Professional Documents
Culture Documents
This manuscript has been reproduced from the microfilm master. UMI films
the text directly from the original or copy submitted. Thus, some thesis and
dissertation copies are in typewriter face, while others may be from any type of
computer printer.
The quality of this reproduction is dependent upon the quality of th e
copy submitted. Broken or indistinct print colored or poor quality illustrations
and photographs, print bleedthrough, substandard margins, and improper
alignment can adversely affect reproduction.
In the unlikely event that the author did not send UMI a complete manuscript
and there are missing pages, these will be noted. Also, if unauthorized
copyright material had to be removed, a note will indicate the deletion.
Oversize materials (e.g., maps, drawings, charts) are reproduced by
sectioning the original, beginning at the upper left-hand comer and continuing
from left to right in equal sections with small overlaps.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
NORTHWESTERN UNIVERSITY
A DISSERTATION
SUBMITTED TO THE GRADUATE SCHOOL
IN PARTIAL FULFILLMENT OF THE REQUIREMENTS
for the degree
DOCTOR OF PHILOSOPHY
Field of Marketing
By
Vishal P. Singh
EVANSTON, ILLINOIS
June 2003
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Copyright 2003 by
Singh, Vishal Pratap
All rights reserved.
UMI
UMI Microform 3087980
Copyright 2003 by ProQuest Information and Learning Company.
All rights reserved. This microform edition is protected against
unauthorized copying under Title 17, United States Code.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
ii
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
A B ST R A C T
Vishal P. Singh
The supermarket industry has undergone dramatic changes in the past few
years. Alternative retail formats such as mass merchandisers, price clubs, and
supercenters have encroached upon supermarket sales to pose a serious threat.
A concurrent trend in the supermarket industry, in part driven by growing com
petition. is the movement towards developing customer databases. The impetus
behind these data collection efforts is a hope that data can be used to improve mar
keting decisions and thereby improve retailer position vis-a-vis new competitors.
However, a common refrain in industry reports is that most retailers are strug
gling to leverage this information. This thesis presents a series of three essays to
demonstrate how the information contained in retailers databases can be used to
guide marketing decisions, such as pricing and customer retention strategy, when
confronted with new competition.
iii
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
The thesis makes contributions to both marketing theory and practice. Past
research on supermarket competition has primarily focused on stores th at are iden
tical in terms of product offerings, cost structure, and pricing policies. Little
attention has been given to the growing competition from mass-discounters and
supercenters. This thesis contributes to this emerging area by addressing how
consumer behavior changes when a low-priced competitor enters the market, and
how traditional supermarkets can use the information in their database to better
compete with the new entrant. The research is also salient to the growing body of
literature focusing on database marketing. The first essay of the thesis shows how
a supermarket manager can use point-of-sales data to develop profitable pricing
policies. Similarly, the last chapter of the thesis demonstrates how a retailer can
exploit the information contained in its frequent shopper database to understand
and interact with their most valuable customers.
hr
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
C ontents
ABSTRACT
iii
List of Tables
vii
List of Figures
Chapter 1. Introduction
2.1.
2.2.
2.3.
Response by Supermarkets
11
13
3.1.
Introduction
13
3.2.
Model
21
3.3.
Estimation
30
3.4.
D ata
42
3.5.
Results
48
3.6.
53
3.7.
Conclusions
61
V
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 4.
63
4.1. Introduction
63
71
82
86
4.5. Results
92
4.6. Conclusion
101
103
5.1. Introduction
103
109
117
Chapter 6.
Conclusion
128
References
Appendix A.
Appendix
133
Tables and Figures for Chapter 3
B.
Tables and Figures for Chapter 4
Appendix
142
155
C.
Tables and Figures for Chapter 5
vi
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
168
List o f Tables
A .l
Descriptive Statistics
143
A.2
143
A.3
143
A.4
144
A.5
145
A.6
147
A.7
147
A.8
148
A.9
148
A. 10
148
A. 11
149
A. 12
149
A. 13
154
vii
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
B.l
156
B.2
156
B.3
157
B.4
157
B.5
158
B.6
158
B.7
Descriptive Statistics
159
B.8
160
B.9
161
B.10
162
B .ll
163
B.12
163
B.13
164
B.14
165
B.15
166
B.16
166
B .l7
167
C .l
169
vtii
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
C.2
C.3
169
170
C.4
171
C.5
171
C.6
171
C.7
172
C.8
172
C.9
173
C.10
174
C .ll
174
C.12
175
C.13
175
C.14
176
C.15
177
C.16
178
C.17
178
ix
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
L ist o f Figures
A .l
146
A.2
146
A.3
A.4
150
A.5
151
A.6
152
153
B .l
160
B.2
164
B.3
165
B.4
165
B.5
166
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
C .l
xi
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
CHAPTER 1
Introduction
Supermarkets operate in an increasingly competitive environment. Rapid growth
of alternative retail formats, in the form of mass discounters, price clubs, and su
percenters. has transformed not only the competitive structure of the industry but
also the way in which consumers shop. While mass discounters like Wal-Mart and
K-Mart do not offer a full array of perishable and nonperishable products, they do
compete with supermarkets in specific high-volume categories such as paper prod
ucts. dry grocery, health and beauty care products, and other general merchandise
items. More recently, the line separating supermarkets from mass discounters has
blurred as supermarkets have added more and more general merchandise items to
their shelves while stores like Wal-Mart have transitioned towards the food sector
through their supercenter format. Adding to the pressure from these new formats
is the increased competition among the grocery retailers themselves, which leaves
the industry facing perhaps its most difficult competitive situation.
Another trend in the supermarket industry, in part driven by growing compe
tition, is the drive towards developing loyalty programs and customer databases.
Supermarket retailers have been at the forefront of adopting bar code sca n n in g and
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
collecting point-of-sale data. They were also one of the first to adopt customer loy
alty cards, which has led to the tremendous increase in information available to the
retailer. Providing more than just store-level sales and local area demographics,
frequent shopper databases allow retailers to track purchase histories for all their
customers. Just as the nature of the data collected has changed over time, its
use has evolved from addressing operational issues on the supply side, to category
management, and more recently to customer management.
However, despite this impetus towards building large data warehouses, a com
mon refrain in industry reports is that most retailers are struggling to leverage
this information (MacLeod 2000). The potential difficulty of converting data into
valuable marketing strategies is illustrated by the recent cancellation of loyalty
card programs by some of the retailers in Europe, who have begun to question
the value of the huge reams of electronic data collected using loyalty cards ( The
Wall Street Journal Europe, May 19, 2000).
hand, the number of stores offering frequent shopper programs has increased to
12,000, accounting for 66% of the all commodity volume (ACV). Further, a study
by ACNielsen found that 70% of American households participate in at least one
frequent shopper program (ACNielsen Frequent Shopper Program Study, 1999).
The wide penetration of these cards indicates that large volumes of data are being
collected. The need to leverage this information becomes even more important with
the growth of formidable competitors like Wal-Mart, which, with its supercenters,
has already become the fourth largest player in the grocery industry.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
level of welfare as under a u n iform chain-wide pricing policy. We find that this
constraint still enables the store manager to capture a large portion of the gains
from unconstrained store pricing. The second practical consideration involves the
potential complexity of computing store-level prices for a chain like D o m in ic k s,
which has a large number of stores. We address this computational complexity
by suggesting an improved zone structure. Using our store-specific price levels, we
cluster the stores into five zones. We find that these zones offer substantially more
profit than the existing zone configuration.
The focus of the second essay is on analyzing competition between discount
stores such as Wal-Mart and Target and a traditional supermarket. This study
develops a conceptual model of how entry by discount stores impacts the sales of
competing products at the grocery store. A critical difference between mass dis
counters and grocery stores is that mass discounters do not offer the full grocery
line found in supermarkets. The conceptual framework models consumer hetero
geneity in terms of the utility they derive from buying all products in a single
shopping trip. Thus, consumers who prefer to buy all their groceries in a single
shopping trip would prefer to shop at supermarkets. Others may prefer to split
their baskets and take advantage of the lower prices of nonperishable goods at the
discounter while buying food items at the grocery store.
The empirical analysis uses the same data as in the first essay. However, the
focus in this study is on demand changes over time due to changes in the compet
itive environment facing the store. During the period of study, we observe entry
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
The study also analyzes the impact of Wal-Marts entry at the product cate
gory level. Retailers increasingly employ category management tools where each
category is treated as a strategic business unit and pricing, merchandising, promo
tions. and product mix are determined at the category level. Thus, it is critical
to analyze the impact of competitor entry on a category-by-category basis and to
develop strategies to foster category-level retention. We estimate a hierarchical
Bayes random coefficient logit model for three product categories. An advantage
of the Bayesian approach is that it allows parameter inference at the individual
level, which allows customers to be characterized in terms of their sensitivity to
marketing mix or preference for specific attributes. This in turn can be used to
develop customized targeting tools.
The rest of the thesis is structured as follows. The next chapter provides a brief
overview of the supermarket retailing industry, focusing primarily on the growth
of new retail formats. Chapter 3 presents the first essay of the thesis. The essay
on competition between supermarkets and discount stores in presented in chapter
4. The third essay, which analyzes the impact of entry by Wal-Mart supercenter,
is presented in chapter 5. We conclude in chapter 6.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
CHAPTER 2
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
supermarkets grew in importance, and by 1970 they had become the primary food
distributor, replacing the smaller grocery stores.
2.2. A lternative R etail Formats
Until recently, supermarkets viewed neighboring supermarkets as their primary
competitors. However, this perspective is changing. A supermarkets biggest com
petitor these days is not only another supermarket, but also include supercenter,
warehouse club, dollar store, mass discounter, or drug store that also happens to be
selling food (Urbanski 2000). We next describe some alternative retail formats that
have emerged in recent years and their degree of competition with supermarkets2.
Mass Discounters: The big three mass discounters (Wal-Mart. K-Mart. and
Target) originated in the 1960s and now account for over 80% share in this retail
format. These stores have only a partial overlap with supermarkets in terms of the
products they sell. The two formats compete in products like general merchandise,
health and beauty aids, and other nonperishable grocery items. At the same time,
each format also sells a unique set of products (e.g., clothes, tires, etc. at mass
discounters; fresh produce and meat at supermarkets). Over time, this overlap has
increased as supermarkets have added more and more general merchandise items
to their shelves while mass discounters have transitioned towards the food sector.
A recent study conducted by Proctor and Gamble (P&G). which analyzed nine
nonperishabie categories, indicates that mass discounters have been able to steal
2This section is drawn from Facts and Figures , Food M arketing Institute web page and from
T he U.S. Retail Food Industry: 2001 Store Form at U pdate W illard Bishop Consulting.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
significant volume from the supermarkets in the dry grocery business (see Table
B .l). There are a number of reasons for this. First and foremost, these formats
operate on a much larger scale, allowing them to bypass the wholesalers and buy
directly from the manufacturers. Economies of scale in distribution, coupled with
buying power with manufacturers, translate into lower everyday prices for con
sumers. Fox et al. (2000) compare the marketing mix across various retail formats
and find mass merchandisers to be least expensive, offering prices 9% below Hi-Lo
grocers. Further, in all of the nine categories reviewed in the P&G study, display
frequency was much higher and twice as long at mass merchants than in super
markets. Finally, mass merchants also create an image of value and convenience
by selling large-pack-size items.
Price Clubs: Price clubs are a membership retail-wholesale hybrid, with a
limited variety of products presented in a warehouse-type environment. The first
store of this format, Price Club Wholesalers, was opened in 1976, while Costco
and Wal-Marts Sams Club opened in 19833. According to Trade Dimension,
supermarket-equivalent sales of Sams Club and Costco in the year 2000 were $15.2
billion and $14.5 billion respectively. Sales of Kroger, the largest supermarket
chain in the country, were $45 billion, generated from about 2,400 stores. In
comparison to Kroger, Sams Club and Costco generated nearly one-third the
sales while operating only a fraction of the stores (Sams = 471, Costco = 253).
Thus, these stores operate on a much larger scale on a per store basis. While price
3Costco and Price C lub later merged in 1993.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
10
clubs are becoming increasingly popular with individual households, their primary
clientele remains small businesses, restaurants, small grocery/convenience stores,
and other institutions.
Supercenter: In the 2000 annual survey by Progressive Grocer, over 75% of
supermarket managers cited supercenters as their biggest concern in the coming
year. Supercenters, which range between 110.000 and 220,000 square feet, combine
full grocery lines with general merchandise under one roof. Many of them also
include services such as a vision center, Tire and Lube Express, a hair salon,
and so forth, providing consumers with a true one-stop shopping experience.
Supercenters have shown dramatic growth in the past few years. For example,
the first Wal-Mart supercenter was opened in 1988, while in 1993 the company
operated only 10 such stores. By 2000, Wal-Mart had over 900 supercenters, and
the company plans to add 180 new such stores every year (a number of these stores
are conversions of existing discount stores to supercenters). Similarly, Target and
K-Mart have started opening their own versions of supercenters. The primary
reason why Wal-Mart and other mass discounters have added food to their stores
is to increase consumer store-visit frequency, hoping th at this increased frequency
will spill over to other general merchandise.
Neighborhood Markets: While the large size of supercenters permits the conve
nience of grocery and general merchandise under one roof, it also limits its ability
to penetrate urban areas. Wal-Mart has recognized this limitation and is testing
scaled-down versions of supercenters ranging in size from 40,000 to 50,000 square
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
11
feet. According to Wal-Mart, its neighborhood market format will charge the same
low price as its supercenters while providing the locational convenience of super
markets and convenience stores. While still in a test phase, this format is already
regarded by supermarket managers as the second biggest threat after supercenters
(Progressive Grocer Annual Survey, 2000).
Other Nutritional/Health Stores: While the mass discounters have made in
roads into the food sector via supercenters, they are not the only ones stealing
away shares from traditional supermarkets. Limited assortment organic or natural
food stores are also growing at a tremendous pace. The $5 billion organic market
is growing at a rate five times greater than the growth of the overall food indus
try. The sales growth of Whole Foods (21%) and Wild O ats (36%). together with
W al-Marts plans to add organic SKUs. underscores the importance of this area.
The past few years have seen some of the biggest mergers and
acquisitions in the industry (e.g., Kroger and Fred Myers, American and Albertson
stores, etc.). Size increases the bargaining power that supermarkets have with
manufacturers while also providing economies of scale in distribution. As seen in
Table B.2. this consolidation wave, coupled with the dram atic growth of alternative
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
12
retail formats, has led to a significant drop in the number of small and independent
grocery stores.
Besides consolidating, supermarkets have also increased their focus on services
and customer satisfaction. For example, supermarkets have added convenience ser
vices that are both food and non-food related. Deli counters, ready-to-eat foods,
salad bars, and cafes have been added to cater to the needs of a wider group of
consumers. Services such as banking, postal services, and video departments are
now commonplace in supermarkets. In terms of overall selection, the number of
SKUs and average store size have also increased. Further, more and more super
markets have introduced frequent shopper programs that deliver price discounts
as rewards to encourage shopper loyalty to a store. The strategy of introducing
loyalty programs, coupled with increasing services, has been to make a customers
shopping in supermarkets a more convenient and rewarding experience.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
CHAPTER 3
13
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
14
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
15
complement the store data, we use competitive information from Spectra marketing
th at characterizes the demographic profile for each stores consumer base as well
as the proximity to local competitors. One of the unique features of the data
is the availability of weekly store margins, which we use to compute a measure
of the wholesale price. Combining wholesale prices with the zone designations
provides another interesting aspect of the data. Unlike previous empirical work on
price discrimination, we are better able to attribute cross-sectional price variation
to discrimination as opposed to alternative explanations, such as cost differences.
Thus, we are in a better position to measure the welfare implications from thirddegree price discrimination.
Using the data for 2 product categories available in the database, refrigerated
orange juice and liquid laundry detergent, we estimate structural demand systems.
Access to retail margins as described above allows us to validate the ability of our
demand system to recover the true underlying data-generating process by com
paring the margins observed in the data to those predicted using the estimated
demand system. This validation is similar to Nevo (2001) and Slade (2002) except
th at we observe a full time-series of margins rather than a single observation of the
approximate mean margin for each alternative. Using a static category manage
ment pricing model, we then simulate the prices and sales consistent with uniform
weekly chain-level pricing. Since the data contain the chains definition of the retail
zones, we are able to compute the price and welfare impacts of zone pricing rela
tive to chain level pricing. Moreover, we are able to assess which local consumer
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
16
segments benefit and which are hurt by this geographic price discrimination and
how these effects vary across product categories. We then repeat the exercise at
the store level by implementing a profit-increasing store-level pricing model and
simulate its impact on prices, profits and consumer welfare. Finally, given the
practical limitations of implementing a store level pricing policy, we propose an
alternative pricing scheme that alleviates this problem.
A key feature of our estimation approach is the ability to identify flexible
substitution patterns while controlling for the endogeneity of prices. We estimate
systems of demand equations, allowing the shape of demand to vary across stores.
As in Berry. Levinson and Pakes (BLP hereafter. 1995), we use an aggregate mixed
logit demand specification. The demand model allows for category expansion by
including a riio purchase option in the specification. The parsimony of the BLP
approach comes, partially, from the use of measurable product characteristics.
This characteristics approach involves projecting consumer preferences onto a
set of product attributes and, thus, using these attributes to help explain aggregate
substitution patterns. As with many of the typical supermarket product categories,
our data do not include an exhaustive set of measurable product attributes. Our
solution is to estimate a richer correlation structure in the distribution of brand
valuations. To preserve some of the parsimony of the mixed-logit specification,
we use a factor-analytic approach. An interesting by-product of this approach is
the ability to produce a map of the brands based on intangible product attributes
(Elrod 1988 and Chintagunta 1994).
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
17
A recent body of empirical research has attem pted to identify and measure the
implications of price discrimination. Existing work has emphasized the identifica
tion of the practice of price discrimination versus cost-based differences (Borenstein
1991. Shepard 1991 and Cohen 2001). Others have looked at the underlying sources
of price discrimination and their magnitude (Verboven 1996 and Goldberg 1996).
Only recently has empirical work emerged documenting the welfare impact of price
discrimination. Leslie (2001) simulates the impact of several counterfactual price
discrimination schemes for tickets at a Broadway theatre. Cohen (2000) measures
the impact of second-degree price discrimination across brands in the paper towel
industry. A related literature on coupons has focused predominantly on the likeli
hood of consumer usage, rather than impact on equilibrium welfare (e.g. Chiang
1995, Gonul and Srinivasan 1996 and Erdem, Keane and Sun 1999). Nevo and
Wolfram (2000) investigate the impact of manufacturers coupons on retail prices
for breakfast cereals. Similarly, Besanko. Dube and G upta (2001a) demonstrate
empirically the potential for profit-enhancing targeted coupons in a competitive
environment.
