Sunteți pe pagina 1din 15

The Relationship Between Customer Loyalty and Purchase Incidence

Stern, Philip;Hammond, Kathy


Marketing Letters; Feb 2004; 15, 1; ABI/INFORM Research
pg. 5

Marketing Letters 15:1, 5-19, 2004


2004 Kluwer Academic Publishers. Manufactured in The Netherlands.

The Relationship Between Customer Loyalty


and Purchase Incidence
PHILIP STERN
Warwick Business School
KATHY HAMMOND*
London Business School, Sussex Place, Regents Park, London, NWJ 4SA, UK

khammond@london.edu

Abstract
Little is known about customer loyalty to brands over many dozens or even hundreds of purchases. In this paper
we describe, and seek to explain, such patterns of loyalty in two very different markets: a consumer market
(laundry detergents), and a more frequently used service (physicians' prescribing of anti-hypertension drugs).
Purchase incidence heterogeneity - a problem in most loyalty studies - is addressed by measuring loyalty at
different rates of category purchase (rather than over time). Using share-based measures we expect that loyalty
will decline as purchase incidence increases, however we clarify the shape of that decline. We find that, as the
number of purchases rises, loyalty initially falls steeply, but after around 15 purchases it starts to stabilize, and
from 60 to 200 purchases there is very little change in observed measures of customer loyalty. A comparison of
the findings with those expected from a stationary market model (the Dirichlet), suggests that the decline in loyalty
seen as the number of purchases rises is largely a statistical artifact, dependent on the number of purchases used
to calculate loyalty. However, we also find that the higher loyalty exhibited by heavier buyers at low purchase
levels is not captured well by the model. The implication here is that, contrary to a central assumption of the
Dirichlet model, brand choice is partially dependent on purchase weight.
Keywords:

1.

customer loyalty, brand loyalty, purchase incidence, Dirichlet model

Introduction

Measuring customer loyalty is important for all firms, but especially for those that have,
or plan to implement, customer loyalty initiatives such as loyalty cards or frequent-user
programs. As Dowling and Uncles (1997) state, operating loyalty programs requires
knowledge of long-run customer loyalty patterns in order to justify and evaluate investments. However, as noted by Dekimpe et al. (1997), even though managers need to
have better information on how buyers might behave over multiple purchases within
a category, there are very few reported studies of consumer behavior over the longer
term (exceptions being East and Hammond, 1996; Johnson, 1984; Mela et al., 1997;
Stern, 1997).
This research explores whether and how customer loyalty varies by the number of purchases. We seek to establish the general patterns of loyalty as the number of purchases
* CoITesponding author.

Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.

STERN AND HAMMOND

bought by a customer increases to 200. This is a full order of magnitude greater than commonly reported (and the equivalent of around 10 years of purchasing in a typical packaged
goods category). Marketing managers who use behavioral measures of customer loyalty to
evaluate the strength of their relationship with customers need to understand the observed
pattern of loyalty as purchases rise. For example, if the share of purchases customers give
to their most preferred brand decreases with increasing category purchase and the level of
this decrease is almost constant, then the strength of the relationship will weaken (in sales
terms) and may require substantial continuous reinforcement. However, if customer loyalty reaches an equilibrium point that is largely independent of further category purchase,
the relationship might need less and perhaps only intermittent reinforcement.
Specific questions are:
(i) Brand loyalty, as calculated using share-based measures, is expected to decline as
more purchases are made, but what pattern does this decline take?
(ii) Brand-level data are relatively easy to collect, and short-run brand data can be used to
predict long-run brand loyalty. How close are such predictions to observed findings?
What additional information do we gain by using different measures of loyalty? How
well does a benchmark model such as the Dirichlet, which is normally used with
brand-level data, perform when used to predict preferences for customers' 'favorite'
brands?
(iii) What are the main factors determining the levels of customer loyalty seen in different
markets?
We measure customer loyalty in three ways. Our first two measures are share-based
measures, the third, non-share based, measure provides an indication of overall brand
switching. For each measure we examine how its value changes as the number of purchases increases. Below we discuss the rationale for our choice of measures.

1.1.

