100% Original, Plagiarism Free, Tailored to your instructions

Order Now!

Predicting Response at Bookbinders: Decision Trees

Predicting Response at Bookbinders: Decision Trees
follow the instructions in the case. Use SPSS to answer the questions in the case. Thanks.
Kenan-Flagler Business School
The University of North Carolina
Professor Charlotte Mason prepared this case to provide material for class discussion rather than to illustrate either
effective or ineffective handling of a business situation. Names and data may have been disguised to assure
confidentiality. The assistance of the Direct Marketing Educational Foundation in supplying data is gratefully
Copyright Ó 2001 by Charlotte Mason.
Predicting Response at BookBinders: Decision Trees
Recursive partitioning algorithms (or decision trees) are a versatile tool for uncovering patterns
or relationships in data. They are especially useful when there is a large set of potential
predictors and when you are not sure which are most important or what the relationships
between the predictors and the target (dependent) variable are. In the case of a binary target
variable, decision tree algorithms iteratively search through the data to find which predictors
best separate the two categories of the target variable.
So far, we have used RFM segmentation and logistic regression to predict the response to the
mailing offer for “The Art History of Florence.” Now we will see how decision trees compare as
an alternative.
Tree using Exhaustive CHAID
We’ll start with a tree using exhaustive CHAID which is one of the algorithms in SPSS’s
AnswerTree software package. Our target variable is BUYER (whether or not they bought The
Art History of Florence) and all other variables will be potential predictors. To see how the tree
grows, let’s take it one step at a time, beginning with the ‘root node’. Because decision trees
are prone to ‘overfitting’, we will split the dataset in two: two-thirds of the observations will be
used to develop the model (the training sample) and the remaining one-third will be used to test
the model (the validation or test sample) to see how well the model performs on ‘new’
Page 2
Exhibit 1 “Root” node for CHAID Tree (Training Sample)
The root node contains all the observations in the training sample. We see that 29981 or
90.88% are not buyers and the remaining 3008 or 9.12% are buyers. Next, we’ll grow the tree
one level. After specifying this, the AnswerTree software searches through the potential
predictor variables to see which one ‘best’ separates the buyers from the non-buyers, and we
see that gender is selected. These results are shown in Exhibit 2.
Exhibit 2 One Level CHAID Tree (Training Sample)
To see how well this one-level tree classifies buyers and non-buyers we can look at the
classification table and ‘risk estimate’. In AnswerTree, the risk estimate is the percent of
customers incorrectly classified. Exhibit 3 shows the misclassification for this one-level tree.
Page 3
Exhibit 3 Misclassification Matrix for One Level CHAID Tree (Training Sample)
Note that the tree predicts a total of zero buyers – meaning, so far, there are no nodes in the
tree where the buyers outnumber non-buyers. Because our sample is so dominated by nonbuyers,
it’s not surprising that the computer predicts non-buyers for many or even all nodes.
Although the classification is correct for 91% of the cases, its predictions are not very useful for
distinguishing good prospects from bad ones. To identify segments worth targeting (i.e., those
with a relatively high probability of responding), we will need to look at the specific response
rates for the nodes. First, though, let’s grow the tree another level. Once again, the
AnswerTree software searches the available predictor variables and selects which variable(s) to
branch on for this level.
Exhibit 4 Two Level CHAID Tree (Training Sample)
Misclassification Matrix
Actual Category
No Yes Total
Predicted No 29981 3008 32989
Category Yes 0 0 0
Total 29981 3008 32989
Training Sample
Risk Estimate 0.091182
SE of Risk Estimate 0.001585
Page 4
Based on 2/3rds of customer records used
for ‘training’ sample
Page 5
Based on 1/3rd of customer records used for
‘test’ sample
Page 6
Note that for females the tree branches on the number of reference books purchased, but for
males the tree branches on the number of art books purchased. Females who have purchased
2 or more reference books had more than twice the response rate (14.44% compared with
6.89%) than females who had purchased one or no reference books. Also, with CHAID
algorithms there can be binary or multi-way splits (e.g. three or more branches from a single
node) as seen above. The tree divides males into three categories or nodes based on the
number of previous art books purchased. The response rate increases substantially with the
number of prior art book purchases.
Exhibit 5 shows a complete tree after an iterative process of growing and pruning branches. So
far, the analysis has used a random sample of 32,989 (equal to 66%) of the 50,000 customers
in the dataset. These 32,989 records comprise the training sample. The remaining one-third, or
17,011 records, form the validation or test sample.
The results using the test sample are shown in Exhibit 6. The tree in Exhibit 6 uses the same
branching rules as the tree we just developed using the training sample – and was used solely
to classify the 17,011 customers in the test sample. Each customer in the test sample is ‘put
through’ the tree starting at the root node and branching until a terminal node is reached. For
example, if the first customer is a female with no past purchases of reference or art books, she
will end up in the node labeled ‘A’ in Exhibit 6. We can see that there were a total of 8078
customers who fit that profile (females with no prior reference or art book purchases) and that
7783 of them did not purchase ‘The Art History of Florence’.
Exhibit 7 shows a summary of the number of customers, the number of buyers and the
response rate in each ‘leaf’ or terminal node for both the training sample and the test sample.
The nodes in Exhibit 7 are sorted by response rate from highest to lowest.
Exhibit 7 Response Rates by Node
Training Sample Test Sample
Customers # Buyers
Customers # Buyers
1016 350 34.45% 513 168 32.75%
1788 438 24.50% 244 62 25.41%
503 119 23.66% 898 208 23.16%
5197 871 16.76% 2749 431 15.68%
2075 298 14.36% 1106 165 14.92%
803 54 6.72% 239 19 7.95%
428 26 6.07% 895 50 5.59%
1689 90 5.33% 401 22 5.49%
501 26 5.19% 1657 85 5.13%
3190 160 5.02% 231 9 3.90%
15799 576 3.65% 8078 295 3.65%
For example, node “A” in Exhibit 5 includes 15,799 customers from the training sample. Of the
15,799 customers in this node, 576 belong to the target category of Yes (i.e. they are buyers),
which is a response rate of 3.65%.
Page 7
The terminal nodes of the tree and the summary statistics in Exhibit 7 are used to identify which
segments to target and which to avoid. An important decision is how ‘deep’ in the customer
base to go. This decision may be based on the number of prospects wanted, a desired
response rate or a desired proportion of potential buyers you want to contact, or profitability.
For the BookBinders’ mailing offering The Art History of Florence, we know the following about
profits for the two groups:
·  Non-responder: -$0.50 for the cost of mailing
·  Responder: $5.50 ($18 revenue less $9 COGS, $3 shipping and $0.50 for the mailing)
We can use this information to determine which nodes are profitable to target. An equivalent
approach is to target customers in nodes with a response rate greater than or equal to the
breakeven response rate.
Validating the Model
Decision trees are prone to overfitting – meaning that the tree is overly tailored or customized to
the dataset used to create the tree. If this is the case, then the tree will do a substantially poorer
job of predicting or classifying on a new set of data. To assess the performance of the decision
tree on ‘new’ data, we use the one-third of the dataset that forms the validation or test sample.
We expect the model to perform slightly worse on the test sample compared with the training
sample – although the difference should be slight. A large discrepancy between the two
suggests that the tree has been overfit and needs to be pruned back.
Case Questions:
1. Using the information in Exhibits 5 and 6, summarize – for the Director of Marketing –
which customer groups should be targeted with this mailing.
2. Use the information in Exhibit 7 to make a cumulative gains chart for both the training
and test samples. Does the tree appear ‘overfit’? Why or why not?
3. Using the same costs as before ($18 selling price, $9 wholesale price, $3 shipping and
$0.50 mailing costs), estimate what the gross profit (in dollars and as a % of gross sales)
as well as the return on marketing would be if the “The Art History of Florence’ offer were
only mailed to those predicted by the CHAID tree results to be good prospects for this
4. Compare and contrast the results and insights from using RFM, logistic regression and
CHAID decision tree analysis for targeting buyers for BookBinders offers.
SPSS functions refresher
To conduct the step-by-step decision tree analysis, you only need a few SPSS functions:
1. To select the appropriate cases for a give node examination, go to Data>Select cases> If condition is satisfied and define the condition (e.g. age = 1 or age = 2; (female = 0) and (age = 1 or age = 2))
2. To create dummy variable, go to Transform>Create dummy variables, under main efects pick a root (e.g. ag for age and inc for income) and a macro name (m1 e.g.) and then SPSS does its magic ?
3. If you want to combine two categories, say age = 1 and age = 2, you can simply go to Transform>Compute variable and sum the dummies for these categories (age12=ag_1 + ag_2, assuming that’s how you names your dummy variables)
4. For cross-tabs, go to Analyze>Descriptive Statistics>Cross tabs and do not forget to click on Chi-square under Statistics. To see more decimals in the output table, activate by double clicking and then right click to format the cell.