Several studies have used the d ata set used in this paper although with different
objectives (e.g. Peltzman 2000 and Chevalier, Kayshap and Rossi 2000). Two
papers using these d ata are closely-related to our work. Hoch et al. (1995) find
th at a large proportion of the variation in category-level consumer price elasticities
across stores is explained by local consumer demographics and, to a much lesser
extent, to local competitive variables. In a follow-up study, Montgomery (1997)
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
18
looks at the profit-implications of zone and store pricing for the supermarket chain
in one specific product category.
Our work differs from these two papers in several ways. Our main objective is
to study the implications of zone pricing on consumer welfare (besides profits), the
results of which should interest policy workers, strategists and marketers. We then
use the underlying structure of the model to propose alternative pricing strategies
that address customer welfare due to store pricing and the computational com
plexity for the store manager. Our work also differs from a practical point of view.
Similar to Montgomery (1997), we find th at estimation of standard demand mod
els (double-log, linear etc.) underpredicts the price sensitivity of demand: implied
margins are much higher than the observed levels. The parsimony of the current
model enables us to solve this problem in two ways. First, we estimate flexible
substitution patterns with a relatively small number of parameters. As a result,
we do not observe "incorrect" signs in our cross-elasticities. 1 At the same time,
we are able to control for the potential endogeneity of prices. Previous w'ork with
various weekly supermarket data and discrete choice modeling consistently finds
that instrumenting for prices leads to noticeably higher magnitudes in the price
response param eter (Besanko, Gupta and Jain 1998, Villas-Boas and Winer 1999
and Chintagunta 2002) . 2 We find our model provides reasonable estimates of the
lNote th a t previous work has often estim ated negative cross-price elasticities of demand in cat
egories which one would expect to consist of substitute products. These incorrect signs will bias
simulated profit-maximizing prices upwards.
^lontgomery and Rossi (1997) do not find noticeable effects from instrum ental variables. How
ever. their specification uses aggregate price indices for a category rather th an shelf prices.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
19
zone-level margins. Finally, we validate the fit of our model using the observed
retail margins.
Consistent with the zone structure, we observe substantial price variation across
stores within a given week in the raw data. In fitting the demand curves, we find
both the price-sensitivity as well as the no-purchase probability vary tremendously
across stores. Interestingly, our estimates of zone-specific prices provide a reason
ably accurate representation of the true prices reported in the data. This finding
gives us confidence in our ability to capture the underlying data-generating pro
cess. As in Hoch et al. (1995), we find that demographic variables consistent with
willingness-to-pay are the most influential for market shares and elasticities. Thus,
we conclude that zone classification is based primarily on discriminating across
consumer types. However, we do find evidence th at measures of competition and
consumer search m atter, but to a lesser extent. To assess the welfare implications
of zone-pricing, we compute the prices and quantities that would prevail if the
retailer adopted a uniform chain-level pricing policy for products. As expected,
category profits are higher under the zone pricing relative to chain pricing. Con
sumer welfare effects vary across stores. We then compute the prices and quantities
that would prevail if the stores adopted a store-specific pricing policy. Now prices
rise substantially more than under the current zone pricing. Interestingly, while
consumer welfare effects vary across stores, we note th at the store-pricing seems to
target higher prices to less-affluent areas with larger ethnic populations and higher
search costs. In contrast, zone-pricing seems to target high prices to more affluent
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
20
areas. VVe also conclude th at shifting from uniform pricing to either zone or storelevel pricing may be better suited in some categories than others. For example, in
liquid laundry detergent, we find only modest gains from zo n in g - We also note the
importance of balancing store profits with consumer welfare. For example, in the
orange juice category, we do find sizeable losses in consumer welfare, especially if
the chain moves to store pricing.
Two practical considerations arise with such profit-enhancing pricing policies.
First, a store manager could be concerned about the overall losses to consumers
in stores where prices are expected to rise. To mitigate these losses, we propose a
constrained pricing procedure that makes use of the underlying economic structure
of the model. The approach restricts the new store prices to offer the population
of consumers in each store at least the same level of welfare as under a uniform
chainwide pricing policy.
manager to capture a large portion of the gains from unconstrained pricing in both
categories. The second practical consideration involves the potential complexity of
computing store-level prices for a chain which, like Dominicks, has a large number
of stores. We address this computational complexity by suggesting an improved
zone structure.
zones. We find th at these zones offer substantially more profit than the existing
zone configuration.
The chapter is structured as follows. Section 2 presents the model. Section
3 discusses the main aspects of the estimation procedure. Section 4 provides an
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
21
overview of the d ata we use for estimation. Section 5 presents results, including
the demand parameters, the perceptual maps and finally the welfare implications
of zone-pricing. We conclude in section 6 .
3.2. M odel
To develop a viable pricing framework for a category manager, we begin by
specifying an economic model of individual choice behavior in a supermarket cate
gory. We then derive the expected aggregate demand facing the category manager.
Using the derived demand, we then model the category manager's pricing problem.
This structural approach presents several advantages. From a m o d elin g per
spective, we are able to compute a theory-based measure of consumer well-being
based on their willingness-to-pay. This measure permits us to compute consumers'
monetary valuations of various pricing policies by the retailer. At the same time,
the structured approach simplifies our treatment of the retailers pricing problem.
The derivation of aggregate demand from first principles yields a "well-behaved
specification. Typically, popular approximations of aggregate demand, such as
the log-log model, generate non-concave regions in the profit function. The non
concavity makes it impossible to maximize profits without imposing additional (ad
hoc) constraints (Anderson and Vilcassim 2001) on the set of profit-maximizing
prices. Similarly, the parsimony of the model helps us identify reasonable aggre
gate substitution patterns. In contrast, studies using standard linear or log-linear
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
23
discussion of discrete choice models and their aggregation we refer the reader to
BLP (1995). One of the main advantages of this specification is parsimony. Con
sumer preferences are projected onto a set of exogenous product attributes, which
greatly reduces the dimension of the estimation problem. For industries with a
large number of alternatives, correlations in valuations of products is characterized
by heterogeneous tastes for the attributes. In many product markets, researchers
are easily able to collect a sufficient set of attributes to capture the underlying
market segments in the category. For example, this approach has been applied
to aggregate d ata for automobiles (BLP 1995, Petrin 1999 and Sudhir 2001a),
PCs (Bresnahan. Stern and Trajtenberg 1997), ready-to-eat cereals (Nevo 2001)
and movie theaters (Davis 2000).
one would find in supermarkets, much of the correlation derives from intangible
sources such as brand perceptions. Since intangibles are not easily measured by
the econometrician, we model the joint distribution of brand valuations explicitly.
We use a parsimonious factor-analytic approach to recover the impact of intangible
attributes, as is typically used with individual data (Elrod 1988, Chintagunta 1994
and Elrod and Keane 1995) as well as with aggregate data (Chintagunta, Dube
and Singh 2002 ) . 3
3Additional m ethods exist for identifying flexible substitution patterns using similar data with
aggregate choices. For instance, Nevo (2001) samples consumer dem ographic profiles from the
empirical joint distribution provided by the census a t the MSA level. BLP (1998) construct
additional moments of th e data-generating process by combining additional micro d ata with
their aggregate data.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
24
U hjt
h =
&h]
-f-
- E jt& h
P jt
~F
jt
d"
~ h jti
,...,J .f = l . . . . , r .
4Since we estim ate a full set of product fixed-effects, we do not need to worry about unmeasured
physical product attributes, as in BLP 1995. We are nonetheless concerned with unobserved (to
th e econometrician) weekly in-store product-specific effects.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
25
correlated errors, such as the multivariate normal, giving rise to the probit choice
model (McCullough and Rossi 1994). Below we discuss some of the limitations of
an i.i.d. additive error.
The formulation allows for an outside good, "no purchase", the utility of which
is given by
U-ktO ~ a h0 + QhYh. + ~htOIn the current context, this alternative represents the shopping occasions where
consumer does not purchase in the category. For practical reasons, this outside
good is im portant for the retailer pricing exercise. In the absence of this alternative,
the total category size would be invariant to the prices of all brands increasing or
decreasing by the same amount. Hence, allowing for the outside good allows the
category sales to be influenced by the prices of the inside goods. For identification
purposes,
is normalized to zero.
5In th e following analysis, we do not address formally how households allocate total income to
their weekly shopping budgets.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
26
eh
where the vectors of means,
+ \ v h, u h ~ N (0 .1)
0
to be estimated.
VVe do allow for a richer covariance structure for the vector of brand perceptions:
a h ~ iV (a, E ) .
In theory, we could estim ate the the full (J x J ) matrix E directly. In practice, as
the number of brands grows, E becomes increasingly difficult to identify. Instead,
we use the factor structure:
E = Lwu/'I', u/ ~ N (0 ,1 ) .
One interpretation for this structure is that L is a (J x K) m atrix of latent at
tributes for each of the J brands, and u; is a ( K x 1 ) vector of consumer tastes
for these attributes. The vector of mean brand perceptions, q, and the matrix
of latent attributes, L, consist of parameters to be estimated. In addition to its
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
27
parsimony, this approach allows us to estimate standard errors for the latent at
tributes. In the current context, we assume K 2. For identification purposes,
we do the following (see Elrod 1988):
( 1 ) The outside or "no purchase '1 option is located at the origin of the map
(translational invariance).
(2) One of the brands is located along the horizontal axis (rotational invari
ance).
(3) We set the variances of jj above to
where 6jt = aj-rxJtJ 8pjt+Jt is common to all consumers and fih]t = x]t\ v h+Lu}h
is consumer-specific6. The unconditional probability qJt that a consumer chooses a
particular product j in week t, after integrating out individual heterogeneity, has
the following form:
(3-i)
<bt
/ = j
exP Vit
+ Phjt)
,,
-o{v)du.
h =
= Q............. =
6N ote that we remove the term BhYh from the equation as it will not be identified in the share
equations below. This term drops out of th e share equation as it is common to all th e alternatives
including no-purchase.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
28
where o () is the pdf of a multivariate standard normal. From the store managers
perspective, (3.1) represents the share of consumers entering the store in week t
that purchase a unit of product j .
QJt = qjtMt.
Our main motivation for using this random coefficients specification, as opposed
to a simpler conditional logit (or homogeneous logit), is the need for flexibility.
The homogeneous logit generates the IIA property (the independence of irrelevant
alternatives property) at the consumer level. The IIA property could manifest
itself into our aggregate analysis in several ways. First, it can be shown that the
assumption leads to aggregate cross-elasticities that are driven by market shares
(see BLP 1995 for a thorough discussion).
set a uniform margin for each of the products in their line (Besanko, Dube and
G upta 2001a). Since our analysis focuses on category management, this property
would imply that all of the products in a category have the same mark-up over
their wholesale prices. To alleviate the restrictive substitution patterns generated
by the IIA property, we allow for consumer-specific deviations from mean tastes.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
29
As a result, the demand function in 3.2 above does not suffer from restrictive
substitution patterns.
We use random taste coefficients to generate correlations in utilities for the
various alternatives and, thus, relax the restrictive substitution patterns generated
by the IIA property. As discussed above, one could also use a correlated additive
error, such as the probit model, which avoids the IIA property directly. In general,
the probit strictly dominates the logit as it allows for freely-varying covariances.
The random coefficients probit also enables one to disentangle heterogeneity from
simple non-IIA behavior at the consumer level. We use the "mixed' logit instead
of a multinomial probit due to the relative ease of estimating the former versus
the latter. McFadden and Train (2000) show that the mixture of normals with
the type I error, the mixed logit, is sufficiently flexible to approximate a broad set
of parametric indirect utility functions, including the probit (see Dalai and Klein
1988 for a related finding). In practice, the flexibility depends on the restrictions
placed on the correlations in the random coefficients, . W ith aggregate data, the
ability to integrate out the logit disturbance, as in (3.1), vastly increases the ease
of implementation versus a multinomial probit. Thus, we choose the mixed logit
due to its relative ease of use.
One of the complications of the mixed logit specification (3.1) is the lack of an
analytic form for the multidimensional integral. While it is true th at for a simple
model with fewer than three random parameters one could solve the expression
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
30
numerically (Hausman and Wise 1978), most categories consist of more than three
alternatives. Instead, we use direct Monte Carlo simulation, as in Nevo (2001).
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
31
in terms of <f>. Since the inverse of (3.1) does not have a simple analytical form, we
resort to numerical inversion using the contraction-mapping procedure suggested
in BLP (1995). The approach requires, for each store-week f, picking some ini
tial guess of the mean utility vector 6t and iterating (3.3) until the following J
expressions converge:
(3.3)
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
32
where the superscript n refers to an iteration, and qjt is the observed market share
for brand j in period t.
The advantage of using 6]t for estimation is that the prediction error. <5j J'
is simply the unobserved product characteristic. Jt. The fact that Jt enters (3.3)
linearly facilitates instrumentation. Moreover, with some intuition for the source of
the unobserved attribute, we are able to impose reasonable covariance restrictions
to set up our method of moments procedure.
We now set up a generalized method of moments (GMM) procedure to estimate
the system of mean utilities. Let t be the (J x 1 ) matrix of unobserved attributes
for each of the products in store-week t. Similarly, we define our instruments, Z t,
an /-dimensional vector including the exogenous product characteristics as well
as other potential covariates that may be correlated with pJt, but not with Jt.
Our key identifying condition is the conditional mean-independence assumption
E (t S Z t \Zt) = 0 and E (t |Z t) = ^ a finite (J x J) matrix. We are now able
to construct our moment conditions:
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
33
Our goal is to find values of 0 close enough to 0o to set the sample moments as
close as possible to zero. We estimate by m in im iz in g the following quadratic
expression:
G (0) = (/iyT( ))'W (/lJT( 0 )) .
The matrix W is a (J T x J T ) weight matrix. Hansen (1982) shows that the most
efficient choice of W is a consistent estimate of the inverse of the variance of the
moment conditions:
w =
=
{(m ))(M))'}
E {& Z tZ[} .
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
34
1,
teristics in the vector Ds. These terms shift both the price sensitivities, 0, and the
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
35
1 ,....
83 stores as follows:
6ha @+ D ' 7 -+- A i/fc
(3.5)
where 6 is the mean marginal utility of income across stores, weeks and households;
D s ~i
is the mean marginal utility of income across consumers and weeks in store
s: and AV/, is the idiosyncratic component of marginal utility of income for some
consumer h. Similarly, a j is the mean brand preference for alternative j across
stores, weeks and households: D's6 is the mean across consumers and weeks in
store s: and Ljuihj is the idiosyncratic component for some consumer h. The
vectors 5 and
6 is common across brands so that store characteristics will shift the category
size (brand qualities relative to no purchase). In principle, one could estimate a
separate 6 and a parameter for each of the 83 stores in our sample. However, this
approach is infeasible since it would require estimation of 166 parameters. The
proposed decomposition has additional potential benefits, such as the ability to
forecast demand in a new store based on its characteristics.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
36
eral. the zip-code and census block level data collected for the Current Population
Survey are too thin to provide credible representations of the true underlying de
mographic distribution. So we use mean demographics in each store area instead.
To the best of our knowledge, no study has modeled store competition explicitly
in determining aggregate demand for a retailer. In modeling demand for yogurt,
Berto Villas-Boas (2001) treats the same brands in different stores as substitutes.
However, she finds extremely small cross-price elasticities across stores, suggest
ing th at store competition has little impact within the yogurt category. Due to
d ata restrictions, most applications of store-level data treat retailers as local mo
nopolists (e.g. Slade 1995, BGJ 1998). Slade justifies her assumption on phone
interviews with store managers in a given market who claim that consumers do not
shop across stores on a product-by-product basis. We also conducted telephone
interviews with Chicago area store managers and our findings were consistent with
this claim. Stores do condition on their competitors actions in a limited way by
collecting a weekly sample of half a dozen SKUs from the local competitors entire
store offerings. However, this behavior seems more consistent with competition
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
37
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
38
Vh (expected utility) and their marginal utility of income as Oh- Suppose a zonepricing policy is introduced that changes consumer valuations for each alternative
from Vrhdu,m to Vh*. As derived in Small and Rosen (1981), assuming individual
marginal value of income is held constant, individual h's associated change in
welfare can be computed as:
(3.6)
CVh =
The numerator of (3.6) captures the expected change in utility. Dividing through
by the marginal utility of income. 9h, makes this change money-metric. Integrating
across the distribution of consumer preferences, we can compute the expected
aggregate change in consumer welfare:
(3.7)
In the next section, we model the supermarket category managers pricing decision.
Using variable profits as the measure of the managers valuation, we are able to
compare the dollar value of gains of various pricing policies both to the supermarket
and to consumers. Similarly, we can compute the change in customer value in going
from zone to store level pricing.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
39
3 .3 .3 . R etailer Pricing
We now describe our model of retail behavior.
Besanko, G upta
and Jain 1998. Sudhir 2001b and Villas-Boas and Zhao 2001). Typically, these
studies do not have access to retail margins and, thus, use the channel structure
to help identify a time-varying wholesale price (see for example, Berto Villas Boas
2001). Kadiyali et. al. (2000) use information on retail margins. However, their
objective was to identify the nature of interactions between manufacturers and a
single retailer.
We assume that each week the retailer jointly sets the profit-maximizing prices,
Pj , of each of the J products in a category (see for example, Sudhir 2000b, Kadiyali
et. al. 2000). Based on our phone conversations with local chain managers, we
believe that weekly price decisions are made by category managers rather than by a
store-wide manager9. In contrast, most promotional decisions (newspaper features,
in-aisle displays etc.) are determined at the store-level.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
40
funded almost entirely by manufacturers, the timing and format are determined
by the retailer. Typically, the promotional calendar is determined in advance so
that category pricing decisions are made conditional on the promotion. Therefore,
we treat the promotion level as exogenous to category pricing10.
We also assume that the retailer's variable costs consist solely of wholesale
prices. w3. We treat all store and/or category-related overhead as fixed costs, Ft.
Thus, in week t . a retailer 11 solves the following optimization problem:
j
max II = y^ (p Jt Ip j I j -
j t ) Q
j t
F t -
Note th at Qjt is the expected demand form (3.2)above. The first-order condition
for a typical brand i is:
( P j t
o.
. J in m atrix form
as:
(3.8)
ft(p - w) + Q = 0,
10O ur phone conversations revealed th a t category managers may in fact request additional pro
m otional funds if they feel the performance of th e category or a specific brand therein is sluggish.
However, we were informed th at the incidence of such endogenous (to the category manager)
promotions are quite unusual.
11To simplify- notation, we drop the subscript for the retailer and the category. In our empirical
work, a retailer may be the store-manager, the zone manager or the chain m anager depending
on the context.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
41
where
P it w \ t
pw =
P jt ~ w Jt
j* l
k
&p,rJi = K
22a
=
Q=
Q jt
J Jxl
This represents a system of J equations, one for each brand. The optimal set of
prices for the retailer are determined by solving:
(3.9)
p = w Q q
where fl-1q is the retail mark-up. By checking the second order sufficient con
ditions. one can verify th at the solution to (3.9) represents an optimum for the
retailer. The actual value of Qjt depends on the level of aggregation considered.