Measures of Customer Loyalty

We initially use the share of category requirements measure (SCR) to capture the relative
share of category purchases that individual households give to each brand they buy. SCR
is a useful measure of brand loyalty as it is easily understood and widely used by brand
managers, and we build on previous studies where SCR has been applied over specific time
periods (e.g., Bhattacharya et al., 1996; Bhattacharya, 1997; Fader and Schmittlein, 1993;
Johnson, 1984; Tellis, 1988). However, SCR has two potential weaknesses, and in order to
address these we use two additional measures of loyalty.
The first weakness is that, particularly over the long run, the SCR measure does not
allow us to distinguish between customers who have a most preferred brand, but only
give that brand a fairly low share of their purchases, and a brand with the same share
bought as a secondary or tertiary brand by customers who have another primary brand.
Managers of loyalty programs or other customer relationship management initiatives often
want to be able to differentiate these two types of customer, focusing retention programs on
the former, and aiming development/re-acquisition or even customer 'sacking' programs

Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.

THE RELATIONSHIP BETWEEN CUSTOMER LOYALTY AND PURCHASE INCIDENCE

at the latter. For our second measure, therefore, we follow Deighton et al. (1994) and
calculate the share of category purchases accounted for by the customer's most preferred
brand (SCRpref). We can think of SCR as a brand loyalty measure and SCRpref as a
customer loyalty measure.
A second weakness is that SCR (in common with SCRpref and all share-based measures) is confounded by purchase incidence. We therefore report a third measure of
loyalty, the polarization index, </>, which captures changes in the heterogeneity in consumer choice vectors as purchase incidence changes. 1 </> ranges between zero and one,
where zero indicates pure homogeneity in consumer choice (i.e., all buyers have the same
propensity to buy individual brands), while as</> approaches 1, there is maximum heterogeneity (i.e., each consumer buys only their favorite brand) (Fader and Schmittlein, 1993;
Sabavala and Morrison, 1977). Before describing the calculation of measures and the data,
we first discuss the reason for focusing on purchase incidence rather than time as the measurement interval for studying loyalty over many purchases.

2.

Heterogeneity in Purchase Incidence

Individual consumers exhibit differences in both category purchase incidence and brand
preferences. Fader and Schmittlein (1993) have argued that heterogeneity in brand choice
is the likely cause of the excess brand loyalty (excess compared with predictions from a
baseline Dirichlet model) observed for high share brands. Such heterogeneity has been
addressed partially by the development of models that incorporate latent segments in their
brand choice component (Danaher et al., 2003). Here we are interested in heterogeneity in
purchase incidence; below we discuss how four aspects of purchase incidence heterogeneity can affect loyalty measures.
Infrequent category buyers Infrequent category buyers raise problems for loyalty measures based on share. For example, if a customer makes only one purchase of the category in the time period being analyzed, then the share of purchases allocated to the brand
bought must be 100%. If they buy the category twice and buy two different brands, the
share for each brand can only fall to 50%, etc. One solution in the past has been to exclude
these (often numerous) infrequent category buyers from the analysis of loyalty and include
only buyers who make a minimum number of category purchases (e.g., 3 or 5 or 15 purchases) in the time period under study (Deighton et al., 1994; Krishnamurthi and Raj, 1991;
Tellis, 1988).
Very frequent category buyers Panel data for research purposes are rarely available for
more than two years (and very few analyses have been reported over periods longer than
this), therefore even for relatively frequently-purchased products, analyses tend to contain few cases of very frequent category buyers. It may be that the behavioral loyalty of
very frequent category buyers is substantially different from that of the average buyer in
a six-month to two-year period, the basis for calculating most measures of share loyalty
(Bhattacharya et al., 1996; Fader and Schmittlein, 1993; Tellis, 1988).

Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.

STERN AND HAMMOND

Small portfolio size Across many common grocery markets the average buyer makes
around 12 category purchases per year (Ehrenberg et al., 2004). For this number of purchases customers' brand portfolios tend to be around four (Hauser and Wernerfelt, 1990).
With portfolios of this size, brand SCR values average around 30% (Ehrenberg et al., 2004),
and SCRpref is 65% to 75%, as originally observed by Cunningham (1956), and confirmed
by Deighton et al. (1994). But such findings do not inform us about loyalty over many
dozens of purchases when brand portfolios have the potential to be much larger.
Different purchase frequencies Even allowing for the previous problems, consumers
still exhibit significant heterogeneity in terms of purchase incidence. This makes it difficult
to compare loyalty levels in a specific time period across buyers with different category
usage rates.
We address the first three heterogeneity problems by analysing data from two long-run
(five year) panel datasets. 2 The final measurement issue (different purchase frequencies)
we address by repeating the analyses for different numbers of purchases rather than reporting findings over time.

