Journal of Business Research 57 (2004) 1307 – 1325

Understanding brand performance measures: using Dirichlet benchmarks

Andrew S.C. Ehrenberga, Mark D. Unclesb,*, Gerald J. Goodhardta
a
London South Bank University, London, UK
b
School of Marketing, University of New South Wales, 3rd Floor, John Goodsell Building, Sydney NSW 2052, Australia
Received 1 January 2002; received in revised form 1 August 2002; accepted 1 November 2002

Abstract

Sales of a brand are determined by measures such as how many customers buy the brand, how often, and how much they also buy other
brands. Scanner panel operators routinely report these ‘‘brand performance measures’’ (BPMs) to their clients. In this position paper, we
consider how to understand, interpret, and use these measures. The measures are shown to follow well-established patterns. One is that big
and small brands differ greatly in how many buyers they have, but usually far less in how loyal these buyers are. The Dirichlet model predicts
these patterns. It also provides a broader framework for thinking about all competitive repeat-purchase markets—from soup to gasoline,
prescription drugs to aviation fuel, where there are large and small brands, and light and heavy buyers, in contexts as diverse as the United
States, United Kingdom, Japan, Germany, and Australasia.
Numerous practical uses of the framework are illustrated: auditing the performance of established brands, predicting and evaluating the
performance of new brands, checking the nature of unfamiliar markets, of partitioned markets, and of dynamic market situations more
generally (where the Dirichlet provides theoretical benchmarks for price promotions, advertising, etc.). In addition, many implications for our
understanding of consumers, brands, and the marketing mix logically follow from the Dirichlet framework. In repeat-purchase markets, there
is often a lack of segmentation between brands and the typical consumer exhibits polygamous buying behavior (though there might be strong
segmentation at the category level). An understanding of these applications and implications leads to consumer insights, imposes constraints
on marketing action, and provides norms for evaluating brands and for assessing marketing initiatives.
D 2003 Elsevier Inc. All rights reserved.

Keywords: Brand performance measures; Benchmarks; Repeat-purchase markets; Penetration; Loyalty; Segmentation; New brands; Marketing mix; Dirichlet
model; Double jeopardy

1. Introduction and general answer to such a question: most performance

measures of most brands are just about normal. And in
1.1. Overview general, sales of a brand are largely determined by much the
same predictable patterns of buyer behavior.
Sales of Folgers, P&G’s leading brand of instant coffee Consumers are seen as mostly choosing from personal
in the United States, depended on it being bought three split-loyalty repertoires, typically buying one brand more
times a year on average by 11% of households. Roughly often than another. Within such a framework of steady but
12% of these customers were 100% loyal. Therefore, most divided loyalties, specific purchases then occur in a seem-
customers bought other brands as well. In total, they did so ingly irregular or even ‘‘as-if-random’’ manner. These per-
about as often as they bought Folgers itself—a ‘‘share of sonal repertoires of brands differ from one consumer or
category requirements’’ (SCRs) of only 50%. household to the next. Yet such heterogeneous behavior
Such performance measures are routinely tabulated in aggregates to brand performance measures (BPMs) that
marketing research reports and cited in marketing plans. But follow much the same ‘‘lawlike’’ pattern from brand to brand.
to interpret the figures, benchmarks and norms are needed. Key examples are that any loyalty-related measure is usually
For example, is it that as many as 12% or only 12% are much alike for different brands, that a brand typically has
100% loyal? The thesis of this paper is that there is a simple many light buyers, and that few of its customers are 100%
loyal over a sequence of purchases. Superimposed are regular
* Corresponding author. Tel.: +61-2-9385-3385; fax: +61-2-9663-1985. relationships with market share, e.g., for purchase rates and
E-mail address: [email protected] (M.D. Uncles). for brand switching.

0148-2963/$ – see front matter D 2003 Elsevier Inc. All rights reserved.
doi:10.1016/j.jbusres.2002.11.001
1308 A.S.C. Ehrenberg et al. / Journal of Business Research 57 (2004) 1307–1325

These ‘‘lawlike’’ patterns have been found in over 50 1.2. A new-brand case
varied product and service categories from soap to soup,
motor cars, prescription drugs, media usage, etc., and in Many marketing analysts seem to be puzzled how a
different countries and at different points in time, as is steady state model like the Dirichlet, which contains no
discussed in Section 2. These patterns are in turn closely explicit ‘‘decision variables,’’ might nonetheless be of use in
predictable from a single and parsimonious model, the dynamic situations such as launching a new brand. To
Dirichlet (or ‘‘NBD-Dirichlet’’ in full). BPMs are all deter- illustrate, we outline a simple new brand case based on an
mined in the model just by the brand’s market share and amalgam of practical experience. This shows the implica-
only indirectly by the marketing mix or consumer-related tions of knowing (or not knowing) the common Dirichlet-
factors acting through the market share. This is outlined in type market pattern that brand loyalty varies little, or
Section 3, with the theory being summarized in Appendix A. relatively little, between competitive brands.
The Dirichlet model postulates that each consumer has a
certain propensity—a probability in the model—to buy a 1.2.1. The new Brand X
given brand. This probability is assumed to be steady for the In considering a new instant coffee brand, ‘‘X,’’ a U.S.
time being but differing across heterogeneous consumers. corporation had decided on a market share of 5% or eight
The model is defined for steady state and unpartitioned purchases a year per 100 households. This target had been
markets where market shares are stationary and there is no set by the management after having considered various
clustering of particular brands. However, this does not imply marketing-mix expenditures and estimated revenues. The
that all markets should be stationary and nonpartitioned management was then given the choice of two marketing
(although many often are). The model purports only to policies, A and B.
describe what markets are like when they are near steady Policy A was to position Brand X as a niche brand since
and nonpartitioned. But, even where the market is not quite 16% of consumers had rated it very high in placement tests.
steady, or where there is some clustering, the Dirichlet X was therefore to be targeted at a small heavy-buying
mostly still holds and it provides useful benchmarks. segment who really liked the brand, with some 1% of the
The Dirichlet-type regularities and associated model still population buying it about eight times a year (thus, satisfy-
cause surprise. Can there be ‘‘lawlike’’ patterns or even just ing the sales target of eight purchases per 100 households).
a ‘‘near-steady state’’ in markets that are constantly sub- As a niche brand, X was to be premium priced, with loyalty-
jected to competitive inputs, technological innovations, and building promotions and advertising in up-market media.
environmental changes? Moreover, even if near-steady state Distribution was also to be up-market, without expensive
markets do exist, can they be of any interest to marketing trade deals, relying on product and advertising pull rather
practitioners who are always trying to change that steady than trade push.
state? The answers to both these questions appear to be a Policy B was an extreme mass marketing positioning. X
clear ‘‘yes.’’ Given that steady state Dirichlet norms or would be an add-on or variety brand based on its special
benchmarks occur widely, it is possible to address many product formulation. It would reach the preset sales target of
practical marketing issues (Section 4): eight by being bought by some 8% of the population but
only about once a year. It would be competitively priced,
(i) Auditing a brand’s performance to see whether or not it with awareness-building promotions, ‘‘try-something-differ-
is in fact normal. ent’’ advertising in mass media, and trade support bought by
(ii) Identifying market partitioning and other departures slotting allowances and eye-catching merchandising.
from the basic norms. The management then asked for a check against con-
(iii) Assessing and interpreting dynamic non-steady-state sumer panel data. The results in Table 1 showed that both
situations such as price promotions, new brands, and policies A and B were totally out of step with the market.
sales trends. Both the observed and theoretical purchase rates for the
different brands were very similar at about 3. This meant
The Dirichlet framework—the empirical patterns, the that the new Brand X also should be expected to stabilize at
mathematical model, and the underlying theory—also has about 3.
wider implications for our understanding of consumers, Interpolating more precisely, the analysts concluded that
brands, differentiation and positioning, advertising, sales if the new Brand X did reach its ultimate targeted share of
promotions, etc. (Section 5). 5% (or eight purchases per 100 households), it would be
The empirical Dirichlet-type regularities have seldom bought about 2.7 times a year by its buyers and hence by 3%
been criticized. But the patterns are not always well-known of all consumers (since 8/2.7 = 3). Therefore, a realistic sales
or understood, and their theoretical interpretation and prac- equation (where sales = percent buying
rate of buying)
tical use is at times controversial. Therefore, our aim in this would be 3% buying 2.7 times (a normal brand) compared
position paper is to bring together, update, and synthesize with the hypothesized strategies of policy A where 1% buy
these findings for marketing researchers, practitioners, eight times (a niche brand) or policy B where 8% buy once
teachers/students, and potential critics. (an add-on brand).
 A.S.C. Ehrenberg et al. / Journal of Business Research 57 (2004) 1307–1325 1309

Table 1 brand with a difference (i.e., one with a distinctive built-in

Annual penetrations and purchase rates (leading brands of instant coffee)
selling proposition). She argued that in fact competitive
Instant coffee Market Percent Purchases brands are mostly much alike; yet some are able to achieve
(USA 1992) share buying (per buyer)
greatness.
(%) O T O T We now step back from this introductory case history to
Folgers 24 11 12 3.2 3.1 review the main patterns of buyer behavior that have been
Maxwell House 22 10 11 3.1 3.1 observed for frequently bought products.
Tasters Choice 17 9 9 2.8 3.0
Nescafé 11 6 6 2.7 2.9
Sanka 9 5 5 3.0 2.8
High Point 1 1 1 2.6 2.6 2. Patterns of buyer behavior
Maxim 1 0.3a 0.8 4.5a 2.6
Brim 0.3 0.2 0.2 2.1 2.6 The basic finding about individual consumers’ varied
buying behavior is that whichever way the data are aggre-
Other brands 16 8 8 3.0 3.0
gated, they show regular patterns. These patterns slowly
The new Brand X 5 – 3 – 2.7 came to be recognized over the years, followed by many
Data: Hallberg (1996)/MRCA; O = observed; T = theoretical Dirichlet replications across different products, years, and countries to
predictions. develop their generalizability.
a
Outlier.
2.1. Brand performance measures

The NPD team now queried whether the similar average The variables analyzed are the standard BPMs used by
purchase frequencies of about three in Table 1 were not an large consumer panel operators (e.g., ACNielsen, IRI,
unexpected coincidence, with at least one big exception (for TNSofres, GfK, etc.) and their clients (e.g., Unilever,
Maxim). The analysts countered that the close fit of the P&G, Colgate, Kraft, Nestle, etc.) (Bucklin and Gupta,
well-established theoretical norms (T) showed that the 1999; Sudman and Wansink, 2002). They are of three kinds,
pattern was neither accidental nor unexpected. They added as illustrated in Table 2 for the leading U.S. instant coffee
that to break out of this pattern, X would have to differ far brand.
more from the existing brands than they differed from each
other. Yet the marketing plan for X had claimed nothing of 2.1.1. Brand-size-related measures
the kind. From this, management concluded that the new The two main measures of size, market share and market
brand would not reach its 5% sales target via either the niche penetration, are very different conceptually:
policy A or the mass market policy B. They also felt that X’s
allocated budget was not sufficient for an achievable me-too
Total purchases of the brand
3
2.7 target. As a result, plans for Brand X were dropped. Market share ð%Þ ¼
This forestalled that it would be one of the traditional ‘‘four Total purchases of the category
out of five new brands that fail.’’
Penetration ð%Þ
1.2.2. A postmortem
The number buying the brand at least once
Later, senior management asked their new marketing- ¼
insights director to critique the earlier planning. First, she The total number of potential customers
queried that two totally different policies, niche versus
variety branding, had been put forward as alternatives for Folgers market share (24% in terms of purchase occa-
the same functional Brand X. This struck her as an sions) was much the same in any length time period in a
inappropriate ‘‘anything goes’’ approach to planning. Sec- typically near-steady market. But the number of customers
ond, she questioned whether a me-too launch should have buying Folgers increased greatly, from about 1% in a week
been ruled out as unattractive compared with launching a to 11% in a year.

