Content uploaded by Manoj Hastak

Author content

All content in this area was uploaded by Manoj Hastak on Nov 30, 2020

Content may be subject to copyright.

Segmentation approaches in data-mining: A comparison of

RFM, CHAID, and logistic regression

John A. McCarty

a,

⁎, Manoj Hastak

b,1

a

School of Business, The College of New Jersey, Ewing, New Jersey 08628, United States

b

Kogod School of Business, American University, Washington, DC 20008, United States

Received 1 November 2005; received in revised form 1 April 2006; accepted 1 June 2006

Abstract

Direct marketing has become more efficient in recent years because of the use of data-mining techniques that allow marketers to better segment

their customer databases. RFM (recency, frequency, and monetary value) has been available for many years as an analytical technique. In recent

years, more sophisticated methods have been developed; however, RFM continues to be used because of its simplicity. This study investigates

RFM, CHAID, and logistic regression as analytical methods for direct marketing segmentation, using two different datasets. It is found that

CHAID tends to be superior to RFM when the response rate to a mailing is low and the mailing would be to a relatively small portion of the

database, however, RFM is an acceptable procedure in other circumstances. The present article addresses the broader issue that RFM may focus

too much attention on transaction information and ignore individual difference information (e.g., values, motivations, lifestyles) that may help a

firm to better market to their customers.

© 2006 Elsevier Inc. All rights reserved.

Keywords: Database marketing; Data-mining; RFM; CHAID; Analytical procedures

1. Introduction

Segmentation in direct marketing has become more efficient

in recent years because of the development of database

marketing techniques. These data-mining approaches provide

direct marketers with better ways to segment their current

customers and develop marketing strategies tailored to

particular segments and/or individuals. Over the recent years,

database marketing techniques have evolved from simple RFM

models (models involving recency of customer purchases,

frequency of their purchases, and the amount of money they

have spent with the firm) to statistical techniques such as chi-

square automatic interaction detection (CHAID) and logistic

regression. More recently, neural network models are employed

in the database marketing arena (Yang, 2004).

In spite of recent statistical advances in data-mining,

marketers continue to employ RFM models. A study by

Verhoef et al. (2002) shows that RFM is the second most

common method used by direct marketers, after cross tabula-

tions, in spite of the availability of more statistically

sophisticated methods. There are a couple of related reasons

for the popularity of RFM. As Kahan (1998) notes, RFM is easy

to use and can generally be implemented very quickly.

Furthermore, it is a method that managers and decision makers

can understand (Marcus, 1998). This is an important consid-

eration in that a successful technique for a direct marketer is one

that differentiates likely responders to a particular mailing from

those who are unlikely to respond, yet does so in a way that is

easy to explain to decision makers. However, it has been argued

that the simplicity of RFM has been overemphasized, but its

ability to differentiate, relative to statistical techniques, has not

been considered to the extent that it should be (Yang, 2004).

Although the efficiency of RFM has been questioned, little

research documents its ability relative to newer statistical

techniques. This paucity of research is partly because RFM

refers to a general approach to data-mining; there are a variety

Journal of Business Research 60 (2007) 656 –662

⁎Corresponding author. Tel.: +1 609 771 3220.

E-mail addresses: mccarty@tcnj.edu (J.A. McCarty),

mhastak@american.edu (M. Hastak).

1

Tel.: +1 202 885 1973.

0148-2963/$ - see front matter © 2006 Elsevier Inc. All rights reserved.

doi:10.1016/j.jbusres.2006.06.015

of ways of applying the use of recency, frequency, and monetary

value. Research that has been conducted on the efficacy of RFM

generally focuses on proprietary or judgmental models of RFM

(e.g., Levin and Zavari, 2001; Magidson, 1988) and not on

empirically based RFM models. More recently, research has

moved away from RFM and has focused instead on newer, more

sophisticated approaches to data-mining (c.f., Deichmann et al.,

2002; Linder et al., 2004). The current study evaluates one

popular, empirically based (as opposed to judgmental) approach

to RFM. This RFM approach is compared to CHAID and

logistic regression, in an effort to understand its capabilities as a

database marketing analytical tool.