3.

Benchmarking Against the Dirichlet Model

In seeking to explain the loyalty patterns we observe, we compare our findings with those
predicted using the benchmark Dirichlet model. The Dirichlet is a stochastic model of
buyer behavior developed for the study of branded packaged goods in established competitive markets (Bass et al., 1976; Goodhardt et al., 1984). The theory underlying the Dirichlet
is that there is a small set of interrelated assumptions that describe and predict the patterns
of purchase incidence and brand choice for any market that is approximately stationary
and unsegmented. The Dirichlet is used here since a basic assumption of the model is that,
for individual consumers, brand choice probabilities are independent of category purchase
incidence. If the Dirichlet model is able to predict closely the loyalty patterns observed
as the number of purchases increases, the implication is that managers can use short-run
data to model brand loyalty over the longer term. However, previous researchers have
noted that the assumption of the independence of purchase incidence and brand choice
does not always hold (e.g., Shoemaker et al., 1977). There have also been studies reporting systematic deviations from Dirichlet predictions; Bhattacharya (1997) describes the
under-prediction of brand loyalty for niche brands; Fader and Schmittlein (1993) report
the under-prediction of brand loyalty for high-share brands, and suggest that this market
share premium is a consequence of consumer segmentation that favors large brands.
The usefulness of the Dirichlet model as a benchmark lies in the fact that it has successfully characterized brand loyalty over the medium term across a wide range of categories
and conditions (Ehrenberg et al., 2004; Uncles et al., 1995). Our aim is to provide an
initial description of loyalty patterns over different purchase weights and rates, however
our results also suggest that purchase incidence heterogeneity may provide an additional
explanation for the previously observed under-prediction of brand loyalty by the Dirichlet.

Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.

THE RELATIONSHIP BETWEEN CUSTOMER LOYALTY AND PURCHASE INCIDENCE

4.
4.1.

Data and Methodology


Data

We calculate all loyalty measures for consumer purchasing of laundry detergent. We also
replicate analysis of the SCRpref measure for a service category, physicians' prescribing
of anti-hypertension drugs. Prescribing by physicians has been chosen as a deliberate
contrast to the more common consumer products usually reported in brand loyalty studies.
Physicians' prescribing differs in an important way from the purchasing of detergents the detergent buyer will subsequently use the product whereas the physician is a service
provider who writes the prescription but does not use or pay for the drug. The addition of
this dataset provides a test of the potential generalizability of our findings.
4.1.1. Laundry detergent The data come from a five-year household panel of 1532
continuous category buyers. The top 11 brands were analyzed; these accounted for 92%
of purchases. 3 The average purchase rate was 73 over the five years, with 816 households
making at least 50 purchases, 122 households making 150 or more purchases, and 35
households making over 200 purchases.
4.1.2. Physicians' prescribing of anti-hypertension drugs Data are from a five-year
continuous panel of 202 medical prescribers in the UK. 4 The leading 17 brands of antihypertension drugs were analyzed, accounting for over 70% of total drugs prescribed in
the hypertension treatment area. The average prescribing rate for the category was 250
over five years, with 162 physicians writing at least 50 prescriptions, 73 writing 150 or
more prescriptions, and 48 writing over 200 prescriptions.
4.2.

Calculation of Loyalty Measures

SCR is calculated as the ratio of total purchases of the brand to total category purchases
among those who buy the brand. SCRpref is the share of category requirements given by
a household to its most preferred brand. 5 The Dirichlet model requires three parameters,
M, S, and K. Mis simply the mean purchase rate and K measures buyer heterogeneity
reflecting the extent to which overall purchasing differs from the mean. The S parameter
measures heterogeneity in brand choice. The marginal distributions of the Dirichlet are
beta for each brand. S for the category is calculated by estimating brand specific S values for each marginal distribution, and taking the market share weighted average of these
marginal S values (Uncles, 1989). M, S, and K are calculated from the average one-year
penetration and purchase frequency data for each brand. Using these parameters, simulated data representing 'pure Dirichlet' purchasing patterns for thousands of 'buyers' were
created for both of the markets studied. From these simulated datasets, the average brand
choice probabilities together with product and brand rates of buying were calculated for
different purchase rates, leading to the estimation of values for SCR and SCRpref for purchase rates from 5 to 200. The S parameter is reported as our third measure of loyalty, the
polarization index,</>.

Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.

