Sie sind auf Seite 1von 13

Criteria for Market Segmentation Studies

Author(s): Neil E. Beckwith and Maurice W. Sasieni


Source: Management Science, Vol. 22, No. 8 (Apr., 1976), pp. 892-903
Published by: INFORMS
Stable URL: http://www.jstor.org/stable/2630021 .
Accessed: 25/11/2014 08:22
Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at .
http://www.jstor.org/page/info/about/policies/terms.jsp

.
JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide range of
content in a trusted digital archive. We use information technology and tools to increase productivity and facilitate new forms
of scholarship. For more information about JSTOR, please contact support@jstor.org.

INFORMS is collaborating with JSTOR to digitize, preserve and extend access to Management Science.

http://www.jstor.org

This content downloaded from 193.226.34.227 on Tue, 25 Nov 2014 08:22:16 AM


All use subject to JSTOR Terms and Conditions

MANAGEMENT SCIENCE
Vol. 22, No. 8, April, 1976
Printed in U.S.A.

CRITERIA FOR MARKET SEGMENTATION


STUDIES*tf
NEIL E. BECKWITH?

AND

MAURICE W. SASIENI**

Morrison recently claimed that models which attempt to explain differences in purchasing
behavior in terms of the characteristics of individual purchasers should be expected to have
low R 2,S. Here we generalize his stochastic model and also examine a slightly different
stochastic model which is consistent with the usual regression segmentation studies. Correctly
specified segmentation studies are shown to have higher R 2S (roughly exceeding 0.5) than
typically reported, indicating many previous segmentation studies are misspecified. We also
show how to estimate the correlation between prediction and long run average behavior from
short term observations.

1. Introduction
Morrison recently examined a stochastic model of purchasing behavior [8]. He
found that a "perfect" model which predicts every individual's true average purchase
rate exactly could still have R 2, below unity as an explanation of the number of units
purchased in a particular time period because of the random fluctuations of each
individual's purchases. Purchases will exceed the individual's average rate in some
time periods and be less in others. Thus, even models which well explain the
differences in average purchase rates would only poorly explain the purchases
observed in short time periods. This is due to use of sales observed over a relatively
short period of time instead of the long run average purchase rate. We demonstrate
simple methods of estimating the correlation between the prediction and the long run
average.
Morrison's model does not include the usual ordinary least squares (OLS) regression model as a special case. Thus his results do not extend exactly to segmentation
studies which employ regression models. We generalize Morrison's stochastic model
without any reference to regression models in the next section. We then examine a
revised model of regression estimators in a subsequent section.
* Processed by Professor Donald G. Morrison, former Departmental Editor for Marketing; received
February, 1974, revised September, 1975. This paper has been with the authors approximately 14 months
for revision.
t Professor Donald G. Morrison's Comment. In Morrison [8] a stochastic model for consumer purchases
was developed and the properties of this class of model were explored. Many of these properties involved
2" type statements. While the results were correct (except for an algebraic error on p. 1220-all p2r
should be 2 for all imperfect models), the models were not regression models. Wildt in the article that
follows correctly noticed that Morrison's model was not a regression. Using Wildt's comments as stimulus
for part of their article Beckwith and Sasieni fully explore the econometric and interpretive issues raised by
the use of regression models in segmentation studies.
The delay in publishing these articles was caused by numerous iterations of drafts between all three
authors. The end result of their cooperation and good humor is a solid pair of articles that clarify the issues
somewhat implicitly raised by Morrison in [8]. The readers of Management Science will benefit from the
patience displayed by Beckwith, Sasieni and Wildt.
t The authors acknowledge the help of Albert R. Wildt, whose comment [9] first brought to their
attention the nature of the inconsistency of Morrison's model with the regression segmentation studies. The
authors also benefited from correspondence with Frank M. Bass, G. S. Goodhart, Donald R. Lehmann,
Donald G. Morrison and Leonard J. Parsons, each of whom shared insights toward resolving these issues.
? Columbia University.
** Unilever Limited.
892
Copyright ? 1976, The Institute of Management Sciences

This content downloaded from 193.226.34.227 on Tue, 25 Nov 2014 08:22:16 AM


All use subject to JSTOR Terms and Conditions

CRITERIA FOR MARKET SEGMENTATION

STUDIES

893

2. Purchasing Process
The number of units Xi purchased by individual i in any particular time period is a
random variable with mean EXi and variance Var Xi. Generalized results can be
obtained subject to the restriction: Var Xi = kEXi for each individual, with k constant
across all individuals.' It is convenient to denote the average purchase rate of
individual i by Xi = EXi. Of course, the average purchase rate Xi differs between
individuals. The mixing distribution of nonnegative Xi has finite mean EX and finite
variance Var X. Otherwise, the mixing distribution can have any arbitrary shape. It is
not necessary to make any other assumptions about the type of process which
generates the average purchase rates Xi. The general results depend only upon the
constant ratio k between the mean and variance of the purchase distribution for each
individual.
Summarizing, the general assumptions about purchasing behavior which Morrison
used are:
Al'. The number of units purchased by individual i, Xi, is a random variable with
mean EXi = Xi and variance Var Xi = kXi, where all individuals have the same k > 0.
A2'. The X, are nonnegative random variables with mean EX and variance Var X.
Process Properties
For the general assumptions (A1' and A2' only) the mean and variance of the
unconditional distribution of the number of units purchased by all individuals are
EX= EX,

