Location Strategy of Chain Retailers: The Case of Supermarkets and Drug Stores in an Urban Market
Gustavo Vicentini Northeastern University This Version: October 2012
Abstract This paper presents an empirical investigation of the location strategy pursued by chain retailers over a period of 55 years. We use a unique data set containing the sequential location decisions of supermarkets and drug stores in a medium size U.S. city between 1956 and 2010. The data indicate that chain retailers located new stores within reasonable proximity to their already existing network of stores through most of the sample period. We propose a strategy to empirically identify whether spatial monopolization intent, economies of density, or both of these factors were the reason for such clustering behavior by chain retailers. Using a discrete choice model for location in geographic space, we find that economies of density were partially responsible for the clustering behavior in the supermarket industry, but less so in the drug store industry. Moreover, spatial monopolization intent was a pursued strategy by both supermarkets and drug stores all along. We also find that in some cases up to 33.2 percent of consumers living between two stores owned by the same retailer found themselves “trapped” in space by the retailer, with no closer shopping alternative.
Keywords:
Location decision, spatial monopolization, economies of density, chain retailers, supermarkets, drug stores.
JEL Classifications: C1, C35, L1, L8, R1. Correspondence:
Gustavo Vicentini Department of Economics Northeastern University
[email protected] https://sites.google.com/site/gvicentini/
I thank Victor Aguirregabiria, Marc Rysman, Mo Xiao, and seminar participants at Northeastern University, Bentley University, and the 2011 IIOC Conference for helpful comments. I also thank Helen Snow (Greensboro Public Library), Mark Schumacher (University of North Carolina at Greensboro Library), and Jason Tomberlin (University of North Carolina at Chapel Hill Library) for assistance in collecting some of the industry data used in this paper.
1
Introduction Chain firms have become the predominant players in many retail markets. According to U.S. Census data
shown in the top panel of table 1, although chain retail firms have traditionally comprised less than 10 percent of all retail firms in the United States (first row), they have accounted for an increasing share of all retail stores (second row), and for an increasing share of all retail sales, reaching 68.1 percent in 2007 (third row). Additionally, the average number of stores per chain retailer has increased from 6.5 stores in 1967 to 11.5 stores in 2007 (fourth row), and the sales volume at a typical chain store has been at least 2.5 times higher than in a single-store firm since 1948 (fifth row). Supermarkets and drug stores, the two retail industries analyzed in this paper, have been no exception to this trend, with an even greater dominance by chain firms, as depicted in the middle and lower panels of table 1. A particular feature of chain retailers is that they tend to penetrate a given urban market in a sequential manner, typically opening one or two stores per year in order to gradually create a network of stores over time. If the choice of where to locate a new store carries some degree of irreversibility due to sunk entry costs, then the geographical arrangement of stores over time is an important strategic decision for these firms. Chain supermarkets have long recognized the importance of this sequential location choice. To quote the manager of the store location research department of Kroger Supermarkets in 1961: “An unnecessarily large number of retail businesses fail each year because of inadequate market research and poor location planning. Many supermarkets are included in the toll, some lasting no more than a year. ...How are stores to be arranged in one, three, five or more years to best serve the company in any given area?... What will be the impact of new stores upon existing company stores? Can two stores co-exist in a given area, or will an older store be cannibalized? … A number of supermarket chains have established location research departments for this purpose.” Ransome (1961)
Despite the importance of chain retailers and of sequential spatial positioning in urban markets, most empirical work on firm location has either focused on industries with primarily single-store firms or has treated each chain store as a separate entity. See, for example, the studies of Bolduc et al. (1996) for physician location, Netz and Taylor (2002) for gas stations, Watson (2005) for retail eyewear, Seim (2006) for retail video stores, and Datta and Sudhir (2011) and Orhun (2012) for supermarkets. A notable early exception is West (1981), who used historical location data on chain supermarkets in Vancouver, Canada. He constructed the set of neighbors of a chain store at time of entry, and found that a statistically significant number of a new store’s neighbors were composed of stores owned by the same chain firm. He interpreted this result as an attempt by chain supermarkets to crowd out the market and spatially preempt the competition. More recently, Holmes (2011) studied the spatio-temporal diffusion of Wal-Mart stores since its first opening. He found the discount retailer proliferated its network in space in a gradual manner, often locating new stores close to its existing ones while foregoing profitable markets that were distant from its network. Holmes attributed some of this behavior to economies of density, which are spatial scope economies enjoyed
2
by clustered stores belonging to the same firm. The closer the stores of a firm are to each other in space, the more scope economies there are.1 Jia (2008) and Ellickson et al. (2012) structurally estimated static location games between discount chain retailers such as Wal-Mart, Kmart, and Target, and found to a varying degree that these retailers benefit from placing new stores near own existing stores, potentially due to economies of density. Nishida (2012) estimated a structural location game between convenience stores, and also found that chain retailers benefit from placing new stores near own existing stores. Our paper presents an empirical investigation of the location strategy pursued by chain retailers in an urban setting, and whether this strategy has changed over time. Data have been collected on the sequential location decision of supermarkets and drug stores for the city of Greensboro, North Carolina (United States), from 1955 to 2010. We estimate a discrete choice model where each year retailers decide where to locate their new stores given the most recent observable spatial market structure. We control for location-specific socioeconomic factors using U.S. Census Bureau decennial data from 1950 to 2010. We utilize the implications of a simple theoretical model to implement our empirical identification strategy. The first question we address is whether a chain retailer tends to open a new store relatively close to one of its existing stores. If the answer is positive, we then empirically distinguish whether the reason for such clustering behavior is either an intent to spatially monopolize a local urban neighborhood or an intent to enjoy economies of density from a clustered network of stores, or both. To our knowledge, this distinction has not yet been made explicitly in the existing literature. Both spatial monopolization and economies of density predict clustering of a chain retailer’s network in space. To partially distinguish between the two forces, consider the following scenario with two isolated neighborhoods (the argument is formalized in the next section). Submarket 1 has three separately owned stores, while submarket 2 has a clustered retailer () with two stores alongside store . Active stores compete in prices. Submarket 1
Submarket 2
Both submarkets expect to observe demand increases in the near future. In the absence of density
economies, a new single-store retailer strictly prefers to enter submarket 2. This is because chain retailer
internalizes the cannibalization effect when setting its prices at stores and , and resulting Nash prices are
higher than in submarket 1. However, if density economies are important, the single-store retailer switches its
1
Examples of density economies enjoyed by spatially clustered same-firm stores may include shared advertising, shared inventory, shared management, easier product quality inspection, easier employee monitoring, and reduced distribution costs.
3
behavior and prefers to enter submarket 1 to avoid the clustered retailer. This is because chain retailer will charge lower prices due to its economies of density. Now suppose that retailer wants to open a third store,
. As will be shown in the next section, retailer will prefer to open near one of its already existing
stores, due to the spatial monopolization motive. The question then is in which submarket, 1 or 2, will be
placed. In the absence of density economies, retailer does not enjoy scope economies and therefore retailer is less reluctant to open in submarket 2. However, if retailer enjoys density economies, then retailer
also switches its behavior and locates in submarket 1 to avoid the tougher price competition in submarket 2. In this case, some combination of density economies and spatial monopolization behavior is present, because retailer avoids clustered competitors (such as retailer ) but also wants to cluster its own stores in space.
Our empirical exercise follows this reasoning. We first determine the propensity of a chain retailer to open new stores near its already existing network. If this propensity exits, we then consider whether new stores enter neighborhoods with large presence of agglomerated chain competitors (e.g., submarket 2). If new stores are not averse to such neighborhoods, then spatial monopolization is the sole motive for clustering. However, if economies of density are important, retailers will avoid such locations given the cost advantages of clustered competitors (e.g., retailer in submarket 2), and some combination of spatial monopolization behavior and cost considerations are dictating the clustering. A third potential motive guiding the location of chain retailers is entry deterrence through strategic positioning of stores. Bonanno (1987) studied an incumbent chain firm facing the threat of entry in a spatial market. He found that for some levels of entry costs the chain firm deters the potential entrant by actually dispersing its stores in space, in order to make the market more competitive and less attractive to new entry. Therefore, if the entry deterrence motive is present, deterrence by dispersion may be a pursued strategy of chain retailers.2 Our simple theoretical model, presented below, also bears this result for some levels of entry costs, and we will discuss what this implies for our empirical identification strategy described above. Our results indicate that, for the entire sample period (1955-2010), both supermarkets and drug stores avoided placing new stores in locations where their closest sister store was within one mile, likely due to cannibalization across sister stores. Instead, both types of chain retailers preferred to place new stores within a distance band of three to four miles to their closest sister store. We interpret this result as evidence of an attempt of spatial monopolization by chain retailers. Economies of density seem to have been relevant in the supermarket industry, but less so in the drug store industry. Moreover, we also found that in some cases up to 33.2 percent of consumers who live between two stores owned by the same retailer were “trapped” in space by the retailer, with no closer shopping alternative.
2
Eaton and Lipsey (1979) also analyzed the location choice of an incumbent chain firm under the threat of entry, and found that the firm strategically locates its stores in a timely manner in order to deter entry.
4
Section 2 develops a stylized model of location choice of single-store and chain retailers. Section 3 presents the data and some stylized facts on the location pattern of supermarkets and drug stores in Greensboro, NC. Section 4 presents the econometric model used to estimate the location choice of retailers, as well as the results. Section 5 concludes.
2
A simple model We build upon the circular city model of spatial competition introduced by Vickrey (1964) and formalized
in Salop (1979). We incorporate multiple chain retailers and economies of density into the framework.3 We use the model as guidance for the location strategy pursued by single-store and chain retailers. We do not present general results, but rather a sequence of examples to illustrate our empirical identification strategy.4 Algebraic equilibrium derivations are in the Appendix. The market is the circle of unit circumference. Consumers are uniformly distributed around the circle, and
population size is one. There are stores symmetrically distributed around the circle, and upon new entry the stores adjust their locations accordingly to maintain the symmetry (this assumption is used in Salop (1979), and greatly simplifies the analysis). Once in the market retailers compete in prices. Demand is unitary. A consumer buys from the store with the lowest personalized price (mill plus transportation cost). The indirect utility of a consumer located at ∈ 0,1 buying from store located at ∈ 0,1 is: = − − −
where is the price charged by store , is the transportation cost parameter, and − is the linear distance between the consumer and the store. is the gross benefit from buying the good, common across consumers and assumed large enough so that the entire market is covered. To facilitate discussion, consider the following spatial market structure , ℓ, , , , … , : k j
ℓ
m1
m2 n
3
Levy and Reitzes (1992) and Brito (2003) previously incorporated chain firms into the circular city model to study horizontal merger in localized markets. See also Giraud-Heraud et al. (2003). 4 For a fully dynamic model that incorporates spatial competition between chain firms, see Aguirregabiria and Vicentini (2012).
5
Retailer has two stores, and , and all other retailers have one store each. Given a set of prices,
one can pin down the marginal consumer who is indifferent between two stores. The demand a store faces, , is then the segment between itself and its clockwise right marginal consumer, plus its clockwise left opposite segment. A single-store retailer, such as , chooses the price of its only store to maximize variable profits: !
where " is marginal cost.
= ! − " !
A chain retailer, such as , chooses the prices of its stores jointly to maximize total variable profits: #
= $#% − 1 − &"' #% + $#) − 1 − &"' #)
where & ∈ 0,1 captures the degree of economies of density in variable costs. That is, if two stores that belong to the same retailer are adjacent (close) to each other, they enjoy scope economies and lower marginal costs. We assume that non-adjacent sister stores do not enjoy density economies. If & = 0, there are no
density economies to be enjoyed by adjacent sister stores, and single-store and chain retailers have the same marginal cost.5 Price best-response functions have closed-form analytical expressions, and their solution yield Nash equilibrium prices. Equilibrium demands and profits then follow directly. We now provide four examples to illustrate our empirical identification strategy.
Details of the algebraic derivations are provided in the
Appendix. Example 1: We start with a simple spatial market structure , , that has three independently owned stores ( = 3), as shown in figure 1. Given this market structure, a new single-store retailer would be indifferent as to where to locate. It would earn equal profits at any location given the symmetry of store ownership and locations. In fact, when there are no adjacent stores owned by the same retailer, the equilibrium solution for each store is: ∗ = " +
∗ =
1
∗
=
However, if chain retailer is to open a second store, , its total profits are different if it clusters its stores (locates between and or between and ) than if it isolates them from each other (locates
between and ). Figure 1 plots the equilibrium profits to retailer as a function of density economies (&)
5
As mentioned above, examples of density economies are shared advertising, shared inventory, shared management, easier product quality inspection, easier employee monitoring, and distribution costs. Smith (2004) comments that, in the U.K., about 90 percent of a supermarket’s operating costs are comprised of variable costs (instead of fixed operating costs). Of these, the greatest portion comes from cost of goods, followed by distribution costs and labor. It is thus reasonable to assume that density economies enter a chain retailer’s cost function through variable costs.