10

STERN AND HAMMOND

70
- - Brand 1 (act)
- - - Brand 2 (act)
---+--Brand 3 (act)
--------- Minor Brand (act)
.Brand 1 (est)
X Brand 2 (est)
........ Brand 3 (est)
.. .. Minor Brand (est)

60
50

40

0::

(.)
(/)

30

........ -....... .
. - . .. . . .. . ........ -....... .

.... .. . . . . . .
20

.... ... .-......................... - ....... .

10
0
10

20

30

40

50

60

70

80

90

100

110

120

Number of purchases

Figure 1.

5.

SCR: Actual and Estimated Values for 3 Leading Brands and 1 Minor Brand of Laundry Detergent.

Results

We first report the actual and expected patterns for SCR for laundry detergent, then the actual and expected values for SCRpref for both laundry detergent and physicians' prescribing of anti-hypertension drugs. Finally we report</> for laundry detergent, and describe the
relationship between</> and SCR.

5.1.

Brand Loyalty Measured Using SCR

Figure 1 shows actual and expected values of SCR, for the leading three brands and one
minor brand of laundry detergent, as category purchases increase from five to 125. Brand 2
illustrates the general pattern; SCR averages 53.7% at five category purchases (48.9%
expected), falling sharply to 35% at 15 purchases (33.4% expected), continuing to fall
steadily to 25% (24.4% expected) at 60 purchases and then stabilizing at around this level.
As expected, larger brands exhibit consistently higher levels of loyalty than smaller brands.
The expected values for individual brands are generally close to actual values, except for
the brand leader, where the Dirichlet consistently under-predicts SCR (by 3 to 10 percentage points or 9% to 18%).

5.2.

Customer Loyalty Measured Using SCRpref

Turning to SCRpref, Figure 2 shows the findings for laundry detergent and also, as a comparison, for physician's prescribing. Customers exhibit greater brand loyalty to detergents
than physicians do to anti-hypertension drug prescribing, but the pattern of loyalty is sim-

Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.

11

THE RELATIONSHIP BETWEEN CUSTOMER LOYALTY AND PURCHASE INCIDENCE

--Detergent: SCRpref (act)


- - - - - Detergent: SCRpref (est)
- - Physicians: SCRpref (act)
701---;:---~"'=--------------i___

......................____________

_:_:~~-~-~P~hy~sic~ia~ns~:~SC~R~pr~ef~(e~st~)----I

-- -- -- -- -- --

?!'.

"

a5o+---'------------------------------___,
~
()

rn

15

30

45

60

75

90

105

120

135

150

165

180

195

210

Number of purchases

Figure 2.

SCRpref: Actual and Estimated Values - Laundry Detergent and Physician's Prescribing.

ilar for both categories: average loyalty declines sharply as the number of category purchases increases from five to 15, then the decline is less steep as purchases increase to
around 60, and from 60 purchases onward there is very little change in loyalty. 6 Our
observed findings for SCRpref are consistent with previous research by Deighton et al.
(1994), who reported mean SCRpref levels of 73% for powder detergent (where the average number of purchases was 11.3). The pattern of decline for SCRpref is similar to
that seen for SCR (the steep decline from 5 to 15 purchases, a less steep curve from 15 to
around 60 purchases), but there are some key differences. If we compare values for SCR
and SCRpref for laundry detergent we see that, not only is the starting point for SCRpref
at a much higher level, but the slope of the SCRpref line is also shallower.
If we compare the actual and expected values for SCRpref, the correlation is high (0.976
for laundry detergent and 0.944 for physicians' prescribing), however, in both categories,
particularly at low purchase rates, the model under-predicts SCRpref At low purchase rates
(which, to be as consistent as possible with previous research, we take as being up to 15
purchases), the under-prediction averages 7.2 points or 10.0% for detergent and 6.0 points
or 13.1% for physicians' prescribing. At medium purchase rates (16-60 purchases), the
under-prediction is 3.8 points (5.5%) for detergent and 4.4 points (11.2%) for physicians'
prescribing. For high purchase rates (61-200 purchases), values for SCRpref are closer
to those predicted: for detergents the model under-predicts by an average of 0.8 points
or 1.2%, while for physicians' prescribing the mean absolute deviation from the model is
0.7 points (2.2%). 7
The fit between the actual findings and expected results for SCRpref (and for the brand
leader for SCR) would suggest that the assumption that purchase incidence is independent
of brand choice, although a good approximation, is not entirely valid. In order to explore
this issue in more depth, we compare our findings for groups of buyers segmented by purchase weight. We divide laundry detergent buyers into three discrete segments according
to their rate of category purchase and plot SCRpref levels over the first 20 purchases for

Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.