(1)

Var X = kEX + VarX,

(2)

and the autocorrelation Pxx between purchases by individuals in independent periods


is
Pxx = E[(Xi -EX)(Xi2

EX)]/Var X = Var X/Var X

(3)

as can be verified in Appendix A. The two periods used to estimate Pxx should be
sufficiently separated so that each individual's purchases are independent between the
two periods. The population Pxx will be positive because of the differing average
purchase rates of the individuals, even though each individual's purchases are independent between the two periods.
Aggregation
The use of a longer time interval for collection of the observed purchase data will
improve the fit of regression model predictions to the unobservable average purchase
rate Xi because the random fluctuations of the observed purchase rate will be lessened.
Lengthening the data collection period by a factor of m is equivalent to the aggregation of m independent single-period observations. This is equivalent to simply replacing k by k/m in all the results throughout this paper, as may be easily verified.2
1
Reported empirical evidence [4] suggests that k < 1 for time periods of a week or more, as an
individual's purchases in successive periods tend to be more regular than suggested by the Poisson
assumption (or k = 1).
2 If the individual's purchases are independent between purchases then over m consecutive periods the
m
individual purchases an average of Yi =
1Xil/m units per period. The mean and variance of this
average are EY = Xi and Var Yi = Xik/m which are identical with Al' except for replacing k by k/m. If
there is autocorrelation between successive periods (for individual's purchases) then k would be replaced by
k/w where w depends on the amount of autocorrelation and the number of periods.

This content downloaded from 193.226.34.227 on Tue, 25 Nov 2014 08:22:16 AM


All use subject to JSTOR Terms and Conditions

894

NEIL E. BECKWITH

AND MAURICE

W. SASIENI

3. Morrison's Generalized Model


Morrison assumed an estimator X1of individual purchases such that the error of the
estimate E = i- Xi is a random variable with zero mean, independent of Xi. Even
though the errors Ei are independent of Xi they might or might not be correlated with
the observed purchase Xi. Morrison's algebraic results reflect his additional assumption that Xi and Ei are uncorrelated, cov(X, E) = 0. The last assumption is not met by
models whose parameters are estimated from observations of the Xi. It may be met
when parameters have been estimated independently of the Xi. Specifically, Morrison
effectively assumed:
B2. A model yields predictions X, = X, + E, where Ei is a random variable with zero
mean, variance a2 , and uncorrelated with Xi and Xi.
This assumption B2 is consistent with Al' and A2', even though certain estimation
procedures, such as regression explanations of Xi, are not consistent with B2. However, the model B2 is relevant for other interesting estimation procedures. For
example, if individuals were simply asked to guess their own average purchase rate of
an item, their response error Ei might be assumed to be uncorrelated with both their
average purchase rate Xi and the number of items purchased in some recent period Xi.
We detail the generalized results for this model in this section. Then in ?4 we
examine the results of changing the assumptions to be consistent with OLS regression
estimation of Xi from the observable purchases Xi.
Model Properties
The accuracy of the model can be described either by the error variance a2
E or by
the squared correlation between the predicted Xi and the unobservable Xi in the
population (see Appendix B):
p2 = [CoV(X, X)] /(Var X Var X) = Var X/(Var X +

GE)

(4)

which is obviously always 0? p2 ?1. But we can also interpret Xi as being an


estimate of the number Xi of units purchased in a period. The expected squared error
of the prediction Xi of Xi in the population is denoted MSE:
MSE = E(X -Xi)

= kEX + q,.

(5)

The R 2 (a population parameter and not an observed sample statistic) is defined to be


= 1 - (MSE/Var X).3 For this model:

R2

R2

1 - MSE/Var X = (Var X - a2)/Var X = (2

p2)Var

X/Var X.

(6)

The interesting thing to notice here is that the model has R 2 < 0 where a,2 > Var X or,
as a critical value P2ritsince
equivalently, where p2 < 2. Morrison identified p2 =
2 < I implies negative R2 and p2 > I implies R2 is positive.4 A negative R2 simply
means that the model has a larger mean squared error than does the alternative naive
model which estimates each person's purchase rate by the cross-sectional average: Xi
= EX. The R2 may be either greater or less than p2.
Unlike the case of OLS regression models, for this model the R 2 is not equal to the
squared correlation between AXand the observable number Xi of units purchased [2].
3 Wildt [9] has pointed out that the total variance (Var X) is not equal to the sum of the explained
variance (Var A) plus the unexplained variance (Var(X - X)) for Morrison's model because Cov(X, X - X)
=
-# 0. Thus, R 2 = 1 - (MSE/Var X) is not equal to the fraction of variance explained (Var X/Var X).
4 It is interesting to note that for models with p2 < P2t = 2 the R2 is an increasing function of the mean
purchase rate kEX. However for models with p2 > P nt= =the
R2 is a decreasingfunctionof kEA.