For the store-level problem, this will take the form (3.2). where M t is weekly store
traffic. However, for a zone pricing problem. QJt would be obtained by integrating
across the store-level demand for each of the stores in the zone in week t. Simi
larly, chain-level pricing would involve integrating across all the store-level demand
curves.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
42
tributes are homogeneous, the optimal retail prices satisfying (3.8) become:
P lt = w +
where
= \ J2j=i
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
43
divided by the total weekly store traffic. Effectively, our model implies that each
time consumers visit the store, they either purchase a unit of one of the alternatives
within the category of interest or they elect not to purchase. In the current context,
a unit corresponds to a UPC code (e.g. a 64 oz Tropicana Premium vs. a 96
oz). The promotion variable is an indicator for whether the given product had a
newspaper feature th at week. As mentioned previously, the promotion decision is
assumed to be exogenous to category management within a store or zone.
In table A.2, we summarize the sources of variation in our price and share
data. For each SKU in our data, we run separate regressions of prices and shares
on store and week fixed-effects. In the table, we report the median R-square across
products for each category. We find that cross-week variation accounts for roughly
72% of the variation in a products price, in the laundry detergent category, and
77% in refrigerated orange juice (78% and 76% respectively for shares). Similarly,
cross-store variation explains roughly
11%
gent. and 6 % in orange juice (2% and 5% respectively for shares). As explained
below, we are able to explain a substantial portion of the intertemporal price vari
ation using the wholesale prices. To help capture the cross-store variation, we use
characteristics of a store's trading area.
We supplement our store data with an extensive set of descriptive variables,
from Spectra (see Hoch et.al 1995), characterizing the underlying consumer base
and local competition associated with each store. ZIP code level demographic
d ata was obtained from the 1990 census. To capture heterogeneity across stores in
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
44
terms of the types of households they face, we include the following demographic
variables: INCOME (log median income), AGE60 (% of population over age 60),
ETHNIC (%of population that are Black or Hispanic) and HVAL (mean household
value). We also include SHOPINDX (ability to shop-% of population with car and
single family house) to capture the relative ability of local consumers to travel.
The two competitive variables used in the study are distance from the nearest
Jewel (the largest supermarket in the area), JEW EL, and minimum of the distance
from the nearest Cubfoods and Omni. EDLP, (the two main EDLP operations).
Our preliminary work also included variables on competitor volume but these had
limited explanatory power and were dropped in the subsequent stages of analysis.
Summary statistics for the demographic and competitive variables are provided
in table A.3. We find considerable variation in the demographic and competitive
characteristics across stores. For example. DFF stores cater to market areas with
Black and Hispanic representation ranging from 2 to 99% of the population. In
terms of consumer wealth, income levels range from $19,000 to over $75,000 (note
we report INCOME in logs). Similarly, average house values range from $64,000
to over $267,000. In terms of competition, some stores are located right next to
rival supermarkets. Others locate over 4 miles from the nearest Jewel and 18 miles
from the nearest EDLP store. We expect these differences to generate noticeable
differences in the levels and price-sensitivities of demand across stores. Note that
the demographics explain, on average, 6.7% of the price variation for detergent
and 3% for refrigerated juice. Thus, we are able to explain over half of the price
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
45
variation across stores using these demographic variables. Since the estimation of
store-specific parameters would generate an unmanageable number of parameters
for the given data sets, we are confident that the demographics do a reasonable
job explaining store-specific differences.
As discussed in the model section, we classify the demographics and competitive
variables into three categories: willingness-to-pay, consumer search and competi
tion. The demographic variables are used to capture consumer heterogeneity in
tastes across stores. The extent to which these factors influence shares at a store
are attributed to price discrimination. The Spectra measure SHOPINDX mea
sures the ability of a consumer to shop and, thus, to search for the "best price
in the local market. Finally, proximity to competitors provide a crude measure of
the extent of local competition. We classify the impact of consumer search and
competition on a store's brand shares as sources of price dispersion.
3.4.1. Zones
An important feature of our data is the ability to disentangle price discrimination
and differences in cost. In general, wholesale prices are virtually identical across
stores in a given week. On average, the standard deviation of wholesale prices
across stores in a given week is 0.008. Unlike previous work (e.g. Shepard 1991)
we can rule out explanations for price variation based on wholesale costs. 12
l2One explanation th a t we have not ruled out is differences in th e opportunity cost of shelf space
in areas with expensive real estate. In areas with high property values a n d /o r high property
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
46
Despite the roughly identical unit costs across the stores in our data, prices and
margins vary substantially within most weeks. Using the liquid laundry detergent
data, the average range in prices for a given product across stores in a given week
is 81 cents (prices are measured on $/oz. basis). By contrast, the average range
within a Dominicks-classified zone is about 16 cents. Given that the mean price in
the category is $5.58. the average weekly price range across stores overall is 14%
of the mean price, versus only 3% within a zone. For certain products, the price
differences are significantly higher. For example, in the refrigerated orange juice
category, the price of Minute Maid (64 oz.) is 38% higher in the highest price zone,
compared to the stores that fall in the lowest price zone.
As mentioned earlier, our data set contains an index th at groups stores into
pricing zones. In practice, the chain does not always appear to respect the specified
zones in its weekly pricing decision. Looking across products, we observe that many
items appear to use a coarser zone definition. For instance, prices of small share
items often have a uniform price across stores. Similarly, in some categories, prices
may only reflect three or four zones, rather than the full 16 th at we observe later
in the data. Other studies that have used this data (for example Hoch et al. 1995)
also suggest th at the actual number of zones might be fewer than those provided
in the data. We investigate this issue by looking at the prices of several products
across randomly selected weeks. Consistent with Hoch et al., we find only 3 levels
taxes, stores may have more rigid capacity constraints th at could affect pricing. In this regard,
zone pricing could reflect differences in these shelving costs.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
47
of prices in the early years of the total available data (e.g. 1989-1990). However,
the number of zones increases over time. By the time of our sample, 1992, we
begin to observe substantially more prices for many large-share products in any
given week. For some of the large share items, we often observe between 9 and 16
different shelf prices across stores within a given week. 13
In table A.4 we report the average laundry detergent prices and demographics
by zone. These d ata demonstrate the ability of our store characteristics to explain
some of the observed price differences across zones. For instance, the highest prices
are in zone 7. where we observe the highest household values as well as fairly high
search costs. In contrast, zone 16 has the lowest prices and exhibits fairly high
incomes with very few elderly or ethnic households. Note also that zone
has the
closest EDLP store and, at the same time, has very low prices.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
48
10%
3.5. R esults
We now present the estimated demand systems for each of the two categories.
Demand parameters are reported in tables A.5 and A.6 . We report the corre
sponding latent brand factors in table A.8 . Rather than reporting the implied
brand correlation matrices. . we plot the brand ^locations using the factors as a
coordinate system. We also report the elasticities of store characteristics on each
of the category sizes (probability of purchase) in table A.7.
Before describing the empirical results, we first explain what the parameters in
tables A.5 and A . 6 mean. We use table A.5 (laundry detergent) as an illustration.
The first five parameters correspond to the mean in trin s ic preferences, oJT for the
J brands included in the category. This is followed by the mean price effect, 6, and
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
49
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
50
seems to separate out brands manufactured by P&rG (Tide and Cheer) from those
manufactured by Unilever (Wisk. Surf and All). Further, the two P&G brands,
Tide and Cheer are perceived to be similar to one another whereas the Lever
brands appear a lot more differentiated in the minds of consumers. By positioning
All close to the P&rG brands. Lever may be using it as a ^ g h tin g brand in the
marketplace. We discuss the effects of store characteristics subsequently.
In the orange juice data, table A.6 , we find that TVopicana premium has the
highest mean preference effect, which is consistent with its dominant position in the
category. We also observe a preference for the smaller 64-oz size, versus the larger
96 oz. As before, promotions increase the likelihood of purchase, but at the same
time, they increase consumers sensitivity to price. In figure A.2, we present the
perceptual map for orange juice brands. Once again, we find the map of the brand
preferences to be quite revealing. There appear to be 3 distinct groups of brands
in the market. The first set. consisting of the TVopicana and Florida brands, is
perceived as being different from the other brands. We do however, observe slight
differences between the product not from concentrate (premium) and the product
from concentrate (SB). The second group consists of Minute Maid, a Coca-Cola
product, and the third group consists only of the store brand (Dominicks). This
finding is good news for the national brands as it does appear th at they have
effectively differentiated themselves from the store brand.
In table A. 7, we report the elasticities of the store characteristics on category
size (the probability of purchase). Since characteristics enter the demand systems
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
51
3.5.1. G oodness-of-Fit
Our next objective is to try and assess the ability of our demand system to predict
the store-level margins. Using the observed wholesale prices, we compute the shelf
prices for chain-level, zone-level and store-level pricing by solving the system of
equations defined by (3.1) and (3.9). Since our main objective is to study the
implications of pricing, we need to verify that our demand estimates and our
category management model produce realistic measures when compared to the
observed margins.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
52
The first step involves determining which model seems to come the closest
to approximating the observed margins in the data. One way to think of this
problem would be to solve a minimum distance procedure in which one minimizes
the distance between true and estimated margins, using the covariance of the
observed margins as a metric:
(3.10)
where /* is the estimated margin and $ is the covariance m atrix of the observed
margins. In table A.9, we take the margins implied by the store-level, zone-level
and chain-level pricing policies and compute the corresponding criterion (3.10).
The zone-pricing model seems to provide the best fit according to the minimum
distance criterion.
We now compare the mean predicted margins for each brand under the three
pricing policies considered and compare these to the true margins observed in
the data.
percentage of prices for the laundry detergent and refrigerated juice categories.
These tables capture how well we can reproduce the levels within stores. For the
laundry detergent data, the levels are reasonably close to the true values and the
correlations are fairly high, especially for the largest-share items (Tide and All).
For juice, our mean margins look, for the most part, fairly reasonable. The main
exception is the store brand for which we over-predict the margin and for which
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
53
price estimates are very poor. Our sense is that store-brand pricing may not fit well
with the category management model. In fact, store-brands may well be priced
below the category manager's level if they are set at a store-wide level to generate
store traffic (Chintagunta 2002).
comparing the zone prices and welfare with those computed at the chain level. In
conducting this analysis, we need to make certain assumptions. First, we assume
th at shifting from chain to zone pricing does not alter the nature of promotions
in the category. This assumption is analogous to Nevo (2000) and Petrin (2001)
who assume th at mergers and product-introductions respectively do not impact
(in the short-run) the mix of observed product attributes. Similarly, we assume
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
54
that zone-pricing does not impact the unobserved attribute, f. Finally, we also
assume that changes in pricing for the product categories in question will not alter
the levels of store traffic each week. To support this assumption, we present results
for a simple model of store-choice, in table A. 13. Assuming th at each household in
a stores market makes a single weekly store-choice decision, this model captures
the binary probability that a households visits the local DFF or not. We measure
the share of households that visit DFF as store traffic divided by the total number
of local households. We explain this choice using local market characteristics as
well as a price index for each of the 33 product categories available. Our results
may be biased due to the lack of competitors prices. However, the fact that both
laundry detergent and refrigerated juices are not significantly affecting the store
choice probability makes us more confident in our assumption above.
In table A. 12. we report the total annual chain-wide welfare implications of
various pricing policies for both of the categories studied. As expected, allowing
for more flexible pricing increases a categorys profitability, so th at zone-pricing
and store-pricing both generate gains to the store. In general, the incremental gain
from store-pricing far exceeds the gains from zone-pricing. At the same time, the
impact on consumers varies by category. Recall that one interpretation of these
consumer welfare numbers is the total dollar amount the chain would need to pay
consumers to make them as well off under some new pricing policy as they were
under uniform chain-wide prices. One must keep in mind th at the main reason
why Dominicks did not pursue a store-specific pricing policy was due to the lack of
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
55
technology at the time for implementation. Thus, the current analysis illustrates
ways in which structural econometric methods can be applied to a standard retail
database to execute more profitable pricing strategies.
We begin with the impact from a shift from uniform chain-pricing to the zone
pricing practice th at appears to be generating the observed data. Then, we will
discuss the impact of a hypothetical move towards store-level pricing. For laundry
detergent, we observe fairly small total effects from zone-pricing on the chain profits
(only 0.6% gain) or on consumers; although consumers do gain overall by 2,158.
Given the high-necessity oriented nature of laundry detergent, it seems intuitive
that such an item would not exhibit tremendous variation in willingness-to-pay
across store markets. In contrast, the orange juice category benefits reasonably
well, generating a $52,400 (1.6%) gain in profits. At the same time, total consumer
welfare falls $19,791. Figure A.5 plots the total compensating variation by store,
indicating that the negative impact on total annual consumer welfare is misleading.
In fact, in many stores, consumers benefit from the price change.
Now we demonstrate how our demand system can be used to implement a
more profitable store-specific pricing scheme. As above, we compare the prices
and welfare levels from the zone model with those of the store-level model. As
expected, the store-pricing leads to much larger profit gains than the zone pricing.
We observe an almost $542,000 (16.3%) gain in profits relative to chain pricing
for orange juice, and $109,700 (9.6%) for laundry detergent. At the same time,
we observe $158,100 in losses to consumers in juice, which is small relative to the
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
56
profit gains, but much larger than the losses from zone pricing. For detergent,
consumers do in fact gain overall by $16,215. As before, we can see th at the low
consumer welfare losses are misleading once we look at each store. In figure A.6 ,
we plot the compensating variation for orange juice by store.
The intuition for why consumers in some stores gain value while others lose
welfare, or value, from these more flexible pricing policies relates to the ability
of the store to re-align its product line pricing according to local demand. For
instance, in store 75, most laundry detergent prices are raised, on average, about
2-3% (e.g. 128 oz price rises by 26 cents). At the same time, the price of 64 oz
Tide (the category leader in the store with roughly 15% more share than the 2ndranked alternative) is lowered by one percent. As a result, the conditional share
of 64 oz Tide rises almost 16%. while the conditional shares of most of the other
brands fall by 1-3%.
in which consumers have fairly high search costs and there are no nearby EDLP
stores. As a result, demand is fairly inelastic, which explains why the store would
want to raise its price level.
As a result,
shares become much more equalized in store 128, with the two largest-share goods
falling 2-3%. At the same time, the conditional share of 64 oz Wisk rises almost
9 percent, making it (by a narrow margin) the new category leader, on average.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
57
Note th at store 128 is in zone 11, which caters to households with relatively low
incomes and house values. The area also caters to a higher proportion of ethnic
households with larger families. As a result, demand is much more elastic, which
explains why the store reduces most prices, allowing for a better-value brand to
gain more relative share.
An interesting question for customer relationship management involves which
types of consumers end up better versus worse off after the chain adopts a more
flexible pricing policy.
Similarly,
welfare rises in areas with greater mean household values for all categories. We
also observe welfare falling in areas with larger ethnic populations. In general,
we do not find much relationship between age and welfare changes. Interestingly,
zone-pricing appears to lower welfare in areas where households have a higher
ability to shop. But. store-pricing raises welfare in stores catering to consumers
with greater ability to shop. This outcome is not surprising since the store-pricing
will clearly favor stores with more price-elastic demand. In terms of competition,
the welfare implications of proximity to a Jewel varies across categories and are
quite small. The impact of proximity to an EDLP store is even smaller. Overall,
it would appear that demographics are more influential for driving patterns in
welfare changes than proximity to competitors.
While our results showing greater profits from the store pricing are consistent
with those from Montgomery (1997), as noted before our analysis is different in
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
58
nature.
Naturally,
higher prices could also encourage new entry by a competing chain not currently in
the territory. Lower prices, on the other hand, could evoke competitive retaliation,
although they could also deter entry into the market. These issues are more long
term considerations that need to be balanced carefully with our short run findings
if they are to be used as inputs into chain strategy. Based on our current results, it
would appear th at managers would be better off allowing for price discrimination
in the laundry detergent category than refrigerated juice.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
59
metrics. In the current context, two such metrics are the category managers valua
tion of a pricing policy, variable profits, as well as consumers valuations, Hicksian
compensating variation. In this section, we leverage both these metrics to pro
pose an overall welfare-enhancing pricing strategy for the retailer. Store managers
may be concerned that extracting too much value from consumers could gener
ate store-switching in the long-run. Previous research has experimented with ad
hoc constraints on pricing, such as holding the average price level fixed. In the
current context, we are able to propose pricing such th at consumer value is fixed.
The structural approach generates a natural theory-based constraint - consumer
welfare. We propose profit-maximizing store-level prices th at are constrained to
offer consumers at least as much surplus as a chain-level pricing policy. The prices
under this policy will, by construction, make consumers better off than under
chain-pricing or, at the very least, leave them indifferent between the two policies.
Formally, the problem involves solving the following problem for each store s in
each week t:
j
max n st =
y_(p3Jt -
wjt)Qsjt
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
60
In table A. 12. we report the resulting change in profits and consumer welfare
associated with such a policy in row labeled constr. store. As expected, the
constraint prevents the category manager from generating the same additional
profits as under the unconstrained store-specific pricing policy of the previous
section. However, even with the constraint, the manager is able to generate roughly
half the gains of the store-pricing, an improvement in profits of 5.6% over a u niform
chain-pricing policy, in the laundry detergent category, and 7.4% in the refrigerated
orange juice. At the same time, overall consumer welfare rises, especially in the
refrigerated juice category where unconstrained pricing led to overall losses to
consumers.
A second consideration regarding the unconstrained pricing of the previous
section is the potential complexity of coordinating a store-specific pricing policy
across the 83 stores. Clearly, one advantage of a zone policy is the simplification
of price-determination. The previous section demonstrates the ease with which
an aggregate database can be leveraged to learn about differences in consumer
willingness-to-pay. However, changing item prices in 83 stores could be costly
from an implementation point of view. Therefore, we propose an alternative zone
pricing policy. Using the store prices computed in the previous section, we con
struct share-weighted price indexes for both the refrigerated juice category and
the laundry detergent category. Using the 83 price indexes, we then rim a simple
non-hierarchical cluster analysis 15 to generate 5 zones. Constraining prices to be
15We use the non-hierarchical k-means-based cluster function in S ta ta version 7.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
61
the same across all stores within each of these 5 zones, we then re-compute the
profits and welfare levels that would prevail. In the row labeled cluster in table
A. 12, we find that this simple 5-zone structure still generates substantial profit
gains relative to the 16-zone pricing policy used by Dominicks during the time the
d ata were collected. An interesting point is the fact th at while the clusters offer
notable gains to the retailer (relative to chain-pricing and the actual zone-pricing),
consumers are better-off with 16-zone pricing policy used by D o m in ick s.
