Spatial Competition in the French Supermarket Industry∗ St´ephane Turolla† June 2015

Abstract This paper challenges the conventional wisdom of a competitive grocery retail sector in France. To measure the intensity of competition in this sector, I develop a structural model of spatial competition that accounts for (i) market geography on consumers’ preferences, and (ii) differences in their shopping list. The demand estimates are used to recover stores’ price-cost margin under alternative pricing strategies. I select the best pricing model by applying non-nested tests and show that retailers noticeably distort their offer in highly concentrated markets. Retail competition in the French supermarket industry appears to be highly localized in that margins are very sensitive to the presence of a competitor within a few kilometers. Finally, I perform counterfactual experiments to quantify the expected gain of an additional store on consumer welfare and retail prices. Keywords: Spatial competition, Structural model, Discrete choice model, Differentiated products, Supermarket industry. JEL Classification: C35, L13, L81.

∗

I am very grateful to Marie-Laure Allain, Kristian Behrens, Aur´elie Bonein, C´eline Bonnet, St´ephane Caprice, Alain Carpentier, Claire Chambolle, Carl Gaign´e, Fran¸cois Gardes, Marc Ivaldi, Jean-Louis Monino, Vincent R´equillart and Jacques-Fran¸cois Thisse, as well as two anonymous referees, for very helpful discussions and comments. This paper has also greatly benefited from suggestions of participants at the JMA Conference, the ADRES Job Market Meeting, the ESEM Conference, the J2E Workshop, the INRA-IDEI Workshop, and seminar participants at CORE, INRA ALISS, INRA SMART, University of Rennes, and TSE. The data used in this paper were provided by the Chamber of commerce of Montpellier. Financial support from the Chamber of commerce of Montpellier is gratefully acknowledged. Part of this paper was written while I was visiting the University of California at Berkeley, whose hospitality is gratefully acknowledged. The usual disclaimer applies. † Address: INRA UMR 1302 SMART, 4 All´ee Adolphe Bobierre, CS 61103, F-35011 Rennes Cedex (France). Email: [email protected]. Phone: +33(0)2 23 48 54 00. Fax: +33(0)2 23 48 53 80.

R´ esum´ e : Cet article s’interroge sur l’id´ee habituellement convenue d’un secteur de la distribution alimentaire concurrentiel en France. Dans cette optique, je propose un mod`ele structurel de concurrence spatiale qui tient compte (i) des caract´eristiques spatiales du march´e dans l’´etablissement des pr´ef´erences des consommateurs et (ii) de la diff´erence entre consommateurs dans leurs listes d’achats. Les estimations des param`etres de demande sont utilis´ees afin d’inf´erer les marges brutes des magasins sous diff´erentes strat´egies de tarification. Le meilleur mod`ele de tarification est ensuite s´electionn´e en utilisant une proc´edure de tests non-emboit´es. Le mode de tarification retenu montre que les distributeurs distordent sensiblement leur offre commerciale dans les march´es fortement concentr´es. La concurrence dans le secteur de la distribution fran¸caise apparait fortement localis´ee en raison d’une variabilit´e importante des marges des magasins en fonction de la pr´esence ou non de concurrents dans un rayon de quelques kilom`etres. Enfin, l’article pr´esente des exercices contrefactuels visant a` quantifier le gain attendu de l’entr´ee d’un nouveau magasin sur le bien-ˆetre des consommateurs et les prix de d´etail. Mots-Cl´ es : Concurrence spatiale, Mod`ele structurel, Mod`ele a` choix discret, Diff´erenciation des produits, Grande distribution.

1

Introduction

In recent years, the grocery retail sector has undergone major transformations in Europe (e.g., increase of concentration operations, passing of zoning regulations, consolidation of buyer groups), substantially increasing the power of large retailers. Consumers, manufacturers’ associations, and policy-makers have expressed serious concerns over the practices of retailers. In substance, it is alleged that retailers take advantage of their dominant position to mitigate price competition and to distort their offers. As a result, the grocery retail sector has been put under scrutiny by numerous competition authorities in Europe. For instance, the UK Competition Commission initiated a thorough investigation in 2006 over a broad scope of issues related to the dominance of large retailers, which also concerned both horizontal and vertical aspects of retailers’ activities (Competition Commission, 2008). Recently, Italian and Spanish competition authorities also launched inquiries aimed at examining the anti-competitive effects of the bargaining power of retailers on both sides of the market.1 This issue is particularly topical in France because of an increase in food product prices that is much larger than that observed throughout the rest of Europe.2 Similar to other European markets, the French grocery retail market has experienced a growing concentration and significant structural changes. However, France underwent important regulatory changes in the mid-1990s (i.e., the so-called Galland and Raffarin Acts) which reinforced the dominance of large retailers, generating important detrimental effects on competition (see Allain et al., 2008, for a review).3 Urged on by public opinion to find an answer to the price increases, the French government has almost exclusively focused on restoring the balance of power between manufacturers and retailers. Several amendments to the resale-below-cost law have been implemented during the past few years, with the aim of passing rebates negotiated by retailers on to consumers. Unfortunately, these modifications have failed to restore a fierce level of competition, as prices fell by just 1% two years after the last reform (INSEE, HICP 2010). The goal of this paper is to propose another explanation for the low level of competition encountered in France by measuring the competitive intensity observed 1

Sweden, Denmark, Finland and Belgium are other examples of countries with national competition authorities that have conducted inquiries in the last four years to evaluate the extent to which retailers’ power affects competition in the food distribution sector (European Competition Network, 2011). 2 According to Eurostat, food product prices increased 5.9% faster than the consumer price index in France from 1996 to 2009. 3 The French Parliament passed the Galland Act and the Raffarin Act in the summer of 1996. On the one hand, the Galland Act was principally dedicated to preventing retailers from engaging in below-cost pricing by clearly defining the below-cost price threshold. Technically, this resulted in excluding the conditional and deferred rebates (i.e., the so-called “hidden margins”) from the invoice such that they could not be deducted from the final price. On the other hand, the Raffarin Act was enacted to reinforce the control over market entry. To protect small independent stores from the growing pressure of mass distribution, especially the entry of German mass discounters (i.e., Aldi and Lidl), the legislator toughened the planning system by extending the administrative authorization (i.e., a prerequisite for granting a building permit) to stores with selling areas over 300 m2 (1,500 m2 under the previous regulation).

1

in local markets and by analyzing the determinants of retailers’ market power. To that end, I develop and estimate a structural model of spatial competition to recover stores’ operating margin. I consider various price-setting schemes and select the one that best fits the data, which allows me to examine a broad range of explanations at the origin of retailers’ market power (e.g., anti-competitive behaviors, chain-stores’ pricing or an insufficient number of competitors). The empirical application relies on a unique database that corresponds to a cross-sectional survey of 1,654 households living in a metropolitan area in southern France during the year 2000. The survey conveys detailed information on the stores visited at the product category level, allowing for tracking of the shopping baskets that households purchased at their primary shopping destination. Questioning the level of downstream competition is a recent occurrence in France. For a long time, the concentrated structure of the French grocery retail sector did not raise concerns over retail prices.4 In its first notice on the sector, the French Competition Authority (henceforth CA) argued, for instance, that: “Concentration in the retail food industry has little effect on the downstream market because competition is fierce among retail chains” (Competition Authority 1997, p.28, personal translation). However, during the past few years, consumers’ associations have conducted price reporting that has highlighted retailers’ abilities to raise their prices more intensively in highly concentrated markets.5 This empirical evidence calls for a more serious consideration of the level of competition in local markets as an explanation for the inflationary trend observed, and reinforces the need to better understand how retailers may distort their offers locally. This paper is closely related to the empirical literature devoted to the estimation of structural models of demand. Recently, several authors have extended the methodology proposed by Berry (1994) and Berry et al. (1995) to analyze the strategic behavior of firms that are spatially differentiated in retail markets (see, e.g., Thomadsen, 2005; Davis, 2006; McManus, 2007; Chiou, 2009; Manuszak, 2010).6 According to the nature of products and the frequency of store visits, the spatial dimension may dramatically alter market outcomes by influencing consumer preferences and firm strategies, such as pricing. This occurrence is particularly true in the food distribution sector where consumers travel a few kilometers - traditionally every week - to reach their primary shopping destination. As a result, a number of papers have studied various decisions of large retailers (e.g., entry, pricing or selling space expansion) in the light of market geography (see Cleeren et al., 2010; Orth, 2012; Ellickson and Misra, 2008; Ellickson, 2006; to cite a few). 4

The aggregate market share of the five largest retailers increased from 61.8% to 75.6% between 1995 and 2009, placing France second among European countries. 5 A survey conducted by Nielsen showed that food retail prices might vary, on average, up to 9% among the 126 largest French cities (see Libre-Service Actualit´es (LSA), March 6, 2008). In the same vein, the consumer association UFC-Que Choisir revealed that prices might vary up to 20% between two hypermarkets of the same retailer depending on the competition that they face (see UFC-Que Choisir, December 26, 2007). 6 Pinkse et al. (2002) propose a different modeling approach that addresses the magnitude of spatial price competition through the estimates of price reaction functions. This original approach differs from the traditional framework of a random utility model used here.

2

Despite the central position of the food distribution sector in OECD economies, and its impact on household purchasing power, very few studies have evaluated the market power of retailers.7 To my knowledge, Smith (2004) and Dubois and J´odar-Rosell (2010) are the rare exceptions. Both studies developed and estimated a structural model of spatial competition by accounting for the spatial distribution of stores and the disutility of travel by consumers. Their market power estimate reveals an approximate level of profitability between 8 to 14% (depending on chain name and store format). In addition, these two studies were interested in quantifying the effect on profitability of two elements of the business strategy: chain store pricing and differentiation through private labels, respectively. The present paper addresses two other issues not previously accounted for that may lead to a bias in the estimate of retailers’ local monopoly power. One issue is the multiproduct dimension of demand. While obvious, this key feature of food distribution activity is usually circumvented by assuming that consumers buy a representative shopping basket, which implies that its composition does not influence consumer store choice. However, the empirical marketing literature has long demonstrated that consumers’ price responsiveness might differ significantly depending on the size of the shopping basket (Bell and Lattin, 1998), which therefore also affects the estimate of a store’s market power. The other central issue that this paper investigates refers to the so-called “price flexing” strategy. This pricing strategy consists of charging different prices for the same product across stores within a chain. This practice is particularly prevalent in France, as reported by consumers’ associations, and allows retailers to extract more consumer surplus by raising their prices more intensively in highly concentrated markets. Until now, studies have solely considered that retail chains set uniform prices within large regional markets, implying identical profitability levels among retail chain stores (at constant cost). As a result, the price-raising effect of local market concentration is absent from these analyses and may lead to an understatement of retailers’ market power. To adequately measure the local monopoly power enjoyed by large grocery stores, I therefore model consumer store choice as a two-step process: (i) a consumer first decides whether to purchase different product categories in large grocery stores, and (ii) then selects the store to visit conditional on its shopping basket. The probabilities of purchase incidence derived from the first stage are computed using bivariate probit models. Subsequently, these probabilities are plugged into a flexible discrete choice model of demand among spatially differentiated grocery stores that accounts for consumers’ preferences over store characteristics and geographic proximity. Using the estimated demand parameters, I recover stores’ price-cost margin under different pricing scenarios (including a price-flexing strategy), and I determine the preferred one by applying a non-nested testing procedure. The results show that the substantial market power enjoyed by retailers in some local areas does not arise from anti-competitive practices but mainly results from a high level of concentration. I find that a significant level of heterogeneity exists 7

In contrast, several papers have conducted appraisals of retailers’ market power for a single product in the context of vertically related markets (see, e.g., Villas-Boas, 2007, or Cohen and Cotterill, 2011).