This content downloaded from 193.226.34.227 on Tue, 25 Nov 2014 08:22:16 AM


All use subject to JSTOR Terms and Conditions

CRITERIA FOR MARKET SEGMENTATION

STUDIES

895

That squared correlation is


= [Cov(X, A)] /(Var X- Var X) = (Var X/Var X)p2
R2

p2/(l

+ kEX/Var X) = (p2/(2

- p-2))R2

(7)

which is necessarily positive and which cannot exceed p2 for this model.5 Similarly,
the squared correlation p2 between Xi and Xi cannot be interpreted as the fraction of
variance of Xi explained by Xi
Inspecting the previous results for p2, R2 and R 2, we see that increasing the data
collection period by a factor of m would not change p2 since it is not a function of
parameter k, but would increase R2 and R2: R2 = [(2- p2)Var X]/[E(X)k/m +
Var X], R2 = p2/[l + E(X)k/m Var X]. Neither R2 nor R2 approach unity as m
becomes infinite (unless a,- = 0). R2 approaches only 2 p 2 and R2 approaches only
m,
2
p2
m
p , and p is not changed by m, (unless m effects a42).
-

Special Cases
The preceding results have all been obtained from the three general assumptions
Al', A2' and B2. These results can be reduced to special cases such as:
Al. The number of units purchased is distributed Poisson for each individual
(k= 1).
A2. The Xi are distributed gamma across individuals with parameters r and t
(EX = r/t, Var X = rlt2).
Bi. A perfect model yields predictions Xi = Xi with no error (oa2= 0).
The results which Morrison reported were for special cases and can be obtained by
substituting the appropriate values of k, EX, Var X, X, or a,2 into the preceding general
results.6 Together, Al and A2 imply a negative binomial unconditional distribution of
purchases across all individuals. The previous results for perfect models with Xi = Xi
reduce to: p2 = 1, R2 = j2 = Var X/Var X = 1/(I + kEX/Var X) for any estimation
procedure B1 which predicts Xi= Xi, as long as the assumptions Al' and B1' are
maintained. For this special case, R 2 is nonnegative and approaches unity as
kEX/m Var X goes to zero, as would happen if the length m of the data period were
increased without bound.
Estimation of p2
The squared correlation p2 between X and X can be estimated by either of two easy
procedures.7 If two observation periods are available for each individual such that the
individuals' purchases are independent in the two periods, then:
p2 = pXX/(2pxx-R

2)

(8)

can be used to estimate the correlation p2 between X and X from estimates of R2 and
the autocorrelation p, for the entire sample of individuals. Or the estimation can
5 The results which Morrison reports in [8] as R 2are all based upon R 2 = l-(MSE/Var
X), and are not
the squared correlation Rj2 between A and X.
6 Morrison's article [8] contains an obvious typographical error: his result (10) should be R2AP = 1 + kt) and Pc2t = o respec(EX + a,2)/Var X. Also his results (16) and (17) should be R,2Gp = (2-p2)/(l
tively. Consequently his Table 2 of "Values of P.t for the Imperfect, Generalized Model" should have all
values of P2rit= 2 also. Morrison's R2 results for the other special cases (Al, A2, or B1) all agree with the
general results presented here.
7 If the two estimates of R2 are significantly different, then either the Al' assumption of constant k, or
the model assumption B2, or the independence of individuals' purchases in the two periods used to estimate
Pxx must be rejected.

This content downloaded from 193.226.34.227 on Tue, 25 Nov 2014 08:22:16 AM


All use subject to JSTOR Terms and Conditions

896

NEIL E. BECKWITH

AND MAURICE W. SASIENI

proceed from single period observations on the n individuals if the statistic X'(A - X)
is collected. The statistic X'(A - X)/n is an estimate of EX(X - X) so:
p2

1/ I+

- X) + E(X - X) )/(EX(\

(EX(A

-X)

+ Var X)

(9)

can be estimated from observable estimates of EX(X - X), Var X, and the expected

squaredresidualE(- X)2.
Morrison's Model Is not a Regression
The model B2 is appropriate for estimation procedures which have errors Ei
uncorrelated with both the long run averages Xi and the short run observed purchases
Xi. But the deduced results do not apply to estimation procedures based upon
observations Xi of purchase behavior. Wildt demonstrated that the special case of the
model with Poisson and gamma distributions (Al, A2, and B2) is not consistent with
the usual regression segmentation studies [9]. Here we show that the contradiction
holds for the general case (Al', A2', and B2). The cross moment E [X(X- X)] for this
model is
E[X (

x)] = E [(Ei + X1)(Ei+ Xi- Xi)] =


X,

2.

(10)

But it is well known [7] that this expectation is necessarily zero for every OLS
regression estimator A = Z,e = Z(Z'Z)-1Z'X which attempts to predict the observed
number of purchases Xi, regardless of what explanatory variables Z are included in
the regression:

E[X(X -Xi)]

A((-X)/n

= X'Z(Z'Z)

Z'(Z(Z'Z) 'Z'X - X)/n = 0.