6
for the two cases (we normalize = 1 and " = 1 in this and all examples to follow). As is seen in the figure,
retailer strictly prefers to cluster its stores in space and monopolize the consumers “trapped” between and
, for any level of density economies. We call this the spatial monopolization effect. By internalizing the
cannibalization effect between its two stores when setting its prices, retailer charges higher prices and exploits the consumers located between its two stores. 6 As economies of density increases (& > 0), the
incentive to cluster becomes even greater. ■ Example 2: Suppose retailer has indeed clustered its 2 stores as suggested in Example 1, and the new
market structure is , , , , as shown in figure 2. Suppose now that single-store retailer , is to choose where to locate its store given this market structure. If retailer , chooses to locate between and (or
between and ), it prefers to be near the clustered chain competitor (retailer ). If retailer , locates between and , it avoids neighborhoods with a clustered competitor. Figure 2 plots the profits to retailer ,
as a function of & for these two location options. As is seen in the figure, in the absence of density economies
(& = 0), retailer , strictly prefers to locate between and . In fact, store , would earn the highest profits
amongst all stores when it does so, while the single-store retailer furthest from the clustered chain retailer, , would earn the lowest profits. Therefore, when & = 0, single-store , wants to be near the clustered chain
retailer because of its higher prices.7 However, when density economies are present (& > 0), the advantage of
locating next to clustered retailer declines to the point where retailer , switches its behavior and chooses to
locate between and , avoiding the clustered competitor as it charges lower prices due to its density
economies and lower marginal costs. We call this the density economies effect. ■ Example 3: Suppose again that retailer has clustered its 2 stores as suggested in Example 1, and the market structure is , , , , as shown in figure 3. However, now suppose that retailer , is a chain retailer that is
choosing where to locate two stores, , and , , given this market structure. Retailer , has four different options, each yielding different profits. First, retailer , may choose to cluster its stores together next to the
already clustered retailer , yielding the market structure , , , , , , , . Second, retailer , may choose to cluster its stores between and , far away from the already clustered retailer , yielding the market
structure , , , , , , , . Third, retailer , may choose to disperse its stores by locating them on each side of the already clustered retailer , yielding the market structure , , , , , , , . Finally, retailer ,
may choose to disperse its stores by locating one store near the already clustered retailer and the other store
far away from it, yielding the market structure , , , , , , , . Figure 3 plots the profits to retailer , as a function of & for the two options where retailer , clusters its stores. The two options where retailer ,
6
Thomadsen (2005) provides an empirical example from the fast food industry where chain sister stores charge higher prices the closer they are to each other in geographic space. 7 As suggested by the following quote from The Wall Street Journal, coffee shops may exemplify well this incentive to locate near a clustered competitor: “Many coffeehouses have found proximity to Starbucks to be a blessing. A small Seattle chain called Tully’s Coffee Corp. has even developed a strategy of placing its stores near a Starbucks shop.” (The Wall Street Journal, “Counting Beans: Despite the Jitters, Most Coffeehouses Survive Starbucks,” September 24, 2002.)
7
disperses its stores are never profit-maximizing, and therefore are omitted from the figure; therefore, retailer ,
always prefers to clusters its stores, for any level of &, due to the spatial monopolization effect. As is seen in
the figure, for small density economies retailer , is somewhat indifferent between locating its two clustered stores near the already clustered retailer or away from it, with a small profit advantage to locating them next
to retailer . However, as & increases, retailer , prefers to avoid the clustered competitor by locating its two stores between and , due to the density economies effect. ■
Examples 1 through 3 assume a “myopic” behavior by entrants as it does not allow for the possibility of future entry by additional competitors. The next example adds a simple dynamic setting by allowing an additional retailer to potentially enter after a first-moving chain retailer has chosen its location. As we will show, this dynamic interaction introduces the possibility that the first-moving chain retailer may strategically locate its stores in order to block the entry of the potential entrant. Example 4: Consider again the simple spatial market structure , , from Example 1. Suppose retailer
moves first and chooses the location of its second store, , followed by the location choice of single-store
retailer ,. Given entry costs -, which location strategy would retailer pursue under the threat of entry by
retailer ,? Figure 4a shows the equilibrium positioning of for different ranges of - under the assumption
that there are no density economies (& = 0). For comparison purposes, the top panel of the figure shows retailer ’s strategy when no additional retailer may enter the market, while the bottom panel shows firm ’s
strategy under the threat of entry by retailer ,. As the bottom panel of figure 4a shows, for relative large values of entry costs (- > 0.0625), entry is blocked for both and ,. For 0.0544 < - < 0.0625, retailer
is able to deter , just by clustering its stores, yielding the market structure , , , . However, for
0.0400 < - < 0.0544, the threat from retailer , becomes more imminent and retailer needs to actually
disperse its stores in space, , , , , in order to increase price competition and deter , from entering.
We call this the deterrence by dispersion effect.8 Finally, for - < 0.0400, retailer accommodates retailer ,, and , chooses to locate near the clustered chain retailer.9 Intuition suggests that beyond some positive value
of density economies (&) retailer might be able to deter , by always clustering its stores, for any value of
reasonable -, avoiding the need to disperse its stores to deter entry. As an example, figure 4b shows the same
scenario as in figure 4a, but with & = 0.1. As depicted in the bottom panel of figure 4b, for - > 0.0825, entry
is blocked for both and ,. However, for 0.0469 < - < 0.0825, the presence of density economies makes
the market more competitive to the rivals of retailer , and serves as a “natural” entry deterrent, and therefore
retailer does not need to use a store dispersion strategy to deter retailer , from entering the market. Finally, The result that retailer may choose to actually disperse its stores to make the market more competitive in order to deter retailer , from entering is qualitatively similar to the result in Bonanno (1987), although his model uses a line segment as the market (instead of the circle). 9 It is worth noting that we have implicitly assumed that entry is irreversible, which makes the entry deterrence strategy fully credible. As pointed by Judd (1985), allowing for an additional and final stage in the game where firms may opt to exit and recover some of their entry costs may change the results. 8
8
for - < 0.0469, retailer accommodates retailer ,, and , chooses to locate near the clustered chain retailer.
Intuition again suggests that as density economies (&) increases even further, retailer , will, when entry
becomes feasible, avoid being near the clustered chain retailer due to the density economies effect. As an
example, figure 4c shows the same scenario as in figures 4a and 4b, but with & = 0.3. As depicted in the bottom panel of figure 4c, for - > 0.1593, entry is blocked for both and ,. For 0.0367 < - < 0.1593,
the mere presence of density economies again allows retailer to deter retailer , from entering the market by
simply clustering its stores. However, for - < 0.0367, although retailer accommodates retailer ,, the latter now chooses to locate away from the clustered chain retailer because of the density economies effect. ■
Identification strategy If retailers behave in a “myopic” manner and disregard potential future entry by competitors, then examples 1 through 3 provide a set of conditions to partially distinguish between the spatial monopolization effect and the density economies effect. With or without density economies, a chain retailer always prefers to cluster its stores in space (figures 1 and 3), which is an indication of the spatial monopolization effect. In the absence of density economies (& = 0), a single-store retailer strictly prefers to locate near clustered chain retailers (figure 2), and a chain retailer is somewhat indifferent to the presence of nearby clustered competitors (figure 3). As density economies increase (&> 0), the single-store retailer becomes averse to neighborhoods with clustered chain retailers and prefers to distant itself from them (figure 2), while a chain retailer still prefers to cluster its stores in space, but now it also avoids being in proximity to clustered competitors (figure 3). In sum, if a chain retailer clusters its stores in space and competitors do not mind being near it, then the spatial monopolization effect is present, but not the density economies effect (or only to a certain extent). However, if a chain retailer clusters its stores and competitors avoid being near it, then a combination of the spatial monopolization effect and the density economies effect is present. If retailers behave in a “non-myopic” manner and internalize potential future entry by competitors, then,
except for a certain range of entry costs -, the identification strategy to partially distinguish between the spatial monopolization effect and density economies effect is the same as in the “myopic” case. That is, if one
momentarily ignores the range of 0.0400 < - < 0.0544 for entry costs in figure 4a, and compares the
location strategy followed by retailers and , displayed in figures 4a, 4b, and 4c, then the same logic outlined
above applies: If a chain retailer clusters its stores in space and competitors do not mind being near it, then the spatial monopolization effect is present, but not the density economies effect (or only to a small extent); if a chain retailer clusters its stores and competitors avoid being near it, as in figure 4c, then a combination of the spatial monopolization effect and the density economies effect is present. It is only in the case where a chain retailer chooses to disperse its stores that the “non-myopic” predictions are different from the “myopic” predictions. If such store dispersion behavior by a chain retailer is observed, as in figure 4a, then all we can affirm is that (a) the deterrence by dispersion effect is present, (b) firms do not behave in a “myopic” manner,
9
as store dispersion is never an outcome under a myopic setting, and (c) the density economies effect is not relevant, as density economies take away the need of a chain retailer to disperse its stores to deter entry.
3
Data and stylized facts
3.1
Firm data The firm data are comprised of the location history of all supermarkets and drug stores within the city
and suburbs of Greensboro, North Carolina (United States). The data span from 1955 to 2010 for supermarkets, and from 1955 to 2009 for drug stores. 3.1.1
Supermarket data
The data for supermarkets were collected from yearly published Greensboro “city” and “suburban” directories, under the listing of “Grocers – Retail.”10 For each year between 1955 and 2010, the listing provides the name, address, and (after the year 1989) the number of employees at each active grocery store. 11 The listing also includes grocery stores inside large discount retailers, such as Wal-Mart supercenters. Based on this initial information, each grocery store address was geocoded to obtain its latitude and longitude coordinates (in decimal degrees) in geographical space. These initial grocery store data contain about 1,000 unique firms and approximately 1,500 unique stores that were active at some point in time between 1955 and 2010 in Greensboro. Obviously, we cannot assume that all of these firms and their stores are similar to each other or even compete with each other. Instead, we are interested in analyzing the location decision of the major players of this retail industry, i.e., the “supermarkets,” and not the location decisions of the smaller “fringe” grocery stores such as “mom and pop,” “curb market,” or “corner grocery” stores. As such, we proceeded to identify the supermarkets from this large initial population of grocery stores using the following steps. First, we reviewed selected documentation on the history of the supermarket industry in Greensboro and North Carolina. Within these documents we were fortunate to encounter historic descriptions of the
10
The Greensboro “city” and “suburban” directories were published by “Hill Directory Co. Inc.” until 1982, and by “R. L. Polk Directories & Co.” after 1982. The Greensboro “city” directory was first published before 1955, while the “suburban” directory was first published in 1966. The 1964 and 1965 “city” directories were published jointly as one volume, as were the 1981 and 1982 volumes. The “city” directory was not published in 1987 and 1995, and the “suburban” directory was not published in 1983, 1988, 1995, and 1996. The listing “Grocers – Retail” was split into two listings, “Grocers – Retail” and “Convenience Stores,” in 1993. 11 According to the publishers of these directories, their data are “…gathered by an actual door to door canvass and is compiled in a way to ensure maximum accuracy.” (“Greensboro City Directory,” Hill Directory Co. Inc., Richmond, Virginia, 1965, 1975; “Greensboro City Directory,” R. L. Polk & Co. Richmond, Virginia, 1985, 1995.)
10
major supermarkets that were active in Greensboro dating back to the 1930s until the present.12 Using this evidence, we preliminarily separated the grocery stores from our directory data into two groups: A group that we labeled “supermarkets,” which likely comprised the major supermarkets of interest; and another group labeled “fringe,” which likely comprised the non-major (i.e., smaller) grocery stores. Next, we conducted internet searches to seek confirmation, when available, that the grocery stores preliminary classified as “supermarkets” were indeed supermarkets and not a “mom and pop,” a “curb market,” or a “corner grocery” store. For example, we visited the website of firms that were still active in 2012 to confirm that they were properly classified. Additionally, in some cases we were able to locate local newspaper articles from the past that covered, for example, the opening of a new store by a particular major supermarket chain in Greensboro. Next, we utilized the fact that the post-1989 city and suburban directories contain the number of employees at a given listed store to confirm that the grocery stores initially classified as “supermarkets” had a relatively larger number of employees than the stores classified as “fringe.” It was found that, for the year 1989 and beyond, the number of employees in stores that we classified as “supermarkets” was usually at least 50, while the “fringe” stores usually had no more than ten employees. As a final corroborative check, we collected data on the size (in square feet), address, year of built, and building description for each currently-standing building within Guilford County, North Carolina. (Greensboro is located within Guilford County.) These “buildings” data were downloaded from the website of the Geographic Information Services Department of Guilford County, North Carolina.13 We were then able to match the address of many of the grocery stores from the directories data to its corresponding address in the buildings data in order to identify the size of each grocery store. Unmatched stores were primarily located in sites that have been subject to urban re-development and therefore the original buildings from the past no longer exist. For the remaining unmatched store addresses, we utilized historical Sanborn Fire Insurance Maps for Greensboro from 1919, 1966 and 1972 to obtain the size of buildings built before 1973.
Sanborn Fire Insurance Maps provide detailed footprints and
dimensions of buildings within the urbanized area of a city as of the time of their publications. After utilizing both the “buildings” data and the Sanborn Fire Insurance Maps, we were able to identify the size (in square feet) of about 85 percent of the approximately 1,500 unique grocery stores from the original
12
For example, a detailed description of the evolution of the food retail industry in Greensboro, including a discussion of its major supermarkets from the 1930s to the 2000s, can be found at
. 13 The website of the Geographic Information Services Department of Guilford County is
. The “building” data were downloaded in March 2011 from an ftp site accessed from this website. Specifically, the website portal for this ftp site was
, and the downloaded files were “BLDG.zip,” “BLDG_MAIN_ADDITION.zip,” and “LAND_LINE.zip.”
11
directory data. Ultimately, these data on building size revealed that the stores that we preliminarily classified as “supermarkets” were indeed on average larger in size than the stores preliminarily classified as “fringe.” Altogether, these steps to classify grocery stores into a “supermarket” category versus a “fringe” category provided us with 25 unique “supermarket” firms and 184 unique “supermarket” stores. The rest of the firms and stores were classified as belonging to the “fringe” category. Figure 5 shows the size, according to our classification, of the fringe stores (solid circles) and of the supermarket stores, at the year when each store first appeared in the directory data.14 The supermarket stores are further divided into locally owned supermarkets (squares) versus non-locally owned supermarkets (triangles) such as Kroger Supermarkets. We identified the locally owned supermarkets based on a review of selected documents about the history of grocery stores in Greensboro.15 According to our classification, there is an increasing gap in store size over time between supermarkets and fringe stores. Additionally, most of the fringe stores were less than 10,000 square feet in size across the entire time period. Figure 6 displays the identity and number of stores per supermarket firm in Greensboro. The fringe grocery stores were excluded from this figure. As is seen, A&P (Great Atlantic & Pacific Tea Company) was a major player from the 1950s to the 1970s. Colonial and Big Bear were also major players in the earlier years, with the latter selling its stores to Harris Teeter in the 1980s. Kroger entered the market in the 1950s, and exited in the early 2000s when it sold all of its North Carolina stores to Harris Teeter. Winn-Dixie was a major player for at least 50 years, finally exiting the market in the mid-2000s. In more recent years, Food Lion and Harris Teeter have dominated the market. The area shaded in light gray, labeled “Other supermarkets,” represents supermarket firms that had at most four stores in any given year. Finally, the areas shaded in dark gray represent the locally owned supermarkets. In particular, Bi-Rite was a cooperative comprised of independently and locally owned supermarket stores that grew to more than 15 stores by the 1970s, but then split into Bestway and Food Rite.16 It is interesting to note in figure 6 that, with the exception of acquisitions and spin-offs, most supermarket chains tend to slowly grow their number of stores over time, opening at most one or two stores on net per year.
