, affluent group has a response rate that's seven
times higher than average. This, by definition, is synergy.
Two moderately predictive variables have been combined in such a
way that an extremely responsive niche within the prospect universe
has been discovered. This is what statisticsbased predictive
modeling is all about!
Tree analysis can also enhance regression models by providing insight
on how to tailor the promotional message to the characteristics
of a given prospect or customer. In this approach, the regression
model determines whom to promote. Then, the tree analysis
provides insight into what the promotional message should be.
This twostep segmentation strategy is central to many stateoftheart
database marketing programs. We'll discuss how next
month.
Jim Wheaton is a Principal at Wheaton Group, which specializes in
direct marketing consulting and data mining, data quality assessment
and assurance, and the delivery of costeffective data warehouses
and marts. He is also cofounder of Data University.
Jim can be reached at 9199698859, or jim.wheaton@wheatongroup.com.
Top >> 