3

among stores’ price-cost margin both for stores under competing retail chains and, more surprisingly, for stores under the same banner. By disentangling the sources of stores’ market power, I show that the stores’ margin differs sharply according to the concentration of their surrounding environment. For instance, I estimate that a store’s margin diminishes, on average, by 2.12 percentage points from the entry of a new rival within a distance band of 1 km to 5 km. Moreover, the competitive effect of a rival store vanishes beyond 10 km. Competition among large grocery stores appears to be highly localized. Finally, the estimated parameters are used to perform counterfactual experiments to quantify the effects of an additional store on retail prices and consumer welfare. The results reveal that even the entry of a single hypermarket enhances consumer welfare through an increase in price competition and a reduction in distance traveled. These findings have important implications for competition policy, and deepen our understanding of the competition forces prevailing in the grocery retail sector. The results speak in favor of a relaxation of entry regulations to promote entry of new stores with the aim of enhancing competition locally. Because many European countries have restrictive entry regulations similar to those of France, the results are interesting in a broader context. The remainder of the paper is organized as follows. First, I briefly depict the market structure of the French supermarket industry (Section 2). Section 3 provides an overview of the data used for the empirical analyses. Section 4 describes the demand model and the pricing equations that allow me to back out the stores’ margin. Section 5 presents the estimates of the demand parameters, displays the stores’ margin according to the preferred pricing model and reports the results of the robustness tests. I discuss the impact of some counterfactual policy simulations on retail prices and consumer welfare in Section 6. Section 7 concludes.

2

The French Supermarket Industry

In 2009, the French grocery retail sector had sales revenues of e181 billion and represented approximately 580,000 jobs. Since the 1960s, the boom in large grocery stores has completely reshaped the retail landscape and consumer habits in many OECD countries. Over the years, the grocery retail sector has become the preferred distribution channel of the French and today accounts for 70% of food sales and 20% of non-food items, although representing barely 3% of commercial equipment (i.e., 21,000 retail outlets). According to the usual categorization, French retailers sell their products essentially through four store formats: hypermarkets, supermarkets, convenience stores and hard discount stores.8 For historical reasons, hypermarkets 8

Professionals and the national institute of statistics (INSEE) identify four store formats according to selling area and product-mix offered. Hypermarkets are large grocery stores with a selling area over 2,500 m2 whose sales arise from one-third of foodstuffs (i.e., this format is a combination of a supermarket and a department store). They are generally located out of town and operate as an anchor store for other retail facilities that are located closer to consumers. Supermarkets have selling areas that range from 400 to 2,500 m2 . They differentiate themselves from hypermarkets by offering a lower assortment of products and by being located in city centers or in the suburbs of large cities. Convenience stores are proximity outlets that almost exclusively sell grocery items. These stores usually operate with a selling area below 400 m2 and are mostly integrated into the

4

are very active in France and account for one-third of the total food sales (see Carluer-Lossouarn, 2007, for a historical analysis of the French grocery retail sector). The largest retail groups operate under two organizational forms that differ principally by the integration of their wholesale activities and their capital structures. Integrated groups, such as Carrefour, Auchan or Casino, operate either directly (i.e., through a company-owned store) or through affiliated entities (i.e., by franchising). In contrast, a network of cooperative groups, such as Leclerc, Intermarch´e and Syst`eme U, is composed of independent shopkeepers who source from their central purchasing unit, which is managed by the cooperative’s member shareholders. Typically, decisions concerning the retail offer are made by the head of the group (e.g., advertising campaigns, terms and conditions of negotiations with suppliers, product listings, and private label assortments), but most of the retail groups adopt a local pricing policy where stores are the final decision-makers.9 One of the most striking features of the success of large grocery stores in European countries is the low number of retail groups that compete together. The French retail market is not different as it is dominated by six retail groups that, taken together, held 84% of the market share in 2009. The six retail groups include Carrefour (24%), Leclerc (17%), Intermarch´e (13%), Auchan (11%), Casino (10%) and Syst`eme U (9%). If this concentration level does not raise anti-competitive concerns at the national level, the situation is mixed at the local level. To highlight this important pattern, I compute various statistics informing about the level of concentration observed in the 500 largest French cities (see details in Appendix A). The results show that the average level of concentration in local markets is almost twice that observed nationally. Further, almost two-thirds of local markets are highly concentrated according to a standard interpretation (i.e., HHI>2000). This finding questions the competitiveness of these markets and reinforces the importance of analyzing this sector at a fine-grained level.10

3 3.1

Data Presentation

This study uses an original database that surveys the store choices of households living in a metropolitan area in southern France for several food and non-food product categories. The area of study is the French administrative aire urbaine of Montpellier (henceforth, Montpellier AU), which covers a total number of 459,916 people. largest retail groups. Finally, hard discount stores correspond to small supermarkets that carry a limited assortment of low- and medium-range foodstuffs. These stores operate under a dedicated banner and are integrated most of the time into national retail groups. 9 For more information on how stores fix prices, see Lin´eaires, July 5, 2006. 10 It is worth noting that the concerns raised by the level of concentration observed locally cannot be dissipated by pro-competitive industry dynamics. Since the passing of the Raffarin Act, the structure of local markets has hardly evolved because of important barriers to entry. In contrast to other retail sectors, few stores have opened. On average, the hypermarket and supermarket formats have respectively had entry rates of approximately 0.3% and 0.7% per year over the last fifteen years.

5

The survey was conducted jointly by the chamber of commerce of Montpellier and the department of economics at the University Montpellier I during the year 2000. The data were collected at the household level, and the store choices of the households were recorded on a yearly basis. A total of 1,654 households were asked about their shopping habits, regarding 49 food and non-food product categories. For each product category, the household filled in all the stores visited during the year, regardless of the distribution channel visited (e.g., large grocery stores, specialized stores, or farmers’ markets). They were then asked to rank the stores according to their frequency of visit. The survey gives an accurate picture of households’ shopping habits for a large group of product categories. In addition, it gathers information on household characteristics, such as the age and the socio-economic status of the household head, the number of persons per household and their location of residence. A detailed description of the data is provided in Appendix B. For the purpose of this study, I focus on the information collected on food products, and restrict the analysis to the 8 most frequently purchased categories (i.e., fruits and vegetables, meat, cooked meat, cheese, other dairy products, grocery items, alcoholic drinks, and soft drinks). Based on the information reported at the product category level, I determine the primary shopping destination of the households and their top-up stores over these 8 food product categories (see Appendix B for more details). The primary shopping destination thus corresponds to the store at which the household purchases the largest portion of its food products. I supplement these data with information on store characteristics, which I obtained from the Atlas de la distribution, a national survey on French grocery stores. I collect information about store characteristics, such as the chain name, the store location, the store size and the number of employees. To determine the distances traveled to a store by a household, I geocoded the addresses of the stores, the cities that belong to Montpellier AU, and the IRIS of Montpellier city (French geographical unit similar to a block-group) in a geographical information system. Assuming that a household resides at the centroid of its geographical unit, I compute the Euclidean distances that a household may travel to visit each store of its choice set.11 In what follows, I assume that the stores identified as a primary shopping destination constitute a separate market from stores visited for top-up shopping (essentially some convenience stores and a few small supermarkets). This assumption is consistent with previous decisions made by the French CA (see, for instance, Decision No. 10-D-08), where top-up stores are deemed unattractive for primary shopping and thus are viewed as competing for a different demand. I thus remove from the data the stores that are attended solely for top-up shopping. As a result, the choice set of a household may include one of the 62 large grocery stores identified as a primary store (hypermarkets, supermarkets, hard discount stores and large conve11

By specifying a single-address model, I assume that a households’ residence corresponds to the only departure point of the shopping trip, which greatly simplifies the model specification. A more ambitious approach would be grounded in a multi -address model, which would make it possible to account for multiple departure points for shopping trips, such as households’ residence, workplace, or the place where their children attend school (see Houde, 2012). This would affect the disutility of traveling but not the primary shopping destination, as I explain in the robustness section.

6

nience stores with selling areas over 400 m2 ) and one outside option that brings together alternative distribution channels (e.g., specialized stores, farmers’ markets or home made). Additionally, I follow the methodology applied by the EC in previous investigations and restrict the composition of a household’s choice set to stores located within 20 km of its residence.12 Note that one limitation of this definition is that stores visited for primary shopping but located far away from the household’s residence are excluded de facto. For instance, this may arise for households living in a small rural town but who work in Montpellier city.13 Overall, I find 43 households that meet these criteria. Hopefully, this situation concerns only a small part of the sample (i.e., approximately 2.60%) that has to be excluded from the study. After eliminating these households, the database used to conduct the analysis corresponds to a cross-section survey of 1,611 households.