2. Analytical segmentation methods in data-mining

2.1. RFM analysis

Recency, frequency, and monetary (RFM) analysis has been

used in direct marketing for a number of decades (Baier et al.,

2002). This analytical technique grew out of an informal

recognition by catalog marketers that three variables seem

particularly related to the likelihood that customers in their

house datafiles would respond to specific offers. Customers

who recently purchased from a marketer (recency), those who

purchase many times from a marketer (frequency), and those

who spend more money with a marketer (monetary value)

typically represent the best prospects for new offerings.

As noted, RFM analysis is utilized in many ways by

practitioners, therefore, RFM analysis can mean different things

to different people. One common approach to RFM analysis is

what is known as hard coding (Drozdenko and Drake, 2002).

Hard coding RFM is a matter of assigning a weight to each of

the variables recency, frequency, and monetary value, then

creating a weighted score for each person in the database. The

assignment of weights is generally a function of the judgment of

the database marketers with a particular database; for example,

past experience may tell a marketer that recency should weigh

twice as much as frequency and monetary value. Therefore, this

application of RFM is often referred to as judgment based RFM.

The weightings could also vary as a function of the particular

mailing (Baier et al., 2002). The weights can, of course, be

empirically derived based on offerings mailed to database

members in the past, thus relying on previous data rather than

judgments.

Regardless of the way that RFM is utilized, there are two

common characteristics of RFM procedures. First, RFM is used

to segment a house file (i.e., a company's current customers)

using information related to recency, frequency, and monetary

value. RFM is not applicable to the prospecting for new

customers because a marketer would not have transaction

information for prospects. Second, RFM analysis generally

focuses on the three behavioral variables of recency, frequency,

and monetary value. Although these variables are considered

powerful predictors of future behavior, traditional RFM is

limited to these three things.

A well known, empirically based RFM method is a pro-

cedure advocated by Arthur Hughes (2000). Hughes' approach

is applicable in instances when a marketer intends to send a

mailing to customers in its database and would like to find those

in the database who are the most likely to respond to the specific

mailing. Hughes recommends a test mailing to a sample of

customers in the file; then the selection of the members of the

rest of the file is made as a function of the results of the test.

Thus, compared with hard coding RFM, Hughes' method is not

arbitrary with respect to the weighting of recency, frequency,

and monetary value. The importance of each of these is

determined by the test mailing for the particular offer.

The first step in the method is for the marketer to sort the

customer file according to how recently customers have pur-

chased from the firm. The database is then divided into equal

quintiles and these quintiles are assigned the numbers 5 to 1.

Therefore, the 20% of the customers who most recently pur-

chased from the company are assigned the number 5; the next

20% are assigned the number 4, and so on. The next step

involves sorting the customers within each recency quintile by

how frequently they purchase from the marketer. For each of

these sorts, the customers are divided into equal quintiles and

assigned a number of 5 to 1 for frequency. Each of these groups

(25 groups) is sorted according to how much money the cus-

tomers have spent with the company. These sorts are divided

into quintiles and assigned numbers 5 to 1. Therefore, the

database is divided into 125 roughly equal groups (cells)

according to recency, frequency, and monetary value.

Hughes recommends conducting a test mailing to a randomly

sampled subset of each cell (e.g., 10%). After the responses of

the test mailing are received, the proportion of respondents in

each cell can be calculated. The cells can then be ordered as a

function of response percent. The marketer can then elect to

mail to a certain portion of the remaining file (e.g., the top 20%

of the cells). Alternatively, the marketer can elect to mail to the

cells that are above a break even percent, given the cost of the

mailing and the expected revenue for each return. For example,

if a mailing costs $1.50 and the revenue received is $50.00 per

order, the break even percentage would be 3%. Thus, for the

90% of the file that is left after the test mailing, the direct

marketer would mail to the RFM cells that the test mailing

predicted a 3% or better return.