3.7. Conclusions
Using a detailed database including weekly store-specific margins, we are able
to estimate flexible demand systems capable of generating reasonable approxima
tions of the true data-generating process in several product categories. Rather
than impose parameter restrictions, as in Montgomery (1997), we resolve the typ
ical store over-pricing problem by controlling for weekly store-specific endogeneity
in prices due to unmeasured demand covariates. We use the demand system to
investigate the impact of zone pricing on firm profits and consumer welfare. Our
results show that flexible-pricing significantly improves store profits, especially in
the case of store-specific pricing.
category. For a necessity item like laundry detergent, we find conservative gains in
profits with small effects on consumers. However, for categories like refrigerated
orange juice, which exhibits far more demand heterogeneity across stores, we find
fairly large profit implications. At the same time, consumers experience differential
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
62
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
CHAPTER 4
in the form of mass discounters, price clubs, and supercenters, has transformed
not only the competitive structure of the industry but also the way in which con
sumers shop. While mass discounters like Wal-Mart and K-Mart do not offer a
comparable array of perishable and nonperishable products, they do compete with
supermarkets in specific high-volume categories such as paper products, dry gro
cery, health and beauty' care products, and other general merchandise items. More
recently, the line separating supermarkets from mass discounters has blurred as
supermarkets have added more general merchandise items to their shelves while
stores like Wal-Mart have transitioned towards the food sector through their su
percenter format.
Past research on supermarket competition (e.g.. Lai and M atutes 1989, Pesendorfer
2000)
uct offerings, cost structure, and pricing policies. More recently, researchers have
also analyzed competition between supermarkets with different pricing formats,
63
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
64
for example, EDLP vs. Hi-Lo pricing (Bell and Lattin 1998, Lai and Rao 1997.
Messinger and Narasimhan 1997). This stream of research has explored issues such
as the relationship between household shopping behavior and store preference (Bell
and Lattin 1998. Bell et al. 1998), the impact on store sales due to a change in
price format (Hoch et al. 1994, Mulhem and Leone 1990), and theoretical explana
tions for the general movement towards EDLP as a positioning strategy (Lai and
Rao 1997. Messinger and Narasimhan 1997). However, the focus in these papers
has been on the competition between grocery stores, with limited attention given
to alternative retail formats like mass discounters and price clubs.
In this paper, we provide an empirical study of the impact of entry by mass
discounters on the demand for products sold by conventional supermarkets. Using
store-level data from a large supermarket chain, we estimate demand systems for
several product categories and analyze how demand changes following entry by
discount stores. In particular, we investigate how entry impacts the total level of
demand as well as consumer price sensitivity in both the short and the long run.
We then combine our demand estimates with a model of category management to
measure the implications for category pricing and profitability.
A critical difference between mass discounters 1 and EDLP grocery stores (such
as those considered in the literature) is that mass discounters do not offer the full
grocery line found in supermarkets. In terms of store choice, this implies a very
xNote th at the focus in this study is on traditional discount stores like W al-M art and Target
th at do not sell perishables. C om petition between superm arkets and the supecenter format is
analyzed in the next chapter.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
65
different choice set for consumers than, say, a choice between a Hi-Lo or EDLP
supermarket. For example, consumers who prefer to buy all their groceries in a
single shopping trip would prefer to shop at supermarkets. Others may prefer to
take advantage of the lower prices of nonperishable goods at the discounter while
buying food items at the grocery store. This paper extends previous empirical
work in that it examines competition between retailers who are differentiated, not
only in their pricing formats but also in their product selection. In addition, while
the previous work has considered competition between stores and consumer store
choice in a static environment, our study considers both the short- and the long-run
impact of entry in an evolving competitive environment.
The limited academic research on inter-format competition is not entirely sur
prising for two reasons. First, the competition between supermarkets and mass
discounters has traditionally been limited to relatively few product categories.
The movement of mass discounters toward the food sector is a relatively recent
phenomenon. Second, supermarkets have typically priced their products against
other supermarkets and have viewed neighboring grocery stores as their primary
competitor. However, trends in the industry suggest that this view of competition
is flawed.
A recent survey by Supermarket News shows that while consumers point to
supermarkets as their favored destination for perishable items, they are turning to
mass discounters and price clubs to purchase nonperishable product lines. This
finding is also substantiated in a recent study, conducted by Proctor and Gamble.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
66
last decade supermarkets have lost over 21% share ($3.4 billion in sales) in these
nine product categories, while mass discounters and price clubs have more than
doubled their share. This lost share to other retail formats is quite alarming for
supermarkets who already operate with very low margins to stimulate high volume,
with average profit margins of only about one percent (FMI Annual Financial
Review 1999). Further, nonperishables account for 70% of a stores selling space,
driving the bottom line for most supermarket operators.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
67
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
68
facing each store. The long time series also allows us to tease out the short- and
long-term impact of warehouse entry.
To model demand, we use a random coefficient logit specification. This model
has recently become popular in both the marketing and economics literature for
modeling demand for differentiated products (e.g.. Berry, Levinsohn and Pakes
1995. Nevo
2000.
parsimonious than a linear or a log-log demand model. For example with N brands,
a linear or log-log demand requires estimating N 2 own and cross price parameters.
W ith a logit demand specification, consumer preferences are projected onto a set of
exogenous product attributes, which greatly reduces the dimension of the estima
tion problem. Further, by allowing consumer tastes to vary in the population, the
model is not subject to the ^proportional draw property of the homogeneous logit.
Unobserved heterogeneity in consumer tastes plays three im portant roles. First,
since consumers are endowed with different marginal utilities from observed prod
uct characteristics, they react differently in response to promotions and changes
in the marketing mix by different brands. Second, since marginal utility from a
particular product characteristic is constant across alternatives, consumers find
products with similar characteristics to be better substitutes. Finally, in our ap
plication. allowing for consumer heterogeneity provides us with in s ig h ts into not
just the changes in the mean price response over time but also its distribution in
the population.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
69
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
70
significantly lower, even lower than the pre-entry period. In addition, the distribu
tion of the price sensitivity is much tighter. Our results indicate that while DFF
lost significant volume due to warehouse entry, the lost sales seem to have come
at the expense of more price sensitive shoppers. Consumers who continue to buy
paper products at Dominick's are relatively homogenous and less price sensitive.
In the long run, Dominicks is left with lower but less elastic demand for paper
towels and bath tissue. Combining our demand estimates with a category profit
maximization framework, our results indicate that it may be optimal to raise prices
in the long run. Though not entirely obvious, this is consistent with Hauser and
Shugan's (1983) result that if the market is highly segmented, it may be optimal
to raise prices with competitive entry. Similar evidence is found by Ward et al.
(2000). who examine the impact of private labels on price response by national
brands. The authors find that an increase in share of private-Iabel goods leads to
an increase in the price of national brands.
The rest of this essay is organized is follows. In the next section we provide
an overview of the relevant literature. Section 3 presents the demand model and
discusses the main aspects of the estimation procedure. Section 4 provides an
overview of the data we use for estimation. Section 5 presents results. We conclude
in section 6.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
71
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
72
response depends on the position of the incumbent, and that it may be optimal
for the incumbent to raise price when it is located farther away from the entrant.
Similar intuition is provided in Judd (1985). The author uses a model of spatial
competition to analyze the response of a multi-product monopolist to entry by a
single product firm.
closest product to the entrant in order to soften price competition and preempt
a price war. The interesting finding here is that in fact it might be optimal to
concede a segment of the market to the entrant rather than compete for this
segment through lower prices.
Perloff et al.(1996) show th at market entry can cause an incumbent monop
olist's price to rise or fall in a model of spatial competition. The author's show
that the incumbent's optimal price response depends on how close the entrant lo
cates in the product space. The price declines when the entrant locates sufficiently
close to the incumbent, rises when the entrant locates further away, and remains
unchanged if the products are sufficiently distant. In their study of anti-ulcer pre
scription drugs, the authors find empirical evidence for their model - price rises for
an incumbent drug in response to entry by a distant therapeutic substitute, but
falls in response to entry by a closer substitute. Similarly. Davis et al. (2001) ana
lyze entry, pricing, and product design in a model with differentiated products and
show that, under plausible conditions, entry into an initially monopolized market
leads to higher prices.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
73
Evidence also exists in the empirical literature th at market entry can lead
to no response, or in some circumstances an increase in price. In an extensive
study. Robinson (1988) examines the marketing mix reactions by the incumbent
for 115 entrants into oligopolistic markets, and finds evidence of minimal reaction
by the entrant. Ward et al. (2001) study the impact of private label entry and
market penetration on brand name pricing for 32 product categories. The authors
use scanner data on retail prices and find a strong positive effect of private label
market share on brand name prices. Similar phenomena have also been observed
in many pharmaceutical markets. For example, Frank and Salkever (1992. 1997)
find th at name-brand prices increased after entry by generic products. Similarly,
Bresnahan and Reiss (1990) infer that variable profit margins rise modestly when
a second car dealership enters a market initially served by a single dealership.
Finally, in a different setting, Kadyali et al. (1999) find an interesting result where
a product line extension by a brand in a competitive setting leads to an increase
in prices and profits for both the extending firm as well as the rival.
A general consensus in the theoretical papers discussed above is that segmen
tation of preferences softens price competition. This means that when consumers
are heterogenous, entry by a differentiated competitor th at fosters market segmen
tation can lead to an increase in prices. Our focus in this study is on analyzing the
impact of entry by mass discounters on a supermarket chain. A relevant question
th at arises is why entry by a warehouse store is different from entry by any other
supermarket?
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
74
discounter does not offer the frill array of products, there will be a segment of
consumers who will continue to purchase paper towels at the incumbent in order
to avoid multiple shopping trips. Further, the new entrant is likely to attract con
sumers who are different from the segment that continues to shop at the incumbent
supermarket. They may be shoppers who have lower search costs, are heavy users
of paper towels, are more price sensitive, etc. Thus, the post-entry demand for
paper towels and the price response by the incumbent will depend on the size and
the characteristics of the customers th at continue to shop at the store.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
To formalize the above argument, consider the following simple model. Suppose
there are a total of N consumers, th at are uniformly located along a segment of
unit length [0,1]. The grocery store, also called the incumbent, is located at
tj (0 . 1 ). Next suppose a discount store E, also called the entrant, enters the
market, initially monopolized by the grocery store, and locates right next to the
incumbent, i.e. tE tj. Both stores sell multiple products but compete in a single
good g which is sold at prices p/ and Pe respectively. In what follows, we assume
that the incumbent and the entrant are located sufficiently far from the endpoints
of segment [0 . 1 ] so that the endpoint considerations do not enter our analysis.
Both stores do not have fixed costs, and their constant marginal costs of procuring
good g are equal to m >
is. we assume that the discount store has a lower cost of procuring good g than
the incumbent. This assumption is motivated by the tendency of discount stores
to have more bargaining power in negotiations with wholesalers.
We let c denote the transportation cost per unit of distance, and let u be the
value of good g to the consumer. Each consumer has a choice of visiting one store,
two stores or neither store. In addition to transportation costs, the consumer
making a shopping trip to a store incurs a shopping cost s. The shopping cost s
may include the cost of finding the product, time spent in the line, etc. Thus, if
a consumer visits both stores, her toted cost is equal to the total transportation
costs plus 2s. Besides their location, consumers differ along another dimension.
We assume that there are two types of consumers differentiated in terms of their
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
76
utility (in addition to the net utility from good g) of visiting a particular store. A
proportion a of consumers obtain an additional utility
ue
store. We call these consumers type-E to reflect their preference (all other things
equal) for the discount store. Similarly, Type-/ consumers, th at constitute
fraction of all consumers, obtain an additional utility u/ from visiting the grocery
store, regardless of whether they buy good g there or not. For example, these
could be shoppers that have demand for not just product g but also for other
food products sold by the grocery store. Similarly, if a type-E consumer visits
the discount store, she obtains utility
ue
there or not. The incumbent and the entrant are uninformed of the type of each
consumer and they only have information on the fractions of type-/ and type-E
consumers. We also assume th at each consumer is of type E with probability a and
of type / with complementary probability 1 a . T hat is. location of a consumer
and her type are independent.
Each consumer has an option of buying from an outside source instead of
shopping at the two stores. We assume that if a consumer chooses this outside
option, her utility is equal to U = 0. Note that the reservation utility is equal to 0
irrespective of consumer location and type. For the purposes of this example, this
assumption is without loss of generality.
To establish a benchmark case, we start with the scenario where the market
consists of a single grocery store. Note that a consumer may travel to the grocery
store for one or two reasons. If she is type-E, she may travel to the grocery store
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
77
solely for the purpose of buying good g there. While if she is type-/, there is an
additional reason of receiving utility U[ from visiting the grocery store. As long as
the value of good g to the consumer u exceeds the price p
by the incumbent,
any consumer that visits the grocery store buys good g there. The incumbent will
set p7* < u since otherwise it will lose all of its customers. A consumer buys a
unit of a good as long as the net value of the good does not exceed its reservation
price. VVe assume that u m s > 0, i.e.. there are gains from trade between
the incumbent and type-E consumer (and. hence, type-/ consumer) located right
next to the incumbent (t = t[). Let
ve
UJ{t.tl)
=Vj
- p f -c\t-t[\-s.
Thus, type-/ consumer located at t buys from the incumbent if and only if
(2)
XJm =
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
78
(5)
m g 2 . v { - ^ r - - } ( p r - m ).
Solving the first-order necessary conditions, which are also sufficient due to
concavity of the objective function in the choice variable p, we find the profit
maximizing price:
(6 )
p'P = | [u + (1 a)u[ + m s ] .
7Tm
= [u -r (1 - a ) u { - m - s ] 2 .
Now suppose the discount store E enters the market monopolized by the gro
cery store and locates at tg = tj. With the entry of the discount store, the con
sumers have a choice of visiting one of the two stores, both or neither. One would
expect th at for positive values of s the relative values of s, u/, ug- and m will
determine the equilibrium behavior. Given the prices set by the two stores, the
consumers decide how many stores to travel to and from where to buy good g.
Type-/ consumers choose among the following options: visit store I and buy good
g there, visit both stores and buy good g at store E. visit only store E and buy
good g there, and do not travel at all. If type-/ consumer located at t buys good
g from store / . her utility is given by
(8 )
If type-/ consumer visits store / but buys good g from store E, her utility is
given by
(9)
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
79
If ty p e-/ consumer visits store E and buys good g there, her utility is given by
(10)
Finally, if type-/ consumer does not visit any stores her utility is equal to 0.
Thus, ty p e-/ will buy good g from store I if and only if
u + ui - p dE - c \ t - tr | (1 1 )
2 s;
.
Similarly. type-E consumer decides between the following four options: visit
0
store E and buy good g there, visit both stores and buy good g at store I. visit
only store / and buy good g there, and do not travel at all. If type-E consumer
located at t buys good g from store E. her utility is given by
(12)
U( tJr ) = u + uE - p E
l - c \ t - t I\ - s .
If type-E consumer visits store E but buys good g from store /, her utility is
given by
(13)
U{t. ti) = u +
ue
If type-E consumer visits store I and buys good g there, her utility isgiven by
(14)
Finally, if type-E consumer does not visit any stores, her utility is equal to 0.
Thus. type-E will buy good g from store E if and only if
u -H uE p j c\t f/| 2 s:
(15)
u -r ue
u -pf} - c\t - /| - s;
0
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
80
If consumers have zero shopping costs then they buy good g at the store that
offers the lowest price. In this case (where s =
travelled to location t. = tj. always visits store j. Thus, if store j is the one that
offers the lowest price of good j. type-j only visits store j. When s = 0. it is
straightforward to derive the Bertrand Nash equilibrium. The discount store sets
price equal to m, all consumers that travel to location tg = ti buy good g from
the discount store. Type-/ consumers also visit the grocery store but solely for
the purpose of collecting payoff U[. Thus, in this case, entry results in decrease
in price of g and an increase in consumer welfare. Obviously, the grocery store is
negatively affected by entry.
The simple analysis above points to the important role played by the shopping
cost. In our empirical analysis, we find many instances where discount stores
entered next to the incumbent supermarket. While these supermarket stores lose
sales in the competing products, the demand does not go down to zero, which
would be the case if discounters had lower price and consumers had no shopping
costs.
W hat happens when there are non-zero shopping costs (i.e. s > 0)? When u/
is sufficiently large such that it is greater than the difference in price between the
two stores, the equilibrium outcome entails type-/ consumers shopping at store /
and type-E consumers shopping at store E. Under this scenario, both firms do
not have an incentive to undercut the price set by the rival since it would have to
decrease the price by an amount directly proportional to the shopping cost. And
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
81
if the shopping cost is sufficiently large, a firm is better off serving the segment
of the consumers that have a preference for its product than trying to attract all
consumers. The equilibrium price set by the incumbent in this case is:
(16)
pj =
[u + U[ + m s ].
Pe = \ [u + uE - s].
<tf = - v 'V [u + U{ - m - s \ .
(19)
cfe =
[u + uE - s ] .
7rj
('l~2c' 'V (u +
(2 1 )
(1
Q)u i na s )2 and
[u + a u E - s}2 .
respectively.
Thus when u{ is sufficiently large, the impact of discount store entry is to
segment the market. The incumbent serves the segment that derives extra utility
from shopping at its store. The sales and profits depend on the size of this segment,
and the extra value they attach to shopping at the grocery store. Also, comparing
(6 ) and (16). we observe that it is optimal for the grocery store to raise the price of
good g after entry, i.e. pj > p. VVe next turn to empirical investigation of some
of these issues.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
82
1,...,
T), M t
consumers each select one of J brands in the category or opt for the no-purchase
alternative. The probability that consumer h will purchase product j in store-week
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
S3
t is given by:
(4.1)
h =
H .j = 1
J,t =
T.
The parameter a hj captures household h's intrinsic preference for brand j. The
variables pJt and dJt are the price and promotion dummy for product j in storeweek t. while the parameters 3 ht and 6h capture household h's price and promotion
sensitivity. The vector. encompasses the effects of unobserved (to the econome
trician) in-store product attributes, such as advertising, shelf-space and coupon
availability that vary across store-weeks (BLP 1995. BGJ 1998). Finally, the vari
able Zt captures the demographic and competitive environment in store-week t.
The model also allows for the no-purchase option, which allows category demand
to increase or decrease based on the marketing mix in the category.
The prob
ability th at household h will not make a purchase from the category in week t is
given by:
(4.2)
P m = -----------
By making the parameters household specific, the model allows the parame
ters to vary across households. There are two popular approaches to account for
consumer heterogeneity: parametric random effects (see for example, Gonul and
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
84
J*
d + lZ t
Ok
+ A i//,, where
~ N (0. /)
where the means. 3 and 6, the impact of demographics and competitor variables,
7
structure:
Q/t ~ N (a. Z ) .
The E m atrix represents the covariance matrix of preferences across consumers
which are allowed to be correlated across brands. In the estimation, we impose a
factor structure on the covariance matrix (for details see the model section in the
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
85
previous chapter).
E = Lu>hu hL \ u/k ~ N (0 .1 ) .
For both exposition and computation, it is useful to separate Vhjt = exp(ahj
(3 h + 7 Zt )pjt + QhdJt + Jt) into two components: a part common to all households
and one that differs across consumers:
LuJh
qJt
qjt =
h =
7 -o o
exp(<5jt + n hjt)
j----------- -- o { v ) d v .
exp(AZt) + l= 1 exp( 6 tt + p h l t )
H ,j = 0
r
I
J .f = l , . . . . r .
where () is the pdf of a standard normal. The integral in (4.3) represents the
share of consumers entering the store in week t that purchase a unit of product j,
and defines the product level demand curve. Thus, expected demand for product
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
86
j in store-week t is :
(4.4)
Qjt = qjtMf
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
87
We supplement the sales data with an extensive set of demographic and competitive
variables that capture the environment facing each store, at each time period.