3.2

Computing a Store-Category Price Index

Unlike homescan or retail scanner data, this survey does not record information on the products purchased. This raises a challenging issue to correctly estimate the household demand for their primary shopping destination. In particular, the absence of data on prices is problematic because price is usually cited by households as one of the main drivers of their decisions. To account for the pecuniary incentives that underpin the store choice decisions, I propose to construct a hedonic store price index that incorporates a price measure for each product category. I detail in this subsection the procedure applied to construct the store-category price indices. To collect information on product prices, I ran a price survey on a subset of stores for a sample of items. I collected the prices of 91 national brand and first-price products from a total of 27 selected stores such that the sample is representative of the retail chains, formats and locations of the stores operating in the area of study.14,15 The prices were reported over three days to avoid seasonal variations, especially for the fruits and vegetables category. The fact that the sample is composed from a large part of national brand products is essential to control for product heterogeneity across stores. By comparing the prices of strictly identical products, I thus limit the aggregation bias that might arise when computing a price index from products with different characteristics (e.g., brand, packaging or quality). Let us first consider the case of non-hard discount stores. The price index of a product category c sold in store j is simply computed as the average price of the K 12

Depending on the average driving speed, a 20 km distance as the crow flies may correspond to approximately 20 to 30 minutes. 13 Similarly households living at the borders of Montpellier AU but who patronize a store outside the area are excluded following this definition. 14 First-price products are low-quality private labels sold under a specific brand. They correspond to the lowest price range of products available in stores. They were introduced by retailers to compete directly with hard discounters’ private labels. 15 As an example of national brand and first-price products surveyed, I recorded the price of Puget extra-virgin olive oil, in a glass bottle of 1 liter and the price of a Camembert-type cheese under a first-price brand, respectively.

7

items surveyed, (k = 1, . . . , K): PK pbjc ≡

k=1

K

pkjc

,

(1)

and the store with the largest sales area in Montpellier AU is chosen as base 100 of the price indices (i.e., Carrefour 2 ).16 Consider now the case of hard discount stores. The price survey reveals that hard discount chains differ from other chains by fixing uniform pricing. This result is not really surprising regarding the size of the market considered here because hard discount groups are principally organized through regional distribution centers that control the commercial policy of their stores. By setting uniform prices inside their stores, the regional centers minimize distribution costs and control their price image, a key feature of their success. Using the data from hard discount stores that I have surveyed, I thus compute the category-price indices for each hard discount store according to Eq.(1). However, because hard discount chains principally offer firstprice products, I remove the national brand products from the shopping basket of the base outlet to normalize the price indices based on a similar basket of products. At this stage, I only partially observe the pricing of the stores, as I was able to compute the category-store price indices for 39 of the 62 stores. I thus have to address missing prices for the non-hard discount stores that were not surveyed. Solving this issue is not trivial and influences the results. Because prices are one of the main components of the household utility function, how I fill in the missing prices impacts the demand model estimates and, particularly, the estimation of household price sensitivity. In turn, the pricing game played by the stores is also affected because the stores’ price elasticity is a function of the price coefficient. Although missing prices is a common issue with consumption data (e.g., prices are usually unobserved for non-purchased brands in scanner panel data), there is no consensus in the literature on how to infer these data, and very few studies have investigated this topic (exceptions are Erdem et al., 1998 and Einav et al., 2010). Some authors have used ad hoc methods, such as interpolation methods, whereas others have estimated the missing data from a parametric model (e.g., Chiou, 2009). Because the determinants of retailers’ pricing have been studied a number of times in the literature and are well-known, I follow the second approach and estimate the missing store-category price indices using hedonic price regressions. Following Handbury and Weinstein (2015), I propose a theoretically rigorous store-category price index that controls for product, retailer and store heterogeneity, as well as for purchaser heterogeneity and product availability. Specifically, I regress the log of the price index of a product category sold in a store surveyed on a set of store characteristics capturing all of the following dimensions: retailer heterogeneity is captured by retailer fixed effects, while store heterogeneity is controlled by a measure of the Herfindahl-Hirschman Index (HHI ) and the number of competing rivals within 5 km around each store (# stores ≤ 5 km); purchaser heterogeneity across stores is proxied by the mean household income belonging to a store’s catchment area (market 16

Because of a confidentiality agreement, I disguise the names of the stores.

8

9 Yes 0.8148 216

Yes 0.8521

4.1580*** (0.2843) 4.1945*** (0.2870) 4.4550*** (0.2799) 0.0087** (0.0037) 0.0007 (0.0010) 0.0464* (0.0267) -0.0073 (0.0077)

Meat

Yes 0.9178

Cooked meat 4.3082*** (0.2835) 4.3248*** (0.2845) 4.6340*** (0.2790) 0.0087** (0.0037) 0.0007 (0.0010) 0.0464* (0.0267) -0.0073 (0.0077) Yes 0.9176

Yes 0.9544

SUR Model Cheese Other dairy product 4.2697*** 4.2047*** (0.2833) (0.2832) 4.2893*** 4.2385*** (0.2840) (0.2837) 4.5850*** 4.4869*** (0.2789) (0.2787) 0.0087** 0.0087** (0.0037) (0.0037) 0.0007 0.0007 (0.0010) (0.0010) 0.0464* 0.0464* (0.0267) (0.0267) -0.0073 -0.0073 (0.0077) (0.0077) Yes 0.9110

Grocery item 4.2916*** (0.2831) 4.3219*** (0.2834) 4.5813*** (0.2786) 0.0087** (0.0037) 0.0007 (0.0010) 0.0464* (0.0267) -0.0073 (0.0077) Yes 0.9583

Alcoholic drink -4.2472*** (0.2828) 4.2698*** (0.2826) 4.4459*** (0.2784) 0.0087** (0.0037) 0.0007 (0.0010) 0.0464* (0.0267) -0.0073 (0.0077) Yes 0.8254

Soft drink 4.1904*** (0.2841) 4.1911*** (0.2864) 4.6259*** (0.2797) 0.0087** (0.0037) 0.0007 (0.0010) 0.0464* (0.0267) -0.0073 (0.0077)

Notes: Standard errors in parentheses.*, **, *** indicate significance at the 10%, 5%, 1% level, respectively. All regressions include retail chain fixedeffects. The omitted retail chain for the hypermarkets and the supermarkets is Leclerc. The omitted retail chain for the convenience stores is March´ e U. The Herfindahl-Hirschman Index (HHI ) variable is computed by assuming that the stores are positioned at the center of their respective cities. A store’s catchment area is composed of cities located within a radius of 10 km around the store’s city. The HHI is then computed based on selling surfaces at the retail group level. The variable # stores ≤ 5 km counts, for a given store, the number of rivals within 5 km. The variable market income corresponds to the mean household income calculated over the set of cities that belong to the catchment area of a given store. The variable market population is computed as the sum of inhabitants living in cities that belong to the catchment area of a given store. The specification adopted implies that these 4 variables are equal across the product categories. The Breusch-Pagan LM test for error independence supports resorting to a SUR specification, χ2 (28) = 74.4710 and p-value=0.0000. Source: Author’s calculations.

Retail chain FE R2 Observations

log(market population)

log(market income)

# stores ≤ 5 km

HHI

Convenience store

Supermarket

Variable Hypermarket

Fruits & vegetables 4.3474*** (0.2857) 4.3611*** (0.2911) 4.6471*** (0.2813) 0.0087** (0.0037) 0.0007 (0.0010) 0.0464* (0.0267) -0.0073 (0.0077)

Dependent variable: (log) store-category price index

Table 1: Hedonic regressions of Log store-category price indices

income); and product availability is controlled at the store level by format fixed effects and by population size in the store’s catchment area (market population). Finally, as I suspect that unobservable store attributes that influence prices (e.g., quality, distribution costs and amenities) are correlated across the products of a store, I run a seemingly unrelated regression (SUR) equations model. Table 1 reports the estimates. The sensitivity of the results is discussed in the robustness section, but it is noteworthy that the demand estimates vary marginally around the estimated price indices.

4

A Structural Model of Spatial Competition

In this section, I first specify the formulation of the household demand model, and then I derive the pricing equation under the alternative pricing games that are likely to be played by French retailers.

4.1

The Demand Model

Empirical IO studies on the supermarket industry usually rely on discrete-choice demand models where consumers buy an identical basket of goods (see, e.g., Smith, 2004; Beresteanu et al., 2010; Dubois and J´odar-Rosell, 2010).17 This convenient assumption implies that store choices are not conditioned by the products a consumer intends to buy in a store. In contrast, several empirical studies have shown that consumers react differently to store characteristics, depending on their shopping basket. In particular, the marketing literature has highlighted that the price responsiveness of consumers varies according to whether consumers are large or small basket shoppers (see, e.g., Bell and Lattin, 1998).18 For instance, large basket shoppers tend to visit stores less often but spend more during each trip, which makes them more sensitive to prices. Hence, depending on the size of a consumer’s shopping list, large grocery stores do not necessarily compete with each other over all the categories they offer. This shopping component inevitably distorts the competition among stores and influences their marketing strategy. As shown by Ellickson and Misra (2008), retailers may adopt different pricing strategies according to the type of shoppers (e.g., small/large basket shoppers) present in their vicinity. As a result, assuming that households buy an identical basket of goods inevitably leads to biased estimates. Using the store-category information recorded in the household survey, I am able to relax this strong assumption, and I develop a discrete-choice demand model with heterogeneous shopping baskets across households. Specifically, I propose a two-step model in which a household chooses its primary shopping destination conditional on the cost of the bundle of product categories (i.e., a shopping basket) it is willing to 17

Richards and Hamilton (2006) are one of the few exceptions. They model the consumer sequential choices of which products to buy and where the resulting shopping basket is purchased through a nested CES discrete choice model. 18 Using the terminology of Bell and Lattin (1998), I refer to a large basket shopper to define a shopper who has a relatively high probability of purchasing for any given category.