It is important to note that Hughes' method does not assume

a monotonic relationship between the dependent variable

(responded/did not respond) with the variables of recency,

frequency, and monetary value. Each cell is a discreet group that

is considered individually in terms of its performance. Thus,

if middle levels of one of the independent variables (e.g.,

frequency) are more related to response compared with higher

or lower levels of this variable, then the procedure can accom-

modate the non-monotonic nature of the relationship.

2.2. CHAID

Chi Square Automatic Interaction Detector (CHAID) (see,

for example, Sargeant and McKenzie, 1999) is a method of

database segmentation that has been used for a number of years.

Research has shown that CHAID is superior to judgment based

RFM with respect to the identification of likely responders

657J.A. McCarty, M. Hastak / Journal of Business Research 60 (2007) 656–662

(Levin and Zavari, 2001; Magidson, 1988). CHAID is similar to

the RFM approach of Hughes because it creates groupings

(nodes) of database members. The main difference is that these

groupings are not created a priori as is the case with RFM.

Rather, the file is split according to a statistical algorithm after a

test mailing is conducted. After the returns of the test mailings

are received, the procedure starts with a node that includes

everyone in the test file. The procedure then searches for the

independent variable (e.g., number of times purchased) that best

discriminates among the file members with respect to a

dichotomous variable (i.e., purchased/did not purchase on

current mailing). It splits the original node on this independent

variable into as many subgroups as are significantly different

with respect to the dichotomous variable. The procedure then

splits these new nodes according to the variables that

discriminate each of them. The procedure continues until no

other splits are significant. CHAID analysis is often called tree

analysis because a trunk (original node) is split into branches,

then more branches, etc. The terminal nodes are those that can

not be split any further.

The analysis is similar to RFM because the terminal nodes can

be evaluated according to which ones break even with respect to

expected profit and mailing costs. The direct marketer can then

use the rules that define the terminal nodes in the test mailing

(i.e., levels of the independent variables that define each

terminal node) to select the groups of people left in the file

after the test that should receive the mailing. It is also similar

to RFM in that CHAID can accommodate relationships between

the dependent variable and the predictor variables that are non-

monotonic. For example, if the number of times purchased relates

to the dependent variable, CHAID may divide the file members

into three nodes: those who purchase 1 to 3 times, those who

purchase 4 to 8 times, and a third node of those who purchase 9 or

more times. These three nodes represent discreet groupings.

An important difference between CHAID and RFM is that

CHAID can accommodate a variety of independent variables.

The independent variables could include recency, frequency,

and monetary value, but could also include other transaction

variables (e.g., used a credit card or not), as well as individual

difference variables such as demographic and psychographic

variables.

2.3. Logistic regression

Logistic regression is a modeling procedure where a set of

independent variables are used to model a dichotomous criterion

variable. Therefore, it is appropriate for direct marketers who

would like to model the dichotomous variable of respond/don't

respond to a mailing. Logistic regression is particularly useful in

these circumstances in that the actual criterion variable is

dichotomous; however, the predicted variable is the response

probability, which varies from zero to one. Therefore, the model

can provide a probability of response for everyone in the file,

given the estimated parameters for a set of predictor variables.

After a test mailing similar to CHAID, logistic regression can

be used to analyze the response variable as a function of several

independent variables (e.g., number of times purchased) and

provide an equation that can calculate the response probability

for the entire house file. The marketer can then mail to everyone

left in the file (excluding those in the test) who has a probability

higher than the break even percent. Similar to CHAID, the

independent variables are not restricted to recency, frequency,

and monetary value.

Logistic regression differs from both RFM and CHAID in

two important ways. First, logistic regression provides a

response probability for individual members of the dataset

rather than creating discreet groups of people. Therefore, in

theory, each person in the dataset may have a different response

probability. In practice, however, if few independent variables

are used to construct the logistic function and each has a small

number of different possible values, then there would be

a relatively small number of different response probabilities

across the people in the file. Second, for continuous predictor

variables, logistic regression model relationships of the

independent variables with the dichotomous dependent variable

that are monotonic; both RFM and CHAID are distribution free.