Zip level demographics for each store is obtained from the Sourcebook of Zip
Demographics. These d ata are available for years 1990, 1993. 1994. and 1997.
We also use an extensive set of competitive variables. The main source of
competitive information is the database maintained by Trade Dimension Inc. The
company provided us with information on location (latitude, longitude) for vari
ous kind of stores including supermarkets, superettes (small grocery stores), mass
discounters (Wal-Mart. Kmart and Target), price clubs, supercenters, convenience
stores, and other liquor and cigarette outlets. The database also includes descrip
tive information on each store including store format, chain affiliation, store size,
number of employees, number of checkout counters, lines of business including
additional services like pharmacy, bakery and so on.
In addition to the store locations, we also use information on the opening
dates for these stores. The data on opening dates of the competitor stores allow
us to explicitly account for changes in the competitive environment facing each
store. This information was obtained from various sources. First, Trade Dimension
provided us with the dates th at the stores were entered in their database. While
this is not the exact store opening date, it provides us with a good proxy as
the database is updated on a regular basis. In addition, we contacted the major
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
88
competitive chains directly and obtained the opening dates for all their stores
operating in the Chicago area. We were able to obtain this information from
the two major grocery chains: Jewel (the largest supermarket chain in the area)
and Cub Foods (the largest EDLP operator). Similarly, we were able to get the
store opening dates for two main mass discount chains- Wal-Mart and Target.
Incidentally, these two chains entered the Chicago market during the period of
study. For all other stores, we use the dates provided by Trade Dimensions as
proxy for when these stores opened2.
Table B.3 describes the demographic and competitive variables used in the
study. Descriptive data is provided for years 1990 and 1996, the first and the last
years in our data set. We use two demographic variables. Ethnic (% of popula
tion th at is Black and Hispanic) and Income (median income in the store area).
Previous research (for example Hoch et al.) has found these variables useful in
explaining demand differences across regions. These variables vary significantly
across stores. For example, DFF stores cater to market areas with Black and
Hispanic representation ranging from 2 to 99% of the population. Similarly, the
median income ranges from under $2 0 , 0 0 0 to over $75,000. Overall, we find that
the proportion of ethnic population as well as median income has gone up over
time.
2In general, we found the trade Dimension information to be fairly reliable. In most cases, stores
were entered in their databse within three m onths of opening. However, in some intances, the
stores appeared in th e database much after they actually opened.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
89
The next five rows of Table B.3 describe the competitive variables used in
the study. The first two rows show the average distance in miles to the nearest
supermarket and drug store. The competitive environment with respect to the
grocery and drug stores does not seem to have changed much over time. On the
other hand, the competition from mass discounters has changed quite dramatically.
While the average distance to the nearest mass discounter was over 5 miles in 1990.
on an average, there was a mass discounter within 2 miles in 1996. This is mainly
due to entry by two large discount chains, Wal-Mart and Target stores, into the
Chicago area. Further, while there were no price clubs in 1990. by the year 1996
this number had grown to over
20
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
90
between the lowest and highest prices. In week 200. the number of price levels has
increased to 15, and the dispersion has increased to 26 cents. Similarly, in week 10,
the margins ranges between 11 and 25 %, whereas for week 200 it varies between
3 and 26%. Note th at these are prices and margins are for a single UPC so that
there is no aggregation bias causing the observed price dispersion.
We next explore what factors may be critical in defining the zones. In the
previous chapter we saw that the demographic variables explained some of the
observed price differences across zones. Here we focus on the role of the competitive
variables in zone definition. To do this we examine the characteristics of the
stores at the opposite ends of the price distribution. Understanding the differences
between these stores provides insight into why they are priced differently, and
provide some rationale for the zone definitions. Table B.5 shows the distance to
nearest competitors for the four stores charging the lowest prices (Stores 59, 70,
77, and 83) and the five stores charging the highest prices (Stores 14. 32, 52. 53,
95). For the stores with the lowest price, we find that in all cases there are at least
two mass discounters located less than half a mile from the store. In contrast, for
the stores charging the highest prices, there are no stores with a mass discounter
located within one mile of the store. Although not reported in the table, the
stores with higher prices (with the exception of Store 95) have significantly higher
income as well as a higher proportion of elderly population. These comparisons
show th at both demographics and competition, particularly from mass discounters
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
91
and EDLP operator, appear to play a critical role in zone definition. It also appears
that stores under threat from mass discounters are priced most aggressively.
In addition to an increase in the number of price zones, we also observe stores
transition from high to low price zones over time: i.e.. stores that were initially
priced high, may later transition to the low price zones, and vice versa. To explore
how changes in the competitive environment may influence transition between
zones, we consider the case of two stores that were originally priced in the same
zone but later followed very different price paths. Figure B.5 shows the price series
for Stores 14 and 71. For the first 100 weeks, both stores are in the highest price
zone so their regular price and promotions are the same. After week 100. when
the number of price zones increases from three to fourteen, the price series of the
two stores diverge. The price at Store 14 increases as it moves to the highest of
the fourteen price zones. The price for Store 71 decreases to 95 cents, putting
it in one of the lower price zone. The difference in price between the two stores
is 16 cents. With occasional exceptions. Store 71 continues to be priced lower
till about week 280, after which all stores have the same price. Table B.6 shows
the competitive environment of the two stores at the end and beginning of the
observational period. The competitive environment of both stores has evolved
over time. However. Store 14 does not have any discount store located close to it,
while Store 71 has several large discount stores located within a few miles of it.
This indicates that a possible reason for Store 71 to be priced more aggressively
than Store 14 is because it operates in a more competitive environment.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
92
4.5. R esults
We now discuss the results from the three product categories: Paper towels,
Bath tissue, and Refrigerated juice. Dominicks competes with mass discounters
in the paper products category, but not in refrigerated juice. We present the
descriptive statistics for the three categories in Table B.7. These data consist
of means across all store-weeks. Share corresponds to the conditional shares
for each brand. The actual market shares used for estimation are computed as
total brand sales divided by the total weekly store traffic. Effectively, our model
implies that each time consumers visit the store, they either purchase a unit of
one of the alternatives or they elect not to purchase. The promotion variable is
the proportion of units sold on promotion. Margin is the average profit margin
on the brand. The profit margins vary significantly across categories as well as
across brands within the categories. Paper products have lower margins than juice
as they also face competition from non-grocery outlets like mass discounters and
drug stores. W ithin each category, the store brand tends to have higher margins
than the national brands.
We plot the sales for the three categories in Figures B .l, B.2, and B.3. The
figures are for the entire chain, i.e. aggregate across all stores. For paper towels and
bath tissue, we find a strong downward trend while the sales for refrigerated juices
are trending up. This is mainly due to entry by Wal-Mart and Target stores which
compete with D FF in paper products but not in juice. The impact of warehouse
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
93
stores on category sales is also reflected in the regressions presented in Table B.8.
The dependent variable is log of category sales while the price index is the log
of share weighted prices. The competitor variables: Wal-Mart, Target, K-Mart,
and Price Club enter as dummies and represent the presence of these stores in the
trading area (within 3 miles). We also included a dummy for the presence of a
pharmacy department. While the presence of a pharmacy is unlikely to impact the
category sales directly, it can have an indirect influence via store traffic. Finally
the regression also includes a dummy for holiday weeks and store fixed effects.
Results show th at the presence of a warehouse store significantly reduces the
category sales for paper towel and bath tissue but has a negligible influence on sales
for refrigerated juices. In relative terms, the impact seems to be higher on the bath
tissue category than on paper towels. Finally, the presence of pharmacy seems to
significantly increase sales in all categories. As seen in the data description section,
Dominicks has added a pharmacy department to a number of stores over time. In
Figure 5 we plot the store traffic and paper towel sales for Store # 9 in the data set.
DFF added a pharmacy department in week 272 which significantly increases the
store traffic and thus paper towel sales. However, it should be noted th at adding
a new department generally involves some remodeling of the store. Thus some of
the effect on store traffic and paper towel sales could be due to these unobserved
factors related to store remodeling.
We present the results from our demand models in Tables B.9, B.10, and B .ll.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
94
P a p e r Towel: The first set of variables are the intrinsic brand preferences.
Note that all of the brand constants are negative because of the large share of
the no purchase option. In relative terms, we find th at Bounty has the highest
preference level followed by Scott. This is consistent with their high shares. The
next 3 variables pertain to the effects of holiday and competitive variables on
the utility of the outside good. In other words, these variables shift the overall
category demand. Holiday and Price Club variables enter as dummies indicating
a holiday week and presence of a Price Club in the store trading area. However,
both of these variables turned out to be insignificant for this category. The third
category- shifter, warehouse, is the number of warehouse stores in the store trading
area. We find that the presence of warehouse stores in the store trading area
lowers the overall category demand. In terms of the marketing mix variables, we
find promotions to increase the likelihood of purchase. Price sensitivity has the
correct negative sign and we do observe significant heterogeneity in consumer price
sensitivity. Interestingly, we find that price sensitivity rises for promoted items.
The next seven variables describe the impact of store demographic and com
petitive variables on price sensitivity. Price sensitivity tends to be higher in store
areas with large ethnic population, while stores in wealthier neighborhoods tend
to be less price sensitive. We also find that proximity to a grocery store increases
price sensitivity while proximity to a drug store has no influence. Recall th at the
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
95
interpretation here is that as the distance increases, the price sensitivity falls. Fur
ther, we find th at presence of a price club in the store trading area significantly
increases consumer price sensitivity.
The last two variables "Ware SR. and "Ware LR capture the impact of ware
house stores on consumer price sensitivity. Ware SR is a dummy variable that
takes a value one in the first 12 months after a entry by a mass discounter in
the store trading area. The dummy "Ware LR captures the long run impact
of warehouse entry and takes a value one after the first 12 months of warehouse
entry unless another warehouse store enters in the store trading area (in which
case Ware SR takes a value one again for the next 12 months). The parameter
estimates show th at while the short run impact of warehouse entry is to increase
consumer price sensitivity, in the long run entry by a warehouse store significantly
reduces consumer price sensitivity.
Why would entry by warehouse stores reduce overall consumer price sensitivity
at the Dominick's store? Recall th at entry by a warehouse store also reduces
overall category sales. One reason why the price sensitivity goes down in the long
run is th a t this lost sales may be coming at the expense of more price sensitive
consumers. In other words, entry by a warehouse store causes the more price
sensitive consumers to shift their purchases to the warehouse store. Thus, in the
long run, DFF is left with lower but less elastic demand for paper towels. We
examine this issue in greater detail in the subsequent section.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
96
B a th T issu e: In the bath tissue category we find that Scott has the high
est preference followed by Charmin. It is interesting th at Scott has the highest
preference despite its lower share, suggesting that it is important to control for
marketing mix variables. Unlike the paper towel category, we find th at presence
of a Price Club significantly lowers the category demand for bath tissue. While
this is consistent with the regression results presented in Table B.8, it is unclear
why presence of a Price Club would lower sales in bath tissue and not in paper
towel. In terms of other category shifters, we find that warehouse stores lower the
overall category demand while holidays have no impact. As expected, promotions
increase the probability of purchasing a brand and prices lower it. The price sen
sitivity exhibit significant heterogeneity. Also, price sensitivity rises for promoted
items. In terms of price interactions, we find that store areas with a large ethnic
population are more sensitive. Further, presence of a Price Club significantly in
creases consumer price sensitivity, while proximity to conventional competitors like
other grocery stores has no impact. As in the paper towel category, we find that
the short-run impact of warehouse entry is to increase consumer price sensitivity,
while in the long-run, price sensitivity is significantly lower.
R e frig e ra te d Ju ices: Consistent with its dominant position in the category,
Tropicana has the highest mean effect. We also find that store areas with warehouse
stores have higher sales for juices. It is unclear why this may be case, although
it is possible th at warehouse stores attract consumers to neighborhood from a
larger area which causes some spillover to the DFF stores. Holiday weeks have
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
97
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
98
price sensitive consumers, then we should find a change in not just the mean price
response but also its distribution in the population.
To get a better insight into this issue, we divide our data into 3 time periods:
pre-entry, period of entry, and post entry3. Recall that we have data available
from 1990-96. a period which saw entry by two major discount chains - Wal-Mart
and Target stores. Wal-Mart entered the Chicago area in early 1990 and opened
between 5 to 7 stores every year till 1995. Target stores, on the other hand, entered
the market on a large scale, opening 25 stores between March 1993 and October
1994. In what follows, we will call the period between January 1993 to December
1994 as the period of entry. We refer to the time period before this as the pre-entry
period, while the last two years in the data 1995-96 will be called the post-entry
period. We estimate the demand models for the three time periods using data from
all available stores. Table B.12 reports the mean price sensitivity and its standard
deviation for the paper towel category. Results for the full models along with the
factor loadings are reported in the appendix.
Consistent with results in the previous section, we find that entry by warehouse
stores leads to an increase in the overall price sensitivity in the short nm (period of
entry), while in the long nm the price sensitivity is significantly lower, even lower
than the pre-entry period. At the same time, the distribution of price sensitivity
is much tighter. To visually examine this effect, we plot the mean price effect with
unobserved heterogeneity in Figures B.3 and B.4 for pre and post-entry periods.
3T his analysis is only for Paper Towel Category.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
99
The results seem to imply that in the long run, consumers who continue to pur
chase paper towels at DFF are not only less price sensitive, but are also relatively
homogenous.
W hat do these results imply in terms of retailer pricing? To address the pric
ing issue, we need to make some assumption about retailer pricing behavior. Most
studies in marketing assume that retailers set prices for different brands in a prod
uct category to maximize total category profits (for example Kim et al. 1995, Raju
et al. 1995. Vilcassim and Chintagunta 1995 etc.). This objective is also consistent
with the industry trend towards category management.
Formally, we assume th at the retailer jointly sets the profit-maximizing prices,
Pj, of each of the J products in a category. We also assume that the retailers vari
able costs consist solely of wholesale prices, Wj and treat all store an d /o r categoryrelated overhead as fixed costs, Ft. Thus, in week t. a retailer solves the following
optimization problem:
j
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
100
(4.5)
n(p - w) + Q = 0 ,
which represents a system of J equations, one for each brand. The optimal set of
retail prices for the monopolist are determined by solving:
(4.6)
p = w Q - I q
retailer.
We report the predicted margins in Table B.13 . The numbers represent the
median across all store-weeks. Not surprisingly, we find that in the long-run it is
optimal to raise prices. This is of course a direct consequence of low price sensitivity
estimates. Looking at the observed margins, we find that while DFF did increase
margins in some products, they are well below the levels suggested by the model.
There are a number of reasons for this. First, the objective of the retailer may not
be category profit maximization. For example, they could be using the category
for traffic generation (Chintagunta 2001). Similarly, our model does not capture
any cross-category effects. Finally, the category profit maximization model above
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
101
assumes static behavior. In reality, the retailer may be concerned about factors
like store price image which could affect long-term profitability.
In spite of the above mentioned limitations, our results do suggest that under
the assumption of category profit maximization, an optimal response to entry by
warehouse stores should be an increase in price. Though not entirely obvious, this
is consistent with some of the theoretical models discussed in Section 2.
4.6. Conclusion
This paper studies the impact of entry by mass discounters on the demand
for products sold by conventional supermarkets. Alternative retail formats like
warehouse stores, price clubs, and supercenters have shown a dramatic growth in
the past few years and pose a formidable challenge to supermarkets. However,
cross format competition has received limited attention in the marketing litera
ture. We analyze the impact of warehouse entry on sales of a supermarket chain in
the Chicago area. Using a rich scanner database, we develop demand models for
three product categories and analyze the short and long-term impact of warehouse
entry. Results show th at entry by warehouse stores resulted in s ig n ific a n t loss of
sales in competing product categories. In terms of consumer price sensitivity, we
find that the short-term impact of warehouse entry is to increase price sensitivity.
More interestingly, we find th at the long-run mean price sensitivity in competing
products (paper towel and bath tissue) is significantly lower. Combining our de
mand estimates with a category profit maximization objective, our results show
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
102
that it may be optimal to raise prices as a result of competitor entry. There are
of course several caveats to our analysis. First and foremost, we do not have any
pricing information on the competitive stores which limits our ability to explicitly
model retail competition. Second, our demand models do not consider any cross
category interactions. Furthermore, due to the aggregate nature of the data, we
can not determine if the lost sales are actually coming from price sensitive con
sumer shifting their purchases to the warehouse stores. For example, it is possible
that all consumers continue to purchase paper products at DFF stores but some of
them have reduced the purchase quantity. We consider some of these issues with
individual level data in the next chapter of the thesis.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
CHAPTER 5
lW almart plans to add between 150 to 180 new supercenters every year, a num ber of these being
converts of the exisiting discount stores.
103
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
104
2As we discuss in the next section, over 75% of the households in our database own a house and
have lived a t their current residence for 14 years on average.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
105
holder information (if a shopper card is used), total expenditure, unit price, quan
tity. and coupon usage for every UPC sold. Access to rich transaction-level data
allows us to investigate not only the impact of Wal-Mart's entry on overall store
sales but also the changes in store-visit frequency, basket size, and category demand
for each individual. We decompose the observed losses in store sales into compo
nents attributed to customer attrition, loss of store traffic, and reduction in basket
size. More im portant, we combine the transaction data with household-specific
demographic and shopping variables to identify and characterize the consumers
who contribute to the observed demand changes at the store level.
From the retailer's perspective, the ability to study these issues at the customer
level can have a significant impact on store-level marketing policies. For example,
if the lost sales are due to a decline in store traffic while the basket size remains
constant, then the focus should be on strategies to bring customers to the store,
such as via feature advertisements. On the other hand, if the customer count
is relatively constant but basket size decreases, the focus should be on in-store
merchandising to increase basket size. Similarly, identifying and characterizing
consumers who defect or decrease spending allows the retailer not only to target
these customers but also to identify consumers with sim ila r characteristics who
may also be at risk of defection.
We also analyze the impact of Wal-Mart's entry at the product category level.
Retailers increasingly employ category management tools where each category is
treated as a strategic business unit, and pricing, merchandising, promotions, and
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
106
product mix are determined at the category level (Blattberg and Fox 1995). Thus,
it is critical for the category manager to identify which customers defect or reduce
expenditures in the category. We estimate a hierarchical Bayes logit model for three
product categories: bath tissue, milk, and laundry detergent. An advantage of the
Bayesian approach is that it allows parameter inference at the individual level.
Thus, customers can be characterized in terms of their sensitivity to marketing
mix or preference for specific attributes, and targeted coupons can be used for
category-level retention.