(11)

Thus, the contradiction is established between this model and all OLS regression
models which attempt to explain the observed number of units purchased Xi (except
"perfect" models with a2 = 0, which cannot be realized in practice). In the following
section it will be seen that the contradiction is due to assuming EXE = 0 instead of
EX(X - X) = O.8
4. OLS Regression Models
Regression segmentation studies have necessarily attempted to explain the observable number of items purchased, Xi, rather than the unobservable average purchase
rate, Xi. Since for any individual Xiis the expected value of Xi, the best predictor of Xi,
based on time invariant characteristics of the individual, is one which fits the Xi
exactly. The only way of coming closer to the Xi would be to use regressors which
vary with time as do the Xi themselves. Most segmentation studies have not used time
varying regressors.
We will first examine the case of naively specified regression models, where the
regression estimated model may (or may not) be misspecified. Then, we subsequently
examine the special case of the correctly specified regression model.
All OLS Regression Models
It is well known that the cross moment EX(X - X) = 0 for all OLS regressions of
the form A = Z(Z'Z) 'Z'X regardless of Z, as shown in [7] or (11). Thus the
appropriate assumptions for a regression study explaining Xi are the two assumptions
8 The assumption E(Xc) = 0 within Morrison's Generalized Model is consistent with correctly specified
OLS regression models, see Appendix D.

This content downloaded from 193.226.34.227 on Tue, 25 Nov 2014 08:22:16 AM


All use subject to JSTOR Terms and Conditions

CRITERIA FOR MARKET SEGMENTATION

897

STUDIES

Al' and A2' about the process, and RI about the estimation:
Rl. An OLS regression model with n observations and v explanatory variables
(including the intercept) yields predictions A = Z(Z'Z) 'Z'X where u. = i- Xi is a
random variable with zero mean, variance 2, and uncorrelated with Xi.
We denote E(EX) of the regression as A. It can be shown that A = E(EX)=
(kEX_ -2
a2)/2. The exact value of A will depend upon the particular explanatory
variables used in the regression and will be seen to be zero for a correctly specified
OLS regression model. As shown in Appendix C:
-

(12)

VarX=VarX+kEX-a2,
a2 <

kEX + VarX,

A < kEX-

(13)
(14)

a2,

AO<

(15)

for all OLS regressions which explain the individual purchases Xi.
We wish to consider the fit of a regression to the Xi and Xi in some detail, but first
we show that estimates based upon OLS regression fitting of the Xi are necessarily
inconsistent with Morrison's assumption B2 that EX(X1 - ki) = 0. Since Xi =
Xi - (Xi - Xi) then EXj(Xj - Xi) = E {[Xi - (Xi - Xi)](Xi - Xi) + Xi(Xi
EXi(X, - X) = 0 for all OLS regressions this simplifies to:

EXi(X" -Xi)

-2

Xi)}. But since

+kEX.

(16)

There is, in general, no reason why a2= kE) for practical regressions so that
EX1(Xi- Xi) is not zero, contrary to B2. (Indeed, it is shown at (26) in Appendix D
that the expectation EX1(Xi- Xi) is strictly positive for a correctly specified OLS
regression). Thus it is not possible to maintain the assumption that E = X - X and X
are uncorrelated for OLS regression models. Thus, this assumption is the key
difference between Morrison's Generalized Model and OLS regression models.
The fit of a regression to the Xi and Xi is examined in terms of the squared
correlations of Xi with Xi and with Xi. Denote by p the correlation between X and the
regression estimate X. Since:
Cov(X, X)

- X)
Et (X - EX)[ (X - EX) + (X^

Var X+ A

(17)

where A = E(EX)= E(X - X)(X- EX) and E = X-A, then p2 = [Cov(X, X)]2/(Var X X
Var X) = [Var X + A]2/(Var X* Var X). On the other hand the correlation R between
the estimate X and the observed purchase X is given by R2 = 1- al/VarX=
(kEX + Var X - 02)/(kEX + Var X) for all OLS regression models.9 Notice that 0
< R2 < 1 for all OLS regressions since a2 < kEX + Var X for all OLS regressions.'0
Notice also that for a "perfect" model B1 which correctly estimates the mean
purchase rates Xi we have Xi =i,
so Var X = Var X and Cov(X, X) = Var X + A
Var X. Thus A = 0 (by (17)) and p2 = 1. On the other hand R2 may be substantially
-

9 The R2 which we discuss is the population value. It is not the statistic commonly obtained in OLS
regression of a sample for two reasons. First, the observed R2 is a random variable, the realized value of
which depends upon the particular sample drawn. Second, the OLS regression procedure maximizes R2,
biasing the observed statistic upward: E(R2) > R2.
10 Substituting a2 = kEX - 2a,2 we obtain R 2 = [Var X + a,2 + 2A]/Var X, which is the same as the
R 2 for Morrison's Generalized Model (6) except for the + 2A term and the sign of the a,2 term. We notice
that any OLS R2 must equal or exceed the R2 for any Morrison Generalized Model which has the same
error variance af2.