Table 2
Common Brand Performance Measures (BPMs): Folgers
USA 1992 Brand size Loyalty-related measures (annual) Switching (annual)
Folgers Market Percent buying in a: Purchases Percent buying Category Percentage of Folgers buyers who also boughta
share (%) Week Year (per buyer) Once 5+ Purchases SCR b
MH TC Ne Sa HP Ma Br Other
Observed 24 1 11 3.2 46 18 6.4 50 31 24 21 12 1 1 1 21
a
For brand names see Table 1.
b
Share of category requirements.
1310 A.S.C. Ehrenberg et al. / Journal of Business Research 57 (2004) 1307–1325

2.1.2. Loyalty-related measures buying). This tendency to be ‘‘punished twice’’ just for
Folgers customers averaged 3.2 purchases in the year. being small was called ‘‘double jeopardy’’ (DJ) by
But half bought the brand only once. These customers McPhee (1963), who explained it as statistical selection
bought the category (i.e., any instant) 6.4 times on average, effect. The DJ effect is however small unless penetrations
giving the brand an SCR of 50% (i.e., 3.2/6.4). are very high.
 Heterogeneity—Consumers are very heterogeneous. On
2.1.3. Switching-related measures average, one in six of a brand’s annual customers bought
Table 2 also shows which other instant coffee brands it five or more times (with DJ again: smaller brands have
Folgers’ customers bought once or more in the year; e.g., even fewer heavier buyers). But some 50% bought it just
31% also bought Maxwell House, and only 1% also bought once in the year (as for Folgers in Table 2).
Maxim.  The natural monopoly effect—How often customers of a
The question now is what these figures mean. Why are brand bought the whole category increases slightly from
they what they are? Is an 11% annual penetration high or 10 to 13 with decreasing market share. This atypical
low? Is a SCR of only 50% as threatening as it seems? Does (upward) trend was called ‘‘natural monopoly’’ by
the relatively high purchase duplication of Folgers with McPhee in 1963. It implies that large brands slightly
Maxwell House mean that these two brands are highly ‘‘monopolize’’ light category buyers.
competitive—much more so than with Maxim?  SCRs—The average SCR of a brand over the year is
quite low; e.g., about 27% annually (on average 3.2/12).
2.2. Compared with what: regularities from brand to brand Any brand’s customers mostly buy other brands,
showing multibrand buying behavior.
To interpret these numbers for Folgers, we can compare  100% loyals—Only some 12% of a brand’s customers
them with those for other brands. To illustrate, Table 3 were 100% loyal in the year, with even fewer for a small
compares a dozen performance measures across the top brand (DJ again). 100% loyals are in fact light users of
eight brands in 12 varied product categories internationally. the whole category (on average about 4 such purchases
Several patterns are apparent in Table 3 and for each of versus 12 by all of the brand’s customers). They do not
the 12 categories: have many opportunities to be disloyal.
 Which other brands are also bought—Much the same
 Double jeopardy—Market shares and penetrations proportions of any particular brand’s customers also
decrease greatly, by almost 10-fold, from Brand A to bought Brand A (all about 54%). Similarly for C (about
Brand H. In contrast, loyalty- or switching-related 33%) or E (about 19%). The markets in question were
measures either stay broadly the same from brand to therefore nonpartitioned, i.e., with no apparent clustering
brand or decrease far less (by factors of 2 or 3 at of some subsets of brands. These switching levels are
most). Smaller brands therefore not only have far fewer themselves proportional to the brand’s penetrations—the
buyers than the bigger brands but also show somewhat so-called ‘‘duplication of purchase law,’’ as described in
lower average purchase frequencies (i.e., lower repeat Appendix A.

Table 3
Performance measures brand by brand (averages for 8 leading brands in 12 product categories)
Brands Brand size Loyalty-related measures (annual) Switching (annual)
(by share) Market Percent buying Purchases Percent Category 100% loyal Percentage who also bought brand:a
Share Weekly Yearly (per buyer) buying Purchasesb SCR Percent Purchases A C E
(%) (5+ times) (%) buying
A 28 3.6 46 3.9 24 10 39 22 4 100 31 17
B 19 2.5 35 3.6 21 11 33 16 4 51 32 18
C 12 1.5 25 3.1 16 12 26 11 4 57 100 20
D 9 1.2 22 2.8 13 11 24 11 4 55 35 23
E 7 0.9 14 3.2 19 12 26 12 3 53 34 100
F 5 0.7 12 2.7 11 12 22 10 3 56 35 19
G 4 0.6 11 2.7 11 12 22 9 3 56 35 17
H 3 0.4 6 3.2 13 13 24 7 3 52 30 17

Average 11 1.4 21 3.2 16 12 27 12 4 54 33 19

brand
Categories: Catsup, cereals, cheese, orange juice, household cleaners, laundry detergents (2
), paper towels (2
), take-home beer, toothpaste (2
).
Sources: AGB (UK), Nielsen and IRI (USA), GfK (Germany), TCI (Japan).
a
A selection of brands.
b
Per brand buyer.
 A.S.C. Ehrenberg et al. / Journal of Business Research 57 (2004) 1307–1325 1311

Table 4 by a single statistical model—the NBD-Dirichlet (‘‘Dirich-

Varied conditions for Dirichlet-type patterns
let’’ for short)—and that this model is simple, consisting of
Products and services Time, space, and people just a few well-based assumptions. The model was devel-
-Food, drink, cleaners, -United States, UK, Japan, oped by Chatfield and Goodhardt (1975) to account for the
and personal care Germany, Australasia, etc. known empirical patterns, and subsequently by Bass et al.
-Prescription drugs, -Light and heavy buyers:
(1976) on a theoretical ‘‘utility’’ basis. To us, the model’s
OTC medicines demographic subgroups
-Gasoline, aviation fuel, -Household or individual critical feature is that it is able to describe the various
cars, PCs purchases observed brand performance patterns; and in that sense, it
-Retail chains, financial -Different length analysis also helps explain and predict them.
services, B2B periods Table 5 illustrates the Dirichlet model’s close theoretical
-TV episodes, programs, -Different points in time:
predictions of the observed data, which were shown in Table
and channels 1950-2003
3. The fit for each separate measure can be summarized in
Brands and product variants Market conditions two respects: (a) the virtual lack of systematic bias, and (b)
-Large and small brands -Near-steady markets the O and T values for the individual brands also differing
-Pack sizes, flavors, -Dynamic markets little (e.g., with an average deviation, or mean absolute
forms, formats, etc. (for loyalty-related measures) deviation (MAD), of less than 2 percentage points for the
-Private/own labels -Nonpartitioned markets penetration-type percentages).
-Price bands -Partitioned submarkets

3.1. The Dirichlet assumptions

 The length of the analysis period—A brand’s penetration
being much higher in a year than in a week (as in Table Two broad propositions underlie the model. One con-
2) is less than pro rata due to some repeat-buying cerns consumers. The other concerns brands.
occurring. In contrast, loyalty-related measures are Consumers are thought of as highly experienced. They
much lower in a longer period: nearly every buyer is are therefore no longer easily influenced by a further
100% loyal in a week (when at most a single purchase purchase or seeing yet another advert (i.e., no additional
of the whole category is usually made) but few are in ‘‘learning’’ occurs). Hence, consumers can behave as if they
the year. It is usual therefore to examine periods that have, for the time being, steady habitual personal purchase
equal and exceed average interpurchase times; for propensities—stochastic probabilities in the model—for
syrup, appropriate analysis periods might be six months when they buy and what brands they choose on different
and a year, whereas for gasoline—where average purchase occasions. Crucially for the Dirichlet model, a
interpurchase times are shorter—a week and a quarter consumer’s probability of buying a particular brand would
might be more appropriate. Studies where consumers not be affected by what other brands the consumer is also
are labeled as ‘‘loyal,’’ without specifying the period of buying.
observation, are likely to be ambiguous and misleading. Brands are characterized in the model by their purchase
 Other measures—Similar regularities occur for other probabilities and hence their market shares. Beyond that, the
BPMs, e.g., quarter-by-quarter repeat buying is much the model does not specify whether the brands are functionally
same for different brands (about 70% for the average of differentiated or not, or differently marketed—this is all
the 12 products in Table 3 but with DJ effects again). subsumed in the model’s steady state purchase probabilities.
Similarly for duplicated buyer’s average rates of buying. These notions are captured in the theoretical model by
Again, a consumer’s favorite brand typically may five specific distributional assumptions: two assumptions
account for only some 60% of their annual category are about consumer heterogeneity for purchase frequency
purchases. and for brand choice, two about the probabilistic incidence
of specific purchases of the product and of a brand, and the
2.3. Conditions under which the patterns occur final one is about the statistical independence of these two
aspects. The individual assumptions are spelt out in Appen-
The generalizability of each of the above patterns is now dix A, along with procedures for calculating the theoretical
widely established across a great variety of very different BPMs. There is strong theoretical as well as empirical
product and services categories, as shown in Table 4. support for each of these assumptions.
Anyone with access to consumer panel or similar data can To use the model, it has to be calibrated for the chosen
check on such patterns. product category and brand. Only four numerical inputs
are needed, typically the penetrations and the average
purchase frequencies of the category and of one specific
3. The Dirichlet model brand. To estimate the BPMs of any brand of particular
interest its own market share has also to be inputted (see
Given their wide generalizability, it is perhaps unsurpris- Table A1 of Appendix A). The model is therefore very
ing that these patterns of buyer behavior can all be described parsimonious.
1312 A.S.C. Ehrenberg et al. / Journal of Business Research 57 (2004) 1307–1325

Table 5
Annual observed and theoretical performance measures
Brands Brand size Loyalty-related measures (annual) Switching (annual)
(by share) Market Percent Purchases Percent Category 100% loyal Percentage who also bought brand:
share buying (per buyer) buying purchases
(%) 5+ times per buyer
Percent Purchasesa A C E
buying
O T O T O T O T O T O T O T O T O T
A 28 46 46 3.9 3.9 24 25 10 10 22 16 4.1 2.5 100 100 31 31 17 20
B 19 35 36 3.6 3.5 21 20 11 11 16 13 4.4 2.2 51 55 32 31 18 21
C 12 25 25 3.1 3.1 16 17 12 12 11 11 4.4 2.0 57 56 100 100 20 21
D 9 22 21 2.8 2.9 13 14 11 12 11 10 2.7 1.9 55 56 35 31 23 21
E 7 14 16 3.2 2.8 19 14 12 12 12 9 2.9 1.9 53 56 34 32 100 100
F 5 12 11 2.7 2.8 11 12 12 12 10 9 2.9 1.8 56 56 35 32 19 22
G 4 11 10 2.7 2.9 11 12 12 12 9 9 2.6 1.8 56 56 35 29 17 18
H 3 6 7 3.2 2.6 13 13 13 12 7 9 3.4 1.7 52 56 30 30 17 18

Average 11 21 21 3.2 3.1 16 16 12 12 12 11 3.6 2.0 54 56 33 31 19 20

brand
O = Observed as in Table 3; T = theoretical Dirichlet predictions.
a
Consistently low—see discussion of systematic discrepancies in Section 3.2.