This has implications for the performance of logistic regression

in instances where the relationship between a predictor variable

and the response variable is neither continuously increasing nor

decreasing. For example, when the relationship between

recency of previous purchase and purchase on the test mailing

is curvilinear, logistic regression may not be able to capture the

relationship in ways similar to that for RFM or CHAID.

3. The studies

The viability of Hughes' approach to RFM as a method to

segment a marketer's customers is evaluated with two datasets.

The RFM method is compared with CHAID and logistic

regression. For these comparisons, CHAID and logistic

regression are limited to the same information used to create

recency, frequency, and monetary value for the RFM method.

Thus, the studies are designed to assess the discriminating

characteristics of the three segmentation approaches given the

same independent variables.

Both datasets were provided by the Direct Marketing

Educational Foundation. One of these datasets is for a multi-

division mail order company; the other is for a non-profit

organization that solicits contributions from its members.

Therefore, these two files provide somewhat different situations

for which a database marketer would apply segmentation

techniques to find likely responders to a mailing. Both datasets

include information about how recently each person purchased

(or contributed for the non-profit organization), how many

times each person purchased (contributed), and the lifetime

dollar amount of purchases (contributions). Each dataset also

has the results of a recent mailing. These results include the

percentage of the dataset that responded to the mailing. The

mail order company's dataset includes 99,200 people and has a

return rate of 27.4% for the recent mailing. Thus, the two

datasets provide for tests of the sensitivity of the segmentation

procedures at very different levels of response (i.e., less than 5%

responding versus over one quarter of the file responding to the

mailing).

658 J.A. McCarty, M. Hastak / Journal of Business Research 60 (2007) 656–662

The frequently used method of cross validation is employed

to evaluate the three segmentation methods. Each of the datasets

is randomly split in half. One half of each dataset is considered

the test group; the second half of each set is a hold out sample.

RFM (Hughes' method), CHAID, and logistic regression are

applied on the test group of each dataset using the number of

days since last activity, total number of purchases (contribu-

tions), and total dollar amount of purchases (contributions). The

parameters developed from the test groups for each of the

segmentation procedures are then applied to the hold out sample

for each dataset. Thus, one could consider the test group to be

analogous to the test mailing of a typical direct marketing

situation and the hold out sample to represent the rest of a direct

marketer's house file to which the results of the test mailing

would typically be applied.

The viability of each procedure is evaluated in two ways. For

these analyses, the percentages of all respondents who would be

reached if a mailing is sent to only a subset of the individuals in

the file (as opposed to the entire file) are calculated for the test

and hold out samples; the subsets include those file members that

each procedure assesses to be most likely to respond to the

procedure. For example, if a procedure is used to select 10%

(or 20%, or 30%, etc.) of the individuals in a file who are deemed

most likely to respond, what percentage of actual respondents

are included in the group that is selected? Note that if a procedure

performs no better than chance, one would expect a selection of

10% of names in a file to yield 10% of all respondents. Database

marketers typically refer to these proportions as gain percentages.

The researchers evaluate the three segmentation procedures at

four levels of depth in the file—20%, 30%, 40%, and 50%.

One analysis involves comparing the gain percentage for a

particular depth of the file (e.g., 10%) in the test group with the

gain percentage for the same file depth in the hold out sample.

This approach provides information about the reliability of the

model developed in the test group. If the model is reliable, it can

be expected that gain percent in the hold out sample will not

differ appreciably from gain percent in the test sample.

Alternatively, a significant difference between the proportion

in the test group and the proportion in the hold out sample

would suggest that the particular segmentation method may be

misleading at that level of file depth.

The second set of analyses involves comparing the gain

percentage for a particular depth of the file across the three

segmentation procedures. These analyses provide a head-to-

head comparison between the approaches on their ability to

discriminate between responders and non-responders.