This content downloaded from 193.226.34.227 on Tue, 25 Nov 2014 08:22:16 AM


All use subject to JSTOR Terms and Conditions

898

NEIL E. BECKWITH

AND MAURICE W. SASIENI

less than one, even for the perfect model. To see this consider Xi = (Xi- Xi) + Xi so
that:
Var X = Var(X - X) + Var X + 2EX(X
= Gu+ kEX +

Var X + 2E(X

= kEX + Var X -

X)

X)(X - X) + 2EX(X - X)

and as XA~~~~~~~~~~~
approaches X we must have aU = kEX so that R2== Var X/(kEX + Var X)
< 1. In practice we often find kEX < Var X so that RA=Awould be between 0.5 and 1.
On the other hand when the model fits the actual purchases Xi very well it may be a
poor fit to the purchase rates Xi, For such a case au2is small, R2 is close to one,
Var X > Var X, and:
p2 = [Var X + A2/

(Var X* Var X)

= [(Var X + A)/Var X] [(Var X

kEX + A)/Var X]

is less than one since -A< 0, kEX > 0 and they do not approach zero as u2 approaches
zero.
Combining the preceding results, we can find an upper bound on the unobservable
squared correlation as a function of the estimable R2 and the estimable autocorrelation Pxx[
p2 =

[Var X + A]2/[Var X* (Var X-

G)]

= ((Var X + A)/Var X) ((Var X + A)/(R2 Var X)).


Therefore

p2 <

(Var X/Var X)/R2

pxx/R2

(18)
(19)

where the previously defined (3) autocorrelation Pxx exceeds 0.5 in typical segmentation studies. Thus, this bound is only useful if a segmentation study has very high R2,
which most do not.
These results for regression models can be reduced to special cases like the
Poisson-gamma assumptions of Al and A2 in the same manner as shown for
Morrison's model.
It is noteworthy that the usual regression assumptions describing the process
generating the dependent variable values have not been made. Hence the preceding
results are applicable to all OLS regression studies explaining Xi, regardless of the true
relation (if any) between Xi and the explanatory variables Zi.
The Sampling Properties of Correctly Specified Regressions
In the previous section we have considered the population parameters of regression
models; in this section we turn to the results to be expected from finite samples, in the
special case where the mean purchase rates Xi all lie in a hyperplane in the regressor
space. We replace assumption A2 by the more specific assumption A2* which defines
what we mean by a correctly specified model.'2
'l Wildt's alternate model [9] is the special case of A = kEX - a2 where the model is misspecified such
that the OLS estimated model X1= Zip has a different set of independent variables Zi (or a different
functional form) than the process which determined the true purchase rates Xi = Zip. We would also
describe a model as misspecified if Xi = Zip + s,, even though the group average E3i = 0 for individuals with
identical Zi.
12 Notice that the only difference between this model and the usual regression model is that the usual
regression process specifies that the random variable Xi - X1 is independent of the Zip (Var(Xi - Xi)
constant) whereas in this model the random variables are assumed instead to have a variance proportional
to Zip (Var(X1 - Xi) = kXi). Generalized least squares GLS estimation would be a more appropriate choice
for a process with Var(X1- X) = kXi. We examine the OLS regression estimator because most (if not all)
actual segmentation studies have used OLS regression models. The OLS and GLS estimators are both
unbiased, but the GLS estimator would be more efficient.

This content downloaded from 193.226.34.227 on Tue, 25 Nov 2014 08:22:16 AM


All use subject to JSTOR Terms and Conditions

CRITERIA FOR MARKET SEGMENTATION

899

STUDIES

A2.* The average purchase rates Xi of each of n individuals are related to v


explanatory variables Zil, Z .212***, Ziv (where Zil = 1) by: Xi = Zi1 =E>iZi/31.
The Z. have values such that the Xi are nonnegative and are distributed in the
population with mean EX and Var A. Such a model based on a finite number of
observations differs from a perfect model with Xi = Xi because of the sampling errors.
It is easily shown, as in Appendix D, that this correctly specified regression model
(assumptions Al', A2*, RI) is a special case of the previous results with:
a 2 = (n

v)/nkEX

(20)

(v/n)kEX

(21)

EXE= A = 0.

(22)

uf2 =

The important thing to notice is that


kEX, so that as the number of
observations increases, the au2approaches kEX and the au2approaches zero. Indeed the
variance a,2 of the total error e is just the fraction v/(n - v) of the residual variance
u2. The expected fraction of variance in the observed purchases Xi explained by the
correctly specified OLS regression X is:
aU2+ aJ2 =

R2

1- u2/Var X = (kEX + Var X - kEX(n

v)/n)/(kEX

+ Var X)

= (Var X + kEX(v/n))/(Var X + kEX).