Our results show that the incumbent supermarket lost significant volume (19%)
following entry by the Wal-Mart supercenter. Further, these losses were higher in
the dry goods grocery, general merchandise, and health and beauty categories. At
the household level, we find that distance to the stores plays a critical role. In
particular, households living close to the incumbent are less likely to defect or
reduce their spending, while the opposite is true for the households living close
to the new entrant. These fin d in g s are consistent with retail site selection models
(Huff 1964, Brown 1989. etc.), as well as with findings from a recent survey by
Progressive Grocer where location was cited by consumers as the most important
factor in determining store choice. This type of information can also be employed
when d e t e r m in in g which customers to target with retention-oriented promotions.
W ith minor exceptions, our analysis finds that changes in consumer purchase
behavior following Wal-Mart's entry are not related to demographic variables. On
the other hand, shopping-related variables are quite useful in predicting which
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
107
customers will migrate to the discount store. Our results show that big-basket
households are more likely to switch their purchases to Wal-Mart, while customers
who spend a higher proportion of their expenditure on perishables are less likely
to do so. These findings are consistent with Bell and Lattin (1998), who find
th at large-basket shoppers are more likely to choose EDLP stores. Finally, the
brand choice models indicate th at consumers who stop buying in a category are
on average more price sensitive. This is especially true for consumers who stop
purchasing in the category but continue to visit the store. This finding lends strong
support to our results in the previous chapter.
The rest of this chapter is organized as follows. In the next section we discuss
the frequent shopper d ata used in the study. Section 3 investigates the impact
of entry by the Wal-Mart supercenter on store sales and consumer purchase be
havior. Category-level results are presented in section 4. which also describes the
hierarchical Bayes model. Section 5 concludes the chapter.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
108
The frequent shopper database used in this study comes from two supermarket
stores that are located 8.5 miles apart but are in separate towns. The data consist
of all transactions made in the stores and capture information such as time and
date of the transaction, card holder information (if a shopper card is used), and
the dollar volume, unit price, quantity, and coupon usage for every UPC sold.
Both stores have over 20,000 active card members. VVe use the store transaction
data to create purchase histories for all the customers. As seen in Table C .l , the
usage of the shopper card program is quite high, with card holders accounting for
over 85% of total store sales3. Although non-card purchases account for 32% of all
transactions, the average order size is significantly lower (average of $8 compared
to $34 for transactions using the shopper card). The non-card purchases often
tend to come from the coffee or snack shop and the pharmacy. For most grocery
categories, the shopper card penetration rate is well over 90%.
D ata are available for a period of 20 months, from November 1999 to June 2001.
In August of 2000, a Wal-Mart supercenter entered 1.8 miles from one of the stores
(hereafter referred to as the focal store) and 7.1 miles from the other (referred to as
the control store4). We observe reasonably long time series both before and after
Wal-Mart "s entry. Further, since the stores are located in small suburban towns,
3About 2% of these sales are on the employee card and thus cannot be traced back to any
individual card holder.
4Note th at the control store is well within the trading area o f th e W al-M art, which industry
sources put a t 15 to 20 miles. As we shall see in th e empirical analysis, W al-M art entry also
impacted the sales of this store though to a much lesser extent.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
109
the d ata provide us with a unique opportunity to analyze the impact of entry by
a Wal-Mart supercenter in a relatively controlled environment.
In addition to their purchase histories, our database also contains the mailing
addresses for all the card holders. We use the addresses to compute each house
holds travel distance from the focal store and the new entrant5. Further, we use
demographic data from a market research firm, which provided us with block-level
demographic information. Summary statistics for the demographics and the dis
tance variables can be found in Table C.2. On average, consumers live about 3.5
miles from the focal store and 4.5 miles from the entrant. The average income is
over $50,000, although we find significant variation across households. One quarter
of the households live in rented dwellings. Finally, on average, these households
have been living at their current residences for about 15 years. In addition to
the distance and demographic variables, we construct variables th at describe the
shopping behavior of customers. These variables are discussed in the next section.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
110
figure indicates a significant change in the baseline sales of the focal store after
Wal-Marts entry. In Table C.3, we report the percentage change in sales, total
transactions6, and sales per transaction following entry. The first row reports the
overall change at the store level, while subsequent rows show the change at the
departm ent and category level. The numbers presented in Table C.3 are based on
a semi-log regression model which includes a price index7, a dummy for Wal-Marts
entry, and month fixed effects as explanatory variables (to conserve space we skip
the actual parameter values and just report the percentage change). All numbers
in bold indicate significance at the 5% level.
Table C.3 shows th at the focal store lost over 19% sales following Wal-Marts
entry. At the same time, sales per transaction have increased significantly, which
indicates that the loss of sales seems to be due to a reduction in store traffic.
This pattern is repeated in the department and category level regressions. At the
departm ent level, the largest losses occurred in the Grocery, General M e rc h a n d is e
(GM), and the Health and Beauty (HBC) departments, followed by Produce and
Meat. Higher losses in the packaged goods departments are to be expected as
these products are less differentiated across stores. The corresponding numbers
for the control store indicate that the proximity to the competitor plays a major
6At the store level, transaction refers to the number of unique baskets. At the departm ent (cat
egory) level, it refers to th e number of baskets th at contain at least 1 item from the departm ent
(category).
'T o capture the price environment a t the store (departm ent) level, we use the total promotional
volume in the store (departm ent). At the category level, the price index is a calculated as the
share weighted price of the top 12 UPCs.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
I ll
role. The sales in the control store are more or less constant (many of the numbers
are insignificant) or the losses are significantly lower than at the focal store.
Two key observations from the regression results presented in Table C.3 must
be highlighted. First, the volume lost by the incumbent is quite significant. This is
particularly troubling considering that supermarkets generally operate on a prin
ciple of low margins and high volume, with profits of only about 1%. The second
consistent pattern that emerges is that the lost volume is due to a loss in the
number of transactions, and the per transaction sales have actually increased post
Wal-Mart entry. There are two potential sources for the loss in store traffic: 1)
the store has fewer customers post Wal-Mart entry but there is no change in the
visit frequency of the remaining customers, or 2) the store has same number of
customers who now visit the store less frequently. In all likelihood, the answer
is a combination of the two. We investigate this issue further by analyzing the
purchase behavior of the top 10,000 customers (in terms of dollars spent prior to
Wal-Marts entry8). The selected customers account for 77% of card member sales.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
112
(1) Distance: household distance from the focal store and Wal-Mart (DistStore, Dist-WM)
(2) Loyalty: To create a loyalty measure, we classified the households into 4
groups (in descending order) based on their expenditures prior to WalM arts entry. The base is the lowest quartile- Loyal Q4 (i.e. households
who spent the least prior to Wal-Mart?s entry) (Loyal Q l, Loyal Q2, Loyal
Q3)
(3) Basket Size: Similar to the loyalty measure, households were assigned to
quartiles based on their average basket size prior to entry. The base is the
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
113
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
114
since industry reports indicate that Wal-Mart supercenters generally have signif
icantly lower quality of perishables9. Finally, for the demographic variables we
find th at higher income households have a lower probability of defecting, while
the number of children in the family have no significant impact. Overall all the
parameters seem reasonable and intuitive.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
115
households. For example, some consumers may shift their purchases to the new
entrant and hence reduce expenditures. At the same time, entry by Wal-Mart may
induce price reductions at the incumbent retailer, which may lead to an increase in
expenditures by some households. In our sample, we find th at approximately onethird of the households increased expenditures after W al-Marts entry. Further,
we suspect households distance from the respective stores to explain part of this
heterogeneity. In all three models reported in Table C.6, we assume that the
parameter for Wal-Mart is distributed normal, with a mean that is a function of
households distance to the focal and the competitive store.
The results show that entry by Wal-Mart significantly lowered monthly expen
ditures in the population. The impact is magnified for consumers living further
away from the incumbent store and for those living close to Wal-Mart. However,
we also find significant unobserved heterogeneity in the population as indicated
by the large and significant standard deviation parameter. Similar effects are ob
served for the Poisson regression model, where Wal-Mart has a significant negative
impact on the mean store visit frequency in the population. For the basket size,
the coefficient for Wal-Marts entry is insignificant, indicating th at the basket size
in the population remained constant. However, the unobserved heterogeneity pa
rameter is quite significant indicating th at the impact of W al-Marts entry varied
across households.
To develop further insights we consider the purchase history of two households
and compare their shopping behavior. The first row in Table C.7 shows that
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
116
both households have reduced their monthly expenditure by about 30% after WalM arts entry. However there are some critical differences in the purchase behavior
of these households. For hhl, the loss in expenditure seems to come entirely from
lower store visits, while the average basket size and the number of items in the
basket have not changed. More importantly, the basket composition in terms
of expenditure over food and non-food items has not changed. In contrast, the
number of store visits for hh2 has remained relatively constant, while the basket
size and the number of items in the basket have gone down quite significa.nt.ly.
Moreover, expenses on the fresh food items have actually increased over time,
while dollars spent on dry good groceries have fallen by half. This household
seems to have divided the shopping basket strategically, buying fresh produce and
meat at the supermarket while shifting dry good expenditures to Wal-Mart.
The example presented above can have important implications for the retailer's
strategy. For households like h h l, the retailer is losing entire baskets due to com
petitor entry. However, the basket size and composition has remained unchanged.
Thus, the focus for these households should be on devising strategies to increase the
store visits. For households, like hh2. which are abandoning particular products
in the store, the focus should be on designing direct-marketing instruments that
bring them back to the more profitable departments and categories. The category
level model presented in the next section can be used to design such direct mar
keting campaigns by learning about the price sensitivity and attribute preferences
for each individual.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
117
Uhjt =
(5.2)
h =
T.
term (for estimation procedures when the error terms are normally distributed see
McCulloch and Rossi 1994). The coefficient vector @h is known to the consumer
but not to the researcher and is assumed to be distributed in the population as
multivariate normal with a mean vector of b and a variance-covariance matrix W ,
i.e.. 0 h ~ N (b .W )lK
Note th at while the above specification uses a parametric approach to model
household heterogeneity, a more common approach with individual level choice
models is the finite mixture or the latent class model (Kamakura and Russell,
1989; Chintagunta et al. 1991; Wedel and Kamakura, 1999, etc.). Latent class
10T his section draws from Allenby and Lenk (1994), Allenby and Rossi (1999), and Train (2001
a.b)
l l Note th a t any demographic or other household specific variables could be incorporated in the
random coefficients as in the previous chapter.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
118
models axe based on the idea that there are a fixed, finite number of segments in
the market. Households belong to each of these segments with some probability
which are assumed to be. a priori, invariant across households and are interpreted
as representing segment sizes. Although appealing, finite mixture models do not
capture the tail behavior of the heterogeneity distribution adequately and the
posterior distribution of the individual-level parameters is constrained to lie within
the convex hull of (generally few) mass points (Allenby and Rossi. 1999).
On each purchase occasion, the household chooses the alternative that provides
the greatest utility. We do not observe the latent utility directly, but rather the
choice th at represents the maximum utility. Let y M denote the household h ' s
chosen alternative on purchase occasion t. and let y* = (y/u,......., V h t) denote the
household's sequence of choices. Also, let the choices of the entire sample H be
labeled as Y = (y i,
(5-3)
W
^ n (s S )
(5.4)
L(yh\b,W ) =
L(yh\3)0(3\b,W )d3
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
119
where <p(3\b. W ) is the normal density with mean b and covariance W . Note
th at this is the same mixed logit probability as in the previous chapter. However,
unlike the classical approach where the parameters b and W are considered fixed
representing the true mean and covariance of 3 h's in the population, in the Bayesian
framework the b and W are considered stochastic from the researchers perspective
and a prior distribution is specified for 6 and W :
(5.5)
k(b. W ) =
k(b)k(W )
(5.6)
k(b) -N (bo,S0)
(5.7)
k ( W) ' I W ( K . I )
Thus, the prior on b is normal with a mean of 6o and a (large) variance S0, while
the prior on W is inverted Wishart with K degrees of freedom and scale matrix
I. a /^-dimensioned identity matrix12. The posterior distribution of the population
param eters is:
(5.8)
12W hile these axe the prior distributions used most frequently in the hierarchical Bayesian choice
models, more flexible priors can be specified using procedures described in McCulloch and Rossi
( 2000 ).
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
120
Our objective is to make draws from this posterior distribution which summarizes
everything we know about the parameters including information from the sample
and the prior. Analysis of such hierarchical models has been made feasible by the
development of Markov chain simulation methods which directly exploit the hierar
chical structure (Tanner and Wong 1987, Gelfand and Smith 1990 etc.). One direct
approach to making draws from the posterior is to use the Metropolis-Hastings (MH) algorithm. However, doing so would be computationally burdensome as each
iteration of the M-H algorithm would require calculating the choice probability
L(yh\b. W) which does not have a close form. A more common approach is to
use Gibbs sampling which allows the researcher to take draws of one parameter
(or a subset of parameters), conditional on other parameters (Casella and George,
1992). The basic idea is to construct a Markov chain defined on the support of
the parameters such that the Markov chain has a stationary distribution equal
to the posterior distribution we are interested in. In the model described above,
the Gibbs sampler works by treating each j3h also as a parameter along with the
population parameters b and W. The Markov chain is constructed by repeatedly
sampling from the conditional posteriors for the three set of parameters b, W , and
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
121
(5.9)
(5.10)
(5.11)
K{W \b, 3 M
- I W (K + N . {K I + N S) / (A' + JV)),
(5.12)
K W i. W. * ) oc n
Given our priors, the conditional posteriors for the population parameters are
known distributions. However, there is no simple way to draw from the posterior
of each person's 3. so the M-H algorithm is used in step 3. Note however that the
right hand side of equation 5.12 is easy to calculate and just involves a product
of logits and o (3 h\b. IV') which is a normal density. The M-H algorithm to make
draws from equation 5.12 works as follows:
(a) Start with some initial value ,3% and make K independent draws from a
standard normal density, r?1
(b) Create a trial value for 3}, as 3h = 3%. + P ^ t)1. where p is a scalar specified
by the researcher and L is the Choleski factor of W
(c) Draw a standard uniform variable t^and calculate the ratio
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
122
(513)
L(yh\0l)<t>(plh\b.W )
L (y M )o t3 ? h\b,W )
(d) If u 1 < F, then accept 3 h as the new value, i.e.; 3 i = 3h- If u 1 > F,reject
3k and put 3 lh = 3%.
(e) Repeat the process many times and for high enough t. 3h is a draw from
the posterior.
Thus, it is relatively straightforward to make draws from the conditional poste
rior of each parameter. The Gibbs sampler works as follows. Start with any initial
values for 6. W. 3% V/,. Take a draw for each parameter from its conditional
distribution:
(1) Take a draw of b conditional on values U'. 30h
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
123
(5.14)
Once the converged draws from the posterior are obtained, the mean and stan
dard deviation of the draws can be calculated to obtain estimates and standard
errors of the parameters.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
124
The product categories studied are bath tissue, laundry detergent and milk.
As is typical of most consumer packaged goods, these categories contain more
than 60 unique UPCs. Such a large number of products dictate some aggregation
of products. Further, in selecting brands one needs to balance between category
representation and aggregation bias. Our approach to selecting and aggregating
products was to run a correlation of prices across weeks, and then b u n d lin g those
UPCs that shared a common set of attributes and had a price correlation of 0.8
or more. In other words, we only aggregate across UPCs th at had the same mean
price, and whose prices co-move highly enough to believe th at they are priced
jointly.
Summary statistics for the three categories are presented in Tables C.8. C.9 and
C.10. In the milk category, there are 12 products which account for 84% of total
category sales. The category is dominated by the private label which accounts for
over 75% of the category volume. The selected 12 products differ in terms of their
brand name. size, and fat content. In the bath tissue category, we use data from
15 products th at account for 66% of the total category volume. These products
differ in terms of their brand names and number of rolls. Scott has the highest,
market share, capturing nearly one-third of the total category volume. Finally,
for laundry detergent we use data from 16 products th at capture 81% of the total
category volume. Tide is the market leader in this category with over 30% market
share. We also report the prices for each brand before and after Wal-Marts entry.
Surprisingly, we do not find much difference in the prices in any category.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
125
Our sample includes all households that have made at least 5 purchases in the
category. Further, we use data from the entire 20 months rather than before and
after time periods (as in the previous chapter). Since our primary objective is to
obtain parameters for individual households, we want to have as long a purchase
history as possible. For our final sample, the average number of purchases in
the three categories are 17.3 for bath tissue, 23.0 for milk, and 10.6 for laundry
detergent.
The results for the three categories are presented in Tables C .ll, C.15, and
C.13. Note that we have used an attribute structure where each product is defined
as a combination of attributes.
attributes which vary within brand greatly improves the fit of traditional choice
models. In addition, this hedonic approach also gives us insight into the household
preference for specific attributes which can then be used for targeting purposes.
In the milk category, we find that the preference in the population is higher
for the smaller size. On average, consumers prefer 2% milk and the store brand.
However, the large standard deviations attached with all the parameters suggest
th at households are very heterogeneous in their price sensitivity and preference for
these product attributes. The covariance and correlation m atrix of the parameters
presented in Table C.12 also shows a number of significant off diagonal elements
suggesting th at allowing for correlation across parameters is necessary.
The results for bath tissue and laundry detergent categories are presented in
Tables C.13 and C.15. In both categories, consumers have a higher preference for
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
126
the smaller sizes. Also, brands Scott and Tide, which have the highest market
share in their respective categories, also have the highest preference. All brand
and size parameters show significant heterogeneity across consumers. Similarly,
households are quite heterogenous in their price response.
We now turn our attention to the individual level param eter estimates. Our
hypothesis from the previous chapter is that entry by the discount store causes
more price sensitive customers to shift purchases to the discount store. We examine
this issue by analyzing the individual level price parameters. In Table C.16, we
report the price sensitivity estimates for four groups of customers. The first group
labelled 'defect consists of households that purchased in the category before WalM arts entry, but stopped doing so after entry. The second group is a subset of
defect' but only includes households that stopped purchasing in the category, but
continue to visit the store. The third group consists of all households that have
increased or maintained their expenditures in the two time periods. Finally, the
fourth group labelled Hi Before Hi After consists of households th at are heavy
buyers (spend more than average) in both before and after time periods.
Table C.16 shows that customers who stop purchasing in the category, par
ticularly those who continue to visit the store but do not buy in the category,
are much more price sensitive. This phenomenon holds up in all three categories.
This lends strong support to our hypothesis in the previous chapter. However, one
needs to keep in mind th at 30 to 35% of these households fall in the right tail of
the distribution, i.e. they are less price sensitive than the population in general.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
127
Thus, price sensitivity is not the only explanation for why people stop purchasing
in the category. Finally, from a managerial perspective, Table C.16 also points
out the importance of households that fall in the fourth group. Not only do these
customers spend more in the category, they are also less price sensitive.