Our Service Charter

  1. Excellent Quality / 100% Plagiarism-Free

    We employ a number of measures to ensure top quality essays. The papers go through a system of quality control prior to delivery. We run plagiarism checks on each paper to ensure that they will be 100% plagiarism-free. So, only clean copies hit customers’ emails. We also never resell the papers completed by our writers. So, once it is checked using a plagiarism checker, the paper will be unique. Speaking of the academic writing standards, we will stick to the assignment brief given by the customer and assign the perfect writer. By saying “the perfect writer” we mean the one having an academic degree in the customer’s study field and positive feedback from other customers.
  2. Free Revisions

    We keep the quality bar of all papers high. But in case you need some extra brilliance to the paper, here’s what to do. First of all, you can choose a top writer. It means that we will assign an expert with a degree in your subject. And secondly, you can rely on our editing services. Our editors will revise your papers, checking whether or not they comply with high standards of academic writing. In addition, editing entails adjusting content if it’s off the topic, adding more sources, refining the language style, and making sure the referencing style is followed.
  3. Confidentiality / 100% No Disclosure

    We make sure that clients’ personal data remains confidential and is not exploited for any purposes beyond those related to our services. We only ask you to provide us with the information that is required to produce the paper according to your writing needs. Please note that the payment info is protected as well. Feel free to refer to the support team for more information about our payment methods. The fact that you used our service is kept secret due to the advanced security standards. So, you can be sure that no one will find out that you got a paper from our writing service.
  4. Money Back Guarantee

    If the writer doesn’t address all the questions on your assignment brief or the delivered paper appears to be off the topic, you can ask for a refund. Or, if it is applicable, you can opt in for free revision within 14-30 days, depending on your paper’s length. The revision or refund request should be sent within 14 days after delivery. The customer gets 100% money-back in case they haven't downloaded the paper. All approved refunds will be returned to the customer’s credit card or Bonus Balance in a form of store credit. Take a note that we will send an extra compensation if the customers goes with a store credit.
  5. 24/7 Customer Support

    We have a support team working 24/7 ready to give your issue concerning the order their immediate attention. If you have any questions about the ordering process, communication with the writer, payment options, feel free to join live chat. Be sure to get a fast response. They can also give you the exact price quote, taking into account the timing, desired academic level of the paper, and the number of pages.

Excellent Quality
Zero Plagiarism
Expert Writers

Instant Quote

Single spaced
approx 275 words per page
Urgency (Less urgent, less costly):
Total Cost: NaN

Get 10% Off on your 1st order!