(23)

This parameter within the entire population (it is not an observed statistic) is
necessarily 0 < R 2 < 1. Notice that as n increases, the expected R 2 does not approach
unity, but increases only toward 1/(1 + kEX/Var X).
In practice, k is about one, or a little less, and kEX is typically much less than
Var X, so R 2S of roughly one-half or larger should be expected for correctly specified
OLS regression models with many observations n > v.
As the length m of the data collection period increases the R2 approaches 1 since
we would replace k by k/m which would go toward zero. Together these results imply
R2 should exceed 0.5 (roughly) for actual segmentation studies (with large n > v) if
the OLS models are correctly specified. These results imply the high R2 because of the
assumption Al' that the individual's purchase variance, Var Xi, is proportional to the
mean rate EXi. If this assumption is tenable, it implies that OLS regression models
with R 2 significantly lower than about 0.5 are misspecified, in our sense. Of course,
the exact test value would depend upon the particular values of Var X, v, n, kEX and
m, but for most practical studies these will imply R2 well above one-half, and
certainly much higher than 0.05 or 0.10 which are found in typical OLS regression
segmentation studies.
The squared correlation between the individual's unobserved average purchase rate
Xi and the prediction Xiis found by setting A = 0 in (17), and substituting (2), (12), and
(20):
p2 = [COV(X, )]/(Var

X *Var X) = Var X2/(Var X *Var X)

=Var X/(Var X + v kEX).

(24)

Notice that p2 iS only slightly less than one since v < n for practical segmentation
studies.'3 Also, setting A = 0 in (18):
p2 = Var X/(R2 Var X) = pxx/R2
(25)
13 Essentially the same result holds for the fraction of variance of
Xi explained by A,, which also
approaches unity for v <<n. For the correctly specified OLS regression model, the fraction of variance of X
which is explained by the estimates A is easily found from the squared correlation between X and A:
1 -(a7 /Var A) = (Var A -kEAv/n)/Var
1)]/Var A = 2 _p-2.
A = [Var A -Var X(p-2The reciprocal function is p2 = 1/[1 + as2/Var A].

This content downloaded from 193.226.34.227 on Tue, 25 Nov 2014 08:22:16 AM


All use subject to JSTOR Terms and Conditions

900

NEIL E. BECKWITH

AND MAURICE W. SASIENI

where Pxx is the autocorrelation (3) between an individual's purchases in two


independent periods. Comparing the results (24) and (25) we see
Pxx = p2R2 =

(Var X/(Var X + v/nkEX))R 2

R2.

This equality (for v < n) provides a powerful test of the model. If the hypothesis
R 2 = PXX can be rejected, then the OLS regression model is misspecified. Since
Pxx = Var X/Var X = Var X/(Var X + kEX) is normally somewhere around 0.5 or
greater for typical segmentation studies, we should expect that this test is equivalent to
the previously discussed test based upon (23). However the estimation of Pxx provides
the statistic necessary to implement the test. If the estimate of Pxx is significantly
greater than the estimate of R2 then the OLS model specification or Al' must be
rejected.
Perfect Models
Practical segmentation models usually have many more observations n than
variables v. If the models were correctly specified then Gu2
would approach zero, by
(21). Since = i- Xi, this would mean that the estimates are nearly identical to the
case of the "perfect model" Bi of 2i = Xi, with p2 = 1, au = kEX, and R2
= Var X/(Var X + kEX) < 1. Thus, the perfect model B1 is simply the special case of
a correctly specified OLS regression model as the number of observations n goes to
infinity.
5. Conclusion
High R2' Sshould be expected for correctly specified OLS segmentation studies
explaining purchases, even in relatively short time periods, such as a week or month.
Increasing the length of the time periods will necessarily increase the expected R2
since the random purchasing behavior is averaged across the longer time interval. The
expected R 2 for any correctly specified OLS regression segmentation study is R2
= (Var X + kEX(v/n))/(Var X + kEX) which is somewhat less than unity for practical studies with n >> v. For practical segmentation studies, the Var X is usually
somewhat larger than kEX. Also the value of k is generally less than 1 corresponding
to slightly more regular than Poisson purchasing by each individual.'4 Hence the
above value of R 2 typically should exceed 0.5 if the OLS regression models are
correctly specified. Most previous segmentation studies have had empiric sample
statistics R2 much lower than 0.5, often in the range of 0.05 to 0.10.'5 We must
conclude that either those regression models were misspecified or that the stochastic
behavior assumption Al' is incorrect. A likely cause of the misspecification is that
individuals with the same explanatory variables Zi do not have identical average
purchase rates Xi, possibly because of omitted variables.
A test of the specification of OLS regression segmentation studies is based upon the
autocorrelation Pxx of purchases among individuals in two independent periods. All
correctly specified OLS regression models have Pxx R 2, (where v < n), and incorrectly specified models have Pxx > R2, where the same length of time periods is used
for estimating both R2 and Pxx.'6 If the observed estimates of R2 and Pxx differ
significantly, then the OLS regression model is misspecified or Al' is not met.
14 See [5] for an empirical example with the Var X; 10 EX or Var X + kEXz lOEX. Since we expect k
to be slightly under unity [4], we expect Var X > kEX.
15 For example, Frank, Massy and Boyd attempted to explain the annual household purchases of 57
different product classes by 14 socioeconomic variables. The highest of the 57 R 2S was 0.26 for one
product class, and about half the product classes had R 2'S less than 0.10, see [6].
16 Of course, if the time periods are of different lengths then the statistic Pxx or R2can
simply be
corrected to obtain estimates for the same length time periods.