3.2. Deviations from the model  For some market leaders, annual purchase frequencies
are a unit or so higher than predicted. But this occurs
Other variables such as marketing-mix inputs or con- only occasionally (Ehrenberg et al., 1990; Fader and
sumer attributes do not have to be explicitly specified. This Schmittlein, 1993; Reibstein and Farris, 1995). It may be
is because the model is for the steady state and assumes due to large brands being more likely to be kept in stock
that these effects will usually already have been subsumed and may not be as important as is sometimes thought.
in the brands’ market shares, which in turn affect the  An early systematic discrepancy was the so-called
brand’s other performance measures. Alternatively, they ‘‘variance discrepancy.’’ This was found to reflect a
will show up as discrepancies in the model predictions, shortfall of very frequent buyers, which was ‘‘explained
which then need to be explained (e.g., Bhattacharya et al., away’’ as an artifact because few people buy a typical
1996; Bhattacharya, 1997; Ehrenberg, 1988; Kahn et al., grocery product more than once a week or more than 13
1988). times in a quarter (Chatfield, 1967; Ehrenberg, 1959).
Given the large consumer panels that are used nowadays  The annual purchase rates of 100% loyal buyers are
(e.g., 10,000+ households), many of the reported discrep- consistently underpredicted (e.g., by a purchase or two
ancies from the model will be ‘‘significant’’ (i.e., not just a for each of the 100 brands in Table 5). This is so far
sampling error). Moreover, a 5% difference from an average unexplained. But the discrepancy usually varies rela-
buying rate of 3.0 as in Table 5—e.g., 2.8 or 3.2—will of tively little from brand to brand (except with market
course matter in terms of sales. But the modeling goal here is share) and has not been of diagnostic (i.e., ‘‘differ-
not to predict the sales volume of individual brands—these entiating’’) marketing value so far.
are already known—but to elucidate market structure. Name-  The distribution of light, medium, and heavy buyers of a
ly, how very similar the brands’ average purchasing rates are total category (or a subcategory) is sometimes a little
in Table 5 compared with the eightfold variation (800%) in ‘‘flatter’’ than predicted by the NBD part of the Dirichlet
their penetrations. model (i.e., there are slightly too many ‘‘medium’’
Nevertheless, some systematic discrepancies have been buyers in the model). The reasons are not yet understood.
reported, for example: The more limited ‘‘empirical Dirichlet’’ model can then
be fitted instead, although at the cost of not being able to
 Quarter-by-quarter repeat buying can be overpredicted predict theoretical length-of-time period effects, such as
(by 5 to 10 percentage points—Ehrenberg, 1988). In penetration growth (Ehrenberg, 1988).
addition, repeat rates over nonadjacent periods tend to
erode somewhat by eight percentage points over a year Such known discrepancies seldom curtail the model’s
(East and Hammond, 1996). This is a real departure from practical use. Attempts to improve or elaborate the Dirichlet
a ‘‘steady state’’ and can be regarded as a ‘‘slightly leaky have so far not resulted in major gains in either predictive
bucket.’’ Follow-up work is needed. power or parsimony (see Appendix A).
 A.S.C. Ehrenberg et al. / Journal of Business Research 57 (2004) 1307–1325 1313

4. Practical applications  For private labels, the Dirichlet consistently shows

somewhat higher buying frequencies. This was probably
We now review various practical uses to which the due to their inherent nonavailability in other chains, as
Dirichlet framework has so far been put. Examples are was tested in two ways: by grouping PLs together as a
auditing the performance of established brands, predicting megabrand and by analyzing a PL within its own store
and evaluating that of new brands, and checking the nature chain. In the outcome, private labels seemed to show
of unfamiliar markets, of partitioned markets, and of dy- normal repeat-buying loyalty (see Uncles and Ellis,
namic market situations more generally. In essence, the 1989; Bound and Ehrenberg, 1997), but more work is
Dirichlet provides theoretical benchmarks for assessing needed.
price promotions or advertising, say, one factor at a time,
instead of using all-in-one marketing mix ‘‘decision-orien- 4.2. Extensions to new conditions
tated’’ modeling (e.g., contrast Leeflang and Wittink, 2000
with Ehrenberg et al., 2000). The Dirichlet also provides benchmarks when analyzing
data for another year, country, or category. Instead of
4.1. Brand performance audits unfocused data mining, it is simpler to check whether the
Dirichlet patterns recur. This took place in all the analyses
A common use of the Dirichlet is to assess how existing listed in Table 4, some for very different kinds of markets or
brands are performing. Thus, it was not clear in Table 2 market conditions. Examples are:
whether any of the figures observed for Folgers were high or
low (or ‘‘good’’ or ‘‘bad’’). In practice, the Dirichlet norms  Geographical extensions—to Australasia and Japan
showed that all but one of the measures in the table was (Wright et al., 1998; Kau et al., 1998) and to markets
normal, as shown in Table 5. In another study, 34 product in ‘‘developing’’ countries (Bennett, 2000).
categories were systematically assessed to see if BPMs for  Unusually frequently bought categories—gasoline, for
P&G brands were as expected (Uncles et al., 1994, 1995). example, which is bought on average as often as once a
Isolated deviations occur but often tend to have fairly week and has distinctive solus-site retail distribution
simple explanations, as in the following five cases: (Scriven and Ehrenberg, 1994).
 Doctors’ medical prescriptions—where the doctor nei-
 In Table 1 for U.S. instant coffee, Maxim’s very high ther buys, pays for, nor consumes the product (Stern,
annual purchase rate of 4.5 was found to be due to two 1995; Stern and Ehrenberg, 1995).
‘‘outliers’’ (two households making over 30 purchases  Durables and services—repurchase of PCs (Long and
each, as against the average of about 3—perhaps the Ehrenberg, 1998) and cars (Ehrenberg and Bound, 2000;
households were running a bridge club!). Colombo et al., 2000) and frequent flying by business
 Brim also had a somewhat low average purchase travelers (Harris, 2003).
frequency of 2.1 (Table 1). This had not occurred the  ‘‘Impulse’’ purchases—made for immediate personal
year before (the first thing to check). It was therefore not consumption (McDonald and Ehrenberg, 2002).
a general characteristic of the brand but a one-off effect  Television—repeat viewing and switching for programs
due perhaps to a fire in a warehouse (‘‘bad’’) or an and channels (Barwise and Ehrenberg, 1998; Goodhardt
exceptional promotion with an extralarge bonus of once- et al., 1987).
only buyers (‘‘good’’ presumably). An analysis exclud-  Industrial purchasing and procurement—isolated cases
ing the promoted month would have shown this more so far, such as chemical additives, paper and packaging
explicitly (but the raw data were no longer available). (Easton, 1980), aviation fuel contracts (Uncles and
 For instant coffee in the UK, the market leader Nescafé Ehrenberg, 1990b), and ready mix concrete (Pickford
has over the years had a somewhat high buying rate, and Goodhardt, 2000).
even given Nescafé’s high 34% share (possibly this  Store choice—repeat visits and switching between
results from the occasional brand leader effect already grocery outlets (Kau and Ehrenberg, 1984; Kau et al.,
mentioned). 1998; Uncles and Ehrenberg, 1990a; Uncles and
 A remarkably high rate of new medical prescriptions Hammond, 1995; Wrigley and Dunn, 1984, 1985),
occurred in 1986 for Squibb’s cardiovascular drug womenswear retailing (Brewis-Levie and Harris, 1999),
Capoten: 10 new prescriptions a year per prescribing and recently on-line electronic shopping (Danaher et al.,
doctor instead of the norm of about 5. This was traced to in press; Moe and Fader, 2001).
doctors who were provided with a ‘‘free’’ PC if they  SKUs—customer retention and switching between stock-
prescribed Capoten often enough for medical monitoring keeping units, such as specific combinations of pack
purposes (this might seem a highly successful loyalty sizes, flavors, and model specification (as for cars or
inducement but would soon become financially crip- PCs). Whereas brands are often similar (‘‘the commodity
pling). The prescribing rate fell back to normal when the with a name’’), SKUs unquestionably differ functionally
inducement was withdrawn (Stern and Ehrenberg, 1995). (e.g., large versus small pack sizes). Early work shows
1314 A.S.C. Ehrenberg et al. / Journal of Business Research 57 (2004) 1307–1325

that the Dirichlet-type patterns nonetheless still predom-  Low at about 5% between two different brands of either
inate (Singh et al., 2000). leaded or unleaded in the southwest and northeast
quadrants, again decreasing with the penetrations. The
4.3. Market partitioning figures are not zero mainly because of households with
two cars, one older and still requiring leaded gasoline.
Markets for directly substitutable brands are usually not  Middling at about 15% for the same brand unleaded and
partitioned; that is, they show no special clustering of leaded (in the southwest and northeast diagonals), again
particular brands. But some subcategories exist; for exam- because of 2+ car families buying both variants mainly at
ple, coffee by ground and instant, decaffeinated and regular. the same local gas station.
Each of these functional attributes may then attract a special
‘‘segmented’’ following, with consumer choice showing Here is a case where having prior knowledge of the
marked clustering for the more similar items (partitioning). duplication patterns can be expected to work much better
Such clustering is brought out by relating the observed than a statistical multivariate ‘‘discovery’’ technique. Sim-
duplications of purchase between pairs of brands to these ilar points emerged from an exhaustive study of car repurch-
brands’ penetrations, with the nonpartitioned Dirichlet-re- asing in Europe, where a considerable number of techniques
lated ‘‘duplication of purchase law’’ as a well-established were applied and compared and where a focus on expected
norm. This allows for the penetration of each brand. The duplication patterns greatly assisted the analysis (Colombo
‘‘law’’ usually continues to apply directly not only within et al., 2000).
each separate partition but also between the partitions, but
with differing ‘‘duplication coefficients’’ (i.e., the ratios of 4.4. New brands
the duplication levels and the penetrations—see Table A4 of
Appendix A for the calculations). Using the Dirichlet for initial new-brand planning has
A striking case of partitioning has been reported for the already been rehearsed in Section 1 (the case of Brand X)
automobile market, with ‘‘luxury cars’’ not unexpectedly (see also Ehrenberg, 1991). Another application is in eval-
showing up as the most distinct of several clusters (Ehren- uating a new brand some time after its launch. Campbell’s
berg and Bound, 2000). Table 6 illustrates another instance newish premium priced ‘‘Tastes of the World’’ (TotW) soup
of complex-yet-simple partitioning, this time for the UK in Britain for example showed low loyalty-related measures,
gasoline market between unleaded and leaded variants compared with its bigger competitors. TotW seemed doomed
(Scriven and Ehrenberg, 1994) (for simplicity, just four until their marketing advisors made Campbells aware of the
brands are reported in the table). DJ phenomenon, i.e., that repeat-buying levels are predict-
The purchase duplications between pairs of brands were ably lower for smaller brands. This showed that TotW’s low
relatively high, low, or middling but always decreased with loyalty was in fact normal for its size. TotW’s problem was
decreasing penetrations: that it had too few customers.
The general view for new brands is that loyalty grows
 High between the unleaded brands in the northwest slowly (e.g., as implied in analyses by ‘‘depth of repeat’’).
quadrant (and also the leaded ones in the southeast But no generalizable results of this have been reported (e.g.,
quadrant) and decreasing from some 30% to 10% in line Hardie et al., 1998). In contrast, Wright and Sharp (1999),
with the brands’ penetrations. and Ehrenberg and Goodhardt (1968a) much earlier, have
reported two isolated cases of near-instant loyalty occurring
Table 6 unexpectedly for two small new brands in Australia and the
Market partitioning: unleaded versus leaded gasoline (percent of buyers of UK. A more extreme exception was Unilever’s new toilet
X who also bought Y. Typical brands) soap Shield some years ago. This achieved a remarkable
Britain Who also bought 20% share in a very traditional UK market almost immedi-
(Quarter I, Unleaded Leaded ately at launch, with repeat buying and switching also very
1990)
Es BP Mo Gu Es BP Mo Gu
high immediately, and indeed as for an established 20%
brand in that market (Wellan and Ehrenberg, 1988). Each of
Buyers of Esso % 100 25 18 11 18 6 2 2
unleaded BP % 36 100 23 14 6 15 2 2
the three cases was regarded at the time as unusual and
Mobil % 32 30 100 10 9 7 15 1 unexpected.
Gulf % 29 25 14 100 6 4 3 17 However, a new empirical finding across 22 cases is that
average purchase frequencies for successful new brands
Buyers of Esso % 19 4 5 2 100 23 13 9 were in fact normal almost instantly; that is, they were at
leaded BP % 7 13 5 2 28 100 16 13
Mobil % 4 4 16 2 25 25 100 14
the Dirichlet level as for any brand in that market (Ehren-
Gulf % 5 4 2 17 22 27 18 100 berg and Goodhardt, 2000). One instance is in Table 7 for
doctors’ prescriptions for the then new antidepressant Pro-
Penetrationa % 20 14 11 8 19 15 10 8 zac (2.3 prescriptions per prescribing doctor). It follows that
a
Percent buying at least once in the quarter. the previous isolated cases of near-instant loyalty were not
 A.S.C. Ehrenberg et al. / Journal of Business Research 57 (2004) 1307–1325 1315