14
Stores that changed ownership due to mergers, acquisitions, or spin-offs are displayed only once in figure 5, at the year of their original opening. Supermarkets that were inside supercenters of discount retailers, such as Wal-Mart, were excluded from the figure. 15 See, for example, the detailed discussion in . 16 See, for example, the discussion in , , and .
12
As an example of the sequential location pattern followed by chain supermarkets, Figure 7 depicts the sequential spatial positioning of Winn-Dixie supermarkets over time.17 As is shown, Winn-Dixie opened its first three stores within about two miles of each other in the central part of the city in the 1950s. It then opened its fourth store (labeled store “#4”) in 1961 on the Northeast area, slightly isolated from the other stores, but then it opened its fifth store (labeled store “#5”) in 1962 almost in between stores number one and three. These three stores (number one, three, and five) then co-existed in such a fashion for nine years. Next, stores number six and seven opened slightly isolated from then-existing stores, but store number eight opened relatively close to store number four, and the two co-existed for many years. Most of the remaining stores opened in more suburban location, as the city expanded further into the suburbs over time. 3.1.2
Drug store data
The data for drug stores span from 1955 to 2009, and were collected from the “Hayes Drug Store Directory,” a directory specific to the retail pharmacy industry published yearly since 1912 by “Hayes Directories, Inc.”18 For each year, the directory provides the name and address of each drug store in every town of the United States.19 Drug stores inside supermarkets or large discount retailers (such as WalMart) are also included in the directory.20 Each store address was then geocoded to obtain the latitude and longitude (in decimal degrees) of the store.
We also collected the drug store listings for towns
immediately outside of Greensboro in order to have the equivalent of a “suburban” listing. These drug store data contain a total of 106 unique firms and 279 unique stores over the 1955-2009 period. Unlike the grocery store data, it is reasonable to assume that there is no clear separation of “major players” versus “fringe” competitors in the drug store retail industry, at least at the store level. That is, each drug store is a reasonable potential competitor to other nearby drug stores, irrespective of whether a store belongs to a chain firm, is an independent store, is inside a supermarket, or is inside a discount retailer. As a corroborative inspection of this assertion, we also matched the addresses of the drug stores in our data to the addresses of the buildings in the “buildings” data in order to identify the size of each drug store, similar to what we did in the grocery store data. Additionally, we again utilized historical 17
The solid gray line in figure 7 represents Greensboro’s urbanized area in 2000, and the dotted gray line represents Greensboro’s city limits in 1950, both according to the U.S. Census Bureau’s geographic boundary files. 18 See . We were not able to identify the Hayes Drug Store Directory for the year 1998. 19 Between 1955 and 1972, the Hayes Drug Store Directory provided the name and address of only one single store for firms that had more than five stores within Greensboro. Therefore, to identify all drug stores of all firms in Greensboro between 1955 and 1972, we complemented the Hayes Drug Store Directory listing with the stores contained in the listings “Pharmacy – Retail” and “Drug Stores” within the Greensboro “city” and “suburban” directories for those years. 20 According to the publisher of this directory, its data contain “…the most accurate and complete list of retail drug stores ever compiled. … Every effort is made to secure complete and reliable data.” (“Greensboro, North Carolina,” Hayes Drug Store Directory, Hayes Directories Inc., 1955-2009.)
13
Sanborn Fire Insurance Maps for Greensboro in order to obtain the size of drug stores in the past. Ultimately, we were able to identify the size of approximately 72 percent of the 279 unique drug stores.21 Figure 8 displays the size of the drug stores (triangles) for which we were able to obtain store size data, at the year when each store first appeared in the directory.22 As is shown, the size of drug stores experienced a small upward trend over the years, increasing from about 5,000 square feet in the 1950s and 1960s to about 11,000 square feet in the 1990s and 2000s. More importantly, except for a few drug stores that opened with a size of less than 6,000 square feet after 1985, there is no clear separation of stores into distinct categories based on store size. Figure 9 displays the identity and number of stores per drug store firm in Greensboro. For example, Eckerd entered in the late-1950s, gradually expanded its number of stores over the years, then changed owners in the 1990s and 2000s (first to JCPenney, then to the Jean Coutu Group), and then was sold to Rite Aid in the late 2000s. Franklin and Edmonds each grew considerably in the first 20 years. Franklin was bought by Rite Aid in the early 1970s, which was then bought by Kerr in the late 1990s. Edmonds slowly phased out of the market. Revco entered in the 1970s and by the 1990s was the largest chain; it was then bought by CVS. Walgreens has slowly entered the market in the 2000s, choosing to open its own stores instead of acquiring a pre-existing chain. The area shaded in light gray, labeled “Other drug stores,” represents the drug stores of all firms with at most four stores in any given year. Finally, the area shaded in dark gray represents drug stores located inside a supermarket (such as Kroger) or large discount retailer (such as Wal-Mart); this type of drug store has become more prominent in the 1990s and 2000s, and seems to have taken some of the share (in terms of number of stores) from the firms that have a small number of stores (i.e., the “Other drug stores” category). Again, it is interesting to note in figure 9 that, despite some acquisitions (as in the case of CVS), many drug store chains tended to slowly grow their number of stores over time in the market, opening at most one or two stores on net per year. Figure 10 depicts, as an example, the sequential spatial positioning of Revco drug stores over time, from its first opening in 1970 until its eventual sale to CVS in 1998.23 As depicted, Revco opened its first store in the Southwest portion of town. It then opened two additional stores in 1973, one (labeled store “#2”) almost five miles to the Northeast of the first store, but the other (labeled store “#3”) within about two miles Northwest of the first store. The fourth store continued Revco’s expansion towards the 21
Using store size as a corroborative check to the assertion that there is no clear competitive separation between “types” or “categories” of drug stores is very likely an imperfect procedure, as (for example) regular drug stores may not be close substitutes to (for example) drug stores inside supermarkets for any given store size. However, we lack data on demand substitutability patterns between types of drug stores, so we believe this to be a reasonable check given our data. 22 Drug stores that changed ownership due to acquisitions or spin-offs are displayed only once, at the year of their original opening. Drug stores inside supermarkets or large discount retailers (such as Wal-Mart) were excluded from the figure. 23 As before, the solid gray line in figure 10 represents Greensboro’s urbanized area in 2000, and the dotted gray line represents Greensboro’s city limits in 1950, both according to the U.S. Census Bureau’s geographic boundary files.
14
Northeast of town, while the fifth store opened in 1979 to the North, isolated from the other then-current Revco stores. Store number six then opened within two miles of the third store, and these two stores coexisted for 18 years. Later stores opened primarily in more suburban locations, although it is interesting to note how close store number 16 opened to store number ten in 1996, just prior to the CVS acquisition. 3.1.3
Retail activity data
In addition to the supermarket and drug store data, we utilized the “buildings” data collected from the Geographic Information Services Department of Guilford County, North Carolina, to re-construct the spatial distribution of retail activity in Greensboro’s county over time. As mentioned above, these data contain information on the size (in square feet), address, year of built, and building description for each building structure within Guilford County as of 2011. For our purposes, the “building description” variable provides information on the functionality of a building, such as whether it is currently used as a family residence, an industrial site, a warehouse, an office, a bank, a discount store, a strip mall store, or some other function. Based on this building description variable, we categorized buildings into whether they are likely used for a retail activity or not. For example, buildings that are described in the data as being sites for banks, discount stores, strip mall stores, and other likely retail-related activities, were categorized as “retail,” while buildings that are sites for family residence, warehouses, industrial sites, etc, were categorized as “non-retail.” Together with the variables on “year of built” and “address” of each building, which are also available in the “buildings” data, we were then able to re-construct the spatial distribution of retail activity in Guilford County for each year between 1955 and 2010.24 This spatial distribution of retail activity is likely relevant for the location decision of retail stores. For example, locations with little retail activity may indicate that zoning laws impose strict restrictions on the ability of retailers to open a store, even if the location offers other attractive features to retailers such as a high population density or easy accessibility. Additionally, locations with large retail activity may carry “agglomeration” economies and provide benefits to co-location with other types retailers, especially if consumers prefer to shop in locations where they can obtain all their desired goods (e.g., food, dry cleaning, clothing) within close proximity. See, for example, the discussion in Datta and Sudhir (2011).
24
One drawback of using these data and methodology to re-construct the spatial distribution of retail activity over time is that it assumes that a building has had the same functionality since it was first built, as the “building description” variable provides the current (2011) functionality of the building. That is, the data do not provide snapshots of the functionality of each building in different points in time, but only as of 2011. Therefore, buildings that switched functionality over time from “retail” to “non-retail,” or vice-versa, would be incorrectly categorized. We feel that such events are likely not predominant and do not justify not using these data. An alternative would be to use the US Census’ “County Business Pattern” data to obtain information on business activity at the zip code level. However, these data were first collected at the zip code level in 1994 (see ), and extrapolating such data back to 1955 would likely be unrealistic.
15
3.2
Socio-economic data We complemented the firm data with historical tract-level socio-economic data from the U.S. Census
for Guilford County, North Carolina. These data were obtained from the U.S. Census Bureau website25 and from the National Historical Geographic Information System website at the University of Minnesota Population Center.26 For each decennial year from 1950 to 2010, we obtained tract-level data such as population count, age distribution, census tract area, and others. Table 2 lists these tract-level variables and shows their mean, standard deviation (in parenthesis), and minimum and maximum values (in square brackets) across all Guilford County tracts. As depicted in the table, most variables have large variation across census tracts, regardless of which decennial year is considered. Additionally, we also obtained the tract-level geographic boundary files for Guilford County from 1950 to 2010. Figure 11 shows the evolution of these census tract boundaries over time. Greensboro is located in the middle of the county, while the smaller city of High Point is located in the Southwest corner. The gray color scheme represents the distribution of population density in the county within each decennial year, where the areas with the lightest shade of gray represent tracts in the first quintile of population density (0th to 20th percentile), the second-lightest shade represents tracts in the second quintile (20th to 40th percentile), and so on. As is seen, over the last 60 years Greensboro grew mostly to the West, Northwest, and Southwest, with little growth to the East.
4
Estimation
4.1
Econometric model Our main econometric model is a conditional logit discrete choice model, where each year a retailer
observes the spatial market structure of its retail industry at the previous year, and then chooses where to locate a new store. More specifically, the market is the city and immediate suburbs of Greensboro, and it
is comprised of 8 locations, indexed by ℓ. These locations were constructed by laying a 12-by-12 grid of
points on top of the map of Greensboro, with adjacent points (i.e., locations) equally spaced by one and a half miles from each other. This grid provided us with a square of 16.5 miles by 16.5 miles containing a
total of 144 (= 8 = 12 ∗ 12) points. Each of these points is a location ℓ wherein a retailer can choose to open a new store. Figure 12 depicts this setting, showing each location as a point in the map. The figure also shows Greensboro’s urbanized area boundary in 1950 (dotted line) and in 2000 (solid line).
25 26
www.census.gov. www.nhgis.org.
16
For each year between 1956 and 2010, a retailer observes the most recent spatial market structure
across all 8 locations, and then chooses a location ℓ to install (if any) a new store. If retailer opens a
new store at location ℓ at year 9, then the contemporaneous store-specific profit associated with the new
store that accrues to retailer at year 9 is: ℓ: ;, <
DEFGHG: FI@ >?@A
= = ;,>?@A ∗ B ℓ,:C >?@A
DF#K G:FLHG
+ ; ∗ JC,ℓ,:C
This contemporaneous store-specific profit,
ℓ9 ;, <,
Q DF#K + < ∗ ℓ: + R@ + Sℓ: + ; ∗ MNOP C,ℓ,:C
is unobserved by the econometrician and thus a
latent variable. Instead, we only observe the location ℓ chosen by the retailer. Therefore, estimation will be
based on the observed location choice by a retailer for its new store. Specifically, we estimate a conditional logit discrete choice model where in each year 9 a retailer chooses in which location ℓ to place a new store
(conditional on the retailer having chosen to open a new store that year). We note that the location choice of each store in our sample is used only once in this discrete choice estimation, at the year the store entered the market (i.e., at the year the store was opened). We now discuss the explanatory variables of the model. First, we discuss the set of “distance band” DEFGHG: FI@ >?@A
dummy variables labeled B ℓ,:C
. The coefficient estimates on these dummy variables will capture
whether a chain retailer has a tendency to locate a new store within close proximity of its already-existing DEFGHG: FI@ >?@A
network of stores. Specifically, we define B ℓ,:C
as indicating whether retailer ’s closest own
store to location ℓ at time 9 − 1 was within a certain distance band from the center of location ℓ. We
specify several distance bands (i.e., several distance ranges for “TU V”) of one mile each from the center of location ℓ, up to the end of the 5th mile. That is, we specify a “0–1” mile distance band (i.e., “TU V”
takes the value “0 to 1 miles”), a “1–2” miles distance band (i.e., “TU V” takes the value of “1 to 2 miles”), and so on, up to a “4–5” miles distance band (i.e., “TU V” takes the value of “4 to 5 miles”). Additionally, we specify a “5–plus” miles distance band (i.e., “TU V” takes the value of “5-plus miles”).
(In other words, “TU V” is an index that starts at the “0–1” mile distance band from a location ℓ, then moves up to the “1–2” mile distance band, then to the “2–3” mile distance band, and so on.) For example, if retailer had opened a new store at location ℓ in year 9, and the closest own store that
retailer had from that location ℓ in year 9 − 1 was 2.5 miles away from location ℓ, then the indicator
variable for the “2 –3” miles distance band (with respect to location ℓ) would have a value of one, and all
other distance band indicator variables with respect to that location ℓ and retailer at year 9 would receive a value of zero. In the actual estimation, we will set the “0–1” mile distance band as the baseline category and omit it from the regression to avoid perfect multicollinearity among the distance band dummies.