10

buy in a large grocery store. This cost is proxied by a store price index that differs across households according to the composition of their shopping basket. Given that the choice decision relies on the primary shopping destination over a yearround period, the construction of the store price index is simpler compared with the case of a weekly shopping trip. In the present case, households are assumed to have perfect information on prices due for instance to the repetitive trips they have made during the year or due to advertising. The formulation of a household-specific store price index rests on the idea that consumers choose to visit a store depending on their shopping plans. Before going to shop, consumers make a shopping list of the product categories they need and then decide which stores to visit by accounting for the price of the ex ante shopping basket. The shopping list being unobserved, I predict the composition of the shopping basket of a household at the product category level as in previous research in the marketing literature (see, e.g., Briesch et al., 2009). Note that the composition of the shopping basket abstracts from time considerations, and especially from intertemporal substitution effects. The shopping basket is supposed to be representative of the product categories usually purchased by the household.19 In the first stage of the demand model, I then estimate the probability that a household will purchase a product category in a large grocery store. These purchase probabilities are used to construct the household-specific store price indices such that a household only accounts for product categories belonging to its consideration set. I define the price index of store j for household h, pehj , as the sum of the store-category price indices, pbjc , multiplied by the category purchase probabilities, Pr (Ihc ), and weighted by the share of the mean expenditure in category c, denoted $c , such as: C X Pr (Ihc ) pbjc $c . (2) pehj ≡ c=1

The category purchase probabilities, Pr (Ihc ), are obtained by running Bivariate Probit Models (BPMs) to control for selection issues due to household location.20 The details of the estimations are reported in Appendix C. The store price index provides a measure of the cost of the shopping basket. However, one has to keep in mind that the absence of quantity (or expenditure) data at the household level prevents us from correctly predicting the costs of products purchased. The second part of the demand model is more familiar with respect to the literature on structural models of demand (see, e.g., Berry, 1994; Nevo, 2000). Given the discrete nature of a household’s decision, I follow the standard approach of random utility models and specify a discrete choice model to assess the determinants of a household’s store choice. Households have heterogeneous preferences regarding their 19

Obviously, the assumptions of perfect information on prices and a time-invariant shopping basket must be relaxed if one wishes to analyze store choice decisions at the shopping trip level (see, e.g., Bell et al., 1998). 20 In a previous version of this paper (see Turolla, 2012), I proposed a modeling framework that addresses the issue of unobserved cross-category effects in the composition of the shopping basket. The results emphasize the importance of common (unobserved) factors across categories (e.g., quality, shopping costs) to explain the distribution channel chosen by a consumer.

11

primary shopping destination due to differences in their shopping basket, location, and tastes over store characteristics. To account for this flexibility, I define a random coefficients logit model (or mixed logit model), which allows me to estimate more realistic substitution patterns than a simple “logit-type” model. Concretely, the mixed logit model yields flexible estimates of own- and cross-price elasticities by avoiding the problematic independence of irrelevant alternatives (IIA) assumption involved in discrete choice models where heterogeneity is solely captured through the idiosyncratic term.21 Let us assume that a household h chooses its primary shopping destination (conditional on a shopping basket) based on the highest utility rule, which is derived from visiting one of the stores j (j = 1, . . . , J) that are included in its choice set Jh , or by choosing an outside option j = 0. Recall that a household’s choice set contains all stores located within 20 km of its residence. Thus, according to the typical notation for discrete choice models of demand, the indirect (latent) utility that a household h that resides in location Lh derives from visiting store j ∈ Jh located in Lj is: Uhj = α0 pehj +

6 X

αg pehj dage hg + λh DIST (Lh , Lj ) + γh SU RFj

g=2

+ φCASHj +

XX s

ϕsq κq dfs ormat + ξr + εhj

(3)

q

where pehj is the household-specific shopping basket cost incurred in store j, dage hg is a dummy variable equal to one if the head of household h belongs to age group g, DIST (Lh , Lj ) is the Euclidean distance between a household’s residence (Lh ) and store location (Lj ), SU RFj is the store size of store j and CASHj is the number of cash registers per one hundred square meters in store j. I also include a dummy variable dfs ormat for one of the S store formats (s=hypermarket, supermarket, hard discount store, large convenience store) and interact this dummy variable with Q variables that represent a set of household and store characteristics denoted by κ. Finally, ξr is an index of unobserved (to the econometrician) retail chain attributes and εhj is the idiosyncratic term that is assumed to be i.i.d. according to a type I extreme value distribution. Price sensitivity varies among the six age classes of the household head (with the youngest taken as the reference). Thus, the coefficient α0 corresponds to the marginal utility of price of a “representative” household, and a deviation from this mean depends on the coefficient of the interaction of the price with the age class of the household head. Similar to the price coefficient, the distance and store size coefficients (λh , γh ) are supposed to vary across households. Rather than using interaction terms, I specify a random coefficient on distance and store size, which is more appropriate as a way of accounting for the diversity of the household location and tastes. Specifically, the random coefficient λh represents the household’s marginal disutility of traveling, and γh is an unknown household taste parameter for store 21

See Train (2003) for further insights.

12

size. These parameters are allowed to vary across households according to     λh λ = + ΠZh + Γνh γh γ

(4)

where Zh is a z × 1 vector of observed household characteristics (i.e., the number of cars, the type of residence and whether a household’s residence is in a rural town), while unobserved household attributes are contained in νh . The coefficients λ and γ correspond to the mean of the distribution of the random coefficients, Π is a 2 × z matrix of coefficients that captures deviations from the mean in the population of household preferences for travel and store size due to household characteristics Zh , and Γ is a diagonal matrix of coefficients that captures unobservable heterogeneity due to random shocks νh . These unobserved household characteristics νh are assumed to be normally distributed and independent of the idiosyncratic term ε. Unobserved store characteristics (e.g., shelf display or assortment) are captured by the fixed effects ξr . I argue that these unobserved characteristics reflect essentially national strategies enacted by the retailers for their chains. Thus, these common shocks are captured by fixed effects defined at the retail chain level. As usual, I assume that the households value these unobserved characteristics in an identical manner. Similar to the “outside good” in classical demand models, households may decide to visit distribution channels other than large grocery stores (e.g., small convenience stores, specialized stores, and farmers’ markets) or not to purchase those food categories at all, which is resumed through the outside option j = 0. Without additional information on the characteristics of this alternative, I decide to normalize the characteristics of the outside option to zero. According to the highest utility rule, household h visits store j with the following probability: Z dF (εh )dF (νh ) Phj = Ahj

where Ahj = {(εh , νh ) |Uhj > Uhl ; l 6= j}, and F (·) denotes the distribution function. Because the store choice probability is a multiple integral, I use a simulation method to accurately approximate its value. Please refer to Appendix D for details on the estimation strategy.

4.2

Supply Models

At the end of the 1990s, retail prices were known to have been heavily affected by the introduction of the Galland Act. This law was enacted to prevent retailers from engaging in below-cost pricing in order to protect both small manufacturers and retailers. By clearly defining the below-cost price threshold as the price paid by the retailer at the time of delivery, the Galland Act prohibited retailers from passing on to consumers any conditional rebates (i.e., end-of-year rebates). Unfortunately, retailers and manufacturers have twisted the law by fixing arbitrarily high wholesale prices, resulting in lower intra-brand competition and higher profits for the industry (see Canivet, 2004, for instance). The negotiations between manufacturers 13

and retailers then progressively shifted from upfront margins to hidden margins. By allowing manufacturers to set industry-wide price floors, and in some extreme cases resale-price maintenance (RPM), the Galland Act has unambiguously caused significant price increases (Biscourp et al., 2013). While vertical relationships are clearly an important component of retail prices in France, the multidimensional nature of the demand model forces me to simplify the price-setting mechanism. Indeed, with a shopping basket composed of nearly 100 products, it appears very complicated to explicitly introduce the role of manufacturers, except by considering a “representative” manufacturer that will apply the same markup to all products. Moreover, some of the products included in the shopping basket have been known to be less impacted by the price-effects of the Galland Act (due to the low bargaining power of manufacturers), and manufacturers have been less able to control retail prices (e.g., fruits and vegetables, meat, cheese; see Canivet, 2004). Consequently, I assume that retailers set prices by only accounting for strategic interactions in the downstream market, and abstract from vertical strategic behaviors between manufacturers and retailers (see Villas-Boas, 2007 and Bonnet and Dubois, 2010 for dedicated analyzes on this subject). Concretely, this means that retailers are free to set prices (e.g., there is no RPM), take wholesale prices as exogenous and maximize their profits under the coercive pressures of spatial competition. Ignoring the role of manufacturers, it cannot be excluded that the recovered stores’ margin differs slightly from the true values if manufacturers adopt certain two-part tariff contracts. Nevertheless, this framework offers a convenient way to accurately proxy retailer market power, while accounting for a representative sample of the products they sell. I now describe the different pricing models that retailers may use while considering wholesale prices as exogenous. I assume that stores compete in prices and set their prices simultaneously, conditional on their characteristics supposedly chosen prior to this decision (e.g., location, store size or quality). To simplify, I assume that store managers try to maintain the price competitiveness of their store across all product categories. They adjust the overall price level of their store, which applies uniformly across all categories. In other words, they choose the price of the shopping basket and do not adopt category management. Thus, the prices that result from this behavior are a Nash equilibrium of the game. By deriving the pricing equations from the first-order conditions of the profit maximization problem, I recover the stores’ marginal cost and then compute their price-cost margin. Rather than assuming an arbitrary pricing behavior, I estimate the stores’ marginal cost under alternative pricing models that reflect different degrees of potential coordination across the stores. I then determine which model fits the data best. I depart from the most competitive case, where prices are set at the store level (model 1), meaning without considerations of ownership. Next, I consider the case in which integrated stores set their prices at the retail chain level, while stores belonging to cooperative groups fix their prices individually (model 2). I then regard the scenarios in which the pricing decision occurs at the retail chain and retail group levels (models 3 and 4, respectively). I also examine the possibility of spatial collusion by assuming that stores located near to one another maximize their joint profits. 14

Specifically, the definition of spatial collusion covers three cases corresponding to a store’s cooperative behavior with its nearest, two nearest or three nearest rivals (models 5 to 7). Finally, I consider the extreme case in which the stores in the area of study collectively behave as a monopolist (model 8).22 Each pricing model is solved as a function of the demand parameters and leads to a specific estimation of the stores’ marginal cost. In the following, I derive the expression of the stores’ marginal cost for the general case of multi-store retailers. The stores’ marginal cost formula for the other pricing models are derived by simply adopting a different definition of the ownership matrix T defined below. Consider the problem of a retailer R that sets its prices in its stores j = 1, . . . , JR . The profit function of retailer R is: X ΠR = (pj − cj ) Mj sj (p) − Cj (5) j∈JR

where cj denotes the constant marginal cost of selling a unit of a shopping basket for store j, Mj is the size of the market for store j, sj (p) is the market share of j and Cj is a fixed cost. If a pure-strategy Nash equilibrium in prices exists, the first-order condition for a typical store j belonging to JR is: sj (p) +

X l∈JR

T (pl − cl )

∂sl (p) = 0, j = 1, . . . , JR ∂pj

(6)

where T corresponds to the ownership matrix, with general element T (j, l) equal to one if both stores j and l belong to the same retail group and zero otherwise. This gives a system of JR equations. After defining ∆ as the retailer’s response matrix with element (j, l) = ∂sl (p)/∂pj , I can express the stores’ price-cost margin of retailer R in matrix notation by stacking up the first-order conditions and rearranging the terms: (p − c) = − [T ⊗ ∆ (p)]−1 s (p) (7) where ⊗ corresponds to the Kronecker product. It follows that the estimated stores’ marginal cost depends exclusively on the parameters of the demand system and the market conduct assumption: ˆ c = p + [T ⊗ ∆ (p)]−1 s (p)

(8)

For each pricing model, I obtain a specific set of estimated stores’ marginal cost. To determine which pricing model explains the data best I then conduct pairwise non-nested tests (Rivers and Vuong, 2002). The details of the procedure are provided in Appendix E. 22

One important feature of the considered market is that the hard discount chains set uniform pricing. Consequently, for all of the pricing models, I ensure that the prices of the hard discount stores belonging to the same retail chain are set so as to maximize the joint profits of its members. As a result, the definition of the pricing models presented here does not apply to hard discount stores.