3.1. Study 1

As noted, the first study involves data from a multi-division

catalog marketer. The marketer had made a mailing to the entire

dataset and a portion of the file responded to the offer in the

mailing. The dataset includes 96,551 members; it is randomly

split into a test sample of 48,275 people and a hold out sample

of 48,276. The overall response rate to the mailing is 2.46%.

The response rate for the test group is 2.44%; the response rate

fort the hold out sample is 2.47%.

3.1.1. Results

3.1.1.1. Reliability of the segmentation methods. Table 1

shows the proportion of respondents captured for 10%

increments of file depth from 20% to 50% of the file for each

of the segmentation methods (RFM, CHAID, and logistic

regression) for the test and hold out groups. The table also

presents the difference in proportions for each method at each

depth between the test and hold out samples. The difference

measure provides an indication of the extent to which each

segmentation method produces results in the test sample that

can be reliably replicated in the hold out sample.

As Table 1 shows, the reliability of RFM is questionable at

both the 20% and 30% levels of the file. For example, the 20%

of the test sample that RFM indicates would be the most likely

to respond captures 39.2% of actual responders. However, when

the parameters of this test are applied to the hold out sample, the

top 20% of the sample only captures 34.6% of all respondents.

Therefore, the proportion of respondents captured in the hold

out sample is significantly lower than the proportion captured in

the test sample. A significant drop in proportion (4.7%) is also

evident at the 30% of depth level. At the 40% and 50% levels of

the file, the difference in proportions of test and hold out

samples are not significantly different for the RFM segmenta-

tion method. As the table shows, the proportions of respondents

in the test and hold out samples are not significantly different for

the CHAID and logistic regression segmentation methods at all

four levels of depth. Therefore, these analyses suggest that if a

marketer elected to mail to a relatively small portion of their

house file, a test mailing using RFM may result in over

prediction of the number of respondents when the results of the

test are applied to the rest of the file.

Table 1

Percent of total responses for various levels of depth of total file

Data-mining technique

RFM (%) CHAID (%) Logistic (%)

20% depth of file

Test sample 39.2

a

38.5

a

36.5

a

Hold sample 34.6

a

37.7

b

35.8

ab

Difference 4.6 ⁎⁎ 0.8 0.7

30% depth of file

Test sample 51.3

a

49.9

ab

47.9

b

Hold out sample 46.6

a

49.9

b

47.6

ab

Difference 4.7⁎0.0 0.3

40% depth of file

Test sample 61.1

a

60.0

a

58.1

a

Hold out sample 58.0

a

58.5

a

57.1

a

Difference 3.1 1.5 1.0

50% depth of file

Test sample 71.2

a

68.8

a

67.6

a

Hold out sample 67.3

a

67.4

a

65.4

a

Difference 3.9 1.4 2.2

Note. In any row for the test or hold out sample, percentages that do not share a

common subscript are significantly different at pb.1 (two-tailed).

⁎⁎ pb.01, ⁎pb.05 (one-tailed).

659J.A. McCarty, M. Hastak / Journal of Business Research 60 (2007) 656–662

3.1.1.2. Gain percent for the segmentation methods.

Although reliability of a segmentation method is a crucial

consideration, an equally important evaluation is the relative

predictive performance of the three segmentation methods. As

Table 1 shows, the gain percentages indicate that the three

segmentation methods perform similarly within each of the four

levels of file depth in the test samples. The one significant

difference is at the 30% depth level where RFM captures a

higher proportion of respondents than logistic regression did.

When the parameters estimated in the test sample are applied

to the hold out sample, CHAID captures a significantly higher

proportion of respondents than does RFM at both the 20% and

30% levels of depth. However, the three segmentation methods

do not differ significantly with respect to performance in

gain percent at the 40% and 50% depth levels for the hold out

sample.