17
Taken together, the coefficient estimates on these distance band indicator variables will reveal how close chain retailers prefer to locate a new store from their closest own already-existing stores. DF#K G:FLHG The variable JC,ℓ,:C captures the intensity of competitors’ stores that firm would face around
location ℓ at time 9 − 1. Its intention is to capture a pure “store-level competition effect” that store “ℓ”
would face at location ℓ from stores owned by firms other than firm , and that is why this variable has a
subscript “−.” This variable was constructed as the sum of the inverse of the squared distance between the
center of location ℓ and the location of each store that belonged to a competitor at time 9 − 1. That is, DF#K G:FLHG
JC,ℓ,:C
= ∑C
)
$AXY% C,ℓ'
, where V:C −, ℓ represents the distance in miles between location ℓ and
a competitor’s store at time 9 − 1. Therefore, locations where firm had a large number of nearby DF#K G:FLHG competitors’ stores will likely be assigned a high values for JC,ℓ,:C , while locations with few nearby
competitors’ stores are assigned low values. Q DF#K captures the extent of clustering by chain competitors (and therefore potential The variable MNOP C,ℓ,:C
density economies by competitors) that retailer would face around location ℓ if it were to place a new store
at location ℓ in year 9. It is defined as a slightly modified version of the typical Herfindahl index of market concentration. Specifically, if none of firm ’s competitors had more than one store within a three-mile radius
Q of the center of location ℓ at time 9 − 1, then MNOP C,ℓ,:C is set to zero. This is because if no competitors to DF#K
retailer had more than one store around location ℓ, then none of them had the potential to enjoy density
economies around that location; in other words, none of them displayed within-firm spatial clustering around that location. Similarly, if firm had no competitors within a three-mile radius of the center of location ℓ at
Q DF#K is also set to zero. In both of these cases, there is no possibility of density time 9 − 1, then MNOP C,ℓ,:C
economies coming from firm ’s competitors around location ℓ, as there are no clustered competitors (i.e., no competitor has more than one store around the location). However, if at least one of firm ’s competitors had
Q more than one store within a three-mile radius of the center of location ℓ at time 9 − 1, then MNOP C,ℓ,:C is DF#K
equal to the typical Herfindahl index.
For example, suppose that at time 9 − 1 location ℓ had five stores within a three-mile radius that
belonged to firm ’s competitors. If these five stores were owned by five separate firms, then none of these five competitors displayed same-firm clustering, and there was no potential for density economies Q DF#K equals zero. On the other coming from the competition around that location. Therefore, MNOP C,ℓ,:C
hand, if two of the stores were owned by a single competitor, and the remaining three stores each
Q DF#K = 0.4 + 0.2 + 0.2 + 0.2 = 0.2, reflecting belonged to separate different competitors, then MNOP C,ℓ,:C
some degree of clustering by competitors (primarily from the one competitor that owns two stores). If
18
two of the stores were owned by a single competitor, and the other three stores were owned by another
Q DF#K = 0.4 + 0.6 = 0.52, reflecting an even greater degree of clustering single competitor, then MNOP C,ℓ,:C
by competitors, and greater potential for density economies. In the extreme case that all five stores were Q DF#K = 1. owned by a single competitor, then there is a high degree of clustered competition, and MNOP C,ℓ,:C
Taken altogether, the parameters ;,>?@A , ; , and ; allow for the implementation of the empirical
identification strategy described at the end of Section 2. Everything else constant, we expect the “storelevel competition effect” parameter ; to be negative, as more competitors’ stores around a given location
indicate a tougher competitive environment and less profitability for a given store. For a given distance band, the parameter ;,>?@A captures whether chain retailers prefer to place new stores in locations where
their closest own existing store is within that distance band. Finally, the parameter ; captures whether
retailers avoid locations with already clustered chain retailers. Most important for our identification strategy, the parameters ;,>?@A and ; should be analyzed in combination when making inference regarding the potential presence of spatial monopolization and/or density economies. For example, suppose ;,>?@A (for a given distance band) and ; are both positive. This would indicate that retailers prefer placing new stores in locations where they have a closest-own existing store within that distance band (as reflected by the positive ;,>?@A ), and at the same time do not mind being in proximity to
clustered competitors (as reflected by the positive ; ). If the distance band under consideration represents a relatively “close” distance in space, this would indicate spatial monopolization behavior without the presence of density economies. On the other hand, if ;,>?@A is positive but ; is negative, then retailers are placing new stores in locations where they have a closest-own existing store within that distance band, but are now avoiding locations with clustered competitors. If, again, the distance band under consideration represents a relatively “close” distance in space, this would indicate spatial monopolization behavior (as reflected by the positive ;,>?@A ) combined with the presence of density economies (as reflected by the negative ; ).
The vector : = : , : , … , Z: ′ contains exogenous socio-economic variables at each location ℓ at
time 9, such as median family income, percent of the population within a certain age group, and others. These variables were constructed as weighted averages of the tract-level socio-economic census variables (listed in table 2), where the weights capture the proximity and relative contribution of each census tract to a given
location ℓ. For example, the percent of the population that was aged zero to 19 years old in location ℓ at time 9 was calculated as:
]C^ _LG FEA
U\Nℓ:
=
]C^ _LG FEA ∑! `!ℓ: U\N!: ∑! `!ℓ:
19
]C^ _LG FEA where U\N!: is the percent of the population that was aged zero to 19 years old in census tract at KFKaX
time 9, and `!ℓ: =
$A !,ℓ'
more
and
)
, where b!: is population count in census tract at time 9, and V , ℓ is the
distance in miles between location ℓ and the center of census tract . Therefore, census tracts that are populated
]C^ _LG FEA
U\Nℓ:
closer
to
location ℓ
receive
higher
weights
when
constructing
for location ℓ than census tracts that are less populated and further away from location ℓ.
This weighting method was used to construct the socio-economic variables to be used in the estimation for each location ℓ and each decennial year.27 We used linear interpolation to fill in non-decennial years and to fill in decennial years for which a particular variable was not collected from the census. : also contains an index for the intensity of retail activity around location ℓ at time 9, defined as: ON9Ucℓ: = = L
UONUL:
$V: O, ℓ'
where UONUL: is the area (in square feet) of retail building O at time 9, and V: O, ℓ is the distance in miles between retail building O and location ℓ. Therefore, locations with numerous and large retail buildings in proximity will have a high retail activity index. With respect to the error structure in the store-specific profit function latent variable (
ℓ9 ;, <),
R@ is a
“neighborhood”-specific fixed effect that is unobserved to the econometrician and constant over time, and is
common to all stores placed in locations that are within “neighborhood” . We define a neighborhood as having nine locations (or nine “points” in figure 12), with the locations within a particular neighborhood forming a three-by-three square of adjacent points starting from the southwest corner of figure 12. Therefore, there are 16 (= 144 / 9) “square” three-by-three neighborhoods, and 16 such fixed effects, each intended to capture neighborhood-specific factors that affect a store’s profitability and that do not vary over time. Finally, Sℓ: is a location- and store-specific random term at time 9 that has a type I extreme value density.
Therefore, we estimate a conditional logit discrete choice model where retailers choose where to locate their new stores in space.
4.2
Results Table 3 shows the results of the conditional logit location discrete choice model for the supermarket
industry. To address whether the coefficients ;,>?@A , ; , and ; may have varied over time, and therefore whether the nature of competition has changed over time, we split the overall supermarket sample into 27
We note that for each socio-economic Census variable to be used in estimation we applied a proper numerator in the construction of the weights. For example, to construct the percent of occupied housing units with no automobile at location ℓ at time 9, we used the count of total occupied housing units in census tract at time 9 (instead of b!: ) as the numerator of `!ℓ: , because occupied housing units are the underlying units of observation used to measure this variable.
20
three overlapping blocks of 30 years each, starting with the 1956-1985 sample, then the 1968-1997 sample, and then the 1981-2010 sample. We excluded the location choices of supermarket stores that were within large discount retailers (such as Wal-Mart’s proprietary grocery stores) and the location choices of the locally owned supermarket chain that was a cooperative of independently and locally owned stores (Bi-Rite; see figure 6). On the other hand, each of these stores was incorporated in the DF#K G:FLHG construction of the regressors that capture the industry’s competitive environment (i.e., JC,ℓ,:C and
Q DF#K ). We again note that the choice of where to locate a store is used only once in this discrete choice MNOP C,ℓ,:C
estimation, at the year the store entered the market.
The first three columns in table 3 contain the results for our “base specification,” which contains only the key dependent variables of interest and excludes any type of control variables. As is seen, the estimates from the first three columns likely suffer from omitted variable bias, as it probably omits regressors that are likely correlated with the included regressors and that have an impact on the choice of where to locate. In other words, at least some of the regressors in the “base specification” are endogenous and therefore correlated with the error term. This bias is reflected, for example, in the fact that the estimate for the coefficient ; is positive and significant, which is counter-intuitive; this would indicate that, holding the other included regressors constant, increasing the number of competitors’ stores near your store increases the profit of your store. This positive sign is likely capturing the effect of other location-specific factors that have a positive effect on store-level profitability (and therefore affect the number of competitors’ stores around a given location) and that are omitted from the regression, ultimately causing the estimates for the coefficient ; to have a positive bias. We therefore augment our base specification by incorporating the neighborhood-level fixed effects described above and also some of the socio-economic demographic variables as controls. Specifically, the three columns on the right side of table 3, labeled “base specification plus controls,” show the results of the model with these additional control variables whose inclusion is intended to alleviate some of the potential bias. For example, as discussed above, “retail activity” controls for the existence of potential agglomeration economies or restrictive zoning laws, and “age 0 to 19 years old” controls for the fact that drug stores likely prefer to locate in areas where the elderly population are more prevalent. If these control variables are omitted from the estimated model, and if they are correlated with some of the included regressors, then the coefficient estimates in the “base specification” may be capturing some of these effects. Some of the other socio-economic control variables, such as median housing value and median contract rent (see table 2), were omitted from the regression due to high multicollinearity with some of socio-economic regressors included in the “base specification plus controls” specification, such as family
21
income. Additionally, as displayed in the bottom of table 3, we are unfortunately working with a somewhat small sample size (i.e., between 50 and 60 observations for each 30-year block sample), and including additional and likely redundant control variables in the model (such as median housing value) quickly exacerbates the multicollinearity issue and the precision of the estimates without making any meaningful contribution to the reduction of the potential bias in the estimates of the key parameters of interest (i.e., the parameters ;,>?@A , ; , and ; ). As is shown in the last three columns of table 3, for each of the 30-year block samples, the magnitude of the ;,>?@A coefficient on the “3–4” miles distance band is one of two “local maxima” amongst all the distance band coefficients ;,>?@A . That is, the level of the coefficient estimates on the distance bands initially rise to a peak at the “3–4” miles distance band, then decline, and then rise again to a second local peak at the “5–plus” miles distance. This is indication that supermarkets preferred to open new stores in locations where they already had an existing sister store within a “3–4” miles distance band. It also seems that supermarkets avoided placing new stores within the “0–1” mile distance band (which is the omitted dummy among all the ;,>?@A dummies), as the coefficient estimates on all distance bands are positive and many are statistically significant. This is likely due to sales cannibalization that may occur across samefirm stores within such a close distance. The coefficient on the intensity of competitors’ stores (; ) is statistically significant at the ten percent level for the first two 30-year block samples and, as expected, negative across all three samples – the more intense the competition, the lower the profitability. The coefficient on the intensity of clustered chain competitors (; ) is negative across all 30-year samples, and statistically significant at the one percent level during the 1968-1997 period, indicating that density economies were likely relevant in the supermarket retail industry during that period.28 Altogether, if one is willing to accept that the “3–4” miles distance band can be categorized as a “relatively close” distance, then these results are evidence that both the spatial monopolization effect and the density economies effect were dictating the location decision of new supermarket stores. Obviously, a contentious point is whether the “3–4” miles distance band can indeed be accepted as a “relatively close” distance band so that the claim that supermarkets were engaging in spatial monopolization behavior can be credibly made. A study by Fox e al. (2007) may shed some light on the issue. Using a dataset for a major metropolitan market in the Southeastern part of the U.S. for the 2002-2004 time period, those 28
We also considered a radius of two miles and a radius of four miles (instead of three miles) as the criterion for the Q DF#K . For each of these two alternative specifications, the inclusion of competitors’ stores in the variable MNOP C,ℓ,:C coefficient estimates on ; were all non-significant, and they had a combination of positive and negative signs across the different 30-year sample periods. We believe that a two-mile radius is too small and likely fails to properly capture the clustering behavior of competitors around a location – as we saw in table 3, retailers do not have a preference to open new stores within two miles of their existing stores, probably due to cannibalization; on the other hand, a four-mile radius would probably be too large of a geographical area to properly depict clustering behavior by competitors with any kind of precision.