Table 7 segmentation would affect the timing of advertising and

A new brand: near-instant loyalty (doctors’ ‘‘new’’ prescriptions of Prozac)
any attempts to spread peak demand.
Prozac Quarters Average  Price-related promotions—A wider use of Dirichlet-type
a
(I) II III VII VIII QII-VIII norms in analyzing 150 price promotions in Britain,
Percent 1 3 8 17 21 14 Germany, Japan, and the United States has shown that
prescribing (i) before-to-after repeat buying is unaffected (i.e., no
Average 1.0 2.3 2.2 2.9 2.3 2.3 ‘‘learning’’), (ii) no before-to-after sales increase occurs,
prescription rateb
and (iii) there are virtually no new buyers (Ehrenberg et
a
Prozac was launched at some time during this quarter. al., 1994a). The conclusion was that price promotions
b
Average new prescriptions a quarter per prescribing doctor.
are mainly taken up by existing customers of the brand:
promotions just bring forward purchase timing where
exceptional after all. This finding of near-instant loyalty the brand is already in the purchase repertoires of
could not have become apparent if the close similarity of consumers.
loyalty measures for established brands (as in Table 5) had  Long-term trends—The 1992 U.S. instant coffee market,
not already been well established. as in Tables 1 and 2, had changed radically from 10 years
before in two ways: (a) category sales had slumped by
4.5. Analyzing market dynamics two thirds, and (b) Folger’s brand share had doubled.
Nonetheless, the loyalty levels were predictably Dirichlet
The steady-state Dirichlet norms have also been used to both in 1992 and 1981 (Ehrenberg, 1997). Only the
diagnose dynamic situations where there are marked market shares and penetrations had changed. Loyalty was
changes in market shares and/or in total market size. The unaffected by the market trends. In another long-run
steady-state norms can then be used to identify and interpret study, the UK toothpaste market was similarly found to
those measures that are still normal and the measures that be Dirichlet both in 1967/1968 and 1990/1991 despite a
have in fact changed. Some isolated cases published so far major decline in sales of Gibbs and the introduction of
include the following: successful new brands—Crest, Aquafresh, and Senso-
dyne (Ehrenberg et al., 1994b).
 A short-term trend—A few years ago, the total sales of
U.S. laundry detergents dropped by as much as 15% These various cases seem to suggest that penetration—
from one quarter to the next. Repeat buying was tracked the number of customers—is the main factor that changes
in a ‘‘conditional trend analysis’’ (CTA) as a diagnostic when sales increase. This is strongly supported by a large
tool. This gives predicted repeat-buying norms from QI systematic study, which shows that although penetration
to QII conditional on each consumer’s QI purchasing and loyalty both grow when share grows (in line with the
level (Goodhardt and Ehrenberg, 1967; Morrison, Dirichlet/DJ cross-sectional results rehearsed here), in-
1969). Table 8 shows that the sales drop was hardly creased penetration is the key factor (Baldinger et al.,
due to a loss of loyalty (neither light nor heavier QI 2002). More work into market dynamics is planned.
buyers bought much less than expected in QII). It was
mainly due to too few ‘‘new’’ buyers coming in only 4.6. Further applications
25% QI nonbuyers buying in QII rather than the ex-
pected 34%. Dirichlet norms have been used as benchmarks in ex-
 Stockouts—Temporary out-of-stocks need not lead to a ploring many other marketing issues. Empirical examples
longer-term loss of sales: CTA showed experimentally are: (i) cannibalization (Lomax et al., 1996), (ii) price
that before-to-after repeat buying was normal, as if there sensitivities (Ehrenberg et al., 1997b; Scriven and Ehren-
had been no out-of-stock. Consumers had not learnt to berg, 1995; Scriven, 1999), (iii) consumer loyalty programs
like their ‘‘new’’ brands during the forced switching (Sharp and Sharp, 1997, 1999), and (iv) subscription mar-
(Charlton et al., 1972). kets (Sharp et al., 2002). Our general conclusion is that the
 Extra product promotions—In an isolated analysis, the well-grounded steady-state norms reflected by the NBD-
giant pack of Unilever’s leading UK laundry detergent
was promoted with extra product. Management expected
Table 8
this to appeal to the brand’s heavy buyers. But the CTA Conditional trend analysis (CTA): powder detergent
norms showed the opposite—the offer of extra product
A 15% sales drop Buyers of powder detergent in QI
attracted recent nonbuyers and light buyers (Ehrenberg, from QI to QII
1988). Nonbuyers Once only Two plus
 The seasonal soup market—The CTA norms showed that O (%) T (%) O (%) T (%) O (%) T (%)
the winter peak of soup sales was about half due to all- Percent buying 25 34 57 63 88 90
year-round buyers buying more and half due to peak- powder in QII
season-only buyers (Wellan and Ehrenberg, 1990). This O = Observed measures; T = theoretical NBD predictions.
1316 A.S.C. Ehrenberg et al. / Journal of Business Research 57 (2004) 1307–1325

Dirichlet successfully provide benchmarks for an unusually et al., 1995; McAlister and Pessemier, 1982); in choice
varied range of situations. modeling (e.g., Horowitz and Louviere, 1995); in qualita-
tive studies (e.g., Fournier and Yao, 1997; Gordon, 1994);
and in the context of limited problem solving (e.g., East,
5. Marketing implications 1997; Foxall and Goldsmith, 1994; Olshavsky and Gran-
bois, 1979). However, the varying behavior of consumers
The Dirichlet framework also has implications for our is sufficiently idiosyncratic and irregular to be successfully
broader understanding of consumers, brands, and the mar- modeled mathematically as being quasi-random, especially
keting mix. These implications follow logically from the collectively.
preceding patterns, models, and applications, as we now
outline. 5.1.2. Brand segmentation
Consumers are widely expected to fall into relatively
5.1. Understanding consumers homogeneous and recognizable subgroupings. But such
brand segmentation is not directly allowed for in the
5.1.1. Polygamous consumers Dirichlet framework, which nonetheless successfully pre-
Over any sequence of purchases, consumers tend to buy dicts most BPMs.
several brands with largely steady habitual propensities, at Indeed, the vast segmentation literature is surprisingly
least for the time being. Typically, consumers are polyga- lacking in explicit empirical cases where directly competitive
mous rather than either promiscuous or monogamous (ex- brands do appeal to different kinds of people. Much of this
cept possibly in ‘‘subscription’’ markets). They usually have literature is focused on techniques, not empirical results (e.g.,
several steady partners—a repertoire—with one or two Wedel and Kamakura, 2000). In contrast, there is now much
usually being favorites (Hammond, 1997). systematic empirical evidence that the user profiles of sub-
For a consumer to be ‘‘loyal’’ to a brand, the Dirichlet stitutable brands seldom differ (Collins, 1971; Hammond et
model does not presuppose any unique or explicit ‘‘com- al., 1996; Kennedy and Ehrenberg, 2000). In practice, the
mitment’’ to that brand, nor that the consumer be either a customers of similar brands are very similar, as would tend to
particularly heavy or exclusive buyer of it. Indeed, as noted follow if nearly each of them uses several brands.
in Section 2, 100% loyals over any extended sequence of Nor do brand users generally differ in their attitudes to
purchases are rare and are light buyers anyway. the brands they buy. For Dirichlet-type markets, the data
Because consumers are generally highly experienced, show that buyers of Brand A feel about A much as buyers of
making a further purchase of a brand would not affect a Brand B feel about B (Barwise and Ehrenberg, 1985;
Dirichlet consumer’s probability of buying the brand (the Dall’Olmo Riley et al., 1997; Franzen, 1994). This similar-
model’s crucial ‘‘zero-order’’ assumption—see Appendix ity of brand-users’ attitudes is consistent with the small role
A). Any ‘‘learning’’ would already have occurred in the brands play when consumers actually consume the product:
past. Although purchase feedback is at times hypothesized in offering a coffee, the common questions are ‘‘milk or
in the literature—especially in discussions of choice models sugar?’’ or perhaps ‘‘decaffeinated or regular?’’ but not the
(e.g., Seetharaman et al., 1999)—only limited empirical choice of brand ‘‘Maxwell House or Nescafé?’’ Neverthe-
support has been cited for this in steady frequent-purchase less, consumers might come to identify with their habitual
markets; and even then, this may have been due to some brands (‘‘I use it, therefore I like it’’—see Barnard and
overlooked nonstationarity rather than real feedback (Shoe- Ehrenberg, 2000).
maker et al., 1977). The notion that a consumer’s past Despite this lack of segmentation between brands, there
purchasing behavior or current ‘‘state dependence’’ will can be strong segmentation at the category or subcategory
invariably change the consumer’s brand choice propensities level. Here, consumer habits, needs, and customer types
seems to us to put the cart before the horse. Thus, the often differ (e.g., Day et al., 1979; Ehrenberg, 1959, 1988;
experienced consumers’ steady purchase propensities can be Bock and Uncles, 2002). For example, larger households are
thought as the outcome of years of past experience. This is usually heavier buyers of the product as a whole. In
the underlying supposition that leads to the Dirichlet model. addition, customers of functionally distinct subcategories
Precisely when a consumer buys the product—e.g., differ: cat food buyers have cats and dog food buyers have
whether this week or next—and which particular repertoire dogs; users of leaded gasoline had older cars; and presweet-
brand is then chosen are both assumed in the model to be ened cereals are eaten more by children. But such subpat-
as-if-random. This is not to say consumers literally toss terns then tend to hold about equally for all the substitutable
mental pennies. Instead, specific choices tend to be gov- brands in that category or subcategory.
erned by a variety of reasons, motives, and feelings—such
as out-of-stock situations, promotions, special displays, and 5.2. Understanding brands
consumers’ own diverse habits, needs, moods, the mother-
in-law coming, etc. Details are extensively discussed in The Dirichlet model is a theory about choice between
standard consumer behavior texts and papers (e.g., Engle competitive entities such as brands. Such brands can be near
 A.S.C. Ehrenberg et al. / Journal of Business Research 57 (2004) 1307–1325 1317