Overall, the results suggest that RFM underperforms, relative

to CHAID, when a marketer elects to mail to a relatively small

portion of the file. At first glance it may appear that the

differences in gain percent between CHAID and RFM are

relatively small and perhaps unimportant at a practical level even

in instances where they are statistically significant. To put these

differences in perspective in terms of potential profit, hypothet-

ical profit and cost figures can be applied to the response

information. Assume that a mailing for a direct marketer costs

$1.50 and the expected revenue per response is $100. Further

assume that the test sample represents a 10% test mailing and the

hold out sample is the rest of the house file; therefore, the results

of the hold out sample would be multiplied by a factor of nine.

Applying these assumptions to the performance of RFM and

CHAID segmentation methods with this file, the RFM method

used in the test sample would predict a profit of $284,903 in the

full file (the hold out sample multiplied by a factor of nine)

whereas CHAID would predict a profit of $277,488. The actual

mailing to the likely members of the full file as determined by the

two methods would show a very different pattern. RFM would

secure a profit of $241,469 while CHAID would provide a profit

of $274,781. Thus, CHAID would outperform RFM by $33,313.

Obviously, small differences in proportions can make a big

difference when applied to large house files.

3.2. Study 2

The second study involves data from a non-profit organiza-

tion that had made a recent solicitation for a donation from

members in its house file. The dataset includes 99,200

members; it is randomly split into a test group of 49,600 people

and a hold out sample of 49,600. The overall response rate to the

solicitation is 27.4%. The response rate for the test group is

27.3%; the response rate for the hold out sample is 27.6%.

3.2.1. Results

3.2.1.1. Reliability of the segmentation methods. Table 2

shows the proportion of respondents captured for 10%

increments of file depth from 20% to 50% of the file for each

of the three segmentation methods for the test and hold out

groups. The table also presents the difference in proportions

between test and hold outs samples for each segmentation

method at each depth.

As the table shows, the difference in proportions between

test and hold out samples for the three methods are very small

(less than a percentage point for all methods at all four levels of

depth). Moreover, none of these differences are statistically

significant. Therefore, this set of analyses suggests that for this

particular file with a rather large response rate, all three methods

are generally able to provide an accurate prediction of the

response rate when the results of a test mailing are applied to the

full file.

3.2.1.2. Gain percent for the segmentation methods. Table 2

also presents the gain percents for the three different seg-

mentation methods for the four levels of depth of the test and

hold out samples. As the table shows, there are no differences

between the performance of RFM and CHAID at all four levels

of depth for either the test or hold out samples. Therefore, for

this dataset, RFM appears to be as accurate as CHAID in

capturing likely responders when CHAID is confined to the

same independent variables as RFM.

A rather unexpected finding with respect to gain concerns the

performance of logistic regression. CHAID performed signif-

icantly better than logistic regressions at both the 20% and 30%

levels of depth for both the test and hold out samples. The RFM

method significantly outperforms logistic regression at the 30%

depth. Therefore, at least for this dataset, logistic regression may

have difficulties when a marketer would elect to mail to a small

portion of the file.

The results of study 2 suggest that RFM may perform on a

par with more sophisticated statistical techniques when the

response level is fairly high. At all four levels of file depth

Table 2

Percent of total responses for various levels of depth of total file

Data-mining technique

RFM (%) CHAID (%) Logistic (%)

20% depth of file

Test sample 36.3

ab

36.6

a

35.7

b

Hold sample 35.6

ab

36.1

a

35.2

b

Difference 0.7 0.5 0.5

30% depth of file

Test Sample 49.1

a

49.0

a

47.7

b

Hold Out Sample 48.9

a

48.8

a

47.7

b

Difference 0.2 0.2 0.0

40% depth of file

Test Sample 60.4

a

60.4

a

60.0

a

Hold Out Sample 59.9

a

60.2

a

60.4

a

Difference 0.5 0.2 0.4

50% depth of file

Test Sample 70.3

a

70.4

a

70.7

a

Hold Out Sample 70.5

a

70.2

a

70.8

a

Difference 0.2 0.2 0.1

Note. In any row for the test or hold out sample, percentages that do not share a

common subscript are significantly different at pb.1 (two-tailed).