22
authors report that consumers took on average 8.1 minutes to drive from home to their grocery store of choice. In another study that uses data for a different major U.S. market, Fox et al. (2004) report that consumers took on average 11 minutes to drive from home to their grocery store of choice, and 9.6 minutes to their drug store of choice. If we were to assume a conservative driving speed of 15 miles per hour (i.e., 0.25 miles per minute) and a travel time of eight minutes to the consumer’s preferred store, then the average distance travelled to the grocery store is 2 miles. Therefore, if supermarkets are locating new stores within a “3–4” miles distance band from their sister stores, then at least we can argue that consumers located in between the two sister stores should be willing to drive to at least one of the two sister stores (because any consumer located between two sister stores that are no more than four miles from each other would be willing to drive to at least one of the sister stores). If in addition there is no other reasonable substitute (e.g., a competitor supermarket store with similar attributes) located between the two sister stores, then the “3–4” miles distance band can be regarded as a reasonably close distance that justifies the existence of a spatial monopolization motive by chain supermarkets. We will revisit this issue below, when we analyze the percentage of consumers who found themselves “trapped” in geographic space between two sister stores. A question remains on how to interpret the second local peak in the estimates of ;,>?@A , which is at the “5–plus” miles distance band. Distances beyond five miles seem to be far enough such that an attempt of spatial monopolization would not be effective, as competitors would have enough space in between two samefirm sister stores to profitably break up the spatial monopoly. And, as discussed earlier in the theoretical model, if density economies are relevant, as indicated in some of the empirical results from table 3, then the forward-looking strategy of deterrence by dispersion is never a pursued strategy, as a chain retailer is able to deter entry by simply clustering its stores in space and charging more aggressive prices due to its lower costs (because of its density economies). A potential explanation for this second local peak beyond 5 miles is that a portion of the chain supermarkets pursued a spatial monopolization strategy, while another portion (which perhaps does not enjoy density economies) pursued the more forward-looking strategy of deterrence by dispersion. If this was the case, then the two local peaks in the ;,>?@A coefficients are picking up this
heterogeneity in firm behavior. Another potential explanation is suburbanization of the city. As suburbs emerge in the more distant outskirts of the city, a chain supermarket may find it more advantageous to enter such up-and-coming suburban locations before its competitors do, irrespective of the layout of its current network of stores. Using the data on the variable “population density,” we attempted to control for this possibility by creating a location-specific “population growth rate” variable and including it as a control variable in the model, but the results were highly insignificant. A final potential explanation is that at the time a retailer chose to open a new store there were no physically available sites within the desired proximity to its network of stores, and therefore it had to open its new store in available sites that were further away from its
23
network. This is a reasonable possibility given that supermarket stores are typically at least 15,000 square feet in size (see figure 5), and therefore space and site constraints may play a role in the location decision.
Among the control variables in table 3, the coefficient on the intensity of retail activity is positive and statistically significant across all three 30-year samples. As previously discussed, this coefficient may be capturing several potential factors. First, supermarkets may benefit from “agglomeration” economies and prefer sites that have a relative large number of dissimilar retailers, as consumers themselves may prefer such sites in order to minimize their shopping transportation costs.
Second, even if there are no
“agglomeration” economies, the intensity of retail activity at a particular location may capture, in general, the degree of consumer activity within a particular location. Third, this coefficient may also be capturing the effect of zoning law restrictions, as retail activity is likely low in locations where it is difficult for retailers to locate due to such restrictions. Surprisingly, the coefficients on population density in table 3 are mostly negative and non-significant. One would expect it to be positive and significant, because for a given number of competing stores, it is likely that supermarkets would prefer to place stores in locations of high demand.
One possible
explanation is the fact that population density has a high pair-wise correlation with retail activity (= 0.8), so potential high multicollinearity may be affecting the empirical identification of the parameter on population density, and additionally causing its estimate to swing from positive to negative across the 30year samples. The coefficients on the percentage of young population (i.e., “Age 0 to 19 years old”) have a positive and, in one 30-year sample, statistically significant coefficient. This may be indication that supermarkets enjoy placing stores in locations where there is a high percentage of families with children, as opposed to an older population. Intuitively, families with children are more likely to obtain their food from supermarkets versus dining out or the like. Not surprisingly, family income is a positive and significant predictor for the location choice of supermarkets. Finally, the coefficient on the percentage of the population that is white at a given location is negative. This is an interesting result, especially given the fact that we are already controlling for family income. One potential explanation is that the non-white population is less likely to dine out, and more likely to prepare a meal at home and therefore purchase their food from a supermarket store. In particular, if the opportunity cost of preparing a meal at home is lower for non-white families versus white families, then this negative coefficient may be capturing that fact – a non-white family is more likely to purchase food in a supermarket and in turn prepare a meal at home.29
29
For example, if a higher percentage of the non-white female population is not in the labor force (as opposed to the white female population), then their opportunity cost of preparing a meal is likely lower and therefore they are more likely to engage in such activity.
24
It is worth noting that the location of distribution centers may also play a role in the location choice of retail stores, particularly in the supermarket industry. We do not possess data on the location of either supermarket or drug store distribution centers over time. At the same time, we believe that once a chain retailer has chosen to enter an urban market such as Greensboro, the location of distribution centers would not be a significant factor for the location choice of its stores within the market. That is, the location of the distribution centers for, say, Kroger supermarkets may have been a determinant factor for whether Kroger chose to enter the Greensboro versus the Columbia (SC) market, for example. However, once Kroger has chosen to enter the Greensboro market, it is likely that the location of its distribution centers did not play a role in the choice of where to locate its stores within the city itself. In this sense, the statistically significant result for the presence of density economies in the supermarket industry for the 1968-1997 sample period, shown in table 3, likely reflects factors such as shared advertising across nearby same-firm sister stores, shared inventory, shared management, easier product quality inspection, and easier employee monitoring. It may also reflect the economics and logistics of delivering products to stores, in the sense that once a proprietary delivery truck has arrived from the distribution center into the urban market (in this case, Greensboro), then it is easier to complete all the deliveries if the stores are in proximity to each other. It is also worth noting that there may be remaining omitted variable bias in the coefficient estimates of the key parameters of interest, i.e., the key “competition” regressors in the models in table 3 may still be endogenous. Specifically, there may be unobservable variables that affect location choice and that are also serially correlated over time. If these unobservable variables vary over time, then they are not being captured by the neighborhood fixed effects. Therefore, they would likely cause bias in the key parameters of interest, even though we use lagged values (i.e., time “9 − 1”) for some of the key parameters of interest. Table 4 shows the results for the drug store industry. We excluded the location choices of drug stores that were within supermarkets or discount retailers (such as Wal-Mart), although each of these stores was incorporated in the construction of the regressors that capture the industry’s competitive environment
DF#K G:FLHG Q DF#K ). As in the supermarket industry, the coefficient estimates from the (i.e., JC,ℓ,:C and MNOP C,ℓ,:C
“base specification” (first three columns in table 4) likely suffer from omitted variable bias. For example, the coefficient on competitors’ store is positive and highly significant, a highly counter-intuitive result. Therefore, the last three columns of table 4 show the results for the drug store industry with the addition of the control variables. Overall, the results for the drug store industry displayed in table 4 are for the most part qualitatively similar to the results from the supermarket industry, although a smaller number of coefficients are
25
statistically significant. With regards to the key parameters of interest, and for each of the 30-year samples, the magnitude of the ;,>?@A coefficient on the “3–4” miles distance band is also one of two “local maxima” amongst all the distance band coefficients. That is, the level of the coefficient estimates on the distance bands initially rise to a peak at the “3–4” miles distance band, then decline, and then rise again to a second local peak at the “5–plus” miles distance. It seems that drug stores also avoided, for the most part, placing new stores within the “0–1” mile distance band (which is the omitted dummy among all the ;,>?@A dummies), as the coefficient estimates on all but two distance bands are positive and many
are statistically significant. The coefficients on the intensity of competitors’ stores (; ) is negative across
all three 30-year samples, although they are all statistically insignificant. The coefficient on the intensity of clustered chain competitors ( ; ) is also negative across all 30-year samples, but not statistically significant in any of the 30-year samples, indicating that density economies were potentially less relevant (if relevant at all) in the drug store retail industry compared to the supermarket industry. As argued above, if one is willing to accept that the distance band “3–4” can be categorized as a “relatively close” distance, then based on the signs of these coefficient estimates one may argue that both the spatial monopolization effect and the density economies effect may have been the reasons for drug stores locating new stores near an already existing sister store, although the degree of density economies was likely less important in the drug store industry than in the supermarket industry. As described above, Fox et al. (2004) report that consumers took on average 9.6 minutes to drive from home to their drug store of choice. By building an argument similar to the one we made above, it is possible that the “3–4” miles distance band is a reasonably close distance to justify the existence of a spatial monopolization motive by chain drug stores. Among the control variable in table 4, the coefficient on the intensity of retail activity is again positive and statistically significant across all three 30-year samples, likely due to the same reasons discussed above for the supermarket results. The coefficients on population density are again mostly negative and non-significant, likely due to its high collinearity with retail activity. The coefficients on the percentage of the young population have a negative sign in two of the three 30-year samples and, in one of the samples, a statistically significant coefficient. This result is intuitive, as drug stores likely benefit from placing stores in locations with an older population, as those individuals are more likely to require prescription medication than the youth. Family income once again has positive coefficient estimates, although it is statistically significant in only one of the 30-year samples. Finally, the coefficient on the percentage of the population that is white is negative for the most part. This may be reflecting the fact that the underlying health status of the non-white population might be lower than the health status of the white population, which could translate into the non-white population needing more medication (per person) than the white population.
26
In sum, both chain supermarkets and drug stores preferred to place new stores within a “3–4” miles distance band and also within a “5–plus” miles distance from their existing stores. Density economies appear to have been significant for supermarkets, particularly within the 1968-1997 time period, an indication that both the spatial monopolization effect and the density economies effect were present in this industry. For drug stores, it seems that the spatial monopolization effect was more relevant than the density economies effect. The fact that density economies were more relevant for the supermarket industry than for the drug store industry is intuitive. For example, supermarkets carry more perishable products than drug stores. It is likely that the logistics involved in delivering perishable products are more evolved (and more costly) than the shipment of, for example, prescription medication. Therefore, having sister supermarket stores in close proximity may save in distribution costs, as the overall delivery route can be completed faster (and less costly) which in turn implies fresher produce to customers. Additionally, supermarkets may require a higher degree of product monitoring once the perishable products are in a store (to ensure that customers have fresher produce available to them).
And,
supermarket stores are typically larger and have more employees than drug stores, and therefore employee monitoring is likely a larger and more complex task than in drug stores. This suggests that it may be more efficient for a supermarket firm to have its stores in reasonable proximity to each other, as management at the firm level (in a market such as Greensboro) may be able to monitor product quality inspection and employee performance more efficiently when the firms’ stores are in closer proximity.
4.3
An alternative to “distance bands” As a robustness check to our results, we consider an alternative specification where the distance band DEFGHG: FI@ >?@A
indicator variables, B ℓ,:C
DEFGHG: FI@ , are replaced by a quadratic function of d9 ℓ,:C :
DF#K G:FLHG DEFGHG: FI@ Q DF#K + < ∗ ℓ: + R@ + Sℓ: ' + ; ∗ JC,ℓ,:C + ; ∗ MNOP ℓ: ;, < = = ;,K ∗ $d9 ℓ,:C Cℓ,:C K
Ke
DEFGHG: FI@ where d9 ℓ,:C is the distance (in miles) between location ℓ and retailer ’s closest own store to
location ℓ that was active at time 9 − 1. Therefore, we specify a parametric non-linear function for how the distance to the closest sister store affects the decision of where to locate a new store. Once this function is estimated, we can collect the parameter estimates for ;, and ;, and use them to calculate the distance (in miles) to the closest sister store that maximizes the predicted probability of choosing a location. If this fitted “probability-maximizing” distance is within the distance band(s) that best predicts
27
the location choices of new stores from the previous distance band econometric specification, then the two specifications likely corroborate each other’s findings.30 Table 5 shows the results for supermarkets (first three columns) and for drug stores (last three columns) using a similar “base specification plus controls” specification from tables 4 and 5 and the same three different 30-year sample periods considered above (each column is for a different sample period). The box shaded in gray in the bottom of the table shows the estimated distance (in miles) to the closest sister store that maximizes the probability of selecting a location for a new store. Except for the last column (i.e., drug stores in the 1981-2010 period), the probability-maximizing distances are between four miles and five-and-a-quarter miles, which is slightly larger than the first local “peak” at the “3–4” miles distance band in tables 3 and 4. This is consistent with the results from tables 3 and 4 in the sense that these probability-maximizing distances are picking up both the first local “peak” at the “3–4” distance band and the second “peak” at the “5–plus” distance. Because we are using a quadratic specification, we can only have one “peak” with this polynomial specification, and that single peak is picking up both of the two local peaks from tables 3 and 4. We also estimated a specification with a polynomial function to the power four, which would allow for the identification of two local maxima from the estimated coefficients, but the results were not precise or even within close proximity to the “3–4” miles distance range predicted in tables 3 and 4. This is likely due to the high degree of multicollinearity between the power regressors. The other coefficient estimates in table 5 are qualitatively similar to the results from tables 3 and 4. Density economies are more prevalent (and statistically significant for one of the 30-year sample periods) in the supermarket industry than in the drug store industry, and retail activity is a major predictor for the location of new stores in both retail industries.
4.4
Were consumers indeed spatially monopolized? Our previous empirical results, combined with the predictions from the simple theoretical model laid
out in section 2, indicate that supermarkets and drug stores consistently preferred to open new stores within a somewhat reasonable proximity to one of their pre-existing stores, an indication that the spatial monopolization effect may have been one of the motives for this behavior. As a corroborative check to this claim, we inspect whether spatial monopolization indeed occurred in practice. Specifically, we investigate whether the consumers who found themselves located between two retail stores owned by the
30
Under our logit discrete choice specification, this “probability-maximizing” distance is the distance that also maximizes b` K the function ∑Ke ;f,K $d9 "cbdNd9 ' , where ;f,K are the estimated coefficients from the model. ℓ,9−1
28
same firm were indeed spatially “trapped” by the chain retailer, or if these consumers had other nearby shopping alternatives.31 We implement this consistency check in the following manner. For each new entering retail store, at the year of its entry, we identified its closest existing sister store at that time. This provided us with “entering store”–“closest sister store” pairs of stores owned by a same retailer. For each of these pairs, we then isolated the set of consumers in geographic space that were under highest contention of being spatially monopolized by this sister-stores pair at the year of entry of the “entering store.” Specifically, we defined these at-stake consumers as those inside an ellipse in geographic space that connects the “entering store” to the “closest sister store.” For each “entering store”–“closest sister store” pair, we defined its ellipse as having a height in the middle that is half the distance between the two sister stores
that comprise the pair. For example, the following picture depicts a hypothetical retailer that has
opened a new store, , within a distance of V miles from its closest existing sister store, :
●
●
V 2
V The shaded gray area inside the ellipse depicts the potentially spatially monopolized consumers under consideration for this pair of sister stores.32 Next, for each consumer inside the ellipse, we analyzed whether the two sister stores, and (and in addition any other store belonging to firm ), were
among the “ ” closest stores to the consumer (where may be 2, 3, or more stores). For example, if
there were no competitors’ stores in between the two sister stores or in their surroundings, then it is likely that each consumer inside the ellipse was spatially monopolized and had no other option but to shop at one of the two stores, increasing the ability of the retailer to charge higher prices. In other words, the stores and were the 2 (= ) closest stores to these consumers. However, if there were competitors’
stores either inside the ellipse or in its immediate surroundings, then it is less likely that the two sister
stores and (and in addition any other store belonging to firm ) were among the two or even three closest stores to a given consumer inside the ellipse.