We now discuss how the retailer can use the individual param eter estimates for
customized target marketing. Table C.17 reports the individual parameter esti
mates from the milk category for 2 households. These are the same two households
that were used in Section 3. Recall from our earlier discussion th at h h ls spending
after Wal-Marts entry had gone down only because of store visits, while the basket
size was constant. The retailer can use the preference parameters reported in C.17
to target this household and induce store visits. For example, the most preferred
product for h h l is the store brand, in large size and 1% fat content. Note that
this is quite different from the average preference in the population, where the
most preferred product is the store brand in the small size with 2% fat content.
Further, h h l is quite price sensitive and thus needs only a small price cut to induce
purchase, especially if the price coupon is for her most preferred product.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
CHAPTER
C onclusion
The supermarket industry has undergone dramatic changes in the past few
years. Alternative retail formats such as mass merchandisers, price clubs, and su
percenters have encroached upon supermarket sales to pose a serious threat. One
way supermarkets have reacted to this growing competition is by building large
d ata warehouses. The impetus behind these data collection efforts is a hope that
d ata can be used to improve marketing decisions and thereby improve retailer
positions vis-a-vis new competitors. However, despite the large sums of money in
vested in creating point-of-sale databases, the goal of creating improved marketing
policies has often proven elusive. In a series of three essays, this thesis developed
econometric models th at can help supermarket managers make better use of their
databases, especially when confronted with new competitors.
The first essay demonstrates how a retail chain can utilize its point-of-sale d ata
in setting up a more profitable pricing policy. In addition to developing profitenhancing pricing policies for the focal retailer, the study also develops metrics
to determine the impact of change in prices from the consumer perspective. A
constrained pricing policy, which enables the retailer to capture a large portion of
the gains from the unconstrained pricing scheme while also mitigating consumer
128
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
129
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
130
1 0 ,0 0 0
households
both before and after Wal-Marts entry. Results show that distance to the store
and shopping-related variables are more important than demographics in predict
ing which customers will defect to the new entrant. The brand choice models
indicate th at consumers who stop buying in the category are on average more
price sensitive, confirming our findings from the previous chapter.
This thesis makes several contributions to marketing theory and practice. Past
research on supermarket competition has primarily focused on stores that are iden
tical in terms of product offerings, cost structure, and pricing policies. More re
cently, researchers have also analyzed competition between supermarkets with dif
ferent pricing formats, for example. EDLP vs. Hi-Lo pricing (Bell and Lattin
1998, Lai and Rao 1997, Messinger and Narasimhan 1997). However, the focus of
these papers has been on competition between supermarkets, with little attention
given to the growing competition from mass discounters and supercenters. Like
wise, while supermarket managers realize the threat from discount stores, little
is known about the impact of these stores on store-visit frequency, basket size,
and individual category demand. This thesis contributes to this emerging area by
addressing how consumer behavior changes when a low-priced competitor enters
the market and how traditional supermarkets should alter their pricing strategies
when faced with such competition.
This research is also salient to the growing body of literature focusing on data
base marketing. The first essay of this thesis shows how supermarket managers
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
131
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
132
may be concerned about factors, like store price image, that could affect long-term
profitability.
Although the limitations of the research are quite apparent in and of them
selves. they are not crippling. Supermarkets in general have access to data from
their own stores and make decisions based on what they observe in their stores.
Further, while the assumption of category profit maximization used in the study
may be restrictive, it is consistent with the industry drive towards category profit
maximization. From an academic perspective, these limitations can be overcome
in future research as consumer panel data th at record purchases from all stores
(including mass merchandisers) are being developed by market research firms like
IRI and Nielsen.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
R eferences
[1] Allenby, G. M. and P.E. Rossi [1999j. Marketing Models of Consumer Het
erogeneity. Journal o f Econometrics, 89. 57-78.
[2] Anderson. E. and N. Vilcassim [2001], Retailer Category Pricing, Working
Paper, University of Chicago.
[3] Basuroy. S., Nguyen, D., [1998], Mulinomial logit market share models: Equi
librium characteristics and strategic implications, Management Science 44.
1396-1408
[4] Bell.D. and J. Lattin[1998] Shopping Behavor and Cosumer Preference for
Store Price Format: Why Large Basket Shoppers Prefer EDLP Marketing
Science, ( 1 ) 6 6 - 8 8
[5] Bell. D.. Teck-Hua Ho and Christopher S. Tang (1998), Determining Where
to Shop: Fixed and Variable Costs of SHopping, Journal of Marketing Re
search. 35. 352-69.
[6 ] Berry, S., M. Camall and P.T. Spiller [1997], Airline Hubs: Costs, Markups
and the Implications of Customer Heterogeneity. Working Paper. Yale Uni
versity,
[7] Berry, S. [1994], Estimating Discrete-Choice Models of Product Differentia
tion. Rand Journal o f Economics. 25, 242-62.
[8 ] Berry, S., J. Levinsohn, and A. Pakes [1995], Automobile Prices in Market
Equilibrium. Econometrica. 63. 841-890.
[9] Berto Villas-Boas. Sofia [2001], Vertical Contracts Between Manufacturers
and Retailers: An Empirical Analysis, University of California, Berkeley,
Working Paper.
133
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
134
[10] Besanko. D., S. G upta and D. Jain [1998], 'Logit Demand Estimation Under
Competitive Pricing Behavior. An Equilibrium Framework, Management
Science, 44,11 (part 1 ).
[1 1 ] Besanko. D., J.-P. Dube and S. G upta [2001a], "Heterogeneity and Target
Marketing Using Aggregate Retail Data: A Structural Approach, Working
Paper. Kellogg Graduate School of Management.
[12] Besanko, D., J.-P. Dube and S. G upta [2001b], "Identifying Retail Pass
through Rates for Competing Brands, Working Paper, Kellogg Graduate
School of Management..
[13] Binkley. J and J. Connor [1998] Grocery Market Pricing and the New Com
petitive Environment Journal o f Retailing, 74(2), 273-294
[14] Blattberg, Robert C. [1998], Assessing and Capturing the Soft Benefits of
Scanning, Coca-Cola Retailing Research Council Study
[15] Blattberg, Robert C. and Scott A. Neslin [1990], Sales Promotions: Concepts,
Methods, and Strategies. NJ: Prentice-Hall: Englewood Cliffs.
[16] Blattberg, Robert C. and Edward J. Fox (1995), "Category Management,
Center of Retail Management, Northwestern University
[17] Borenstein. S. [1991]. Selling costs and switching costs: explaining retail gaso
line margins. The Rand Journal o f Economics, 22, 354-69.
[18] Bradlow. E.T. and A. Montgomery [1999], Why Analyst overconfidence
about the functional form of demand can lead to overpricing, Marketing
Science, 18, 569-83.
[19] Bresnehan. T. and Peter Reiss [1990] Entry in Monopoly Markets Review
o f Economic Studies .57, 531-555.
[20] Brooks, N. R. [2000], Agency finds Redlining, Zone Pricing of Fuel En
ergy: Although FTC says th at industry practices on West Coast arent illegal,
some say they inflate prices. The Los Angeles Times, Los Angeles, Calif. ;
Sep 27, 2000, 1.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
135
[21] Brown, Stephen [1989], "Retail Location Theory: The Legacy of Herald
Hotelling, Journal of Retailing, 65, 450-70
[22] Brownstone, D. and K. Train [1998], "Forecasting New Product Penetration
with Flexible Substitution Patterns. Journal o f Econometrics. 89, 109-129.
[23] Bucklin. R. and J. Lattin [1992], A Model of Product Category Competition
Among Grocery Retailers, Journal o f Retailing, 6 8 , 271-293
[24] Carpenter, G. and K. Nakamoto [1990] Competitive Strategies for Late Entry
into a Market W ith a Dominant Brand Management Science, 1 0 , 1268-1278
[25] Chevalier. J. A, A.K. Kashap and P. E. Rossi [2000], W hy dont prices rise
ruing peak periods of demand? Evidence from scanner data. Working Paper,
University of Chicago.
[26] Chiang, J. [1995]. Competing Coupon Promotions and Category Sales. Mar
keting Science, 14. 105-22.
[27] Chintagunta. P. [1994], Heterogeneous Logit Model Implications for Brand
Positioning. Journal o f Marketing Research, 31, 304-11.
[28] Chintagunta, P. [1998], Inertia and Variety Seeking in a Model of Brand
Purchase Timing, Marketing Science, 17, 253-72.
[291 Chintagunta. P. [2000], A Flexible Aggregate Logit Model. Working Paver.
University of Chicagol7, 253-72.
[30] Chintagunta, P. [2002], Investigating Category Profit Maximization at a Re
tail Chain. Journal o f Marketing Research, forthcoming.
[31] Chintagunta. P.. J.P. Dube and V. Singh [2002], Market Structure Across
Stores: an application of a random coefficients model with store level data,
Advances in Econometrics: Econometric Models in Marketing, ed. Philip Hans
Franses and Alan Montgomery (JAI Press: New York), 191-222.
[32] Cohen, A. [2000], Package Size and Discrimination in Paper Towels. Work
ing Paper, University of Virginia.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
136
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
137
[45j Gilbert, C. [2000], Governor attacks dairy price zones Bill would hurt states
family farms, officials say, Milwaukee Journal Sentinel, Milwaukee, Wis., Jim
18, 1 .
[46] Goldberg, P. K. [1996], "Dealer Price Discrimination in New Car Purchases:
Evidence from the Consumer Expenditure Survey, Journal o f Political Econ
omy. 104. 622-654.
[47] Gonul, F. and K. Srinivasan [1996], "Estimating the Impact of Consumer Ex
pectations of Coupons on Purchase Behavior: A Dynamic Structural Model.
Marketing Science. 15, 262-79.
[48] Gruca. T.. D. Sudarshan and K. Kumar [2001], "Marketing Mix Response to
Entry in Segmented Markets International Journal of Research in Marketing
. 18. 53-66.
[49] Gruca. T.. S. Kumar and K. Sudarshan [1992] An Equilibrium Analysis of
Defensive Response to Entry Using a Coupled Response Function Model
Marketing Science 11(4), 348-358
[50] Hauser, J. and S. Shugan [1983] Defensive Marketing Strategies Marketing
Science. 2 (4) 319-360
[51] Hausman. J. A. and D. A. Wise [1978], A Conditional Probit Model for
Qualitative Choice: Discrete Decisions Recognizing Interdependence and Het
erogeneous Preferences. Econometrica. 46, 403-26.
[52] Hoch, S. J., B.-D. Kim, A. L. Montgomery and P. E. Rossi [1995], Deter
minants of store-level price elasticity. Journal o f Marketing Research, 32,
17-30.
[53] Hoch, S. J.. Xavier Dreze and Mary E. Purk [1994], EDLP. Hi-Lo, and Margin
Arithmetic. Journal o f Marketing, 58, 16-27.
[54] Huff. David (1964), Defining and Estimating a Trading Area, Journal of
Marketing, 28, 34-38
[55] Judd, Kenneth [1985], Credible Spatial Pre-emption Rand Journal o f Eco
nomics. 16, 153-166
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
138
[56] Kadiyali, V., P.K. Chintagunta and N.J. Vilcassim[2000], Manufacturerretailer channel interactions and implications for channel power: An empirical
investigation of pricing in a local market, Marketing Science, 19, 127-47.
[57] Kadiyali, Vrinda, Naufel Vilcassim and Pradeep Chintagim ta [1999] Product
line extensions and competitive market interactions: An empirical analysis
Journal o f Econometrics, 89. 339-363
[58] Kamakura, W. and G. Russell [1989], Probabilistic Choice Model for market
segmentation and elasticity structure. Journal o f Marketing Research. 26.
379-390.
[59] Karunakaran. S. [2000], "Structural Analysis of Competitive Pricing in the
Presence of a Strategic Retailer. Marketing Science, forthcoming.
[60] Kim. B R. Blattberg, and P. Rossi [1995] Modeling the Distribution of Price
Sensitivity and Implications for Optimal Retailer Pricing Journal o f Business
and Economic Statistics, 13, 291-303
[61] Kumar, S. and D. Sudharshan [1988] Defensive Marketing Strategy: An
Equilibrium Analysis of Based on a Decoupled Response Function Model
Management Science 23, 405-415
[62] Kumar. V. and Robert P. Leone (1988), Measuring the Effect of Retail Store
Promotions on Brand and Store Substitution, Journal o f Marketing Research,
25. 178-186.
[63] Lai, R. and C. M atutes [1989] Price Competition in Multi-Market Duopolies
Rand Journal o f Economics.Z0, 516-537
[64] Lai. R. and R. Rao [1997] Supermarket Competition: The Case of Everyday
Low-Pricing" Marketing Science, 1 , 60-80
[65] Leslie, P. J. [2001], Price Discrimination in Broadway Theatre. working
paper, UCLA.
[6 6 ] McCulloch, R. and Peter Rossi (1994), An Exact Likelihood Approach to
Analysis of the MNP Model, Journal of Econometrics, 64, 207-240
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
139
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
140
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
141
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
APPENDIX A
142
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
143
C a te g o ry
Refr OJ
B ra n d
MM
MM
Dom
Trop Prm
Trop SB
Trop Prm
Florida
Ldry Detergent
SURF
WISK
WISK
ALL
ALL
CHEER
CHEER
TIDE
TIDE
Table
C o st ( $ /u n it)
1.69
1.99
1.15
1 .8 8
1.62
2.18
1.90
3.01
3.31
3.53
2.41
2.19
3.62
3.42
3.52
3.79
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
144
ZO NE
1
2
3
4
5
6
7
8
10
11
12
13
14
15
16
M EAN
P ric e In co m e Age60 E th n ic
(log $)
(%)
(S)
(%)
3.48
10.56
0 .2 0
0.26
3.27
1 0 .6 6
0.17
0 .1 1
3.38
10.62
0.18
0.03
3.28
10.80
0 .2 1
0.05
3.26
10.59
0 .2 2
0.07
3.20
10.71
0.14
0.08
3.48
10.43
0 .2 1
0.23
3.27
10.56
0.13
0.15
3.46
10.57
0.26
0.15
3.38
1 0 .1 0
0.15
0.46
3.26
10.74
0.16
0.14
3.26
10.72
0.09
0 .1 1
3.26
1 0 .8 8
0.09
0.07
3.25
10.57
0.14
0 .2 1
3.19
10.78
0.06
0.08
3.31
10.62
0.16
0.15
T a b le
A .4 .
D e m o g ra p h ic s
a n d
Demographics
S h o p in d x
(%)
0.62
0.79
0.89
0.80
0.83
0.87
0.42
0 .8 8
0.75
0.25
0.84
0.94
0.76
0 .8 8
0.81
0.76
C o m p e titiv e
HVAL
($0 0 0 )
166.52
147.37
143.83
160.00
111.59
135.32
190.35
121.19
116.41
97.37
152.75
151.07
179.07
100.39
139.06
140.82
V a ria b le s
b y
Jew el E D L P
(miles) (miles)
1.29
7.24
1.19
4.20
1.04
2 .1 0
2.47
1.63
1.76
2.80
1.26
1.14
1.35
8.51
0.34
11.65
1.83
7.28
1.58
9.53
0.93
4.89
3.91
6 .6 8
1.43
3.10
1.80
4.41
1.72
2.60
1.59
5.18
Z o n e
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
145
par am s.e.
Surf
-6.273 2.654
Wisk
-7.216 2.786
All
-5.316 2.554
Cheer
-3.305 2.550
Tide
- 1 . 8 6 6 2.541
price
-7.843 4.082
s.d. price
0.348 0.393
promo
0.749 0.068
price* promo
-1.106 0 . 1 1 2
64-oz
0.742 0.008
holiday
0.227 0.026
income
0.412 0.262
age60
-0.907 0.420
ethnic
0.064 0.207
shopindx
0.436 0.255
hval
-0.009 0 . 0 0 1
Jewel
-0 . 0 0 2 0.029
EDLP
-0 . 0 1 0 0.008
price* income -0.713 0.421
price*age60
3.296 0.672
price*ethnic
-0.641 0.324
price*shopindx -0.273 0.410
price*hval
0 .0 2 0
0 .0 0 2
price* Jewel
0.050 0.046
price*EDLP
0.053 0.013
Table A.5. Demand for Laundry Detergent
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
146
U RF
U i-
1h
0.5 iOrCIDE
CHEER
-0.5 -
-1
-1.5 OMSK
*2 o
o!i
1J
2J
3.5
OOMINICKS
OJh
01-
FLORIDA
-OJ CROP SB
0.4 -
CROP PRM
-0 .0 -0 J 3.4
Figure A.2. Perceptual Map for Refrigerated Orange Juice
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
147
par am
10.858
9.802
11.953
11.008
8.724
-50.389
-0.033
1.632
-3.694
-0.485
0.198
-0.921
-1.933
0.375
-0.062
-0.015
s.e.
MM
1.823
Dom
1.924
Trop Prm
1.817
Trop SB
1.814
Florida
1.969
price
4.443
s.d. price
0.874
promo
0.079
price*promo
0 .2 1 1
96-oz size
0 .0 1 1
holiday
0.031
income
0.186
age60
0.284
ethnic
0.147
shopindx
0.170
hval
0 .0 0 1
Jewel
0 .0 1 2
0 .0 2 0
EDLP
-0.029 0.006
price* income
2.950 0.459
price*age60
8.096 0.693
price*ethnic
-0.436 0.346
price*shopindx -0.004 0.419
price*hval
0.049 0 . 0 0 2
price* Jewel
-0.062 0.049
price*EDLP
0 .1 2 0
0.013
Table A.6 . Demand for Refrigerated Orange Juice
Category Income Age60 Ethnic Shopindx Jewel EDLP
Laundry Detergent -0.107 0.126 -0.018
0.160
0.024 0.067
Refrigerated O J 0.751
0.109 0 . 0 1 0
-0.026
0.003 0.036
Size (purchase prob.)
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
148
Category
Laundry Detergent
Brand
Surf
Wisk
AH
Cheer
Tide
Refrigerated OJ
MM
Dom
Trop Prm
Trop SB
Sunny D
Table A.8 . Latent
store
zone
chain
Tab e A.9.
s.e. dim 2
dim 1
s.e.
1.766 0.396 1.750 0.421
3.005 0.592 -1.689 0.582
1.080 0.252 -0.277 0.387
0 .1 2 0
0.463 -0.344 0.653
0 .1 0 0
0.368
-0.138 0.678 0.763 0.338
0.056 1.207 1.226 0.512
0.008 0.444 -0.341 0.418
-0 . 0 2 0 1.160 0.117 0.591
-0.440 1 . 0 0 2
Brand Factors
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chain
22.3%
26.6%
25.4%
24.3%
26.2%
15.7%
16.4%
15.6%
14.8%
149
Brand
Size (oz) Conditional Share TRUE Store Zone Chain
MM
64 oz
2 1 .2 %
23.7% 32.2% 30.6% 30.5%
MM
96 oz
4.1%
26.2% 30.1% 26.9% 26.5%
Dom
64 oz
24.4%
28.2% 38.7% 38.8% 38.4%
Trop Prm
64 oz
20.5%
28.1% 30.8% 28.4% 27.8%
Trop SB
64 oz
17.7%
30.1% 33.4% 30.7% 30.6%
Trop Prm
96
7.7%
27.0% 28.9% 25.4% 24.7%
Florida
64 oz
4.3%
31.6% 41.8% 42.4% 42.1%
Table A .ll. Predicted Margins for Refrigerated Orange Juice
Category
Aggregation
Profit
Change Change
Change
Profit Profit ($) Welfare ($)
Laundry
Detergent
Chain
$1,148,500
Zone
$1,155,400 0 .6 %
$6,900
$9,058
Store
$1,258,200 9.6%
$109,700
$125,915
Constr. Store $1,212,500 5.6%
$64,000
$103,381
cluster
$1,192,500 3.8%
$44,000
$45,082
Refrig.