660 J.A. McCarty, M. Hastak / Journal of Business Research 60 (2007) 656–662

tested, Hughes' RFM method is able to achieve gain levels of

likely responders as well as logistic regression at a similar level

as CHAID.

4. Discussion

The two studies present some intriguing findings with

respect to the performance of the three segmentation methods.

The two datasets present very different circumstances that may

be presented to database marketers and the results with respect

to these datasets are somewhat different. Study 2 is a non-profit

organization that solicits contributions in their house file; the

recent mailing that is modeled provided for a response rate of

roughly one quarter of the entire file. Given these features of the

dataset and mailing, RFM is as successful as CHAID and

logistic regression in capturing likely responders to the

solicitation at all tested levels of depth of the file (20% to

50%). Furthermore, the parameters of the test for RFM appear

to be as reliable as those of CHAID and logistic regression when

applied to the hold out sample. Therefore, if one were to

consider only the results of study 2, it would be concluded that

RFM is generally a robust procedure that is similar to the other

two segmentation procedures in its ability to segment likely

respondents. It appears that RFM may be successful when the

overall response rate is fairly high.

The characteristics of study 1, however, present a fairly

common scenario for database marketers. This dataset is for a

multi-division mail order company; the response rate for the

offer is under 5%. Given these relatively common character-

istics of a direct marketing situation, the results suggest that

Hughes' RFM may not perform as well as CHAID when a

marketer only mails to a small portion of the file (i.e., 30% or

less). In these instances, CHAID outperforms RFM in terms of

reliability and ability to capture likely responders. The actual

performance of CHAID in the hold out sample is quite similar to

its predicted performance in the test sample. By contrast,

the actual performance of RFM in the hold out sample is

significantly worse than its predicted performance in the test

sample. Also, CHAID captures more respondents than RFM at

both the 20% and 30% depths of the file. The superiority of

CHAID in these instances suggests that the grouping of dataset

members by a statistical algorithm, as in the case of CHAID,

may be superior to the arbitrary and a priori groupings of RFM.

In this study, CHAID creates fewer cells than the fixed and large

number used in RFM. Given the low response rate to the offer

and the large number of cells specified by RFM, there is a

greater likelihood that chance fluctuation rather than systematic

differences play a role in the outcome for RFM compared with

CHAID. Thus, the predicted level of response in the test does

not hold up when the parameters of the test are applied to the

hold out sample.

The results across the two studies allow the researchers to

consider the circumstances where RFM underperforms relative to

CHAID. The findings suggest that RFM may have difficulties

when the response rate is low (as in study 1) and the database

marketer desires to send an offering to a relatively small portion of

the entire file (30% or less). Under these circumstances, RFM may

be less reliable than CHAID. Alternatively, when the response

rate is relatively high (as in study 2) or the database marketer

desires to mail to a relatively large portion of the file, RFM may

provide results similar to CHAID and logistic regression. Overall,

the study can conclude that Hughes' approach to RFM can

perform at an acceptable level in many database marketing

situations when a direct marketer is limited to using basic

transaction variables. Given that statistical modeling can be more

costly than RFM because of the need for highly trained personnel

(Drozdenko and Drake, 2002), RFM can be considered an

inexpensive and generally reliable procedure.

Two caveats or limitations should be considered with respect

to the findings. First, the two datasets represent different sets of

circumstances, both of which are relatively common in database

marketing. The offers that are modeled in these data likely

represent fairly common mailings in direct marketing. There-

fore, the researchers assume that the characteristics of these files

are not unusual circumstances in direct marketing. Having said

this, the study must concede that these two examples may not

generalize to all other database files. House files of different

organizations may have their own peculiar characteristics and

different offers may vary on a variety of dimensions. Therefore,

the conclusions with respect to the performance of the three

segmentation methods must be tempered with the understand-

ing that they may not hold for all database marketing cir-

cumstances. For example, if there is a curvilinear relationship

between a predictor (e.g., recency) and response, this would

likely impact the performance of logistic regression, as this

method models a monotonic relationship between predictors

and response. Future research that tests these segmentation

procedures under a variety of circumstances using simulated

data would be useful. Simulating different possible relation-

ships between predictors and response will allow the research-

ers to further understand the sensitivities of the three

segmentation procedures to conditions that may arise in a

variety of direct marketing situations.