12

STERN AND HAMMOND

--heavy buyers
medium buyers
---light buyers

50+---~--~--~--~--~--~--~--~--~--__,

15

13

11

17

19

21

23

25

Number of purchases

Figure 3.

SCRpref over 20 Purchases for Buyers with Different Purchase Rates.

90

80

--......

;!.

't

a. 70

c::
en

(.)

60

-SCRpref: purchases 5-24


-+-SCRpref: purchases 65-84
--.!r- SCRpref: purchases

105-124

50
5

11

13

15

17

19

21

23

25

Number of purchases

Figure 4. Patterns of Declining Loyalty for Different Purchase Sequences (for the Same Sample of Buyers,
Laundry Detergent).

the three segments (Figure 3). The similarities between medium and heavy buyers can
be clearly seen, as can the consistently lower loyalty exhibited by lighter buyers. We discuss a possible explanation for this pattern and how it affects deviations from Dirichlet
predictions in Section 6.
An important point to note is that this pattern of an initial steep decline in brand loyalty followed by a much more gradual decline is not determined by where in the database
the purchase records are sampled. If we look at the patterns that occur for SCRpref in
three series of 20 purchases from the same buyers, each with a different starting point,
we find that they are very similar (see Figure 4). 8 This suggests that the pattern of ap-

Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.

13

THE RELATIONSHIP BETWEEN CUSTOMER LOYALTY AND PURCHASE INCIDENCE

15

30

45

60

75

90

105

120

135

150

165

180

195

210

Number of purchases

Figure 5.

Laundry Detergent: Brand Switching Parameter,

rp.

parent declining loyalty revealed by share-based measures such as SCR and SCRpref is
largely a statistical artifact, dependent on the number of purchases used to calculate the
measure. This motivates our use of a third loyalty measure, </>, which is not similarly
constrained.

5.3.

Brand Loyalty Measured Using the Polarization Index,</>

Figure 5 plots the observed values of</>, the polarization index or switching parameter, as
the number of purchases increases. The decline in </> suggests that buyers become more
similar in their brand choices as purchase incidence rises. </> influences share-based measures of loyalty, but as </> falls, its impact on SCR and SCRpref declines. This is demonstrated in Figure 6(a) which plots the relationship between SCR and </> at one purchase
rate (75) for a market with three simulated brands (with market shares of 40%, 20% and
1%). For the 40% and 20% brands SCR stabilizes when </> is about 0.25 and shows almost
no further decline. For the 1% brand, SCR continues to decline for values of </> less than
0.25. In practice this means that once</> falls below 0.25 (which equates to an S value of 3)
there is little further incremental change in SCR or other share-based measures, unless the
brand has a very small market share. In order to illustrate in more detail the relationship
between </> and SCR, Figure 6(b) shows expected values of SCR for a 20%-share brand
at different values of</> and at three different purchase rates (5, 15, and 75). We find that
Dirichlet predictions of SCR decline in line with</>. Superimposed on the theoretically generated curves are actual values of SCR for a 20%-share brand at different purchase rates
and different values of</>. In each case the SCR values are higher than expected given the
</>values.

Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.

14

STERN AND HAMMOND

70
,.. 40% share brand
60
50

+--~--------------------------<

"

..

"'

-2--------w ..

....

40

<ft.

20% share brand

1 % share brand

30

20
10

... -- -- -- -- - -- - - - --- -- - - -

- -- - - -

-&- - - - - - - - - - - - - - ....

80
70
60

'$. 50
~

"' 40
30
20
10
0.9

0.8

0.7

0.6

0.5

(b)

0.4

0.3

0.2

0.1

Figure 6. Relationship Between SCR and rp. (a) Illustration for Simulated Brands with Different Shares, Purchase Rate = 75. (b) Illustration for 3 Different Purchase Rates Simulated for a 20% Brand Compared with
Actual Values.

6.
6.1.

Discussion
Loyalty Measures Show a Decline but Overall Loyalty is Largely Stable

We find that customer loyalty as defined by all three measures (SCR, SCRpref and </>)
declines sharply as the number of purchases increases from 5 to 15, declines more slowly
from 16 to around 60 purchases, and from 60 purchases onward there is almost no change
in observed loyalty. The pattern of declining loyalty seen from 5 to 15 purchases (equiva-

Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.