identical in the theory itself—no differentiating attributes But in the Dirichlet, the different measures are all predict-
need to be specified except for the brand names and market able from each other (i.e., they measure the same thing).
shares. This is often so in practice since sales-effective Furthermore, each measure tends to be similar for compet-
product advantages and innovations are usually soon copied ing brands (apart from the market-share-based DJ phenom-
(e.g., Ehrenberg et al., 1997a; Foxall, 1999). Even early enon, which is usually small). Loyalty is therefore a given:
mover advantages tend not to last (Chiang and Robinson, changes in sales do not—or hardly—arise from changes in
1997; Roquebert et al., 1996; Szymanski et al., 1995). It loyalty but from changes in the sheer number of loyal
seems that sustainable product differentiation of a major kind customers.
seldom occurs between competitive brands. This Dirichlet view of brand loyalty is out of kilter with the
Functional differentiation does however exist but usually widely held view that ‘‘brand equity’’ is an idiosyncratic
within brands, as SKU-related product variants: different property of an individual brand (e.g., Aaker, 1996). While
pack sizes and flavors of soft drinks; shampoos for oily, dry, there clearly are big brands and smaller ones, there is no
and even normal hair; interest rates varying by withdrawal evidence—from a consumer behavior viewpoint—that over
notice; cars having 2, 3, 4, or 5 doors; PCs with a great and above this there are ‘‘strong’’ brands and ‘‘weak’’ ones
variety of technical specifications and software add-ons; and (Ehrenberg, 1997; Feldwick, 1996; Goodhardt, 1999). Nor is
so on. As we have noted, consumers spend more time, there evidence that brands that supposedly have high con-
effort, and money choosing between functional specifica- sumer-based equity do subsequently get bigger (Ehrenberg,
tions than between brands as such (Heath, 1999; Moran, 1993a). This also appears to explain the quite limited ability
1990). But competitive brands in a category tend to have of loyalty programs to increase brand loyalty dramatically
much the same range of product variants. In this sense, (Dowling and Uncles, 1997; Sharp and Sharp, 1997, 1999;
nearly all brands are umbrella brands and similar to each Uncles et al., 2003).
other as brands. The nonpartitioned Dirichlet model can An apparent paradox is that in Table 2, the annual
then apply directly, and in practice it usually does so, as we SCRs of a brand like Folgers was only about 50% and
have seen. even lower in Table 3. Folgers’ customers could therefore
In many product categories, some functionally distinct have bought the brand up to twice as often without having
submarkets or partitions have been deliberately created (e.g., had to drink more coffee. They would just have had to buy
decaffeinated and regular coffee, luxury cars, unleaded and more Folgers and less of the other similar brands—insofar
leaded gasoline). In most instances, direct competitors will as competing brands are perceived to be similar. But both
then each launch product variants into these submarkets. in theory and in practice, this does not seem to happen.
Sometimes these product variants are given distinct brand Instead, the normal split loyalty patterns tend to persist.
names (possibly with the same house name such as Kel- The reason for consumers’ persistent polygamy may be
logg’s or Heinz for all of the variants, although not some basic desire for ongoing variety. It could also be that
always—for Unilever and P&G brands, consumers in most having initially tried out many different brands, consumers
countries will hardly know from which company the brands then simply stay with some of them (e.g., Charlton and
and variants come). Ehrenberg, 1976).
Nonetheless, what we call ‘‘minor differences’’ between Theorists may feel that there is a lack of well-grounded
competing brands occur in practice, such as different bottle reasons for consumers having steady repertoires of similar
tops, car door handles, or degrees of sweetness for mueslis. brands. But there is equally a lack of well-grounded
Many of these differences are not featured ‘‘on-pack’’ nor reasons for consumers sticking with just one of several
advertised, and some may appear almost meaningless (Car- near lookalikes. Our inference is this: staying with a brand
penter et al., 1994). However, such minor and nonpublicized repertoire requires less mental effort than making new
differences may be noticed after consumers have started choices, but having a repertoire enables consumers to
using the brand and may lead to longer-term brand prefer- exercise some choice without having to reevaluate all
ences for otherwise similar brands. brands on all their criteria or without somehow weighing
up the expected ‘‘utilities’’ at each new purchase, as is
5.2.1. Brand loyalty supposed in ‘‘rational economics’’ for example (see also
Although competitive brands tend to be similar, any Thaler, 1994).
consumer generally shows loyalty to certain specific ones.
This is reflected in the Dirichlet model where consumers are 5.3. The role of marketing-mix factors
deemed ‘‘loyal’’ to the brand if they go on buying it. This is
so even if the brand’s customers routinely buy other brands A major lesson of the Dirichlet-type findings is that
as well (‘‘split loyalty’’) and if the frequency of buying that brands often differ little in their loyalty-related measures
brand is low (i.e., the customer is deemed ‘‘loyal’’ but buys but vary greatly in their penetrations. This means that
infrequently). varying marketing-mix inputs such as changes in price,
Loyalty has over the years been measured in many product formulation, selling, and distribution can have little
different ways (for a recent review, see Odin et al., 2001). if any impact on increasing loyalty, but they may affect the
1318 A.S.C. Ehrenberg et al. / Journal of Business Research 57 (2004) 1307–1325

brand’s penetration, and in particular its market share and namely, that under near-steady state conditions, traditional
sales volume. loyalty-related measures for competitive brands are unaf-
The observed outcomes of price-related promotions fected by any marketing-mix inputs, other than indirectly
support this view, as noted earlier. Large short-term sales through changes in the brand’s market share.
blips (often 100% up or more) show that promotions While the Dirichlet model itself is explicitly defined for
have a dramatically ability to attract extra buyers. But the steady-state and nonpartitioned zero-order markets, it in no
blip stems almost entirely from past customers—these way implies or predicts that all markets should be near
only need to change their quasi-random purchase timing stationary and nonpartitioned, as we have already stressed.
of the promoted brand rather than more fundamentally The model merely describes what markets are like when they
switch to an unfamiliar brand (Ehrenberg et al., 1994a). are more or less steady and nonpartitioned with no ‘‘purchase
Hence, it need not be surprising that promotional sales feedback.’’ But since many markets or submarkets are like
blips last only while ‘‘giving the product away’’ and do that most of the time, it behooves the management to try and
not affect repeat buying loyalty thereafter (Abraham and understand such steady markets and the factors that deter-
Lodish, 1987; Jedidi et al., 1999). mine their sales. These factors are how many customers a
For brand advertising, the bulk of it occurs between brand has, how loyal they are, which competitive brands
competitive brands, between which there is usually little customers also buy, and how often they do so, together with
sustainable differentiation. Hence, advertising does not, we the Dirichlet type of constraints on all this.
believe, have to try to persuade consumers that Brand A An analogy for the prime importance of the near-steady
really differs from the similar B, or even that it is better than state in the Dirichlet view of markets lies in our use of
B. Consumers just have to choose one of the available automobiles. Dynamics matter: we have to start and stop.
options of the right type: ‘‘Any reasonable brand will do’’ But driving at more or less steady speeds accounts for most
(Heath, 1999). In practice, consumers seem then to find it of the distance covered. Similarly, near-steady state markets
convenient and reassuring to have already developed their generate the bulk of our sales and therefore they are likely to
habitual split-loyalty choice propensities. Advertising an account for most of our revenue.
established brand therefore must mainly publicize the brand,
mostly by reminding experienced consumers—‘‘Here I am,
remember me’’—often in highly creative ways for impact Acknowledgements
and memorability (Ehrenberg, 1974; Barnard and Ehren-
berg, 1997; Ehrenberg et al., 2002). We are greatly indebted to Helen Bloom for much valued
advice and also to Byron Sharp, John Bound, Neil Barnard,
5.4. Marketing management issues Kathy Hammond, John Scriven, Robert East, Tim Bock,
Rachel Kennedy, Cam Rungie, Ban Mittal, and anonymous
The ubiquity of the predictable Dirichlet-type patterns of reviewers for helpful comments.
buying, together with most markets being more or less
steady most of the time, could add to existing doubts about
the role of marketing. For example, are marketing inputs Appendix A. The Dirichlet theory
such as advertising worth it if they do not soon lead to extra
sales? Our view is that there is nonetheless plenty of scope We discuss: (1) the theoretical assumptions of the Dirich-
for marketing (e.g., securing better distribution) since mar- let, (2) how the model is calibrated, (3) how the theoretical
ket shares differ greatly. However, such scope exists also for performance measures are estimated, and (4) alternative
competitors. The outcome is therefore usually competitive models of buyer behavior.
equilibrium rather than actual big gains or losses. This calls
for maintenance of sales and retention of split-loyalty A.1. The NBD-Dirichlet assumptions
customers (both consumer and trade). In this sense, compe-
tition means running hard to stand still, with profitable In the NBD-Dirichlet model (or ‘‘Dirichlet’’ for short),
survival being preferable to the most likely alternative. consumers are seen to have, for the time being, steady
Any dramatic and lasting gain is a rare bonus. personal purchase propensities—or stochastic probabili-
ties—for when they buy the product and what brands they
then choose. This involves five distributional assumptions
6. Conclusion (Goodhardt et al., 1984). Two assumptions are about buying
the product category:
To conclude, the Dirichlet model describes the widely
observed patterns of near-steady state buying behavior (i) A Gamma distribution for consumers’ differing average
that tend to occur. In this, the model consistently predicts purchase rates
fairly complex patterns and BPMs from very simple and Each consumer is assumed to buy the category at some
limited inputs. It also helps to explain the patterns; steady long-run rate (e.g., once a year or 10 times a
 A.S.C. Ehrenberg et al. / Journal of Business Research 57 (2004) 1307–1325 1319

year, i.e., with weekly probabilities of about 0.02 or their own fixed brand choice probabilities. This is
0.20). This reflects what the observed data tend to be the widely used zero-order ‘‘multinomial’’ distribu-
like when tabulated, say week-by-week or quarter-by- tion of brand choice (typically, consumers have
quarter. However, an individual’s purchase rate can ‘‘reasons’’ but behave as-if-random with their
change at times, perhaps with a change in lifestyle or in personal choice probabilities). Crucially, the proba-
the marketing mix. Consumer heterogeneity in these bility of buying Brand A on a particular occasion is
individual average-buying rates is assumed to follow a independent (‘‘zero order’’) of the brands the
smooth ‘‘Gamma’’ type of distribution (e.g., Ehrenberg, consumer has previously bought. Thus, a consumer
1959, 1988). Like the 80:20 rule of thumb, the Gamma with a three-brand portfolio and zero-order proba-
invariably means there are many light buyers and few bilities of 0.6, 0.3, and 0.1 will buy the three
heavy ones, unless the mean is very high. There are brands 60%, 30%, and 10% of the time in a random
strong theoretical reasons for the Gamma relating to the order.
observed near independence of buying the different An equivalent specification is used in choice models
brands (Goodhardt and Chatfield, 1973). (e.g., Luce, 1959; Ben-Akiva and Lerman, 1985;
(ii) Poisson distributions for individual purchases McFadden, 1986). The deterministic part of the
A given consumer’s specific purchases of the choice model might suggest that utility for Brand A
category are assumed to be spread irregularly (‘‘as- is higher than the utility for other brands, but the
if-randomly’’) over time with each consumer’s long- random error terms in the utilities can result in a
run probability (e.g., about 0.02 or 0.20) and independ- consumer choosing a brand other than A. Thus, if
ently of when the previous purchase was made (a ‘‘zero- the deterministic part of the utilities are in the ratio
order’’ process). This specifies ‘‘Poisson’’ distributions of 0.6, 0.3, and 0.1 for a three-brand portfolio, a
for each consumer, as has been widely tested for brand consumer does not always choose the brand with a
purchasing (Bass et al., 1984; Chatfield and Goodhardt, utility of 0.6. In aggregate, the three brands are
1973; Dunn et al., 1983; Schmittlein et al., 1985). The bought 60%, 30%, and 10% of the time.
zero-order assumption makes simple sense with the
typically low weekly purchase probabilities: even for The fifth Dirichlet assumption is about the relationship
someone buying 10 times a year, the chance of buying between product category buying and brand choice:
in two successive weeks is only 0.04 and hence near
zero—no special assumption of ‘‘dead periods’’ be- (v) The independence of purchase incidence and brand
tween purchases has to be made. choice.
The Poisson and Gamma assumptions combine to The Beta distributions of brand choice probabilities
give the negative binomial distribution or ‘‘NBD’’ are taken to be the same irrespective of how often
model (Ehrenberg, 1959). The NBD has been used particular consumers buy. This is in line with observed
widely in repeat-buying studies where each brand is market shares being much the same for light, medium,
considered on its own. In the multibrand Dirichlet and heavier category buyers.
model, the NBD is used for the category distribution
of purchases. In summary, the Dirichlet model specifies how many
purchases each household (or individual consumer) makes
The third and fourth Dirichlet assumptions are about of each of the available (or specified) g brands in a
brand choice: chosen period of length T (not the T for the theoretical
Dirichlet predictions). The model represents this as a g-
(iii) A multivariate Beta distribution of brand choice variate discrete random variable with a joint frequency
probabilities. distribution given by the mixture of the Multinomial,
Heterogeneity in consumers’ brand choice probabilities Dirichlet, Poisson, and Gamma distributions (e.g., Ehren-
is assumed to follow a smooth Beta distribution of a berg, 1988, p.258; Rungie, 1999).
multivariate ‘‘Dirichlet’’ type (after a French-named Having specified the model by these distributional
German mathematician). Strong theoretical backing for assumptions, the parameters of the model must next be
this assumption derives from the observed near numerically calibrated. Then any brand’s theoretical perfor-
independence of brand choice in line with the tradi- mance measures have to be estimated. This we now describe
tional theoretical notion of ‘‘independence from irre- in two stages.
levant alternatives’’ or IIA (Goodhardt et al., 1984;
Leeflang et al., 2000; Luce, 1959; Morrison and A.2. Calibrating the model
Schmittlein, 1988).
(iv) Multinomial distributions for specific purchases To calibrate the Dirichlet model for a given product or
On any one category purchase occasion, a consumer service category, four observed input measures are used, two
is assumed to choose a brand as if randomly with penetrations and two purchase rates, as is shown in Table A1.
1320 A.S.C. Ehrenberg et al. / Journal of Business Research 57 (2004) 1307–1325

Table A1
The inputs for calibrating the Dirichlet parameters (as required for Tables 1, 2, and 5)
Instant coffee Market Percent Purchases Percent Category 100% loyal Percentage who also bought:
(USA 1992) share buying per buyer buying purchases
in year 5+ times per buyer
Folgers Nescafé Maxim
O T O T O T O T O T O T O T O T
Category 31 5.0

Folgers 24a 11 3.2

Maxwell House 22
Tasters Choice 17
Nescafé 11
Sanka 9
High Point 1
Maxim 1
Brim 0.3

Average Brand
a
A brand’s market share (or equivalent) is needed to estimate its theoretical performance measures.