15

5 5.1

Results Mixed Logit Demand Model

This subsection presents the estimates for the household’s store choice model defined in Eq.(3). The simulated maximum likelihood estimates are reported in Table 2. The estimated parameters must be interpreted with respect to the outside option. Almost all of the coefficients are both statistically and economically significant. Overall, I note that shopping patterns differ significantly across the households, store formats and areas of living. As expected, the households express a disutility of price and distance. More precisely, we observe that the households whose heads are between 30 to 39 years old (age group 3) appear less sensitive to price than the youngest age class. This heterogeneity among household preferences is also notable in the expression of the disutility of traveling. As defined in the indirect utility specification (see Eq.(3)), the marginal valuation of distance may vary with observable and unobservable household characteristics. This possibility allows me to reveal that the high disutility of traveling (the estimated mean of the distance coefficient distribution is -2.1602) intensifies for people living in a house and in a rural town. Conversely, the higher the number of cars owned by a household, the lower the household’s sensitivity to distance is. Nonetheless, the statistical significance of the estimated standard deviation of the random coefficient on distance suggests that the household’s willingness to travel must be explained by other individual characteristics that are not observed. Similar to the distance coefficient, the store size parameter is allowed to vary by household. On average, the households positively value the log of a store’s selling area (mean=1.2601), although important heterogeneity around this mean is observed (S.D.=2.4580). Moreover, I note that households seem to pay great attention to the number of cash registers per one hundred square meters, as suggested by the estimated parameter of this variable. I interpret this variable as a proxy for the inverse of waiting time at the checkout lines. I introduce several interaction terms with store formats to more accurately capture the variety in shopping patterns. As shown in Table 2, households living in Montpellier appear more likely to choose the outside option at the expense of a large grocery store. This result suggests that these households are sensitive to the number of alternative options located nearby and are aggregated into the outside option (e.g., specialized stores). In addition, for a given store, I count the number of rivals by format and by distance bands of 1 km and 2.5 km to interact these variables with the store’s format. By doing so, I investigate the nature of the competition among store formats with respect to the distance that separates them and also control for an endogeneity bias by accounting for factors that may influence the store manager’s price-setting behavior. Interestingly, it appears that supermarkets and hard discount stores compete fiercely when they are close together (within 1 km), but this competitive pressure vanishes for the supermarkets as soon as the radius is extended to 2.5 km. In addition, hard discount stores (resp. hypermarkets) take advantage of the commercial attractiveness generated by hypermarkets (resp. supermarkets). This relationship 16

Table 2: Results from the mixed logit model Hypermarket

Interactions with store formats Supermarket Hard discount

# hypermarket ≤ 1 km

41.0358*** (10.9084) –

41.9419*** (10.3335) –

# supermarket ≤ 1 km

–

–

# hard discount ≤ 1 km

0.3959 (0.3555) –

-0.5517*** (0.1975) -0.3528 (0.3282) –

Variable Price Price × age group 2 Price × age group 3 Price × age group 4 Price × age group 5 Price × age group 6 Ln(distance) Mean S.D. Ln(distance) × # cars Ln(distance) × house Ln(distance) × rural town Cash registers Ln(surface) Mean S.D.

-1.6636*** (0.5027) 0.2589 (0.2337) 0.5178** (0.2118) 0.1778 (0.2234) 0.0268 (0.2486) 0.1477 (0.2411) -2.1602*** (0.1214) 0.8460*** (0.0860) 0.1229* (0.0639) -0.2428** (0.1164) -0.7380*** (0.1671) 1.8211*** (0.4536) 1.2601*** (0.3584) 2.4580*** (0.2598)

Constant

# hypermarket ≤ 2.5 km # supermarket ≤ 2.5 km

0.4759*** (0.1245) -0.2483*** (0.0691) -2.2063 (1.9182) -8.7517*** (1.9988) -0.2016 (1.6546) 0.0384 (1.1687)

# hard discount ≤ 2.5 km Single household Montpellier city Rural town Size-adjusted household income Observations Log-likelihood

1,611 -3,415.37

Choice set radius

Sensitivity Price -1.7334*** -1.6481*** -1.8360*** -1.8659***

16 18 22 24

km km km km

Convenience store

-0.0009 (0.1018) -2.4700 (1.7757) -7.0741*** (1.8647) 1.2514 (1.5446) -0.2972 (1.0949)

40.1411*** (9.5232) 1.0162** (0.5061) -1.0373** (0.4217) –

47.3785*** (11.7990) –

0.5839*** (0.2233) -0.5421*** (0.1644) –

–

-2.8844* (1.6912) -5.7526*** (1.7729) 2.6948* (1.5765) -0.5048 (1.0332)

-1.3908 (1.8320) -5.7357*** (1.9946) –

– –

– –

-0.8524 (1.1970)

of estimated coefficients to households’ choice set definition Ln(distance) – Mean coef. (0.5310) -2.1467*** (0.1270) (0.5171) -2.1417*** (0.1208) (0.4777) -2.1856*** (0.1207) (0.4729) -2.2238*** (0.1211)

Notes: Standard errors in parentheses.*, **, *** indicate significance at the 10%, 5%, 1% level, respectively. The store-price index is divided by 10. The results with 100 Halton draws and retail chain fixed-effects. The omitted retail chain of hypermarkets is Leclerc. The omitted retail chain of supermarkets is Casino. The omitted retail chain of hard discounts is Norma. The omitted retail chain of convenience stores is March´ e U. Source: Author’s calculations.

17

indicates that a unilateral complementarity effect may exist between these formats, depending on the distance between them. One advantage that a mixed logit model has over a simple logit model is that it provides accurate estimates of substitution patterns, as cross-price elasticities vary by the competing alternatives. I determine the elasticity of the market share of a given store j according to the following expression: ( p R j αh shj (1 − shj ) dF (ν) if l = j ∂sj pl sj = (9) pj R − sj αh shj shl dF (ν) otherwise ∂pl sj The second and third columns of Table 3 report the average own-price elasticity and the mean of the cross-price elasticities by retail chain. Overall, the mean of the distribution of own-price elasticities across stores is -7.96, with a standard deviation of 2.65. Upon closer examination, I observe that the hypermarkets appear less sensitive to a change in their price (except for Auchan and Hyper U) than the other store formats, although important heterogeneity within formats and retail chains has yet to be considered.23 Likewise, a given price decrease in a hypermarket is more profitable for its rivals than a price decrease in a convenience store, as suggested by the cross-price elasticities. A sample of the estimates of own- and cross-price elasticities is reported in Appendix F. Note that some cross-price elasticities are null because certain pairs of stores cannot belong to the same choice set.

5.2

Pricing Policy and Stores’ Market Power

Applying the pairwise non-nested tests, I find that the store-level pricing model (model 1) is never rejected in favor of the other models (see Appendix E). Retailers, then, are more likely to use store-level pricing rather than any other pricing model considered in this study. Although quite unexpected given the organizational structure of French retail groups, this result echoes recent retailers’ statements about their pricing strategies.24 This suggests that stores do not internalize the cannibalization effect of rival stores operating within the same retail chain. This finding is explained by several factors that reduce coordination efforts at the retail chain level to soften competition between the stores of a chain. First, the existence of independent shopkeepers and affiliated stores (alongside company-owned stores) naturally implies that the objectives of store managers do not coincide with the objectives of their retail chain: they maximize their own profits and not those of their retail chain. Further, even if the “independent” stores of a retail chain source from a 23

The variation in a store’s market share resulting from a change in its price depends both on the store characteristics and its rivals’ locations. As the stores within a retail chain do not face similar competitive environments, their price variations affect their market shares in different ways (for identical characteristics). 24 For instance, Michel-Edouard Leclerc (the chief executive of the eponymous French retail chain) states in the press: “Nantes is one of the cheapest cities in France because there are 5 Leclerc franchisees, 10 Syst`eme U franchisees and as many Intermarch´e franchisees who are at each other’s throats!” (Libre-Service Actualit´es (LSA), March 6, 2008).

18

Table 3: Retail chains’ price elasticities and (gross) margin Retail chain

Elasticities Own-price Cross-price

Avg.

PCM (in %) S.D. Median

Obs.

Hypermarket Auchan Carrefour G´eant Casino Hyper U Inno Intermarch´e Leclerc

-9.67 -6.00 -7.66 -11.06 -5.59 -8.34 -6.81

0.18 0.57 0.20 0.06 0.20 0.30 0.28

10.34 22.35 13.15 9.04 17.90 13.71 14.67

– 14.15 1.07 – – 4.86 –

10.34 16.11 13.15 9.04 17.90 13.71 14.67

1 4 2 1 1 2 1

Supermarket Atac Casino Champion Intermarch´e Leclerc Monoprix Stoc Super U

-10.69 -10.09 -9.93 -7.33 -9.22 -9.43 -8.67 -5.71

0.17 0.13 0.05 0.14 0.30 0.12 0.19 0.09

9.35 9.93 10.14 16.87 10.85 10.61 12.81 20.95

– 0.49 0.88 9.81 – – 4.78 8.48

9.35 9.72 10.14 12.93 10.85 10.61 10.46 20.95

1 3 2 11 1 1 4 2

Hard discount Aldi Cdm Ed Leader Price Lidl Norma

-7.92 -8.85 -7.01 -7.78 -8.43 -7.41

0.10 0.02 0.07 0.09 0.06 0.11

13.50 11.32 17.22 14.62 12.40 13.74

0.59 0.38 7.52 4.68 1.43 1.86

13.50 11.32 13.48 12.11 12.05 13.74

2 2 4 4 5 2

Convenience store March´e U Shopi

-9.74 -9.33

0.04 0.04

12.26 13.52

4.94 7.23

12.26 8.63

2 3

Notes: Price elasticities and price-cost margins over retail prices (PCM in %) are averages by retail chain. Standard deviations and medians of the PCM are also displayed. The PCM reported correspond to the preferred pricing model and are computed following the expression (p − c) /p. Source: Author’s calculations.

central purchasing unit that provides them with a recommended resale price, they are free to set prices.25 Thus, managers of independent stores are not compelled to cooperate with the other stores belonging to the same retail chain. Second, all stores do not source exclusively from their central purchasing unit. A non-negligible part of their offer is provided by local producers (especially for perishable products). These “local” products have the advantage of offering greater flexibility in negotiations and even greater wiggle room for fixing retail prices. Third, a retail chain may internalize the substitution effects among its stores by using devices other than price (and not accounted for in the model). For instance, advertising, loyalty programs or sales may be more relevant levers for softening the competition among the stores of a chain. Consequently, it can be argued that assuming tacit coordination among 25

The recommended retail price is determined on the basis of the wholesale price and a productspecific margin, and is adjusted according to the degree of local competition.