31
The following spatial monopolization analysis is in a similar qualitative spirit as the one conducted by Stelder (2011). The actual implementation of this analysis involved discretizing Greensboro and its surroundings into adjacent square grids of 0.15 by 0.15 miles, and defining the center of each grid as a “consumer” point.
32
29
To illustrate, consider the following scenario, where competitor has a store slightly to the
Northwest of store :
●
●
h ●
g ●
●
For the case of = 2, retailer is able to monopolize consumer g , but not consumer h , because
consumer h is closer to store than to store . More specifically, all consumers to the right of the
dashed line are spatially monopolized by retailer for the case of = 2. (The dashed line is the perpendicular bisector of the dotted line that connects store to store .) For the case of = 3, both
and are among the 3 = closest stores to all consumers within the ellipse. Therefore, = 2 is a
stronger criterion for consumer “trapping” than = 3.
Table 6 shows the average of the proportion of consumers inside an “entering store”–“closest sister store” pair ellipse (at the year of entry of the “entering store”) whose closest “ ” stores contained the two chain sister stores (or any other store belonging to the firm with the pair of sister stores).33 We separated the results by industry (first column), by two distinct time periods (1956-1980 versus 1981-2010, second column), and by whether the “entering store” entered within a certain distance (in miles) to its closest sister store (third column). As is shown, for the most part, the closer the two sister stores are to each other at the time of entry of the “entering store,” the more consumers are spatially monopolized, as the percent of consumers “trapped” increases as one moves down the rows within any of the four combinations of industry and sample period, irrespective of which “ ” is considered. This is an intuitive result – the closer the pair of sister stores are to each other, the less space there is for competitors in the middle. Additionally, it seems that supermarkets have been more successful in spatially monopolizing consumers than drug stores, as the percentages in the top two panels (“supermarkets”) are higher than the percentages in the bottom two panels (“drug stores”). Looking at the strongest form of spatial monopolization, i.e., the “ = 2” category, we observe that supermarkets who located a new store within two miles of an existing sister store during the 1981-2010 period were able to spatially monopolize, on average, 33.2 percent of the at-stake consumers. This is a
33
The proportion of consumers for each ellipse is weighted by the population density at each discretized consumer point.
30
relative high percentage, especially if one considers the fact that it takes only one competitor store between the two sister stores to break the spatially monopoly for all consumers within the ellipse under
the “ = 2” category. By slightly relaxing the requirement for spatial monopolization and looking at the
next column to the right, i.e., the “ = 3 ” category, the average of the proportion of spatially
monopolized consumers increases to more than half, at 56.8 percent. This category (“ = 3”) may still be very relevant with respect to spatial monopolization.
As shown by Datta and Sudhir (2011), a
supermarket of a particular “type” (e.g., a supermarket that brands itself as “organic”) may compete more fiercely against other supermarkets of the same type (another “organic” supermarket), and less intensely with a supermarket of a different type (e.g., a low-cost, low-price supermarket). Therefore, it could potentially be realistic to consider the “ = 3” as indication of meaningful spatial monopolization,
especially if the third store (i.e., the competitor’s store) is of a different type than the pair of clustered sister stores.34 By relaxing the “distance” constraint and looking at the supermarkets that entered within four miles of sister stores during the 1981-2010 period, we still observe some degree of spatial monopolization,
particularly for supermarkets in the 1981-2010 time period. For example, for the “ = 2” category, the
proportion of spatially monopolized consumers is 13.1 percent, and for “ = 3” the proportion is 27.2 percent. More broadly, if one is willing to accept the “ = 3” category as a reasonable criterion of spatial
monopolization, then both supermarkets and drug stores were successful in “trapping” and spatially monopolizing anywhere from 14.0 percent to 56.8 percent of consumers across different time periods, depending on which row within the “ = 3” category column one considers. Such percentages of “trapped” consumers seem to be high enough to motivate chain retailers to engage in spatial monopolization behavior.
5
Conclusion This paper presented an empirical investigation of the location strategy pursued by chain retailers over a
period of 55 years. We used a unique data set containing the sequential location decisions of supermarkets and drug stores in Greensboro, NC. Using a discrete choice model for location in geographic space, we found that chain retailers preferred to open new stores within a “3–4” miles distance band from their closest existing sister store. Economies of density were partially responsible for such location behavior in the supermarket industry, although this motive was less intense in the drug store industry. Moreover, spatial monopolization intent was a pursued strategy by both supermarkets and drug stores throughout the entire sample period. Additionally, we
34
Unfortunately, our data do not contain details on the “type” of a supermarket. Otherwise, this spatial monopolization analysis could be refined further.
31
also found that in some cases up to 33.2 percent of consumers located between two stores owned by the same retailer were “trapped” in space by the retailer, with no closer shopping alternative. The paper also proposed and implemented an empirical strategy to identify the presence of density economies in retail industries. The strategy is simple to implement and does not require data on store prices or costs. Rather, it only requires data on the location and firm ownership structure of stores. Such empirical strategy could potentially be used, for example, in merger analysis to investigate whether cost economies are relevant within a particular retail industry, and whether a particular merge may benefit consumers via increased cost efficiencies of merged firms.
32
Appendix Algebraic Derivations of Model Recall the initial spatial market structure discussed in the text, , ℓ, , , , … , , where retailer has
two stores, and , and all other retailers have one store each. If we denote the location of store as point 0, the consumer who is located between stores and ℓ and is indifferent between them is:
!ℓ = The demand for store is then:
1 1 ℓ − ! + 2 2
! = $1 − i! ' + !ℓ ! =
1 1 $i + ℓ − 2! ' + 2
where i! is the consumer indifferent between stores and . In fact, the demand any store faces is given by the general rule: =
1 1 C + j − 2 + 2
where “-1” is the adjacent store to its clockwise left and “ +1” is the adjacent store to its right.
A single-store retailer, such as , chooses the price of its only store to maximize its variable profits: !
where " is marginal cost.
= ! − " !
A multi-store retailer, such as , chooses its prices #% and #) for its stores jointly to maximize its total
variable profits: #
= $#% − 1 − &"' #% + $#) − 1 − &"' #)
where & ∈ 0,1 captures the degree of economies of density in variable costs. The best-response function of a single-store retailer such as is:
1 1 !kl = $i + ℓ ' + m" + n 4 2
Note how price best-response functions depend only on prices of adjacent stores. The best-response function of a chain store such as that has an adjacent sister store is:
1 1 1 1 − &" kl # = ℓ + #) + o + p % 4 2 2 2
33
Therefore a chain retailer with adjacent sister stores internalizes the cannibalization effect amongst them when chooses prices. Equilibrium prices are obtained by solving the system of price best-response functions, and facilitated by imposing symmetry with respect to the relative location and type (i.e., single- or multi-store) of store. Equilibrium demands and profits then follow by straight forward calculation. We now provide the equilibrium solutions for prices, demands, and profits for the spatial market configurations considered in the examples in the text. Market structure , , : By imposing symmetry in the price best-response functions (q∗ % = k∗ = r∗ ), equilibrium prices, demands, and profits at each store are: ∗ = " +
∗ =
1
∗
=
Market structure , , , : Retailer has its two stores isolated from each other, and therefore no density economies can be enjoyed. Equilibrium prices, demands, and profits are the same as in market structure , , , but with = 4.
Market structure , , , : Retailer has its two stores adjacent to each other, and therefore it enjoys density economies if & > 1 .
By imposing symmetry given the location and firm profile we have in
equilibrium: q∗ % = q∗ )
k∗ = r∗
Solving for the system of price best-response we obtain equilibrium prices, demands, and profits: q∗ % = q∗ ) = " + k∗ = r∗ = " +
1 m24 − 9&"n 15
q∗% = q∗) =
1 m18 − 3&"n 15 =
∗ q% ∗ k
=
∗ q)
=
∗ r
=
k∗ = r∗ =
1 m24 + 6&"n 450
1 m24 + 6&"n 30
1 m36 − 6&"n 30
1 m36 − 6&"n 30
Market structure , , , , ,: By imposing symmetry given the location and retailer ownership profile we have in equilibrium: q∗ % = q∗ )
k∗ = s∗
Solving for the system of price best-response we obtain equilibrium prices, demands, and profits:
34
q∗ % = q∗ ) = " + k∗ = s∗ = " +
1 m19 − 7&"n 12
q∗% = q∗) =
1 m14 − 2&"n 12
r∗ = " +
k∗ = s∗ =
1 m13 − &"n 12 ∗ q%
=
∗ q)
∗ k
=
∗ s
∗ r
=
= =
r∗ =
1 m19 + 5&"n 24
1 m28 − 4&"n 24
1 m26 − 2&"n 24
1 m19 + 5&"n 288
1 m28 − 4&"n 576
1 m26 − 2&"n 576
Market structure , , , ,, : This market structure has the same equilibrium solution as the above
market structure , , , , ,, except that now becomes ,, and vice-versa.
Market structure , , , , , ,: By imposing symmetry given the location and retailer ownership profile we have in equilibrium: q∗ % = q∗ ) = r∗ % = r∗ )
k∗ = s∗
Solving for the system of price best-response we obtain equilibrium prices, demands, and profits: 1 q∗ % = q∗ ) = r∗ % = r∗ ) = " + m5 − 2&"n 3
q∗% = q∗) = r∗% = r∗) =
1 k∗ = s∗ = " + m4 − &"n 3 ∗ q%
=
∗ q) ∗ k
= =
∗ r%
=
∗ s
=
k∗ = s∗ = ∗ r)
=
1 m5 + &"n 6
1 m8 − 2&"n 6
1 m5 + &"n 18
1 m4 − &"n 9
Market structure , , , , ,, : By imposing symmetry given the location and retailer ownership profile we have in equilibrium: q∗ % = q∗ )
k∗ = r∗ )
s∗ = r∗ %
Solving for the system of price best-response we obtain equilibrium prices, demands, and profits: q∗ % = q∗ ) = " +
1 m30 − 11&"n 19
k∗ = r∗ ) = " +
q∗% = q∗) =
1 m22 − 3&"n 19
k∗ = r∗) =
35
1 m15 + 4&"n 19
1 m22 − 3&"n 19
r∗ % = s∗ = " +
1 m20 − &"n 19 ∗ q%
=
∗ q)
=
∗ k
=
∗ r)
=
∗ r%
=
∗ s
=
r∗% = s∗ =
2 m15 + 4&"n 361
1 m22 − 3&"n 361 1 m20 − &"n 361
36
1 m20 − &"n 19
References Aguirregabiria, V. and Vicentini, G., 2012, “Dynamic Spatial Competition Between Multi-Store Firms,” working paper, University of Toronto. Bolduc, D., Fortin, B., and Fournier, M.-A., 1996, “The Effect of Incentive Policies on the Practice Location of Doctors: A Multinomial Probit Analysis,” Journal of Labor Economics, 14(4), 703-732. Bonanno, G., 1987, “Location Choice, Product Differentiation and Entry Deterrence,” Review of Economic Studies, 54(1), 37-45. Brito, D., 2003, “Preemptive Mergers under Spatial Competition,” International Journal of Industrial Organization, 21, 1601-1622. Datta, S., and Sudhir, K., 2011, “The Agglomeration-Differentiation Tradeoff in Spatial Location,” working paper, Yale University. Ellickson, P., Houghton, S., and Timmins, C., 2012, “Estimating Network Economies in Retail Chains: A Revealed Preference Approach,” working paper. Eaton, C. and Lipsey, R., 1979, “The Theory of Market Pre-emption: The Persistence of Excess Capacity and Monopoly in a Growing Spatial Market,” Economica, 46, 149-158. Fox, E., Montgomery, A., and Lodish, L., 2004, “Consumer Shopping and Spending Across Retail Formats,” The Journal of Business, 77(S2), pp. S25-S60. Fox, E., Postrel, S., and McLaughlin, A., 2007, “The Impact of Retail Location on Retailer Revenues: An Empirical Investigation,” working paper. Giraud-Heraud, E., Hammoudi, H., and Mokrane, M., 2003, “Multiproduct Firm Behavior in a Differentiated Market,” Canadian Journal of Economics, 36(1), 41-61. “Greensboro, North Carolina,” Hayes Drug Store Directory, Hayes Directories Inc., 1955-2009. “Greensboro, North Carolina,” Sanborn Fire Insurance Maps, The Sanborn Map Company Inc., 1919, 1966, 1972. “Greensboro City Directory,” published by Hill Directory Co. Inc., Richmond, Virginia, 1955-1982. “Greensboro City Directory,” published by R. L. Polk Co., Richmond, Virginia, 1983-2010. “Greensboro Suburban Directory,” published by Hill Directory Co. Inc., Richmond, Virginia, 1966-1982. “Greensboro Suburban Directory,” published by R. L. Polk Co., Richmond, Virginia, 1983-2010. Holmes, T., 2011, “The Diffusion of Wal-Mart and Economies of Density,” Econometrica, 79(1), 253-302. Jia, P., 2008, “What Happens When Wal-Mart Comes to Town: An Empirical Investigation of the Discount Retailing Industry,” Econometrica, 76(6), 1263-1316. Judd, K., 1985, “Credible Spatial Preemption,” The RAND Journal of Economics, 16(2), 153-166.