THE RELATIONSHIP BETWEEN CUSTOMER LOYALTY AND PURCHASE INCIDENCE

15

lent to 4-12 months' purchasing for laundry detergent, and a very common base period for
loyalty analyses), is very different from that observed at high levels of purchase incidence.
However, when we model customer loyalty using the Dirichlet, which assumes that
brand choice probabilities are independent of purchase incidence, we find similar patterns
of declining loyalty. This suggests that:

The pattern of apparent declining customer loyalty is largely a statistical artifact,


dependent on the number of purchases used to calculate the measure.

6.2.

But, Heavier Buyers Are More Loyal

However, we do find that purchase weight (i.e., whether a buyer is a light, medium or
heavy purchaser of the category), affects share-based measures of loyalty. This seems a
real effect rather than one determined by statistical artifact.
The commonly reported finding that high loyalty tends to be driven by light buyers (who
are perhaps 100% loyal because they make few purchases, as suggested by Ehrenberg et al.,
2004), holds true only if the comparison is between light and heavier buyers in the same
time period.

If we compare brand loyalty across an equal number ofpurchases for light and heavier buyers (and therefore control for the small-number effect) we find that heavier
buyers are more loyal.
This may be an important consideration when designing customer relationship management programs (where customers are often segmented by weight of purchase).

6.3.

The Modeling of Customer Loyalty

What do our findings imply for the modeling of loyalty? Managers who wish to predict
long-run brand and customer loyalty can gain considerable insight by using a simple stationary market model such as the Dirichlet, and operationalizing the model with parameters
derived from short-term aggregate purchase data. The Dirichlet closely predicts the general
pattern of loyalty, but consistently under-predicts SCR for the brand leader. In addition, for
the share of category requirements satisfied by the most preferred brand, the model predictions, while describing a pattern similar to that observed, have a lower starting point and a
shallower slope, so that loyalty is under-predicted at low purchase rates.
Do buyers of different purchase weights contribute equally to this excess loyalty? In
Figure 7, we compare observed SCRpref for light, medium, and heavy buyers, with the
overall Dirichlet model predictions (which are the same for buyers of all purchase weights).
We find that:

Dirichlet predictions are closer to the actual values for light buyers than they are
for heavier buyers. The contribution of heavier buyers to brand sales could partially
explain the under-prediction of brand loyalty previously reported for some high share
brands.

Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.

16

STERN AND HAMMOND

~heavy

buyers

-----.>-----medium buyers
---e--- light

buyers

predictions

80

0~

~c.

70

0::

(.)

en

--....
- -- --- -------- _,.. ____ __ --------....-....

---

...

60

10

20

30

40

50

60

70

80

90

100

110

120

Number of purchases

Figure 7. SCRpref: Heavy, Medium, and Light Buyers - Actual vs. Dirichlet Model Predictions, Laundry
Detergent.

Why are heavy buyers more loyal than lighter ones at low levels of purchasing? We
observe from the laundry detergent data that, where a buyer has a two-brand portfolio
consisting of brands A and B, the sequence of brand purchases, 'A, A, A, A, A, B, B, B,
B, B' is far more common than 'A, B, A, B, A, B, A, B, A, B.' A tentative explanation for
this observation is that lighter buyers may have less opportunity to incorporate feedback or
learning effects into their purchase decisions, in practical terms they 'forget' more easily
than more frequent buyers. This could explain why a zero-order model such as the Dirichlet
is better at predicting loyalty for light buyers than for heavier buyers.

6.4.

Future Research

In terms of future research, the relationship between the measures of loyalty used here and
other loyalty measures, such as portfolio size (the number of different brands utilized by a
customer), requires investigation. Preliminary research indicates that similarly predictable
patterns will be found. We expect that just as SCR and SCRpref decline rapidly and then
stabilize, portfolio size will increase and then stabilize. We also expect that light buyers
will have larger brand portfolios than heavier buyers at low purchase rates but that there
will be little difference by purchase weight as the number of purchases increases. There
is also a need to simulate loyalty measures for a range of different markets characterized
by varying structural parameters. We would expect that the shape of the customer loyalty
curve would be determined largely by the number of purchases, with the level of loyalty
depending on </>.

Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.