19

the stores of a retail chain is not necessarily an appropriate hypothesis if one focuses on local competition. In addition, the results of the Rivers and Vuong tests yield valuable information on the determinants of a store’s market power. As shown by the non-nested tests, the pricing models built upon the assumption of spatial collusion (models 5, 6, 7) are rejected in favor of model 1. This means that tacit collusion among stores located near one another may be excluded from possible explanations of a retailer’s market power and, therefore, as a consequence of a low degree of competition. Given the preferred pricing model, I only comment on the recovered margins under the store-pricing scenario (model 1). The right hand side of Table 3 reports the estimated price-cost margins (in % of retail prices) by retail chain. It is worth noting that the estimated marginal cost of a supermarket located at the limit of the area of study gives a negative outcome (and thus, a margin higher than 100%), which led me to exclude it from the rest of the analysis.26 Not surprisingly, one first observes that the outcome of the Bertrand-Nash competition among the stores is far from corresponding to a perfect competitive environment, as the store margins are substantially higher than zero. Depending on the retail chain, the store’s price-cost margin represents, on average, between 9.04% and 22.35% of the retail price.27 The mechanism by which some large grocery stores exert significant market power is easily comprehensible. According to standard oligopoly theory, a firm’s markup is inversely proportional to its price elasticity. If I refer to the estimated elasticities, I note that the retail chains less sensitive to price variations are more likely to set higher prices relative to their marginal cost. Thus, the shopping basket offered by those retail chains is more differentiated than the shopping basket offered by their rivals because of factors such as assortment, advertising, store amenities, and location for instance. Considering that, I find that the retail chains Carrefour and Super U offer the most profitable product-mix (i.e., store characteristics, price strategy, and location). More generally, what conclusion can be drawn from these estimates? Principally, that the average level of profitability in Montpellier AU is not too excessive, especially when compared with the profitability levels of other retail sectors (e.g., department stores (36.4%), homewares (38.6%), clothing and shoes (43.5%), from INSEE’s retail census). Looking abroad, the UK grocery retail market provides an interesting benchmark owing to its similar market structure. Following the gross margins reported by Smith (2004, from 4% to 14%, see Table 3), I note that the levels of profitability are close in these two countries, albeit slightly in favor of the 26

This outlier results from the definition of the geographical boundaries of the survey. This definition truncates the choice set of households living in the same city as this store. Hopefully, this misspecification in the households’ choice set appears only for this city. 27 The relevance of these figures must be assessed by comparing them with the accounting margins published by the retail groups. For instance, Carrefour (Carrefour, Stoc, March´e Plus and Ed, for the most important retail chains) and Promod`es (Continent, Champion and Shopi) respectively reported an average gross margin of 21.1% and 19.3% for their group in 1999. These values can be considered as an upper bound of the true price-cost margins because they are computed from average variable costs. However, they are calculated at the national level and encompass all of the store formats. Under these limits, these figures inform the relevance of my estimates.

20

French retailers. Nonetheless, the average profitability masks important disparities within store formats as well as retail chains, as shown in Table 3. For example, the gross margins recovered for the Intermarch´e supermarket chain (i.e., the most represented retail chain in this area) lies between 9.46% and 43.98%. Beyond the differences in the store characteristics and marginal costs, which are small within a chain, one of the main reasons for this heterogeneity is the inclusion of market geography in the model. Depending on their location, stores face different demand and competitive environments, which implies heterogeneous price elasticities. As a result, the local monopoly power enjoyed by large grocery stores varies substantially depending on their location. Table 4 further investigates this relationship by regressing store’s price-cost margin (in %) on various measures of spatial concentration, while controlling for store characteristics. This standard structure-conduct-performance approach may suffer from an endogeneity bias due to the existence of local planning authorities that control entry, and hence the level of concentration locally. Since the inception of the Royer Act in 1973, the opening or extension of a large retail store has first to be approved by a regional zoning board before retailers can apply for a building permit (see Bertrand and Kramarz, 2002, for further insights). Depending on the strength of entry deterrence, a regional zoning board may foster or prevent the entry of new retail stores, which directly influences the structure of competition encountered locally by retailers and, indirectly, their market power. To account for this fact, I adopt a two-stage least squares estimator and use the log of the sum of approvals and refusals (in square meters) since the introduction of the Raffarin Act (in 1996) against concentration measures in the first stage. Approvals and refusals are counted at the canton level (a French administrative level grouping several cities) to reflect the decisions of local elected officials who compose half the members of the regional zoning boards. I also use as an additional instrument-the density of workers-to account for the concentration of employment at the commune level. Standard errors are clustered at the canton level to control for correlation in the disturbance term at an aggregate geographical level. Based on the exogeneity tests, I only detect an endogeneity issue when using distance-based measures of spatial concentration. I thus only report the 2SLS estimates for these specifications. The estimated coefficients confirm that the spatial distribution of rivals noticeably impacts a store’s market power, regardless of the chosen measure. I adopt a standard measure of local concentration (HHI) in Column (ii), whereas Columns (iii)-(vi) use crude measures of spatial concentration. For instance, the store’s pricecost margin in models (iii)-(vi) is regressed on the number of competitors (defined as stores or retail chains) counted by distance bands of 5 and 10 km. The coefficients show that the greater the number of rival stores (or retail chains) located within 10 km, the lower the store’s price-cost margin is. Using model (iii), I observe that each store entry between 1 and 5 km reduces a store’s price-cost margin by 2.12 percentage points on average. The decrease becomes more significant (3.19 percentage points) when a store competes with a new retail chain, as shown by model (v). Because there are 1.18 stores per retail chain on average, these figures indicate that a store suffers more when competing with a store belonging to a rival retail chain. The 21

22

Yes 0.0289 – –

(i) 0.0051 (0.0419) -0.0217 (0.1018) 0.2285 (0.2286) 0.0560 (0.1702)

Yes 0.2395 – –

(ii) 0.0172 (0.0344) -0.1453* (0.0776) 0.3523 (0.2004) 0.4690*** (0.1515) 0.2776*** (0.0407)

Yes 0.1749 – –

-0.0212*** (0.0047)

(iii) 0.0339 (0.0383) -0.2054* (0.1119) 0.2186 (0.2625) -0.0792 (0.1420)

Yes 0.2436 – –

-0.0135*** (0.0026) -0.0066 (0.0062)

OLS (iv) 0.0172 (0.0390) -0.1753* (0.0821) 0.3462 (0.2044) 0.3996* (0.2046)

Yes 0.1670 – –

-0.0319*** (0.0071)

(v) 0.0271 (0.0405) -0.1751 (0.1051) 0.1614 (0.2471) -0.0481 (0.1375)

Yes 0.2259 – –

-0.0289*** (0.0052) 0.0053 (0.0159)

(vi) 0.0093 (0.0426) -0.1533* (0.0759) 0.2884 (0.1999) 0.3273 (0.2271)

Yes 0.1622 – –

-0.0115*** (0.0034)

(vii) 0.0296 (0.0379) -0.1778* (0.0973) 0.2877 (0.2359) -0.0061 (0.1592)

Yes 0.1366 0.4699 0.0154

0.2294*** (0.0539)

Yes 0.1668 0.4310 0.0408

0.0441*** (0.0067)

2SLS (viii) (ix) -0.0139 0.0262 (0.0363) (0.0329) 0.0636 -0.1114 (0.1654) (0.1201) 0.4173* 0.3778* (0.2452) (0.2243) -0.5167** -0.0514 (0.2427) (0.1621)

Notes: This table reports the estimates of the second-stage regressions. The first-stage estimates are available from the author upon request. OLS and 2SLS regressions run with 61 observations. Clustered standard errors (at the canton level) are reported in parentheses. *, **, *** indicate significance at the 10%, 5%, 1% level, respectively. Source: Author’s calculations.

Retail format FE R2 Hansen J (p-value) Exogeneity test (p-value)

Sum of selling areas (’000) (1, 5) Dist. to the nearest store Cum. dist. to the 3 nearest stores

# retail chains (10, 15)

# retail chains (1, 10)

# retail chains (1, 5)

# stores (10, 15)

# stores (1, 10)

# stores (1, 5)

HHI

log(market income)

Mall

Gas station

Variable Fixed effects ξbr

Dependent variable: (log) price-cost margin (in %)

Table 4: Impact of rivals’ locations on store’s price-cost margin (in %)

importance of the location of rivals is again supported if the market power of a store is explained by the distance that separates it from its closest competitor (model viii) or by the cumulative distance between the store and its three nearest competitors (model ix). The same conclusion may be drawn if, instead of the number of rivals between 1 and 5 km, one accounts for the sum of the selling areas (model vii). The findings indicate that competition in the French grocery retail industry is highly localized. The competitive pressure exerted by a large grocery store is limited to a few kilometers.28 By thoroughly analyzing the intensity of the competition in this sector for a given metropolitan area, I show that consumers may suffer locally from strong positions held by retailers. Hence, a quarter of the stores of the survey compete, at most, with three rivals within 5 km. This limited level of competition allows stores to inflate their margin ,compared to the other stores, by six percentage points on average. This inflation results in a worse retail offer, which means higher prices and (potentially) a lower offer, lower store quality (e.g., freshness of products and cleanliness) and fewer services offered.