37
Levy, D. and Reitzes, J., 1992, “Anticompetitive Effects of Mergers in Markets with Localized Competition,” The Journal of Law, Economics and Organization, 8(2), 427-440. Minnesota Population Center. National Historical Geographic Information System: Pre-release Version 0.1. Minneapolis, MN: University of Minnesota 2004. (www.nhgis.org) Netz, J. and Taylor, B., 2002, “Maximum of Minimum Differentiation? Location Patterns of Retail Outlets,” Review of Economics and Statistics, 84(1), 162-175. Nishida, M., 2012, “Estimating a Model of Strategic Network Choice: The Convenience-Store Industry in Okinawa,” NET Institute Working Paper No 08-27. Orhun, Y., 2012, “Spatial Differentiation in the Supermarket Industry: The Role of Common Information,” Quantitative Marketing and Economics. Ransome, J., 1961, “The Organization of Location Research in a Large Supermarket Chain,” Economic Geography, 37(1), 42-47. Salop, S., 1979, “Monopolistic Competition with Outside Goods,” Bell Journal of Economics, 10, 141-156. Seim, K., 2006, “An Empirical Model of Firm Entry with Endogenous Product-Type Choices,” The RAND Journal of Economics, 37(3). Smith, H., 2004, “Supermarket Choice and Supermarket Competition in Market Equilibrium,” Review of Economic Studies, 71(1), 235-263. Stelder, D., 2011, “Spatial Monopoly of Multi-Establishment Firms; An Empirical Study for Supermarkets in the Netherlands,” Papers in Regional Science, forthcoming. Thomadsen, R., 2005, “The Effect of Ownership Structure on Prices in Geographically Differentiated Industries,” The Rand Journal of Economics, 36(4), 908-929. U.S. Census Bureau, “American Fact Finder,” http://factfinder.census.gov/home/saff/aff_transition.html, accessed July 15, 2011. U.S. Census Bureau, Economic Census of Retail Trade, Subject Series Tables, “Single- and Multi-Unit Firms for the United States,” 1948-2007. Vickrey, W. S., 1964, “Microstatistics,” Harcourt, Brace and World, New York, NY. Watson, R., 2005, “Entry and location choice in eyewear retailing,” working paper, University of Texas at Austin. West, D., 1981, “Testing for Market Preemption Using Sequential Location Data,” The BELL Journal of Economics, 12(1), 129-143.
38
Table 1: Prevalence of Chain Firms in Retail Markets in the United States 1948
1958
1967
1977
1987
1997
2007
All retailers Chain firms as % of all retail firms Chain stores as % of all retail stores Chain store sales as % of all retail sales Number of stores per chain firm (Sales per chain firm store) / (Sales per single-unit store)
9.2 29.6
10.2 33.7
2.2 12.5 39.8
2.8 17.9 48.0
6.5 33.1 56.5
6.8 38.6 62.1
5.5 40.1 68.1
-
-
6.5
7.6
7.1
8.6
11.5
4.2
4.5
4.6
4.2
2.6
2.6
3.2
8.2 41.3
9.6 53.0
1.5 15.4 61.1
2.5 25.7 71.4
6.7 39.5 77.8
6.3 37.3 83.9
4.8 38.5 85.7
-
-
12.0
13.5
9.0
8.8
12.5
7.9
10.6
8.6
7.2
5.4
8.7
9.5
10.8 27.7
12.7 28.8
3.6 16.5 39.1
5.2 29.6 58.6
7.1 40.4 68.7
7.6 50.9 74.5
7.1 57.9 76.3
-
-
5.3
7.6
8.9
12.7
17.9
3.2
2.8
3.2
3.4
3.2
2.8
2.3
Supermarkets and other grocery stores Chain firms as % of all retail firms Chain stores as % of all retail stores Chain store sales as % of all retail sales Number of stores per chain firm (Sales per chain firm store) / (Sales per single-unit store) Drug stores Chain firms as % of all retail firms Chain stores as % of all retail stores Chain store sales as % of all retail sales Number of stores per chain firm (Sales per chain firm store) / (Sales per single-unit store)
Notes: 1. Data on number of chain firms were not available for 1958 and earlier; 2. Pre-1987 data include retailers with and without payroll, while data for 1987 and later exclude retailers without payroll; 3. The category "Supermarkets and other grocery stores" excludes convenience stores after 1987; 4. SIC industry code 541 (years 1948-1987) and NAICS industry code 44511 (years 1997-2007) were used for the category "Supermarkets and other grocery stores"; 5. SIC industry code 591 (years 19481987) and NAICS industry code 446110 (years 1997-2007) were used for the category "Drug stores." Source: US Economic Census of Retail Trade, Subject Series Tables, “Single- and Multi-Unit Firms for the United States ,” 1948-2007.
39
Table 2: Summary Statistics of Tract-Level Socio-Economic Variables - Guilford County, NC - 1950-2010 1950
1960
1970
1980
1990
2000
2010
45
64
78
87
94
98
118
14.6 (16.2) [0.5 - 38.9]
10.3 (14.5) [0.4 - 44.1]
8.4 (12.6) [0.2 - 40.1]
7.6 (12.0) [0.4 - 47.7]
7.0 (11.0) [0.4 - 47.7]
6.7 (10.6) [0.4 - 47.7]
5.5 (9.2) [0.4 - 48.1]
4,246 (1,837) [1,606 - 8,747]
3,852 (1,692) [1,041 - 7,574]
3,700 (1,589) [358 - 9,073]
3,645 (1,667) [703 - 9,190]
3,696 (1,614) [542 - 10,456]
4,296 (2,245) [550 - 15,822]
4,139 (1,647) [1,300 - 9,324]
Population density (persons per square mile)
2,508 (2,695) [44 - 8,557]
2,418 (2,480) [48 - 10,343]
2,322 (2,197) [47 - 9,273]
2,100 (1,686) [66 - 8,418]
1,991 (1,519) [76 - 8,055]
2,108 (1,509) [91 - 7,919]
2,110 (1,397) [94 - 6,640]
Population 0-19 years old (as percent of tract population)
36.6 (5.9) [20.1 - 46.9]
40.3 (4.4) [23.5 - 48.7]
36.8 (7.5) [10.3 - 58.4]
29.7 (7.2) [8.5 - 47.9]
25.7 (5.8) [5.6 - 41.0]
26.3 (6.5) [4.5 - 49.9]
26.5 (6.4) [7.6 - 49.7]
Population 20-54 years old (as percent of tract population)
52.0 (5.6) [42.2 - 63.8]
46.6 (2.8) [41.1 - 53.1]
46.5 (5.1) [34.0 - 67.5]
49.1 (5.4) [39.6 - 69.8]
52.6 (6.0) [40.1 - 67.6]
52.8 (6.5) [36.5 - 73.2]
49.4 (7.2) [33.0 - 76.5]
Population 54-plus years old (as percent of tract population)
11.4 (2.3) [6.9 - 17.4]
13.1 (3.9) [3.7 - 24.8]
16.8 (7.2) [6.6 - 37.9]
21.2 (8.7) [5.0 - 47.1]
21.7 (7.1) [2.7 - 40.4]
20.9 (7.2) [1.7 - 43.7]
24.2 (7.8) [3.6 - 47.4]
Male population (as percent of tract population)
48.5 (3.6) [30.1 - 53.9]
48.5 (3.3) [26.7 - 52.6]
47.7 (4.8) [22.7 - 66.5]
47.1 (3.2) [30.3 - 56.6]
47.4 (2.9) [34.0 - 56.2]
48.2 (3.8) [36.4 - 75.0]
47.7 (2.6) [37.8 - 60.9]
Non-white population (as percent of tract population)
18.7 (23.1) [0.1 - 94.8]
17.0 (26.2) [0.0 - 99.7]
20.6 (30.4) [0.0 - 99.8]
25.0 (29.0) [0.0 - 99.7]
29.1 (28.6) [0.6 - 99.7]
38.4 (29.3) [0.3 - 99.6]
43.5 (27.7) [1.8 - 98.8]
Persons without high-school diploma (as percent of persons 25+ yrs old)
71.3 (19.6) [18.3 - 95.1]
61.8 (18.6) [12.6 - 90.8]
56.2 (18.8) [13.0 - 89.3]
40.0 (18.1) [4.4 - 78.3]
25.8 (15.0) [1.3 - 62.7]
19.8 (13.2) [0.0 - 55.5]
14.2 (10.2) [0.0 - 48.6]
Unemployed persons (as percent of civilian labor force)
2.3 (1.3) [0.0 - 5.3]
2.4 (1.3) [0.0 - 5.4]
2.2 (1.3) [0.0 - 5.9]
5.1 (3.1) [0.6 - 13.7]
4.7 (3.6) [0.7 - 22.9]
6.0 (6.9) [0.7 - 50.1]
9.5 (6.8) [0.0 - 36.3]
Housing units at most 10 years old (as percent of all housing units)
28.8 (18.9) [1.4 - 73.4]
38.7 (18.3) [4.5 - 85.3]
31.2 (21.8) [0.0 - 88.1]
23.4 (17.1) [0.0 - 57.5]
21.6 (17.0) [0.7 - 71.4]
19.9 (17.5) [0.0 - 75.3]
15.5 (13.9) [0.0 - 63.3]
Housing units 11-20 years old (as percent of all housing)
18.6 (7.7) [2.9 - 35.8]
20.2 (9.3) [4.8 - 46.5]
26.8 (12.5) [0.0 - 65.8]
25.9 (13.8) [3.1 - 77.3]
20.1 (12.1) [0.9 - 48.2]
14.5 (9.9) [1.4 - 55.2]
18.3 (14.4) [0.0 - 61.4]
Housing units 21-plus years old (as percent of all housing units)
52.6 (22.0) [8.7 - 91.2]
41.1 (21.4) [3.4 - 90.6]
42.0 (22.6) [6.3 - 100.0]
50.7 (24.6) [8.5 - 91.8]
58.4 (24.2) [3.7 - 96.5]
65.6 (23.1) [11.1 - 97.5]
66.2 (23.6) [9.4 - 99.3]
n/a
17.9 (14.5) [0.0 - 66.2]
16.8 (16.8) [0.0 - 79.6]
11.3 (12.7) [0.0 - 70.2]
10.8 (11.8) [0.0 - 60.9]
9.3 (9.9) [0.0 - 52.7]
8.1 (8.5) [0.0 - 42.7]
2,722 (984) [458 - 5,659]
5,506 (1,681) [2,638 - 11,742]
9,705 (2,957) [4,535 - 20,494]
18,937 (6,412) [6,914 - 43,542]
36,683 (13,275) [14,037 - 103,601]
51,446 (19,810) [17,892 - 137,595]
60,592 (27,376) [13,882 - 161,818]
Median value of housing units (among owner-occupied housing units)
6,218 (2,915) [3,138 - 15,015]
9,677 (4,099) [4,856 - 27,160]
15,650 (6,340) [6,838 - 37,705]
41,588 (17,194) [14,453 - 103,172]
80,966 (41,180) [32,100 - 298,900]
112,751 (53,561) [47,700 - 336,300]
155,842 (69,902) [56,000 - 440,600]
Median contract rent of housing units (among renter-occupied housing units)
27 (15) [7 - 76]
47 (14) [17 - 91]
75 (24) [29 - 162]
166 (59) [88 - 367]
344 (101) [172 - 675]
497 (164) [219 - 1,250]
628 (214) [199 - 1,910]
Observations Tract area (in square miles) Population
Occupied housing units with no automobile (as percent of occupied housing units) Median family income
mean (st dev ) [min - max ]
Note: The value "n/a" in 1950 indicates that the variable was not collected at the tract-level for Guilford County in the 1950 Census. Sources: U.S. Census Bureau and National Historical Geographic Information System.