THE RELATIONSHIP BETWEEN CUSTOMER LOYALTY AND PURCHASE INCIDENCE

17

In this research, the aim was to explore the general pattern of customer loyalty and we
did not have access to additional explanatory variables such as price, promotion, availability, or who buys versus who consumes. We believe that the relationship between customer
loyalty and purchase incidence revealed by this research is likely to generalize. However,
the inclusion of additional explanatory variables may improve the model fit where the forecasting of long-run loyalty for individual brands is the focus of attention.
If the deviations from Dirichlet model expectations reported here prove to be systematic,
the model could be adapted to take account of the short-run under-prediction of loyalty or
a model could be developed for each purchase weight segment. Finally, in the simulation
procedure used to generate expected values for SCR and SCRpref we held S constant,
however as our findings suggest that S increases (i.e. </> falls) as the number of purchases
rises, this may be an additional factor contributing to deviations in model fit. A future
development could be to introduce dynamic parameters into the model to reflect these
empirical findings.

Acknowledgements
The authors are grateful to TN AGB for providing access to the detergent data and to Colin
Maitland at ISIS research for the Jigsaw data on anti-depressant prescribing. We are also
indebted to the editor and two anonymous reviewers for their thoughtful and very helpful
comments on earlier versions of this paper.

Notes
1. </J, expressed as S, is the switching parameter from the Dirichlet model; <P = 1/(l+S). The use and comparison
of multiple measures (SCR, SCRpref and S) arose from suggestions from two anonymous reviewers on the
most useful measures for brand and customer loyalty, and for the need to identify clearly the effect that any
statistical artifact plays in commonly reported loyalty measures.
2. All continuous buyers on the two panels we analyze made at least one purchase per year - the panels cover
five years of continuous repo1ting and no continuous buyers are excluded from the analysis. This helps address the problem of infrequent category buyers. The length of the analysis pe1iod (five years), together with
the relatively high frequency of purchase in the two categories studied, means that over half of each panel
purchased fifty or more times; we are therefore able to observe the extent to which customer loyalty changes
as the number of purchases increases to very high levels, so addressing the issue of whether very frequent
category buyers are more or less loyal than the average buyer. The variation in portfolio size can be directly
observed as category purchasing increases, therefore the problem of a possibly aitificially-limited portfolio
size and subsequent over-estimation of the level of brand loyalty is also directly addressed. The markets analyzed here are mature and, in the aggregate, near-stationary (i.e., total category purchases and brand shares
change little from quarter to quarter over the five years).
3. UK, TN AGB consumer panel. The leading brand had an average market share of 28%, the market shares
for the other brands studied were: 21 %, 11 %, 9%, 8%, 5%, 3%, 2%, 2%, 2%, 1.3%. Eight households had
a minor brand (outside the top eleven) consistently as their most prefelTed brand and were excluded from the
analysis, giving a sample size of 1524 households.
4. Data are from the "Jigsaw" panel (ISIS Research). Each week physicians on the panel list all new and changed
prescriptions they have written. The sample closely matches the UK General Practitioner population in terms
of practice size and geographical location. Here we analyze just one type of drug physicians prescribe - anti-

Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.

18

5.

6.
7.
8.

STERN AND HAMMOND

hype1tensives - used to reduce blood pressure and prevent and treat heart disease. The market shares of the
top five brands were: 14.2%, 12.1 %, 8.8%, 6.9%, 6.6%.
We also studied 2nd and 3rd prefeITed brands in order to obtain an indication of the impo1tance of the most
prefeITed brand. For laundry detergent, on average, the brand most prefeITed by individual buyers accounts for
between 60% and 90% of purchases and there is relative! y little difference across brands. The brand prefeITed
second accounts for between 17% and 27% of purchases. Third prefeITed brands always made up less than
10% of purchases.
When repo1ting the findings for drug prescribing, for simplicity we refer to purchases rather than prescriptions
written.
t-tests gave p = 0.86 for physicians' prescribing at high purchase rates, for all other rates for physicians'
prescribing and for all rates for detergent purchasing, t-test p values were <0.05.
How Figures 2, 3, and 4 relate to each other. The 3 lines plotted in Figure 3 are drawn from 3 populations
which, when averaged, give the first 20 points on the solid line for SCRpref for detergent shown in Figure 2.
The 3 lines shown in Figure 4 are different purchases sequences, again of 20 purchases, but now for the same
sample of buyers, in this case they are all heavy buyers. The solid line in Figure 4 is the same as the solid line
in Figure 3.