5.3

Robustness Checks

In this subsection, I check the robustness of the results by assessing the goodness of fit of the model and the sensitivity of the results to several hypotheses. First, I note that the predicted market shares do not differ by more than 0.6 percentage points, on average, from the observed market shares (with the highest deviation equal to 4.9 percentage points). Furthermore, I ensure that the model realistically allocates consumers to stores. To that end, I plot in Figure 1 both the observed and predicted travel distances for each household. I observe that the shapes of the distributions are similar, which confirms the predictive power of the model. I also check that the model does not suffer from identification problems that may arise because of a poor approximation of the random coefficient distributions (Chiou and Walker, 2007). Thus, I re-estimate the model with different numbers of draws (200 and 1,000 Halton draws) and starting values to ensure that the coefficients are identified by checking their stability. The stores’ margin reported in the previous section depends critically on the estimate of the price coefficient. In the following paragraphs, I examine how this coefficient varies depending on the different assumptions made. Measurement error in prices. The price variable suffers from a measurement error because some of the store price indices correspond to estimated values. Because the price sample is clustered by retail chain and is limited in size, I cannot resort to a bootstrap method to replicate the predictions of the SUR model and fully address this issue. Instead, I merely examine the sensitivity of the results by drawing 100 observations from the 95% prediction interval of each predicted price index. For each replicated sample, I then re-estimate the mixed logit model and re-compute the stores’ margin under the different pricing models. Overall, both the 28

A similar result was shown by Ellickson and Grieco (2013) for the US.

23

Figure 1: Distance to chosen store (in km)

demand estimates and the estimated margins vary slightly. The standard deviation of the price coefficient is 0.12 over the 100 replications. The stores’ price-cost margin for pricing model 1 varies, on average, by 9.43% with regard to the values reported in Table 3. Table 5 displays the original average margins by retail chain and the average margins calculated from the 100 replications. Household-specific price of the shopping basket. By incorporating between-household heterogeneity into the computation of the shopping basket’s price, I recover a price coefficient that may differ from the one I would obtain with a less flexible model. To evaluate the effect of the first stage of the model on household price sensitivity, I re-estimate the model by assuming that all households necessarily buy the same shopping basket composed of all the product categories at the same price (i.e., by setting P r (Ihc ) = 1 in Eq.(2)). The results of this alternative specification are reported in Appendix G. I observe little difference in the demand estimates, except a slight decrease of the disutility of price that must be interpreted with caution given the automatic rise of the shopping basket’s price (because now P r (Ihc ) = 1) with respect to the outside option. Using these estimated parameters, I derive higher price elasticities and lower margins, while reaching an identical conclusion for the preferred pricing model (store-level pricing). This result shows that the introduction of a household-specific store price index decreases the price sensitivity of the households. By defining more realistic shopping baskets (i.e., a narrower range of products), I make households less sensitive to price changes over the entire set of products sold in the stores. Consequently, ignoring the between-household heterogeneity in the composition of the shopping basket may underestimate stores’ market power. Boundaries of choice sets. The manner in which I define households’ choice sets directly impacts the distribution of the prices faced by households and can 24

Table 5: Original and replicated margins for pricing model 1 Retail chain

PCM (in %) Original 100 replications

Hypermarket Auchan Carrefour G´eant Casino Hyper U Inno Intermarch´e Leclerc

10.34 22.35 13.15 9.04 17.90 13.71 14.67

11.31 24.42 14.38 9.89 19.56 15.04 16.05

Supermarket Atac Casino Champion Intermarch´e Leclerc Monoprix Stoc Super U

9.35 9.93 10.14 16.87 10.85 10.61 12.81 20.95

10.24 10.89 11.10 18.45 11.90 11.58 14.04 22.89

Retail chain

PCM (in %) Original 100 replications

Hard discount Aldi Cdm Ed Leader Price Lidl Norma

13.50 11.32 17.22 14.62 12.40 13.74

14.74 12.38 18.85 15.96 13.56 15.04

Convenience store March´e U Shopi

12.26 13.52

13.41 14.91

Notes: The figures reported for the “100 replications” column correspond to the average margins over the 100 replications. Source: Author’s calculations.

therefore significantly affect the price coefficient. To test this link, I re-estimate the model under alternative definitions of the households’ choice set. The corresponding estimated price and distance coefficients are reported at the bottom of Table 2. Overall, because the price parameter varies between -1.6481 and -1.8659, there are small variations in the margins, whereas the mean of the distribution of the distance parameter varies between -2.1417 and -2.2238. Single-address model. I assume in the demand model that households only leave from their residence to go shopping. However, household travel surveys reveal that a significant share of households chain their trips and go shopping during their lunch break or after work, for instance. If a consumer leaves from another place than its residence (like its workplace) to visit a grocery store, it may either still reach a store included in its choice set or go to a store outside its choice set. In both cases, it travels a different distance than it would have done if it had gone from its residence, and I estimate an incorrect disutility of traveling. As I demonstrated in Section 3, only 2.60% of consumers have their primary shopping destination beyond a radius of 20 km of their residence, so the secondary concern is not really relevant. Further, based on the French census, I note that the average distance traveled to work in Montpellier AU is 14 km (so a little less in Euclidean distance). Because in the data I calculate an average of 3 km (in Euclidean distance) to go shopping, the bias on the distance coefficient introduced by trip-chaining behaviors does not seem too important.

25

6

Counterfactual Experiments

Many European countries are interested in enhancing competition in the grocery retail sector. Among the possible levers, both the UK and the French competition authorities have proposed amending the planning regimes of their respective countries to lower barriers to entry encountered locally and thereby promote the opening of new stores. Using the estimates obtained previously and the same NashBertrand equilibrium assumption, I assess the benefits that can be expected from opening a hypermarket by quantifying the effects of a hypermarket’s presence on other stores’ prices and welfare. Assuming that the stores may react to these new market structures by changing their prices, I perform several counterfactual experiments by alternatively removing each hypermarket and computing the new market equilibrium for each scenario. To measure the competitive effect of an existing hypermarket, I proceed as follows. I first eliminate the corresponding alternative from the households’ choice set. Then I use the demand estimates and the marginal costs recovered under the preferred pricing model to numerically compute the new market equilibrium. Given Eq.(8), the predicted equilibrium prices p∗ solve the system of J − 1 equations: p∗ = ˆ c − [T∗ ⊗ ∆ (p∗ )]−1 s (p∗ )

(10)

where cˆ are the estimated marginal costs recovered from the store-level pricing model (model 1) and T ∗ is the updated ownership matrix.29 This system of J − 1 equations has a unique solution if the price variable p∗ corresponds to the price of the shopping basket and not to the prices of the product categories. It follows from this restriction that I cannot compute a household-specific price of the shopping basket, as I did previously. However, to correctly approximate the gain associated with the presence of an alternative, I drop the first stage of the demand model, which entails identical prices across households. Based on the discussion in the previous section, this decision might soften the price changes in the simulations because the margins are slightly underestimated. I then compare the consumer surplus, the sector’s total profit and the equilibrium prices of the baseline situation with the results obtained from the removal experiments. In what follows, I retain the compensating variation as a measure of the change in consumer welfare. Hence, assuming that the marginal utility of income for each household remains constant after the removal of an alternative, Small and Rosen (1981) have shown that the amount of money a household would need to be compensated can be derived from: hP hP
i
i Jh Jh pre post ln − ln j=0 exp Vhj j=0 exp Vhj (11) CVh = |αh | where Vhjpost and Vhjpre correspond to Eq.(3) using the pre- and post-removal predicted prices, and net of the idiosyncratic term. Thus, the total change in consumer surplus is obtained by adding this measure for all households. These calculations assume 29

The prices of the outside option and the aforementioned outlier are assumed to remain constant. Moreover, I do not restrain hard discounters from setting uniform prices in their chain.

26

Table 6: Equilibrium outcomes following the removal of a hypermarket Hypermarket removed Auchan 1 Carrefour 1 Carrefour 2 Carrefour 3 Carrefour 4 Geant Casino 1 Geant Casino 2 Hyper U 1 Inno 1 Intermarche 8 Intermarche 13 Leclerc 2

Initial Mkt. Sh. 3.72 4.59 10.86 12.29 5.28 3.29 2.86 0.56 2.55 1.55 0.37 3.72

%∆wp

%∆p3

%∆ profit

0.19 0.06 0.31 -1.93 0.12 -0.11 0.03 0.02 0.05 0.13 0.07 0.03

0.64 3.47 0.96 -7.15 1.07 0.32 0.05 0.12 3.03 2.14 0.29 -0.09

1.45 -5.78 1.29 -13.22 0.78 0.22 0.55 0.09 0.14 -0.66 0.21 0.48

Decomposition and total change of %∆ CV Removal effect Price effect Total change -0.18 -0.17 -0.35 -0.55 0.09 -0.46 -0.68 -0.24 -0.92 -0.96 1.54 0.58 -0.19 -0.05 -0.24 -0.15 0.00 -0.15 -0.27 -0.05 -0.32 -0.02 -0.02 -0.04 -0.20 -0.08 -0.28 -0.13 -0.02 -0.15 -0.02 -0.04 -0.06 -0.20 -0.04 -0.24

Note: The variations reported are computed with respect to the pre-removal equilibrium. (%wp) corresponds to the store size-weighted average price change, %∆p3 corresponds to the average price change for the 3 nearest stores of the removed alternative, and %∆ profit is the total change in the sector’s profit. The removal effect captures the “gross” effect on the consumer surplus from eliminating the alternative, whereas the price effect quantifies the impacts of the price variations on the consumer surplus. The last column (%∆ CV) reports the total change in consumer surplus. Source: Author’s calculations.

that stores do not respond to the shock by changing their characteristics (observed and unobserved) and that the utility of the outside option remains unchanged. Table 6 displays the outcomes of the new equilibrium for each hypermarket removal. Columns (5) and (8) give the total change in the sector’s profit (%∆prof it) and the total change in consumer surplus (%∆CV ). As expected, in almost all cases, the simulations reveal a negative impact on the consumer surplus due to the elimination of a hypermarket. Depending on the hypermarket removed, the total change in consumer surplus varies between -0.92 and 0.58. In other words, the total amount of money the households would need to be compensated at each period of purchase for the closing of a hypermarket equals the variation multiplied by the total consumer surplus derived from the baseline situation and expressed in monetary units. The magnitude of the changes in total consumer welfare indicates that consumers noticeably value the presence of an additional hypermarket. Moreover, the largest hypermarkets (in terms of market share) do not necessarily contribute the most to consumer welfare. Along with the valuation of the store characteristics, the consumer surplus also accounts for the competitive pressure exerted by the hypermarket removed on the other stores. Because the total change in consumer welfare is likely to mask some opposite effects, I decompose the total variation into two components. One component captures the raw effect of eliminating the alternative (i.e., without price adjustment, see column 6), while the other component quantifies the price effect on consumer welfare (column 7). This decomposition indicates that the magnitude of the effects differs significantly across the simulations for these two components and notes that the expected effect of an additional hypermarket on consumer welfare depends greatly on its characteristics (e.g., location, chain name, and store size). In addition, by decomposing the effect, one can explain why households benefit from the removal of a hypermarket in one simulation because, in this particular case, it can be seen that the price effect more than offsets the removal effect. This finding suggests that this simulation generates an important price decrease. Finally, in column (5) it can 27

be seen that eliminating an alternative benefits more often to the industry, as the industry’s profits rise significantly. As indicated by the values of the price effect on consumer welfare, the changes in retail prices are moderated for a large portion of the simulations. Column (3) gives the store size-weighted average price change (%wp). At most, I denote a price increase of 0.31%, which is quite significant following a removal of a single hypermarket. However, I more frequently obtain an increase of below 0.10%. These low variations in price may be explained by the traditional trade-off between the “market share effect” and the “price competition effect” faced by stores. To illustrate this trade-off, I report in column (4) the average price change for the 3 nearest stores to the closing hypermarket. The fact that the nearest rivals increase their prices more intensively than the mean of all of the stores indicates that the competitive effect of an additional hypermarket falls rapidly with distance. Hence, the stores located closer to the removed alternative can achieve a greater profit by raising their prices than by cutting them to attract new customers. Conversely, those stores located farther away adopt the opposite strategy. The general message from these experiments is that reducing the market concentration level by promoting the entry of a new competitor is almost always beneficial to consumers. However, to ensure a significant price decrease, it appears necessary to favor the entry of several stores and to pay attention to the product-mix that they offer.