This content downloaded from 193.226.34.227 on Tue, 25 Nov 2014 08:22:16 AM


All use subject to JSTOR Terms and Conditions

901

STUDIES

CRITERIA FOR MARKET SEGMENTATION

Of course, the main purpose of segmentation studies is not to predict individual


purchase rates Xi. For the usual purpose of determining the characteristics of heavy
users it seems sufficient only to predict the mean purchase rates of similar individuals
or groups [1]. If the differences in average purchase rates between groups are
explained by Zi3 then the estimated models may be very useful managerially, even
though they would be misspecified in our sense of not having Xi = Zi3.
Regression segmentation studies may be useful descriptions of the true purchase
rates (or the group means for identical individuals) even though the models are
misspecified by omitted variables or functional forms different from the true process.
However the estimated model coefficients will be biased, except in special circumstances. We hope that the misspecification tests developed here will facilitate the
search for correctly specified segmentation models. Additional commentary concerning some of the results developed herein is contained in [3].
Appendix A
Proofsfor the PurchasingProcess
We assume Al' and A2' and we use the trivial fact that Xi consists of the sum of two
independent variates, Xi and (Xi - X1).Since E(Xi - Xi)= 0 it follows that
(1)
EX= EX
Var X = Var X + kEX.

(2)

The autocorrelation between purchases in two distinct periods 1 and 2 is given by Pxx
= Cov(X1, X2)/Var X. Assuming that the purchases in the two periods are independent for
each individual:
Cov(X1, X2) = Cov[Xi + (Xi, - Xi), Xi + (X2 - Xi)]

= Cov(X, X) + terms with zero expected value


= Var X.

Therefore,
(3)

Pxx = Var X/Var X.

Appendix B
Proofsfor Morrison'sGeneralizedModel
We now assume Al', A2' and B2, so that the estimates X1each consist of the sum of two
independent variables Xi and (i - Xi)= Ei.Then
Cov(X, A) = Cov(X, X) + Cov[X,(X - X)] = Var X

Var X = Var X + Var(X - X) = Var X +

a,2.

Hence the squared correlation between X and X is given by


p2 = [Cov(X, A)] /Var X Var X-Var X/(Var X + a2).

(4)

The error of prediction of Xi is X, - Xi = (X - Xi) - (Xi - Xi). The two terms are independent,

by B2, so that the mean squared error is:


MSE = Var(X - X) = Var(X -X)

+ Var(X - X) =

Thus R2 = 1 - MSE/Var X = (Var X -af2)/Var


and substituting yields:
R2=

Var X(l

- (1 - p2)/p2)/Var

a,2

+ kEX.

(5)

X. Solving (4) for a2 = (1 - p2)Var X/p2

X = (2

- p-2)

Var X/Var X.

(6)

Similarly the squared correlation between Xi and Xi is found using Cov(x, X) = Var X:
R 2 = [Cov(X,

)] /(Var X Var A) = p2R2/(2 -

p2).

(7)

Substituting Pxx = Var X/Var X (3) into (6) and solving for the squared correlation p2

betweenX and X yields:


p2

Pxx/(2Pxx

R 2).(8

This content downloaded from 193.226.34.227 on Tue, 25 Nov 2014 08:22:16 AM


All use subject to JSTOR Terms and Conditions

902

NEIL E. BECKWITH

The covariance between X and

AND MAURICE W. SASIENI

X- is:

= EX(X -X) - EX(X-X)

EX(A -X)

= EXE - EX(X-X)

ExXi(EiX, - X) - kEX

= 0-

E(X -

kEX

-EAX,(0)

-kEX.

- kEX, and using the previous result for


Solving (5) for a2 = E(-X)2
squared correlation (4) between X and A:

p2 = Var X/(Var

X+ a2)

= 1/[1 + (EX(A

-X)

kEX we obtain the

1/[1 + a2/(Var X - kEX)]

+ E(X

-_

-X)

))/(EX(A

+ Var X)].

(9)

The cross moment E [A(A - X)] is:


E[X#(Xi - Xi)] = E[(Ei + Xi)(ci + Xi - Xi)]
= EE7 + 2E(EiXi) =

a2 + 0-0

+ EXA - E(XiXi)

E(EiXi)

+ EXA2 - EA(XE1Xi) = ag.

(10)

Appendix C
ProoJs for all OLS Regression Models
We now assume Al', A2' and RI, and will establish the results (12) through (17). Since
i= ( - Xi) + Xi:
Var A = Var(A - X) + Var X + 2 Cov[(A - X), (X
=a2

+ kEX+VarX-2a2

- A

= kEX + Var X-a2

+ A)]
> 0.

(12)

Therefore,
a2< kEX + Var X.