References
Bass, Frank M., Abel P. Jeuland, and Gordon P. Wright. (1976). "Equilibrium Stochastic Choice and Market
Penetration Theories: Derivation and Comparisons," Management Science, 22(June), 1051-1063.
Bhattacharya, C. B. (1997). "Is Your Brand's Loyalty Too Much, Too Little, or Just Right?: Explaining Deviations
in Loyalty from the Dirichlet Norm," International Journal of Research in Marketing, 14(5), 421-435.
Bhattacharya, C. B., Peter S. Fader, Leonard M. Lodish, and Wayne S. DeSarbo. (1996). "The Relationship
between the Marketing Mix and Share of Category Requirements," Marketing Letters, 7(1), 5-18.
Cunningham, R. M. (1956). "Brand Loyalty - What, Where, How Much?" Harvard Business Review, 34
(January-February), 116-128.
Danaher, P. J., I. Wilson, and R. Davis. (2003). "A Comparison of Online and Offline Consumer Brand Loyalty,"
Marketing Science, 22(4), 461-476.
Deighton, John, Caroline M. Henderson, and Scott A. Neslin. (1994). "The Effects of Advertising on Brand
Switching and Repeat Purchasing," Journal of Marketing Research (February), 28-43.
Dekimpe, Mamik G., Jan-Benedict E. M. Steenkamp, Martin Mellens, and Piet Vanden Abeele. (1997). "Decline
and Variability in Brand Loyalty," International Journal of Research in Marketing, 14(5), 405-420.
Dowling, Grahame R. and Mark D. Uncles. (1997). "Do Customer Loyalty Programs Really Work," Sloan Management Review, 38(4), 71-82.
East, Robert and Kathy Hammond. (1996). "The Erosion of Repeat-Purchase Loyalty," Marketing Letters, 7(2),
163-171.
Ehrenberg, A. S. C., M. D. Uncles, and G. J. Goodhardt. (2004). "Understanding Brand Pe1formance Measures:
Using Dirichlet Benchmarks," Journal of Business Research (forthcoming).
Fader, Peter S. and David C. Schmittlein. (1993). "Excess Behavioral Loyalty for High-Share Brands: Deviations
from the Dirichlet Model for Repeat Purchasing," Journal of Marketing Research, 30( 4), 478-493.
Goodhardt, G. J., A. S. C. Ehrenberg, and C. Chatfield. (1984). "The Dirichlet: A Comprehensive Model of
Buying Behaviour," Journal of the Royal Statistical Society A, 147, 621-655.
Hauser, J. R. and B. Weme1felt. (1990). "An Evaluation Cost Model of Consideration Sets," Journal of Consumer
Research, 16, 393-408.
Johnson, Tod. (1984). "The Myth of Declining Brand Loyalty," Journal of Advertising Research, 24(February/
March), 9-17.
Krishnamurthi, Lakshman and S. P. Raj. (1991). "An Empirical Analysis of the Relationship between Brand
Loyalty and Consumer Price Elasticity," Marketing Science, 10(2)(Spring), 172-183.
Mela, Carl F., Sunil Gupta, and Donald R. Lehmann. (1997). "The Long-Term Impact of Promotion and Advertising on Consumer Brand Choice," Journal of Marketing Research, 34(2), 248-261.

Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.

THE RELATIONSHIP BETWEEN CUSTOMER LOYALTY AND PURCHASE INCIDENCE

19

Sabavala, Darius J. and Donald G. Monison. (1977). "A Model of TV Show Loyalty," Journal of Advertising
Research, 17(6), 35-43.
Shoemaker, Robert W., Richard Staelin, Joseph B. Kadane, and F. Robe1t Shoaf. (1977). "Relation of Brand
Choice to Purchase Frequency," Journal of Marketing Research, 14(4), 458-468.
Stem, Philip. (1997). "Long Run Loyalty -An Exploratory Study," Marketing and Research Today, 25(2)(May),
70-78.
Tellis, Gerard J. (1988). "Advertising Exposure, Loyalty and Brand Purchase: A Two-Stage Model of Choice,"
Journal of Marketing Research, 25(2), 134-144.
Uncles, Mark D. (1989). BUYER. Buyer Behavior Software.
Uncles, Mark D., A. S. C. Ehrenberg, and Kathy Hammond. (1995). "Patterns of Buyer Behavior: Regularities,
Models and Extensions," Marketing Science, 14(3), Part 2 of2, G71-G78.

Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.

S-ar putea să vă placă și