For some chosen ‘‘base period’’ (which will be greater are aggregated in the model to give a theoretical estimate
than the average interpurchase interval), such as a year, the or ‘‘prediction’’ of any chosen performance measure, such
inputs are as follows: as a particular brand’s penetration or its average purchase
frequency.
 B, the category penetration (i.e., households who bought The algebra for the model’s theoretical estimate of a
the category of instant coffee at least once in the year, chosen performance measure is again complex, as is illus-
31%). trated in Table A2. In practice, one can use computer
 bF , the penetration of one particular brand, Folgers say software (contained in the same packages as for the model
(11%). calibration) to give numerical estimates such as those in
 W, the average frequency of buying the category per Tables 1 and 5 earlier. It is also possible to simulate
category buyer (5.0). individual Dirichlet-type consumers and then tabulate any
 wF , the average frequency of buying Folgers (3.2). required measure, as one would with observed data (Good-
hardt, 1995). This can be useful for theoretical measures,
(We note that penetrations are usually expressed as propor- which are not covered by the available estimation software.
tions algebraically, but as percent numerically). A method based on using the Juster scale to replace panel
If the model fits the observed data exactly, any brand data has been described too (Wright et al., 2002).
could be used in this calibration process and would give
the same results. In practice, there are deviations. It is then A.4. Simplifying approximations
usual to fit the model for each of a number of leading
brands and to average or smooth the results (in the Alternatively, simple verbal descriptions can be used to
computer program). The selection of brands to use in such communicate what the theoretical estimators say—like
model calibration can be deliberate. In particular, any those in Section 2 for the observed patterns in the data.
unusual brand can be left out of the calibration process Thus, for ‘‘DJ,’’ we said that far fewer people buy a small
(e.g., the unusual Maxim in Table 1 or brands with ab- brand and buy it somewhat less often (this is like the
normal availability like a private label or regional brand). wording commonly used to describe the normal distribution
But their estimated performance measures would still be
checked subsequently. The mathematics of this calibration
Table A2
process are fairly complex, but software is available (e.g., The Dirichlet’s theoretical formulae for the penetration
Uncles, 1989; Hewitt, 1990; Kearns, 2000; Rungie et al.,
An illustration
2003).
The Dirichlet theoretical formulae for the penetration of Brand X with
market share zx is indirect. It requires first calculating SCn, i.e., how
A.3. Estimating the theoretical BPMs many consumers do not buy X but do buy the category any number of
times. Here:
By calibrating the model, each possible purchase made K þ n  1 ½Sð1  zx Þ þ n  1 A
Cn ¼ Cn1
by any consumer is implicitly specified as a ‘‘stochastic’’ or n ðS þ n  1Þ 1þA
probabilistic representation. These purchase probabilities where K, A, and S are the parameters of the fitted Dirichlet model.
 A.S.C. Ehrenberg et al. / Journal of Business Research 57 (2004) 1307–1325 1321

in statistics, where we say ‘‘5% lie beyond F 2 standard Table A4

The duplication of purchase law: an illustration
deviations’’ rather than spell out the pdf).
Furthermore, several close quantitative approximations Instant coffee Percentage who also bought: Average
to the complex Dirichlet-type estimators are also available. (USA, 1992) Maxwell Nescafé Sanka Maxim Brim (8 brands)
These are much simpler than the exact Dirichlet expres- House
sions and therefore more insightful. The two most used Folgers buyers % 31 21 12 0 1 13.9
simplifications are: (i) the ‘‘DJ’’ relation wx(1  bx)Swo, a 2.8 penetrationa % 28 17 14 1 1 13.9
Exact Dirichlet % 27 15 13 1 0 12.5
constant for that set of brands and chosen time period (this
estimate
relates Brand X’s average purchase frequency wx directly
to its penetration bx), and (ii) the ‘‘Duplication Law’’ Penetration % 10 6 5 0.3 0.2 4.9
relation bx/ySDbx (this relates the proportion bx/y of buyers Estimated duplications of purchase for Folgers Instant Coffee as in Table
of Brand Y who also buy Brand X in the analysis period, 2 = D
Penetration.
a
to X’s overall penetration bx, and where D is defined D is estimated as (average duplication)/(average penetration) = 13.9/
below). 4.9 = 2.8.

A.4.1. The ‘‘w(1b)=constant’’ approximation simply proportional to bx, Brand X’s penetration in the
This says that w varies with 1/(1  b). Thus, the whole population (e.g., Ehrenberg and Goodhardt,
smaller the b (as a proportion), the smaller is 1/(1  b) 1968b, 1970; Ehrenberg, 1988; Goodhardt, 1966). This
and hence the smaller is w. This is McPhee’s DJ notion approximate relationship, or ‘‘duplication of purchase
that was noted earlier (but is not readily apparent from the law,’’ is illustrated in Table A4 for the duplication of
exact but complex Dirichlet-type algebra such as in Table purchase of Folgers with the other instant coffee brands.
A2 where wx and bx/y would depend on all the observa- The duplication D is very simple to estimate as the
tions for all the brands and not just on bx) (Ehrenberg, average of the observed duplications for all pairs of
1988). brands (leaving out any extreme outlier) divided by their
The approximate formula predicts well, as illustrated in average penetration.
the final column of Table A3. For brands with fairly low Expressing brand penetration in ‘‘relative’’ terms as the
penetration, the average purchase rates hardly differ (e.g., percentage of category buyers who buy the brand (i.e., b/B)
for bs of 10% or less as in Table A3, the ws are all about 3). probably leads to further simplifications. Thus, the duplica-
Only for brands with high penetration are the differences in tion coefficients D would tend to be close to 1 (implying
average purchase frequencies quite large. Relatively high independent brand choice, a very simple result). But the
penetrations can occur both for large brands, and for small base B then varies with the length of the analysis period.
ones over long time-periods. Remarkably, the w(1  b) More work is needed.
formula copes with both cases.
A.5. Elaborating the Dirichlet model
A.4.2. The ‘‘bx/y=Dbx’’ approximation
This says that bx/y , the percentage of buyers of Brand The Dirichlet and related stochastic approaches to mod-
Y who also buy X in the chosen analysis-period, is eling buyer behavior have been elaborated in two main
ways. First, case-specific ‘‘mass points’’ have been used to
Table A3 model consumer heterogeneity rather than the Dirichlet’s
Exact and approximate estimates of the average purchase frequencies smooth Gamma and Beta distributions (e.g., Colombo and
Instant coffee Market Percent Average purchases Morrison, 1989; Reader and Uncles, 1988; Fader and
(USA 1992) share buying (per buyer) Hardie, 1996). In a one-off study of saltine crackers, a
O T O T A semiparametric random effects model was found to capture
Folgers 23 11 12 3.2 3.1 3.2 heterogeneity in brand preferences across households, but
Maxwell House 22 10 11 3.1 3.1 3.1 the predictive gains were slight and came at the expense of
Tasters Choice 17 9 9 2.8 3.0 3.1 added complexity (Chintagunta et al., 1991). It appears that
Nescafé 11 6 6 2.7 2.9 3.0
Sanka 9 5 5 3.0 2.8 2.9
with these alternative models, there is some scope to
High Point 1 1 0 2.6 2.6 2.8 improve goodness of fit, but only by forsaking parsimony
Maxim 0.9 0.3a 0.2 4.5a 2.6 2.8 and generalizability. Nor so far have these alternatives
Brim 0.3 0.2 0.5 2.1 2.6 2.8 reported the same range of BPMs that can be derived
routinely from the Dirichlet.
Other Brands 16 8 8 3.0 3.0 3.0
Second, consumer heterogeneity has been related to
Average brand 11 6 6 3.0 2.9 3.0 various possible causal sources such as sociodemographic
O = Observed as in Table 1; T = theoretical Dirichlet predictions; A = wo/
and attitudinal factors (Allenby and Lenk, 1994; Bass, 1993;
(1  b). Bhattacharya et al., 1996; Ehrenberg, 1959; Fader, 1993;
a
Outlier. Fader and Lattin, 1993; Jones and Zufryden, 1980; Russell
1322 A.S.C. Ehrenberg et al. / Journal of Business Research 57 (2004) 1307–1325

and Kamakura, 1994; Vilcassim and Jain, 1991; Wrigley changes in buying rates or loyalty)? Do sales decreases
and Dunn, 1985). A variety of factors have proven to be simply show the reverse pattern from sales increases or are
statistically significant in isolated studies, but systematic they more complex? Do increased sales for Brand X come
effects and generalizations have eluded analysts. Moreover, from competitive brands pro rata to their size (as in the
few major gains in predictive power have been either steady-state duplication of purchase law), or do gains and
reported or even claimed for these more elaborate model losses operate quite differently from that? These are empir-
specifications. ical questions for which it seems to us useful to know
whether some answers generalize before theorizing exten-
A.6. Other modeling approaches sively about them.

The Dirichlet model, and its underlying descriptive and A.6.1. Other structural models
benchmarking uses, differs from many other modeling The Dirichlet model is purely structural for steady-state
approaches. The alternative stochastic and/or econometric markets, as we have stressed, with no explicit ‘‘explanato-
response models for buyer behavior in the literature are ry’’ variables specified (e.g., advertising or price changes).
mostly dynamic. Many aim to predict changes in market An alternative structural model is the first-order Markov
shares from changing marketing-mix inputs (e.g., advertis- process. This was popular in the 1960s and 1970s (e.g.,
ing or price). In this, they are generally attempting to model Kuehn, 1962; Massy et al., 1970) and has occasionally been
nonstationary market conditions (e.g., Guadagni and Little, mentioned ever since (e.g., Bronnenberg, 1998; Leeflang et
1983; Leeflang et al., 2000; Lenk et al., 1993; Kannan and al., 2000). In its simplest form, the Markov approach also
Yim, 2001; Vilcassim and Jain, 1991; Wagner and Taudes, has no ‘‘causal’’ inputs but assumes (i) homogeneous
1986; Yim and Kannan, 1999). consumers and (ii) fixed switching and repeat-purchase
The parameters of such econometric-type models are probabilities for each brand. These assumptions run counter
mostly respecified on each occasion. Some of the models to the Dirichlet-type empirical findings (and to much of the
assume turbulent (i.e., continually changing) choice proba- response modeling literature more generally). It is now
bilities for each consumer (e.g., Erdem, 1996; Erdem and known that (i) consumers are heterogeneous and (ii) that
Keane, 1996). Such models usually require many—possibly switching and repeat buying are independent of the brands
hundreds—of parameters to be estimated and interpreted, as such but vary with market shares. It is seldom that a
with conceptual and at times computational complexities modeling theory and the facts clash so starkly.
(e.g., multicollinearity), which seem largely unresolved. Another purely structural model is the Hendry system
Specific results have been reported but with few generaliz- (Butler, 1966) and the derivatives of it (e.g., Kannan and
able findings and insights. Sanchez, 1994). This uses the Duplication of Purchase Law
Economists and psychologists such as Kahneman and formulation for pairs of purchases. But it makes very
Tversky have carried out very innovative research into different assumptions from the Dirichlet and has different
broader issues of consumer choice, often along experi- outcomes (e.g., Ehrenberg and Goodhardt, 1974). More
mental and/or mathematical lines and frequently with generally, ‘‘pairs of purchases’’ can however be used very
roots in game theory (e.g., Kagel and Roth, 1995; constructively in the absence of continuous panel data (e.g.,
Thaler, 1994). But they seldom touch on the issues of Bennett, 2002; Colombo et al., 2000; Ehrenberg and Bound,
brand choice in highly competitive markets, covered by 2000).
standard scanner panel tabulations and the Dirichlet
approach.
Nearly all such modeling appears to be aimed at disco- References
vering and predicting effects that have not yet been ob-
Aaker DA. Building strong brands. New York: Free Press; 1996.
served. Indeed, these other approaches tend not to mention
Abraham MM, Lodish LM. Promoter: an automated promotion evaluation
any of the ubiquitous empirical patterns reported in Section system. Mark Sci 1987;6(2):101 – 23 [Spring].
2 or their theoretical modeling as described in Section 3 Allenby GM, Lenk PJ. Modeling household purchase behavior with logistic
(see for example reviews by Bucklin and Gupta, 1999; nominal regression. J Am Stat Assoc 1994;98:1218 – 31.
Hanssens et al., 1990; Leeflang and Wittink, 2000; Lilien et Baldinger AL, Blair E, Echambadi R. Why brands grow. J Advert Res
al., 1992). Even the early theoretical derivation of the 2002;42(1):6 – 14 [January/February].
Barnard NR, Ehrenberg ASC. Advertising: strongly persuasive or nudging?
Dirichlet model by Bass et al. (1976) did not refer to any J Advert Res 1997;37(1):21 – 30 [January/February].
empirical patterns. In contrast, the Dirichlet-type approach Barnard NR, Ehrenberg ASC. Consumer attitudes and brand choice. R&DI
first establishes what kinds of empirical changes have in Report. South Bank University; 2000.
fact taken place and then tries to model that knowledge Barwise PT, Ehrenberg ASC. Consumer beliefs and brand usage. J Mark
(Ehrenberg, 1993b). Res Soc 1985;27:81 – 93.
Barwise PT, Ehrenberg ASC. Television and its audience. London: Sage;
For example, do any sales increases for a given brand 1998.
follow the steady-state DJ pattern for large and small brands Bass FM. The future of research in marketing: marketing science. J Mark
(implying mainly penetration growth rather than separate Res 1993;30(1):1 – 6 [February].
 A.S.C. Ehrenberg et al. / Journal of Business Research 57 (2004) 1307–1325 1323