(13)

To demonstrate that A < kEX - a 2we start by establishing that:


= Var A + Var X - 2EXA

a2 = E[(A -)(X-)]

= kEX + Var X- a 2 + Var X - 2EX(X + E)


= kEX-

a2

+ 2Var X - 2(Var X + A) = kEX

- 2A,

and
A=

(kEA-aU2-a

2)/2=

kEX -a 2-(kEX

2-

a2)/2.

Therefore
A < kEX -au.

(14)

To establish (15), where Q is the positive definite matrix Z'(Z'Z) IZ, notice that:
A = E(

= EXi(Ai - Xi) + EXi(Xi -Xi)

= E#iQ

- xi).

Now (A - X)'X = (QX - X)'X = X'MX = [A + (X - A)]'MA = A'MX + (X - A)'M1Awhere M is


the negative semi-definite matrix Q - I. The second term has an expected value of zero since
(Xi - i)X has zero expectation for any fixed i. The first term is never greater than zero.
Hence
A = E E -=E(Xi-X1)Xj

< 0.

(15)

Appendix D
Proofsfor CorrectlySpecified OLS Regression Models
Maintaining Al' and A2*, all the previous results of Appendix C for all regression models
must hold for the special case of the correctly specified regression model Rl. We first find a12
by the usual regression model argument: -u = X -A = X -Z(Z'Z)-1Z'X
= Zj3 + r -

This content downloaded from 193.226.34.227 on Tue, 25 Nov 2014 08:22:16 AM


All use subject to JSTOR Terms and Conditions

CRITERIA FOR MARKET SEGMENTATION

903

STUDIES

+ r) = [I - M]r where M denotes the idempotent matrix Z(Z'Z) 1Z'. Since

Z(Z'Z) -'Z'(Z/

I-M
is symmetric
idempotent,
the sum of squared residuals
is u'u = (-u)'(-u)
= r'[I - M]'[M]r = r'[I - M]r. Since [I - M] has rank n - v and is idempotent, it nec-

essarily has n - v characteristic roots of value 1, and v roots of value 0. There exists an
orthogonal matrix P such that P'[I - M] P is a diagonal matrix of the characteristic roots.
Thus: u'u = r'P[P'MP]P'r = (n - v)r'PP'r = (n - v)r'r and
2=

Eu'u/n

= E(r'r)(n

v)/n

= kEX(n - v)/n.

(20)

We next evaluate E'E:


E

= i - A = Z(Z'Z)
= Zf3 + Z(Z'Z)-

'Z'X
IZ'r - Z

Z(Z'Z)

Z'(Z13 + r) - Zf

= Z(Z'Z) - Z'r = Mr,

,E' = r'M'Mr = r'Mr.

Comparing this result with the above result u'u = r'[I - M]r we conclude
that
a2= E 2=

E' /n = (Er'r - Eu'u)/n

= kEAX-

v kEX.

'E = r'r - r'u, so

(21)

.n

Since EE2 = kEX - 2A - a2 for all regressions, it follows that:


EXE= A = 0

(22)

for correctly specified OLS regressions. Evaluating the Cov(X, E) from (16):
Cov(X,

E) = kEX-

a2

kEX - kEX(n

v)/n = kEXv/n > 0.

(26)

References
1. BASS, FRANK M., TIGERT, DOUGLAS, J. AND LONSDALE, RONALD T., "Market Segmentation: Group
Versus Individual Behavior," Journal of Marketing Research, Vol. 5 (August 1968), pp. 264-70.
2. BECKWITH, NEIL E., "Conditions for which R2 Is Less Than the Squared Correlation (y, ),"
unpublished working paper, Columbia University, November 1973.
3.
AND SASIENI, MAURICE W., "Marketing Segmentation Studies, Uses and Abuses of R 2,
Proceedings of the Business and Economics Statistics Section of the American Statistical Association,
1975, pp. 112-116.
4. CHATFIELD, C. AND GOODHARDT, G. J., "A Consumer Purchasing Model with Erlang Inter-Purchase
Times," Journal of the American Statistical Association, Vol. 68 (December 1973), pp. 828-35.
5. EHRENBERG, A. S. C., Repeat-Buying: Theory and Applications, American Elsevier, New York, 1972, p.
142.
6. FRANK, RONALD E., "Market Segmentation Research: Findings and Implications," in F. M. Bass, C. W.
King, and E. A. Pessemier, Applicationsof the Sciences in Marketing Management, Wiley, New York,
1968, p. 45.
7. HUANG, DAVID S., Regressionand Economic Methods, John Wiley & Sons, Inc., New York, 1970, p. 65.
8. MORRISON, DONALD G., "Evaluating Market Segmentation Studies: The Properties of R 2," Management Science, Vol. 19, No. 11 (July 1973), pp. 1213-21.
9. WILDT, ALBERT R., "On Evaluating Market Segmentation Studies and the Properties of R2," Management Science, Vol. 22, No. 8 (April 1976).

This content downloaded from 193.226.34.227 on Tue, 25 Nov 2014 08:22:16 AM


All use subject to JSTOR Terms and Conditions

Das könnte Ihnen auch gefallen