Second, the analyses compared RFM to CHAID and logistic

regression where each method is constrained to use the same

independent variables of recency, frequency, and monetary

value. This constraint is enforced to achieve a fair test of the

analytical algorithms of the three methods. Therefore, the

conclusions about the relative performance are made with the

understanding that the researchers are considering them only in

the context of these transaction variables. In practice, however,

CHAID and logistic regression are not constrained with respect

to the variables that can be used as predictors. Response to a

mailing can be modeled with a variety of variables using these

two methods. One would assume that more precise modeling

could be achieved using other variables.

This second caveat raises a broader, and perhaps more

important issue. The analysis of recency, frequency, and monetary

value, whether by an RFM model or by a statistical technique such

as CHAID, focuses entirely on the past behavior of individuals.

Although social scientists recognize the power of past behavior as

a useful predictor of future behavior, such a narrow focus likely

limits the direct marketer in their ability to understand their

customers. Zahay et al. (2004) raise this point in their discussion

661J.A. McCarty, M. Hastak / Journal of Business Research 60 (2007) 656–662

of transactional and relational data. They argue that an emphasis

on transactional information is taking a very sales oriented

approach to customers. Such an emphasis may aid sales in the

short run, however, it does not add to the long term relationship

with customers. A consideration of relational data such as

information about the motivations, attitudes, values, and lifestyles

is taking more of a marketing approach to customers. Although

these variables may be less useful than transaction information in

their ability to predict a response to an immediate marketing

activity (i.e., a mailing), they may be enormously useful in

understanding the underlying tendencies in customers. This

consideration would favor analytical techniques such as CHAID

and logistic regression that can accommodate a variety of

personality and individual difference information.

References

Baier M, Ruf KM, Chakraborty G. Contemporary database marketing: concepts

and applications. Evanston: Racom Communications; 2002.

Deichmann J, Eshghi A, Haughton D, Sayek S, Teebagy N. Application of

Multiple Adaptive Regression Spines (MARS) in direct response modeling.

J Interact Market 2002;16:15–27 [Autumn].

Drozdenko RG, Drake PD. Optimal database marketing: strategy, development,

and data mining. Thousand Oaks: Sage Publications; 2002.

Hughes AM. Strategic database marketing. New York: McGraw–Hill; 2000.

Kahan R. Using database marketing techniques to enhance your one-to-one

marketing initiatives. J Consum Mark 1998;15:491–4.

Levin N, Zahavi J. Predictive modeling using segmentation. J Interact Market

2001;15:2-22 [Spring].

Linder R, Geier J, Kolliker M. Artificial neural networks, classification trees and

regression: which method for which customer base? Database Mark Cust

Strategy Manag 2004;11:344–56 [July].

Magidson J. Improved statistical techniques for response modeling: progression

beyond regression. J Direct Mark 1988;2:6-17 [Autumn].

Marcus C. A practical yet meaningful approach to customer segmentation.

J Consum Mark 1998;15:494–501.

Sargeant A, McKenzie J. The lifetime value of donors: gaining insight through

CHAID. Fund Rais Manag 1999;30:22–7 [March].

Verhoef PC, Spring PN, Hoekstra JC, Leeflang PS. December the commercial

use of segmentation and predictive modeling techniques for database

marketing in The Netherlands. Decis Support Syst 2002;34:471–81.

Yang AX. How to develop new approaches to RFM segmentation. J TargetMeas

Anal Mark 2004;13:50–60 [October].

Zahay D, Peltier J, Schultz DE, Griffin A. The role of transactional versus

relational data in IMC programs: bringing customer data together. J Advert

Res 2004;44:318 [March].

662 J.A. McCarty, M. Hastak / Journal of Business Research 60 (2007) 656–662