Bass FM, Jeuland AP, Wright GP. Equilibrium stochastic choice and market NBD model as applied to purchasing at individual stores. J R Stat Soc
penetration theories: derivation and comparisons. Manage Sci 1976;22: Ser C Appl Stat 1983;32(3):249 – 59.
1051 – 63 [June]. East R. Consumer behavior: advances and applications in marketing. Lon-
Bass FM, Givon M, Kalwani MU, Reibstein D, Wright GP. An investiga- don: Prentice-Hall; 1997.
tion into the order of the brand choice process. Mark Sci 1984;3 East R, Hammond K. The erosion of repeat purchase loyalty. Mark Lett
(4):267 – 87 [Fall]. 1996;7(2):163 – 71 [March].
Ben-Akiva M, Lerman SR. Discrete choice analysis. Cambridge (MA): Easton G. Stochastic models of industrial buying behaviour. Omega 1980;
MIT Press; 1985. 8:63 – 9.
Bennett D. Brand loyalty among petrol buyers. Working paper. Academy of Ehrenberg ASC. The pattern of consumer purchases. Appl Stat 1959;8:
Marketing Doctoral Colloquium, University of Derby, July 4; 2000. 26 – 41.
Bennett D. Using two-purchase data. R&DI Report. South Bank University; Ehrenberg ASC. Repetitive advertising and the consumer. J Advert Res
2002. 1974;14:25 – 34.
Bhattacharya CB. Is your brand’s loyalty too much, too little, or just right? Ehrenberg ASC. Repeat-buying: facts, theory and applications. 2nd ed.
Explaining deviations in loyalty from the Dirichlet norm. Int J Res London: Edward Arnold; New York: Oxford Univ. Press: 1988. Re-
Mark 1997;14(5):421 – 35. printed in the J Empir Generalisations Mark Sci 2000;5:392 – 770.
Bhattacharya CB, Fader PS, Lodish LM, DeSarbo W. The relationship Available: http://www.empgens.com.
between the marketing mix and share of category requirements. Mark Ehrenberg ASC. New brands and the existing market. J Mark Res Soc
Lett 1996;7(1):5 – 18. 1991;33(4):285 – 99 [October].
Bock T, Uncles MD. A taxonomy of differences between consumers for Ehrenberg ASC. If you’re so strong, why aren’t you bigger? Admap 1993a;
market segmentation. Int J Res Mark 2002;19(3):215 – 24. 13 – 4 [October].
Bound JA, Ehrenberg ASC. Private label purchasing. Admap 1997;375: Ehrenberg ASC. Theory or well-based results: which comes first? In: Lau-
17 – 9. rent G, Lilien G, Pras B, editors. Research traditions in marketing.
Brewis-Levie M, Harris P. An empirical analysis of buying behavior in UK Dordrecht (The Netherlands): Kluwer; 1993b. p. 79 – 108.
high street womenswear retailing using the Dirichlet model. Int Rev Ehrenberg ASC. B.E. or not B.E? Proceedings of Advertising Research
Retail Distrib Consum Res 1999;10(1):41 – 57. Federation 43rd Annual Conference, New York, April 7 – 9; 1997.
Bronnenberg BJ. Advertising frequency decisions in a discrete Markov Ehrenberg ASC, Bound JA. Turning data into knowledge. In: Chakrapani
process under a budget constraint. J Mark Res 1998;35:399 – 406 C, editor. Marketing research: state of the art perspectives (the Hand-
[August]. book of the American Marketing Association and the Professional Mar-
Bucklin R, Gupta S. Commercial use of UPC scanner data: industry and ket Research Society). Chicago (IL): American Marketing Association;
academic perspectives. Cambridge (MA): Marketing Science Institute; 2000. p. 23 – 46.
1999. Ehrenberg ASC, Goodhardt GJ. Repeat-buying of a new brand. Br J Mark
Butler BF. Hendrodynamics: fundamental laws of consumer dynamics. 1968a;2:200 – 5.
Croton-on-Hudson: Hendry Corporation; 1966. Ehrenberg ASC, Goodhardt GJ. The incidence of brand-switching. Nature
Carpenter GS, Glazer R, Nakamoto K. Meaningful brands from meaning- 1968b;220:304.
less differentiation: the dependence on irrelevant attributes. J Mark Res Ehrenberg ASC, Goodhardt GJ. A model of multi-brand buying. J Mark
1994;31(3):339 – 50. Res 1970;7:77 – 84.
Charlton P, Ehrenberg ASC. An experiment in brand-choice. J Mark Res Ehrenberg ASC, Goodhardt GJ. The Hendry brand-switching coefficient.
1976;13:52 – 60. Admap 1974;10:232 – 8.
Charlton P, Ehrenberg ASC, Pymont B. Buyer behavior under mini-test Ehrenberg ASC, Goodhardt GJ. New brands: near-instant loyalty. J Mark
conditions. J Mark Res Soc 1972;14:171 – 83. Manage 2000;16(6):607 – 17 [Revised version] J Target Meas Anal
Chatfield C. Some statistical models for buying behavior. London Univer- Mark 2001;10(1):9 – 16.
sity: Doctoral Thesis; 1967. Ehrenberg ASC, Goodhardt GJ, Barwise PT. Double jeopardy revisited. J
Chatfield C, Goodhardt GJ. A consumer purchasing model with Erlang Mark 1990;54:82 – 91.
inter-purchase times. J Am Stat Assoc 1973;68:828 – 35. Ehrenberg ASC, Hammond K, Goodhardt GJ. The after-effects of price-
Chatfield C, Goodhardt GJ. Results concerning brand-choice. J Mark Res related consumer promotions. J Advert Res 1994a;34 (4):11 – 21 [July/
1975;12(1):110 – 3. August].
Chiang J, Robinson WT. Do market pioneers maintain their innovative Ehrenberg ASC, Uncles MD, Carrie D. Armed to the teeth: an exercise
spark over time? Working paper. Krannert School, Purdue; 1997. in brand management. Cranfield (UK): European Case Clearing
Chintagunta PK, Jain DC, Vilcassim NJ. Investigating heterogeneity in House; 1994b [reference M94-005:594-039-1/594-039-4/594-040-1/
brand preferences in logit models for panel data. J Mark Res 1991; 594-040-4].
28:417 – 28 [November]. Ehrenberg ASC, Barnard NR, Scriven JA. Differentiation or salience.
Collins MA. Market segmentation—the realities of buyer behavior. J Mark J Advert Res 1997a;37(6):7 – 14 [November/December].
Res Soc 1971;13:146 – 57. Ehrenberg ASC, Scriven J, Barnard N. Advertising and price. J Advert Res
Colombo RA, Morrison DG. A brand-switching model with implications 1997b;37(3):27 – 35 [May – June].
for marketing strategies. Mark Sci 1989;8:89 – 99. Ehrenberg ASC, Barnard NR, Sharp B. Decision models or descriptive
Colombo RA, Ehrenberg ASC, Darius S. Diversity in analysing brand- models. Int J Res Mark 2000;17(2 – 3):147 – 58.
switching tables: the car challenge. Can J Mark Res 2000;19:23 – 36. Ehrenberg ASC, Barnard N, Kennedy R, Bloom H. Brand advertising as
Dall’Olmo Riley F, Ehrenberg ASC, Castleberry SB, Barwise PT, Barnard creative publicity. J Advert Res 2002;42(6):7 – 18 [July/August].
NR. The variability of attitudinal repeat-rates. Int J Res Mark 1997;14 Engle JF, Blackwell RD, Miniard PW. Consumer behavior 8th edition.
(5):437 – 50. Orlando (FL): Dryden Press; 1995.
Danaher PJ, Wilson IW, Davis R. A comparison of online and offline Erdem T. A dynamic analysis of market structure based on panel data. Mark
consumer brand loyalty. Mark Sci [in press]. Sci 1996;15(4):359 – 78.
Day GS, Shocker AD, Srivastava RK. Customer-oriented approaches to Erdem T, Keane MP. Decision-making under uncertainty: capturing dynam-
identifying product-markets. J Mark 1979;43:8 – 19 [Fall]. ic brand choice processes in turbulent consumer goods markets. Mark
Dowling GR, Uncles MD. Do customer loyalty programs really work? Sci 1996;15(1):1 – 20.
Sloan Manage Rev 1997;38(4):71 – 82 [Summer]. Fader PS. Integrating the Dirichlet-multinomial and multinomial logit mo-
Dunn R, Reader S, Wrigley N. An investigation of the assumptions of the dels of brand choice. Mark Lett 1993;4:99 – 112 [April].
1324 A.S.C. Ehrenberg et al. / Journal of Business Research 57 (2004) 1307–1325