Chain
$3,336,000
OJ
Zone
$3,388,400
1 .6 %
$52,400
$32,609
Store
$3,878,200 16.3% $542,200
$384,100
Constr. Store $3,582,400 7.4%
$246,400
$295,013
cluster
$3,623,400 8 .6 %
$287,400
$130,500
Table A.1 2 . Welfare Implications of Pricing Policies
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
400 -
0
to
*
20
30
40
--
50
90
70
nor*
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
151
600 (-
* -200 r
-*00 -
-600 r
60
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
152
10000
5000 r
5000 L-
30
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
153
10000
5000
5 0 0 0
40
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
154
p a ra m
d e m o g ra p h ic s
L og
P ric e
s .e .
t-s ta t
0 .0 9
-4 .0 9
- 4 .7 3
in c o m e
-1 .6 9
0 .3 6
0 .0 2
0 .0 0
8 .9 1
ag e6 0
-1 .1 1
0 .6 8
- 1 .6 3
e th n ic
0 .2 3
0 .4 6
0 .5 1
s h o p in d x
2 .0 9
0 .3 7
5 .6 2
je w e l
0 .3 1
0 .0 3
8 .9 0
e d lp
- 0 .0 4
0 .0 1
- 4 .0 7
A n a lg e s ic s
1 .9 5
0 .3 3
5 .9 3
B a th
so ap
0 .5 8
0 .1 3
4 .3 0
B eer
- 5 .1 2
0 .7 9
- 6 .4 5
B o ttle d ju ic e s
- 3 .5 7
0 .8 6
- 4 .1 3
(liq )
0 .4 8
0 .7 2
0 .6 6
so u p s
- 1 .8 7
0 .5 0
- 3 .7 6
F ro n t-e n d
c a n d ie s
0 .5 0
0 .5 2
0 .9 7
F ro zen
3 .9 1
D is h
d e te rg e n t
C a n n e d
e a tin g
d in n e rs
1 .5 1
0 .3 9
F ro z e n e n tre e s
1 .7 3
0 .3 4
5 .0 6
F ro z e n ju ic e s
-0 .7 4
0 .3 4
-2 .1 4
F a b ric s o fte n e rs
- 1 .2 4
0 .9 6
- 1 .2 9
p ro d u c ts
- 1 .1 5
0 .4 2
-2 .7 3
d e te rg e n ts
G ro o m in g
L a u n d ry
N o n -s lic e d
D is h
- 0 .3 8
h v a lm e a n
In d e x
C a n n e d
h o lid a y
d e te g e n t
- 1 .1 0
0 .8 2
- 1 .3 4
ch eeses
1 .5 0
0 .4 9
3 .0 5
- 1 .8 2
(p o w d e r)
-0 .9 2
0 .5 0
s a lm o n , c ra b s , e tc
1 .0 6
0 .3 6
2 .9 8
O a tm e a l
- 2 .8 3
0 .6 6
-4 .3 0
P a p e r to w e ls
-3 .1 1
1 .2 4
- 2 .5 1
R e frig e ra te d ju ic e s
-0 .2 0
0 .1 6
- 1 .2 1
ch eeses
3 .0 9
0 .6 6
4 .7 0
S o ft d rin k s
2 .0 5
0 -6 4
3 .2 4
S h am p o o s
S lic e d
0 .7 0
0 .7 9
0 .8 8
c ra c k e rs
1 .6 5
0 .5 9
2 .8 1
S o ap s
1 .3 8
1 .3 9
0 .9 9
T o o th b ru s h e s
- 4 .6 0
1 .0 9
-4 .2 2
tu n a
- 0 .4 3
0 .2 6
-1 .6 4
T o o th p a s te s
- 0 .8 3
0 .5 0
-1 .6 4
tis s u e s
- 3 .8 9
0 .8 5
-4 .5 9
c o n s ta n t
4 4 .1 4
5 .1 5
8 .5 7
S n ack
C a n n e d
B a th ro o m
o b s e rv a tio n s
R -sq u a re
1 4 8 6 .0 0
0 .3 7
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
APPENDIX B
155
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
156
N um ber o f Stores
% Change
1990
2000
All Grocery Stores 145,000 127.980
-11.7%
Supermarkets ($2,000,000+) 30,750 31,830
3.5%
Chain Supermarkets 17,460 20,825
19.3%
Independent Supermarkets 13,290 11,005
-17.2%
Other Stores (Under $2,000,000) 58.250 37.310
-35.9%
Convenience Stores 56,000 58,000
3.6%
Table B.2. Consolidation in the Supermarket Industry
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
157
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
158
2.40
10.65
18.93
1.40
10.30
6.45
7.07
5.00
Table B.6 . Distance to Nearest Competitor
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
5.53
2.27
159
B a th T issu e
Angel
Cottenelle
Charmin
Dominicks
Northern
Scott
Green Forest
Share
5.8%
19.8%
26.7%
6 .8 %
21.9%
12.9%
6 .1 %
Price/4 roll
Share
4.1%
11.7%
25.7%
16.8%
8 .1 %
20.9%
1 2 .6 %
Price/roll
0.82
0.75
1.13
1 .1 0
1.52
1.45
1.27
1.26
2.27
0.90
Margin Promotion
36%
21.9%
15.0%
36%
21%
13.8%
24.6%
17%
26%
16.1%
1 0 .6 %
25%
25.2%
31%
P a p e r Towels
Sparkle
Mardi Gras
Bounty
Dominick's
Brawny
Scott
Viva
0 .6 6
1.04
1.03
1.09
Margin Promotion
35%
25.8%
25.6%
32%
13.8%
18%
28.5%
29%
20.4%
25%
14.6%
30%
15.6%
38%
R e frig e ra te d J u ic e
Florida
Minute Maid
Dominick's
Tropicana
Sunny D
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
160
JMOOO .
I
%
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
161
P a p e r Tow els
Param
Se
Sparkle -3.865 0.077
Mardi Gras -2.837 0.075
Bounty -0 . 8 6 8 0.083
Dominicks -3.005 0.070
Brawny -2.653 0.081
B ra n d F E
Scott -1.478 0.081
Viva -1.846 0.081
Warehouse -0.116 0.004
C a te g o ry S h ifte rs
Price Club
0.034 0.026
Holiday 0.008 0.009
Prom
0.600 0.029
M a rk e tin g M ix
Price -2.575 0.130
Price Std.
0.549 0.127
Price* Prom -0.262 0.030
Price*Ethnicitv -0.566 0.017
Price*Wealth
0.117 0.015
P ric e In te ra c tio n s
Price* Drug -0.003 0 . 0 0 2
Price*Groc
0.016 0.003
Price* Pclub -0.926 0.025
Price*WareSR -0.077 0.009
Price*WareLR 0.5346 0.009
Table B.9. Estimation Results: Paper Towels
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
162
B ath T issu e
Par am
Se
Angel -3.971 0.025
Cottenelle -1.729 0.032
Charmin -1.343 0.028
B ra n d F E
Dominicks -2.982 0.025
Northern -1.938 0.025
Scott -0.985 0.040
Green Forest -3.633 0 . 0 2 1
Warehouse -0.059 0.005
C a te g o ry S h ifters
Price Club -0.402 0.017
Holiday -0.026 0.009
Prom
0.576 0 . 0 2 1
M a rk e tin g M ix
Price -1.673 0.019
Price Std.
0.335 0.080
Price* Prom -0.054 0.014
Price*Ethnicity -0 . 1 0 2 0 . 0 1 0
Price*Wealth -0.017 0 . 0 1 0
P ric e In te ra c tio n s
Price*Drug -0 . 0 0 2 0 . 0 0 1
Price*Groc -0 . 0 0 1 0 . 0 0 2
Price*Pclub -0 . 2 0 2 0 . 0 1 1
Price*WareSR -0.033 0.006
Price*WareLR 0.1638 0.006
Table B.10. Estimation Results: Bath Tissue
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
163
O range Ju ice
Par am
Se
Florida -3.251 0.113
MM -2.091 0 . 1 1 2
B ra n d F E
Dominicks -2.477 0.106
Tropicana -1.276 0 . 1 1 2
Sunny D -3.693 0.104
Warehouse 0.055 0.005
C a te g o ry S h ifters
Price Club
0.055 0.046
Holiday -0.033 0 . 0 1 0
Prom
0.167 0.049
M a rk e tin g M ix
Price -1.303 0.090
Price Std.
0.387 0.061
Price*Prom
0.038 0.024
Price*Ethnicity -0.064 0.009
Price*Wealth
0.242 0.009
P ric e In te ra c tio n s
Price*Drug
0 .0 0 1
0 .0 0 0
Price* G roc
0 .0 1 2
0 .0 0 1
Price* Pclub -0.716 0 . 0 2 2
Price*WareSR -0.060 0.005
Price*WareLR
0.015 0.009
Table B .ll. Estimation Results: Orange Juice
Paper Towels
Before (1991-92) Entry (1993-94) After (1995-96)
Number of new stores
8
34
6
Mean Price Sensitivity
-2.34
-3.83
-1.44
0.42
0.61
0.13
Std. Price
Table B.12. Paper Towel Price Sensitivity: 3 periods
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
164
Sparkle
Bounty
Brawny
Scott
Mardi Gras
Viva
Store
Table B.13
Margins
Actual
Predicted
Before After Before After
27%
29%
30%
49%
14%
14%
29%
46%
23%
19%
28%
51%
14%
17%
31%
48%
24%
32%
31%
43%
17%
17%
30%
47%
27%
28%
35%
41%
Margins for Paper Towels
3500
3000
2900
2000
1900
1000
500
4 p ar. M m a * (COUNT) -
* p ar
A g.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
4000
98
-3.6
-3.1
-2.6
-2.1
-1.6
- l.l
-0.6
-18
-13
1.8
1.3
0.8
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
se
0.35
0.37
0.08
0.17
0.24
0 .2 2
0.19
0.04
0 .0 2
166
1.15
PRCE14
1.1
1.05
1
0.95
0.9
0.85
0.8
0.75
0.7
1
31
61
91
121
151
181
211
241
271
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
301
167
Dim 1
se
Sparkle -0.7489 0.3566
Mardi Gras 0.9344 0.3551
Bounty 0.3703 0.2578
Dominicks 0.7381 0.2849
Brawny 0.7581 0.3025
Scott 0.2448 0.2807
Viva 1.2625 0.2025
Table B.17. Latent Attributes for
Dim 2
-0.7497
-1.8398
-0.1662
-0.1605
-0.835
-1.2556
se
0.4463
0.2952
0.2944
0.3926
0.3228
0.2405
Paper Towels: After
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
APPENDIX C
168
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
169
Dist. Store
Dist. WM
Income
Children
Rental
Length of Residence
Table C.2. Summary
graphic Variables
T e M a m S ta
700000
aoooool
(A 9 3
WMt
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
170
Focal Store
Control Store
Sales Transactions
Sales
Transactions
Store
-19.3%
-28.9%
-3.7%
-3.9%
Grocery
-16.8%
-23.0%
-4.0%
-4.7%
Meat
-13.0%
-17.9%
0.5%
- 6 .8 %
Produce
-13.3%
-18.7%
-3.0%
- 8 .0 %
GM
-17.4%
-26.1%
-2 .0 %
-8.3%
Deli
-0.9%
-3.9%
9.3%
9.0%
HBC
-17.8%
-21.9%
-3.6%
-8 .2 %
Pharmacy
-52.2%
-59.3%
-3.7%
1.9%
CSD
-14.8%
-19.6%
4.9%
2.3%
BreadRolls
-18.5%
16.8%
-0.9%
-2 .0 %
Milk
-18.0%
-2 2 .1 %
-2 .0 %
-8.5%
Cereal
-17.9%
-17.4%
-6.4%
-7.5%
Cheese
- 2 1 .0 %
-23.0%
-9.7%
-15.4%
Processed Meat -17.9%
-21.3%
-1.9%
-9.1%
Salty Snacks
-17.4%
-15.0%
-3.9%
-7.2%
Beer
-29.1%
-33.6%
0 .8 %
-3.0%
Coffee
-26.4%
-25.6%
-23.0%
-19.0%
Soup
-17.5%
-17.3%
-8 .1 %
-7.1%
Bot. Juice
- 1 1 .0 %
-14.3%
5.3%
-0.5%
Cookies
-17.5%
-18.9%
3.8%
- 1 .1 %
Rfj. Juice
-3.9%
-6.4%
-4.6%
- 6 .8 %
Bath tissue
- 6 .8 %
- 2 1 .6 %
8 .8 %
-9.4%
Table C.3. Regression Results for Changes in Sales, Transactions.
and Sales per Transaction Following Wal-Mart Entry-
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
171
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
172
hhl
hh 2
Before After Before After
Monthly Expd
490
357
594
410
Monthly Visit
5.2
3.9
7.2
7.9
Basket Size
94
92
57
75
No. of Items in basket
58
56
67
39
Exp on dry Groc
290
210
361
198
Exp on fresh Food
117
81
134
119
% expd on dry Groc
59%
59%
61%
48%
%expd on fresh food
24%
23%
20%
33%
Table C.7. Comparison of Household Shopping Behavior
Brand
Store
Store
Store
Store
Store
Store
Store
Store
Crowley
Crowley
Crowley
Crowley
After
Before
Size
Fat Share Price/unit Share Price/unit
T8%
1 / 2 gallon Full
1.386
1.302
7.0%
1 / 2 gallon
2% 11.5%
1.377
1.287
9.7%
1 / 2 gallon
6.5%
1.370
1.272
1%
5.8%
7.7%
1 / 2 gallon
FF
7.1%
1.283
1.368
Gallon
Full 12.2 %
2.452
13.6%
2.493
Gallon
12.6%
2.424
2.389
1% 11.0%
2.371
Gallon
2% 19.7%
21 .8 %
2.428
Gallon
FF 10.8%
12.2%
2.410
2.398
Gallon
Full 3.7%
3.060
3.026
3.0%
Gallon
2% 3.6%
3.061
3.033
3.1%
Gallon
2.7%
3.052
3.000
1%
2 .0 %
FF
Gallon
2.9%
3.011
3.049
2 .0%
Table C X Summary Statistics: Milk
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
173
Before
After
Size-Roll Share Price/roll Share Price/roll
Brand
Kleenex
4
4.6%
0.623
8.0%
0.651
Charmin
4
9.0%
0.636
9.9%
0.659
Charmin
12
4.7%
0.541
7.6%
0.600
Charmin
24
10.5%
0.262
5.7%
0.282
Store
1
4.3%
0.549
4.0%
0.619
Store
4
5.9%
0.490
5.2%
0.532
Store
12
2.2%
0.476
1.4%
0.497
Northern
4
6.2%
0.691
6.8%
0.691
Northern
12
6.0%
0.325
3.3%
0.325
Northern
24
5.2%
0.302
4.2%
0.288
Marcal
1
7.7%
0.586
7.6%
0.593
Scott
1
10.6%
0.655
6.6%
0.726
Scott
4
3.5%
0.698
3.4%
0.705
Scott
12
10.7%
0.623
19.1%
0.627
Scott
24
8.9%
0.593
7.2%
0.638
Table C.9. Summary Statistics: Bath Tissue
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
174
Before
After
Size Share Price/lOOoz Share Price/lOOoz
100 10.5%
Wisk
6.00
6.28
5.5%
100 7.4%
5.03
4.91
7.7%
All
Pur ex
100 8.2%
4.66
8.2%
4.23
100 5.3%
AH
4.50
4.37
4.6%
Dynamo 100 4.5%
5.41
8.4%
5.39
Era
200 3.7%
4.92
4.4%
5.07
Era
100 12.7%
5.14
4.65
13.7%
100 18.5%
Tide
6.16
16.6%
6.28
200 6.3%
Tide
6.39
11.0%
6.36
128 3.5%
Brtwtr
2.12
2.21
3.3%
Pur ex
200 3.6%
3.34
3.08
3.6%
50
5.6%
Tide
6.98
5.2%
7.55
Brtwtr
50
1.4%
3.44
0.9%
3.73
128 4.5%
Store
2.56
2.95
3.5%
64
Store
1.9%
3.11
3.11
1.4%
Xtra
128 2.4%
2.32
2.28
1.9%
Table C.10. Summary Statistics: Laundry Detergent
Milk
Price
Std-Price
Size (1 gallon)
Std-size
Fat-FF
Std-FF
Fat-1%
Std-1%
Fat-2%
Std-2%
PC
Std-PC
Table C.l]
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
175
Price
Size
FF
1%
2%
Store
Store
0.12
-0.18
-0.07
-0.06
-0 . 0 4
3.83
Price=:Price/Roll
Parameter Standard Error
Price
-16.67
0.23
std-price
11.55
0.20
Single roll
1.62
0.08
std-single
4.50
0.09
12-Pk
-2.84
0.07
std-12pk
3.59
0.07
24-Pk
-5.39
0.18
std-24pk
4.54
0.17
Charmin
-0.85
0.18
std-charmin
3.55
0.18
Store
1.10
0.07
std-store
5.72
0.07
Northern
-1.13
0.15
std-north
3.38
0.13
Marcal
2.35
0.09
std-marcal
6.01
0.07
Scott
3.96
0.08
std-scott
6.26
0.08
Table C.13. Model Estimation Results: Bath Tissue
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
176
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
177
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
178
Category
Customer Type
Median Mean % below population mean
Defect
Milk
-22.2 -22.51
53%
Milk
Defect-visit store
-28.03 -25.75
66%
Milk
Maintained
-20.36 -21.51
49%
Milk
Hi Before-Hi After -17.13 -19.84
42%
Defect
Bath
-22.1
-19.6
69%
Bath
Defect-visit store
-22.9
-20.4
71%
Bath
Maintained
-16.38 -16.25
49%
Bath
Hi Bef-Hi Aft
-15.6
-15.8
46%
Laundry
Defect
-3.22
-3.11
59%
Laundrv Defect-visit store
-3.47
-3.25
66%
Laundrv
Maintained
-2.84
-2.88
51%
Laundry
Hi Bef-Hi Aft
-2.26
-2.43
36%
Table C.16. Estimated Price Sensitivity, By Group
Milk Category
No. of obs
% change in exp
Price
Size
FF
1%
2%
Store
hhl
hh2
45
57
-29% -42%
-41.47 -30.16
1.53
0.25
2.43 -1.20
6.46 -1.70
4.12
1.74
0.56
0.94
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.