40
Table 3: Discrete Choice Model for Location Decision - Conditional Logit With Distance-Band Dummies Supermarkets - Estimates for Blocks of 30 Years Base specification 1956-1985 Closest own store is within 1.0 - 2.0 miles Closest own store is within 2.0 - 3.0 miles Closest own store is within 3.0 - 4.0 miles Closest own store is within 4.0 - 5.0 miles Closest own store is beyond 5.0 miles Competitors' stores Modified Herfindahl index (3 miles)
δ1, 1-2 δ1, 2-3 δ1, 3-4 δ1, 4-5 δ1, 5+ δ2 δ3
1968-1997
Base specification plus controls 1981-2010
1956-1985
1968-1997
1981-2010
-0.501
-0.393
1.633
0.205
0.193
2.558**
(0.444)
(0.491)
(1.046)
(0.456)
(0.503)
(1.061)
-0.525
-0.215
1.466
0.827*
0.964*
2.857***
(0.438)
(0.492)
(1.057)
(0.476)
(0.528)
(1.083)
-1.007**
-0.720
1.467
1.417**
1.358**
3.260***
(0.495)
(0.559)
(1.068)
(0.593)
(0.634)
(1.107)
-2.325***
-1.561**
0.941
1.155
1.004
2.777**
(0.795)
(0.719)
(1.129)
(0.918)
(0.809)
(1.178)
-2.738***
-1.990***
0.479
2.301**
1.213
2.822**
(0.567)
(0.594)
(1.090)
(0.909)
(0.828)
(1.168)
0.0214***
0.0189***
0.00779***
-0.0218*
-0.0175*
-0.00513
(0.00716)
(0.00588)
(0.00241)
(0.0131)
(0.00919)
(0.00452)
1.816**
0.974
1.771***
-0.303
-4.630***
-0.481
(0.756)
(0.876)
(0.684)
(1.796)
(1.757)
(1.123)
1.221***
1.236***
0.820***
(0.223)
(0.220)
(0.204)
Retail activity Population density Age 0 to 19 years old Family income White population
0.650
-0.337
-0.316
(0.467)
(0.520)
(0.567)
2.323
4.703**
1.912
(1.589)
(2.059)
(1.898)
2.159*
3.420***
5.065***
(1.200)
(1.246)
(1.322)
-0.571
-0.961*
-2.399***
(0.482)
(0.533)
(0.727)
Neighborhood dummies
No
No
No
Yes
Yes
Yes
Observations
57
54
50
57
54
50
Pseudo-R2
0.150
0.0829
0.0552
0.342
0.260
0.210
Log-likelihood
-240.8
-246.1
-234.8
-186.4
-198.6
-196.3
Standard errors in parentheses * p<0.10, ** p<0.05, *** p<0.01
41
Table 4: Discrete Choice Model for Location Decision - Conditional Logit With Distance-Band Dummies Drug Stores - Estimates for Blocks of 30 Years Base specification
Closest own store is within 1.0 - 2.0 miles Closest own store is within 2.0 - 3.0 miles Closest own store is within 3.0 - 4.0 miles Closest own store is within 4.0 - 5.0 miles Closest own store is beyond 5.0 miles Competitors' stores Modified Herfindahl index (3 miles)
δ1, 1-2 δ1, 2-3 δ1, 3-4 δ1, 4-5 δ1, 5+ δ2 δ3
1956-1985
1968-1997
-0.122 (0.600)
Base specification plus controls 1981-2010
1956-1985
1968-1997
1981-2010
-0.975*
-0.526
0.0764
-0.587
-0.213
(0.583)
(0.650)
(0.573)
(0.586)
(0.655)
0.397
-0.344
0.244
0.820
0.730
1.393**
(0.583)
(0.514)
(0.587)
(0.561)
(0.537)
(0.613)
0.0902
-0.645
-0.365
1.094*
1.389**
1.732**
(0.626)
(0.549)
(0.651)
(0.619)
(0.588)
(0.705)
-0.809
-1.912**
-1.233
0.715
0.499
1.368
(0.781)
(0.829)
(0.874)
(0.793)
(0.882)
(0.941)
-1.920***
-2.114***
-1.037
0.177
0.635
2.262***
(0.701)
(0.590)
(0.641)
(0.745)
(0.683)
(0.805)
0.0310***
0.00412***
0.00392***
-0.00944
-0.000476
-0.0000776
(0.00357)
(0.00110)
(0.00115)
(0.00740)
(0.00260)
(0.00230)
2.887***
1.654**
1.458**
-0.246
-1.718
-1.013
(0.932)
(0.792)
(0.731)
(2.425)
(2.447)
(1.997)
0.983***
0.787***
0.692***
(0.172)
(0.130)
(0.139)
-0.404
-0.295
-0.0732
(0.349)
(0.411)
(0.514)
Retail activity Population density Age 0 to 19 years old Family income White population
-3.115***
-0.487
1.757
(1.115)
(0.778)
(1.535)
1.686*
0.0245
0.432
(0.987)
(1.006)
(1.138)
-0.945**
-0.424
0.0814
(0.394)
(0.397)
(0.605)
Neighborhood dummies
No
No
No
Yes
Yes
Yes
Observations
83
81
63
83
81
63
Pseudo-R2
0.138
0.0461
0.0350
0.293
0.241
0.248
Log-likelihood
-355.5
-384.0
-302.2
-291.5
-305.5
-235.6
Standard errors in parentheses * p<0.10, ** p<0.05, *** p<0.01
42
Table 5: Discrete Choice Model for Location Decision - Conditional Logit With Quadratic Polynomial Supermarkets and Drug Stores - Estimates for Blocks of 30 Years
Distance from submarket to closest own store Quadratic distance from submarket to closest own store Competitors' stores Modified Herfindahl index (3 miles) Retail activity Population density Age 0 to 19 years old Family income White population Neighborhood dummies Probability-maximizing distance Observations
δ1,1 δ1,2 δ2 δ3
Supermarkets
Drug Stores
(base specification plus controls)
(base specification plus controls)
1956-1985
1968-1997
1981-2010
1956-1985
1968-1997
1981-2010
1.096***
1.156***
0.926***
1.085***
1.038***
1.007***
(0.325)
(0.356)
(0.312)
(0.362)
(0.317)
(0.308)
-0.105***
-0.127***
-0.0890**
-0.131***
-0.106***
-0.0820**
(0.0406)
(0.0463)
(0.0360)
(0.0474)
(0.0360)
(0.0325)
-0.0245*
-0.0184**
-0.00485
-0.0104
-0.000536
-0.000638
(0.0130)
(0.00933)
(0.00466)
(0.00738)
(0.00264)
(0.00238)
-0.690
-4.830***
-0.332
-0.140
-2.199
-1.389
(1.809)
(1.783)
(1.148)
(2.390)
(2.488)
(2.088)
1.238***
1.249***
0.780***
1.003***
0.776***
0.702***
(0.221)
(0.221)
(0.207)
(0.172)
(0.129)
(0.141)
0.803*
-0.178
-0.119
-0.360
-0.199
-0.0409
(0.470)
(0.534)
(0.566)
(0.355)
(0.418)
(0.514)
2.292
4.770**
1.917
-3.294***
-0.421
2.021
(1.623)
(2.079)
(1.852)
(1.140)
(0.746)
(1.537)
2.104*
3.581***
4.729***
1.638*
-0.145
0.506
(1.200)
(1.260)
(1.314)
(0.992)
(1.011)
(1.128)
-0.454
-0.946*
-2.141***
-0.945**
-0.369
0.0757
(0.487)
(0.529)
(0.713)
(0.397)
(0.393)
(0.600)
Yes
Yes
Yes
Yes
Yes
Yes
5.23 miles
4.56 miles
5.20 miles
4.14 miles
4.90 miles
6.14 miles
57
54
50
83
81
63
Pseudo-R2
0.350
0.272
0.198
0.301
0.243
0.242
Log-likelihood
-184.1
-195.4
-199.2
-288.2
-304.7
-237.3
Standard errors in parentheses * p<0.10, ** p<0.05, *** p<0.01
43
Table 6: Average of Proportion of Population Within an Ellipse Who Was Spatially Monopolized by Clustering Chain Stores
Average of proportion of population within an ellipse for which "n " closest stores contained the same-firm clustered stores Retail
Sample
Distance to closest sister
Number of same-
industry
period
store at time of entry
firm store pairs
Supermarkets
1956-1980
Supermarkets
Drug stores
Drug stores
1981-2010
1956-1980
1981-2010
n =2
n =3
n =4
n =5
n =6
5 miles or less
45
9.1%
21.0%
34.0%
44.3%
52.1%
4 miles or less
41
9.9%
22.2%
35.8%
46.3%
54.8%
3 miles or less
35
11.6%
26.0%
41.6%
53.5%
62.2%
2 miles or less
22
18.1%
40.1%
63.8%
80.6%
89.1%
5 miles or less
34
14.8%
30.1%
45.3%
61.8%
71.6%
4 miles or less
29
13.1%
27.2%
42.6%
60.0%
70.6%
3 miles or less
22
17.2%
34.7%
53.9%
74.9%
86.4%
2 miles or less
9
33.2%
56.8%
77.5%
93.8%
99.7%
5 miles or less
33
1.4%
14.0%
21.6%
32.0%
38.7%
4 miles or less
30
1.6%
15.4%
23.7%
35.2%
42.3%
3 miles or less
24
2.0%
19.2%
29.7%
44.0%
52.9%
2 miles or less
12
3.0%
31.9%
44.6%
67.6%
76.6%
5 miles or less
29
6.9%
16.9%
33.6%
48.1%
58.6%
4 miles or less
25
8.0%
19.6%
38.8%
55.3%
66.3%
3 miles or less
20
10.1%
24.5%
48.5%
66.4%
76.0%
2 miles or less
10
20.0%
44.7%
74.9%
86.0%
91.1%
44
Figure 1: Location Choice of ܣଶ 0.8
ܣଵ
Profit to A (clusters) Profit to A (disperses)
0.6 0.4 0.2
ܦ
ܤ 0.0 0.0
0.2
0.4
0.6
0.8
1.0
Economies of density (α)
Figure 2: Location Choice of ܧ ܣଵ
ܣଶ
0.06
Profit to E (next to A) Profit to E (away from A)
0.04
0.02
ܦ
ܤ
0.00 0.0
0.2
0.4
0.6
0.8
1.0
Economies of density (α)
Figure 3: Location Choice of ܧଵ and ܧଶ ܣଵ
ܣଶ
0.5
Profit to E (next to A) Profit to E (away from A)
0.4 0.3 0.2 0.1
ܦ
ܤ
0.0 0.0
0.2
0.4
0.6
Economies of density (α)
45
0.8
1.0
Figure 4a: Location Choice of Aଶ Under Threat of Entry (α = 0) No Entry Threat
A clusters its stores (A1,A2,B,D)
0 0
.0378
A clusters & E enters near A (A1,A2,E,B,D)
Entry Threat
.0555
.0400
A does nothing & E enters (A1,E,B,D)
Entry blocked for A2 (A1,B,D)
.0544
A deters E by dispersing (A1,B,A2,D)
.0625
A deters E by clustering (A1,A2,B,D)
Entry blocked for A2 and E (A1,B,D)
K
Figure 4b: Location Choice of Aଶ Under Threat of Entry (α = 0.1) No Entry Threat
0
A clusters its stores (A1,A2,B,D)
0
.0469
A clusters & E enters near A (A1,A2,E,B,D)
Entry Threat
.0825
Entry blocked for A2 (A1,B, D)
.0825
A deters E by clustering (A1,A2,B,D)
Entry blocked for A2 and E (A1,B,D)
K
Figure 4c: Location Choice of Aଶ Under Threat of Entry (α = 0.3) No Entry Threat
0
A clusters its stores (A1,A2,B,D)
0
Entry Threat
.1593
.0367
A clusters & E enters away from A (A1,A2,B,E,D)
Entry blocked for A2 (A1,B, D)
.1593
A deters E by clustering (A1,A2,B,D)
46
Entry blocked for A2 and E (A1,B,D)
K
Figure 5: Size (ft2) of Supermarket and Fringe Food Retail Stores at Year of Opening 70,000
Supermarkets (non-locally owned)
Supermarkets (locally owned)
Fringe grocery stores
60,000
Store size (square feet)
50,000
40,000
30,000
20,000
10,000
0 1955
1960
1965
1970
1975
1980
1985
1990
1995
2000
2005
2010
Year of store opening Notes: 1. Stores that switched owners due to mergers, acquisitions, or spin-offs are displayed only once, at the time of their original opening; 2. Supermarkets inside discount retailers, such as Wal-Mart, were excluded; 3. Number of observations: 103 non-locally owned supermarket stores, 30 locally owned supermarket stores, and 1130 fringe grocery stores.
47
Figure 6: Market Structure of Supermarket Industry In Greensboro, NC - 1955-2010 50
45
Food Rite
Other supermarkets
40
Bi-Rite (post-cooperative)
Bi-Lo
Number of stores (stacked)
35
Bi-Rite (local cooperative)
30
Bestway Lowes
25
Food Lion A&P
20
15
Colonial
Harris Teeter Big Bear
10
Kroger 5
Winn-Dixie 0 1955
1960
1965
1970
1975
1980
1985
1990
1995
2000
2005
2010
Year Notes: 1. The category "Other supermarkets" (shaded in light gray) contains all non-locally owned supermarket firms with at most four stores in Greensoboro at any given year, as well as supermarkets located inside large discount retailers such as Wal-Mart; 2. Supermarkets shaded in dark grey represent Greensboro locally owned supermarket firms.
48
Figure 7: Sequential Location of Stores of Winn-Dixie Supermarket in Greensboro, NC
#12 (1990-2003)
#14 (2001-2005)
#4 (1961-2005)
#8 (1972-2001) #6 (1964-1980)
#1 (1956-1970)
#5 (1962-1999)
#3 (1958-1970) #2 (1957-1974) #7 (1972-1976)
#10 (1975-1991)
#13 (2000-2003)
#9 (1973-1989)
2 miles
#11 (1980-2005)
5 miles Note: Solid gray line represents Greensboro 2000 urbanized area, and dotted gray line represents Greensboro 1950 city limits, both based on U.S. Census boundary files.
49
Figure 8: Size (ft2) of Drug Stores at Year of Opening 16,000
14,000
Store size (square feet)
12,000
10,000
8,000
6,000
4,000
2,000
0 1955
1960
1965
1970
1975
1980
1985
1990
1995
2000
2005
Year of store opening Notes: 1. Stores that switched owners due to mergers, acquisitions, or spin-offs are displayed only once, at the time of their original opening; 2. Drug stores inside supermarkets or discount retailers, such as Wal-Mart, were excluded; 3. Number of observations: 145.
50
2010
Figure 9: Market Structure of Drug Store Industry In Greensboro, NC - 1955-2009
70
Number of stores (stacked)
60
Drug stores in supermarkets or discount retailers
50
40
Other drug stores (firms with at most 4 stores in any given year) Walgreens
30
Revco CVS
20
Kerr
Edmonds
Rite Aid
10
Eckerd (JCPenney)
Franklin Eckerd
Eckerd (Jean Coutu)
Rite Aid
0 1955
1960
1965
1970
1975
1980
1985
1990
1995
2000
2005
Year Notes: 1. The category "Other drug stores" (shaded in light gray) contains all drug store firms with at most four stores in Greensoboro at any given year; 2. Drug stores shaded in dark grey represent drug stores inside supermarkets or discount retailers such as Wal-Mart.
51
Figure 10: Sequential Location of Stores of Revco Pharmacy in Greensboro, NC
#11 (1990-1997)
#12 (1990-1997) #13 (1992-1997) #5 (1979-1997) #7 (1980-1997) #6 (1980-1997)
#14 (1995-1996)
#4 (1975-1993)
#9 (1986-1997) #3 (1973-1997)
#16 (1996-1997)
#2 (1973-1982)
#10 (1987-1997) #17 (1997-1997) #15 (1995-1997)
#1 (1970-1990)
#8 (1983-1997)
2 miles
#18 (1997-1997)
5 miles Note: Solid gray line represents Greensboro 2000 urbanized area, and dotted gray line represents Greensboro 1950 city limits, both based on U.S. Census boundary files.
52
Figure 11: Census Tracts for Guilford County, NC Choropleth Maps for Quintiles of Population Density
1950 (45 Tracts)
1960 (64 Tracts)
1970 (78 Tracts)
1980 (87 Tracts)
1990 (94 Tracts)
2000 (98 Tracts)
53
Figure 12: Construction of Locations Around Center of Greensboro, NC
1.5 miles
Note: Solid gray line represents Greensboro 2000 urbanized area, and dotted gray line represents Greensboro 1950 city limits, both based on US Census boundary files.
54