Fader PS, Hardie BG. Modeling consumer choice among SKUs. J Mark Kannan PK, Sanchez S. Competitive market structures: a subset selection
Res 1996;33:442 – 52 [November]. analysis. Manage Sci 1994;40:1484 – 99 [November].
Fader PS, Lattin JM. Accounting for heterogeneity and non-stationarity in a Kannan PK, Yim CK. An investigation of the impact of promotions on
cross-section model of consumer purchase behavior. Mark Sci 1993;12: across-submarket competition. J Bus Res 2001;53(3):137 – 49.
304 – 17. Kau KA, Ehrenberg ASC. Patterns of store-choice. J Mark Res 1984;21:
Fader PS, Schmittlein DC. Excess behavioral loyalty for high-share brands: 399 – 409.
deviations from the Dirichlet model for repeat-buying. J Mark Res Kau KA, Uncles MD, Ehrenberg ASC, Barnard NR. Competitive brand-
1993;30:478 – 93 [November]. choice and store-choice among Japanese consumers. J Prod Brand Man-
Feldwick P. Do we really need ‘brand equity’? J Brand Manag 1996;4(1): ag 1998;7(6):481 – 94.
9 – 28. Kearns Z. Dirichlet no solver software. Working paper. Massey University,
Fournier S, Yao JL. Reviving brand loyalty: a reconceptualization within New Zealand; 2000.
the framework of consumer-brand relationships. Int J Res Mark 1997; Kennedy R, Ehrenberg ASC. Brand user profiles seldom differ. R&DI
14(5):451 – 72. Report. South Bank University; 2000.
Foxall GR. The substitutability of brands. Manage Decis Econ 1999;20: Kuehn A. Consumer brand choice as a learning process. J Advert Res
241 – 57. 1962;1(2):10 – 7.
Foxall GR, Goldsmith RE. Consumer psychology for marketing. London: Leeflang PSH, Wittink DR. Building models for marketing decisions: past,
Routledge; 1994. present and future. Int J Res Mark 2000;17(2 – 3):105 – 26.
Franzen G. Advertising effectiveness. Henley-upon-Thames (UK): NTC Leeflang PSH, Wittink DR, Wedel M, Naert P. Building models for market-
Publications; 1994. ing decisions. Dordrecht (The Netherlands): Kluwer; 2000.
Goodhardt GJ. The constant in duplicated television viewing. Nature Lenk PJ, Rao AG, Tibrewala V. Nonstationary conditional trend analysis:
1966;212:1616. an application to scanner panel data. J Mark Res 1993;30 (3):288 – 304
Goodhardt GJ. Simulated Dirichlet customers: a software specification, [August].
R&D I Report. South Bank University; 1995. Lilien GL, Kotler P, Moorthy KS. Marketing models. Englewood Cliffs
Goodhardt GJ. Strong and weak brands. Int J Advert 1999;18:190. (NJ): Prentice-Hall; 1992.
Goodhardt GJ, Chatfield C. Gamma distribution in consumer purchasing. Lomax W, Hammond K, East R, Clemente M. The measurement of canni-
Nature 1973;244:316. balization. Mark Intell Plann 1996;14(7):20 – 8 [Reprinted in] J Prod
Goodhardt GJ, Ehrenberg ASC. Conditional trend analysis: a breakdown by Brand Manag 1997;6(1):27 – 39.
initial purchasing level. J Mark Res 1967;4:155 – 61. Long S, Ehrenberg ASC. The PC market’s retention and switching. R&DI
Goodhardt GJ, Ehrenberg ASC, Chatfield C. The Dirichlet: a comprehen- Report. South Bank University; 1998.
sive model of buying behavior. J R Stat Soc A 1984;147:621 – 55. Luce RD. Individual choice behavior: a theoretical analysis. New York:
Goodhardt GJ, Ehrenberg ASC, Collins M. The television audience. 2nd Wiley; 1959.
ed. Aldershot (UK): Gower; 1987. Massy WF, Montgomery DB, Morrison DG. Stochastic models of buyer
Gordon W. Taking brand repertoires seriously. J Brand Manage 1994;2(1): behavior. Boston (MA): MIT Press; 1970.
25 – 30. McAlister L, Pessemier E. Variety seeking behavior: an interdisciplinary
Guadagni PM, Little JD. A logit model of brand choice calibrated on review. J Consum Res 1982;9:311 – 22 [December].
scanner data. Mark Sci 1983;2(3):203 – 38 [Summer]. McDonald C, Ehrenberg ASC. Personal buying is like family shopping.
Hallberg G. Disaggregating repeat buying patterns. Presentation to GVR R&DI Report. South Bank University; 2002.
using MRCA data. Amsterdam: October, 1996 [Supplementary tabula- McFadden D. The choice theory approach to market research. Mark Sci
tions with David E. Learner]. 1986;5:275 – 97.
Hammond K. Brand loyalty for frequently-bought goods. London Business McPhee WN. Formal theories of mass behavior. Glencoe (NY): Free Press;
School, London University; 1997. 1963.
Hammond K, Ehrenberg ASC, Goodhardt GJ. Market segmentation for Moe WW, Fader PS. Uncovering patterns in cybershopping. Calif Manage
competitive brands. Eur J Mark 1996;30(12):39 – 49. Rev 2001;43:106 – 17 [Summer].
Hanssens DM, Parsons LJ, Schultz RL. Market response models: econo- Moran WT. Brand presence and the perceptual frame. J Advert Res 1990;
metric and time series analysis. Boston (MA): Kluwer; 1990. 30:9 – 16 [October].
Hardie B, Fader PS, Wisniewski M. An empirical comparison of new product Morrison DG. Conditional trend analysis: a model that allows for nonusers.
trial forecasting models. J Forecast 1998;17:209 – 29 [June – July]. J Mark Res 1969;6:342 – 6 [August].
Harris J. An investigation of brand choice in repeat purchase markets: the Morrison DG, Schmittlein DC. Generalizing the NBD model for customer
case of business airline travel. Doctoral Thesis. School of Marketing, purchases: what are the implications and is it worth the effort? J Bus
University of New South Wales; 2003. Econ Stat 1988;6:145 – 59 [April].
Heath R. The low-involvement processing theory. Admap 1999;34(3):14 – 7 Odin Y, Odin N, Valette-Florence P. Conceptual and operational aspects of
[March]. brand loyalty: an empirical investigation. J Bus Res 2001;53(2):75 – 84.
Hewitt I. DIPRED: a programme for predicting buyer behavior from the Olshavsky RW, Granbois DH. Consumer decision making-fact or fiction? J
Dirichlet model, Version 2.6. Port Sunlight (UK): Unilever Research Consum Res 1979;6:93 – 100 [September].
Laboratory; 1990. Pickford C, Goodhardt GJ. An empirical study of buying behaviour in an
Horowitz J, Louviere J. What is the role of consideration sets in choice industrial market. Proceedings of the Academy of Marketing Annual
modeling? Int J Res Mark 1995;12:39 – 54 [May]. Conference, University of Derby, UK, July; 2000.
Jedidi K, Mela CF, Gupta S. Managing advertising and promotion for long- Reader S, Uncles MD. The collection and analysis of consumer data. In:
run profitability. Mark Sci 1999;18(1):1 – 22. Uncles MD, editor. Longitudinal data analysis: methods and applica-
Jones MJ, Zufryden F. Adding explanatory variables to a consumer pur- tions. London: Pion; 1988. p. 45 – 57.
chase behavior model: an exploratory study. J Mark Res 1980;17: Reibstein DJ, Farris PW. Market share and distribution: a generalization, a
323 – 34. speculation, and some explanations. Mark Sci 1995;14(3,):G190 – 202
Kagel JH, Roth AE. Handbook of experimental economics. Princeton: [Part 2 of 2].
Princeton Univ Press; 1995. Roquebert JA, Phillips RL, Westfall P. Markets versus management: what
Kahn BE, Kalwani MU, Morrison DG. Niching versus change-of-pace ‘drives’ profitability? Strateg Manage J 1996;17:653 – 64 [October].
brands: using purchase frequencies and penetration rates to infer brand Rungie C. Heterogeneity in brand choice. Doctoral Thesis. University of
positionings. J Mark Res 1988;25:384 – 90 [November]. South Australia; 1999.
 A.S.C. Ehrenberg et al. / Journal of Business Research 57 (2004) 1307–1325 1325

Rungie C, Goodhardt G, Driesener C. Research note: brand performance Uncles MD. BUYER: buyer behavior software. London: London Business
measures from the parameters of the Dirichlet model. School of Market- School; 1989.
ing Working Paper, University of South Australia; 2003. Uncles MD, Ehrenberg ASC. Industrial buying behavior: aviation fuel
Russell GJ, Kamakura WA. Understanding brand competition using micro contracts. Int J Res Mark 1990a;7:57 – 68.
and macro scanner data. J Mark Res 1994;31:289 – 303 [May]. Uncles MD, Ehrenberg ASC. The buying of packaged goods at US retail
Schmittlein DC, Bemmaor AC, Morrison DG. Why does the NBD chains. J Retail 1990b;66(3):278 – 96.
model work? Robustness in representing product purchases, brand pur- Uncles MD, Ellis K. The buying of own labels. Eur J Mark 1989;23(3):
chases and imperfectly recorded purchases. Mark Sci 1985;4:255 – 66 57 – 70.
[Summer]. Uncles MD, Hammond K. Grocery store patronage. Int Rev Retail Distrib
Scriven JA. Patterns of response to price changes. Proceedings of ANZ- Consum Res 1995;5(3):287 – 302.
MAC Conference. University of New South Wales, Sydney, Australia; Uncles MD, Hammond K, Ehrenberg ASC, Davis RE. A replication study
1999. of two brand-loyalty measures. Eur J Oper Res 1994;76:375 – 84.
Scriven JA, Ehrenberg ASC. The structure of the UK petrol market. Uncles MD, Ehrenberg ASC, Hammond K. Patterns of buyer behavior:
Reports prepared for Shell International Petroleum and BP Oil Eu- regularities, models and extensions. Mark Sci 1995;14(3):G71 – 8
rope; 1994. [Part 2 of 2].
Scriven JA, Ehrenberg ASC. How consumers respond to price. Proceedings Uncles MD, Dowling G, Hammond K. Customer loyalty and customer
of the MRS Annual Conference. London: The Market Research Soci- loyalty programs. J Consum Mark 2003;20(4):294 – 316.
ety; 1995. Vilcassim NJ, Jain DC. Modeling purchase-timing and brand-switching
Seetharaman PB, Ainslie A, Chintagunta PK. Investigating household behavior: incorporating explanatory variables and unobserved hetero-
state dependence effects across categories. J Mark Res 1999;36: geneity. J Mark Res 1991;28:29 – 41 [February].
488 – 500. Wagner U, Taudes A. A multivariate polya model of brand choice and
Sharp B, Sharp A. Loyalty programs and their impact on repeat-purchase purchase incidence. Mark Sci 1986;5(3):219 – 44.
loyalty patterns. Int J Res Mark 1997;14(5):473 – 86. Wedel M, Kamakura W. Market segmentation: conceptual and methodo-
Sharp B, Sharp A. Loyalty programs and their impact on repeat-purchase logical foundations. 2nd ed. Dordrecht (The Netherlands): Kluwer;
loyalty patterns: a replication and extension. Proceedings of 28th 2000.
EMAC Conference, Berlin, 11 – 14 May; 1999. Wellan DM, Ehrenberg ASC. A successful new brand: Shield. J Mark Res
Sharp B, Wright M, Goodhardt GJ. Purchase loyalty is polarised into either Soc 1988;30(1):35 – 44 [January].
repertoire or subscription patterns. AMJ 2002;10(3):7 – 20. Wellan DM, Ehrenberg ASC. A case of seasonal segmentation. Mark Res
Shoemaker RW, Staelin R, Kadane JB, Shoaf FR. Relation of brand-choice 1990;1:11 – 3.
to purchase frequency. J Mark Res 1977;14:458 – 68. Wright M, Sharp A. New brand effects in a Dirichlet market. Proceedings
Singh J, Ehrenberg ASC, Goodhardt GJ. Stock-keeping units: examining of 28th EMAC Conference, Berlin, 11 – 14 May; 1999.
patterns of buyer behaviour. Proceedings of ANZMAC Conference. Wright M, Sharp A, Sharp B. Are Australasian brands different? J Prod
Gold Coast: Griffith University; 2000;1185 – 91. Brand Manag 1998;7(6):465 – 80.
Stern P. Prescriptions for branded and generic pharmaceuticals. J Brand Wright M, Sharp A, Sharp B. Market statistics for the Dirichlet model:
Manage 1995;2(3):177 – 83. using the Juster scale to replace panel data. Int J Res Mark 2002;19
Stern P, Ehrenberg ASC. The market performance of pharmaceutical (1):81 – 90.
brands. Mark Res Today. 1995;285 – 92 [November]. Wrigley N, Dunn R. Stochastic panel-data models of urban shopping be-
Sudman S, Wansink B. Consumer panels. 2nd ed. Chicago (IL): AMA; havior: 2. Multistore purchasing patterns and the Dirichlet model. En-
2002. viron Plann A 1984;16:759 – 78.
Szymanski D, Troy LC, Bharadwaj S. Order of entry and business perform- Wrigley N, Dunn R. Stochastic panel-data models of urban shopping be-
ance: an empirical synthesis and reexamination. J Mark 1995;59:17 – 33 havior: 4. Incorporating independent variables into the NBD and Di-
[October]. richlet models. Environ Plann A 1985;17:319 – 31.
Thaler R. Quasi rational economics. New York: Russell Sage Foundation; Yim CK, Kannan PK. Consumer behavioral loyalty: a segmentation model
1994. and analysis. J Bus Res 1999;44:75 – 92.