7

Conclusion

This paper addresses the issue of the French grocery retail sector’s competitive intensity. I develop and estimate a structural model of demand in which households have preferences over both store characteristics and geographic proximity. The methodology combines previous contributions of the literature on discrete choice models of demand among spatially differentiated firms and an original approach to determine household-specific shopping basket. In addition, the paper extends the existing literature devoted to the appraisal of retailers’ market power by accounting for the store’s ability to set its prices according to local market structure. As shown by recent inquiries, “price flexing” is a key feature of the business strategy applied by French retail chains, and consequently an important dimension of a store’s market power. Using the estimated parameters of the demand model, I recover the stores’ price-cost margin under alternative pricing strategies that differ by the degree of cooperation across stores. I then select the preferred pricing model by applying a non-nested testing procedure. The model is estimated for a French metropolitan area using a cross-sectional household survey containing detailed information on the stores visited for the main food product categories. The results show that stores set prices according to the most competitive scenario (i.e., store-level pricing). This finding rules out any collusive behavior as a cause of local monopoly power. The average estimated level of profitability exhibits no signs of low degree competition in this market. However, a closer look at the results shows important differences among the stores. These 28

differences indicate that a significant proportion of large grocery stores exert excessive market power. In practice, these stores take advantage of a weakly competitive environment to distort their offer and increase their margin. For instance, I find that stores that compete with, at most, three rivals within 5 km set their margin six percentage points higher on average than other stores. Although the survey is based on data covering a single market area, I am confident that the findings can be extended to the rest of France, as this market is less concentrated than the average of local markets (see Table 7). Together, the results contribute to the debate on the level of competition in the French grocery retailing sector. I provide new empirical evidence on the existence of market areas with a low degree of competition based on the low density of the stores combined with a high disutility of traveling expressed by the consumers in the area. The counterfactual experiments show that by promoting the entry of a new competitor, consumer welfare is significantly improved and retail prices almost always decreased. However, to ensure a significant price decrease, it appears necessary to favor the entry of several stores and to pay attention to the product-mix that they offer. More broadly, the simulations suggest that, to decrease retail prices in France, French policymakers should more seriously consider softening the Raffarin Act.

References Allain, M.L., Chambolle, C., Turolla, S., Villas-Boas, S., 2013. The Impact of Retail Mergers on Food Prices: Evidence from France. Working Papers hal-00920460. HAL. Allain, M.L., Chambolle, C., Verg´e, T., 2008. La loi Galland sur les relations commerciales. Jusqu’o` u la r´eformer ? Opuscule du CEPREMAP. ed. de la rue d’ulm edition. Bell, D.R., Ho, T.H., Tang, C.S., 1998. Determining where to shop: Fixed and variable costs of shopping. Journal of Marketing Research 35. Bell, D.R., Lattin, J.M., 1998. Shopping behavior and consumer preference for store price format: Why ”large basket” shoppers prefer EDLP. Marketing Science 17, 66–88. Beresteanu, A., Ellickson, P., Misra, S., 2010. The Dynamics of Retail Oligopolies. Mimeo. Duke University. Berry, S., 1994. Estimating discrete-choice models of product differentiation. RAND Journal of Economics 25, 242–262. Berry, S., Levinsohn, J., Pakes, A., 1995. Automobile prices in market equilibrium. Econometrica 63, 841–890. Bertrand, M., Kramarz, F., 2002. Does entry regulation hinder job creation? evidence from the french retail industry. Quarterly Journal of Economics 117, 1369– 1413. 29

Bhat, C.R., 2001. Quasi-random maximum simulated likelihood estimation of the mixed multinomial logit model. Transportation Research Part B: Methodological 35, 677–693. Biscourp, P., Boutin, X., Verg´e, T., 2013. The Effects of Retail Regulations on Prices: Evidence from the Loi Galland. Economic Journal 123, 1279–1312. Bonnet, C., Dubois, P., 2010. Inference on vertical contracts between manufacturers and retailers allowing for non linear pricing and resale price maintenance. RAND Journal of Economics 41, 139–164. Bresnahan, T.F., 1987. Competition and collusion in the american automobile oligopoly: The 1955 price war. Journal of Industrial Economics XXXV, 457– 482. Briesch, R.A., Chintagunta, P.K., Fox, E.J., 2009. How does assortment affect grocery store choice? Journal of Marketing Research 46, 176–189. Canivet, G., 2004. Restaurer la Concurrence par les Prix. Les Produits de Grande Consommation et les Relations entre Industrie et Commerce. La documentation Fran¸caise, Paris. Carluer-Lossouarn, F., 2007. L’Aventure des Premiers Supermarch´es. Lin´eaires Editions du Boisbaudry. Chiou, L., 2009. Empirical analysis of competition between wal-mart and other retail channels. Journal of Economics & Management Strategy 18, 285–322. Chiou, L., Walker, J.L., 2007. Masking identification of discrete choice models under simulation methods. Journal of Econometrics 141, 683–703. Cleeren, K., Verboven, F., Dekimpe, M.G., Gielens, K., 2010. Intra- and interformat competition among discounters and supermarkets. Marketing Science 29, 456–473. Cohen, M.A., Cotterill, R.W., 2011. Assessing the impact of retailer store brand presence on manufacturer brands in an equilibrium framework. Journal of Industrial Economics LIX, 372–395. Competition Authority, 1997. Avis relatif a` diverses questions portant sur la concentration de la distribution. Report 97-A-04. Paris. Competition Commission, 2008. The Supply of Groceries in the UK Market Investigation. The UK Competition Commission, London. Davis, P.J., 2006. Spatial competition in retail markets: Movie theaters. RAND Journal of Economics 37, 964–982. Dubois, P., J´odar-Rosell, S., 2010. Price and Brand Competition Between Differentiated Retailers: A Structural Econometric Model. mimeo. Toulouse School of Economics. 30

Einav, L., Leibtag, E., Nevo, A., 2010. Recording discrepancies in nielsen homescan data: Are they present and do they matter? Quantitative Marketing and Economics 8, 207–239. Ellickson, P.B., 2006. Quality competition in retailing: A structural analysis. International Journal of Industrial Organization 24, 521–540. Ellickson, P.B., Grieco, P.L., 2013. Wal-mart and the geography of grocery retailing. Journal of Urban Economics 75, 1–14. Ellickson, P.B., Misra, S., 2008. Supermarket princing strategies. Marketing Science 27, 811–828. Erdem, T., Keane, M.P., Sun, B., 1998. Missing price and coupon availability data in scanner panels: Correcting for the self-selection bias in choice model parameters. Journal of Econometrics 89, 177–196. European Competition Network, 2011. ECN Brief Extended Issue 05/2011. Technical Report. Handbury, J., Weinstein, D.E., 2015. Goods prices and availability in cities. Review of Economic Studies forthcoming. Houde, J.F., 2012. Spatial differentiation and vertical mergers in retail markets for gasoline. American Economic Review 102, 2147–2182. Manuszak, M.D., 2010. Predicting the impact of upstream mergers on downstream markets with an application to the retail gasoline industry. International Journal of Industrial Organization 28, 99–111. McManus, B., 2007. Nonlinear pricing in an oligopoly market: The case of specialty coffee. RAND Journal of Economics 38, 512–532. Nevo, A., 2000. A practitioner’s guide to estimation of random-coefficients logit models of demand. Journal of Economics & Management Strategy 9, 513–548. Nevo, A., 2001. Measuring market power in the ready-to-eat cereal industry. Econometrica 69, 307–342. Orth, M., 2012. Entry and Spatial Differentiation in Retail Markets. mimeo. University of Gothenburg, Department of Economics. Petrin, A., Train, K., 2010. A control function approach to endogeneity in consumer choice models. Journal of Marketing Research 47, 3–13. Pinkse, J., Slade, M.E., Brett, C., 2002. Spatial price competition: A semiparametric approach. Econometrica 70, 1111–1153. Richards, T.J., Hamilton, S.F., 2006. Rivalry in price and variety among supermarket retailers. American Journal of Agricultural Economics 88, 710–726. 31

Rivers, D., Vuong, Q.H., 2002. Model selection tests for nonlinear dynamic models. Econometrics Journal 5, 1–39. Small, K.A., Rosen, H.S., 1981. Applied welfare economics with discrete choice models. Econometrica 49, 105–130. Smith, H., 2004. Supermarket choice and supermarket competition in market equilibrium. Review of Economic Studies 71, 235–263. Thomadsen, R., 2005. The effect of ownership structure on prices in geographically differentiated industries. RAND Journal of Economics 36, 908–929. Train, K., 2003. Discrete Choice Methods with Simulation. Cambridge University Press. Turolla, S., 2012. Spatial competition in the French supermarket industry. Working Papers SMART - LERECO 12-01. INRA UMR SMART. Villas-Boas, S., 2007. Vertical relationships between manufacturers and retailers: Inference with limited data. Review of Economics Studies 74, 625–652. Vuong, Q.H., 1989. Likelihood ratio tests for model selection and non-nested hypotheses. Econometrica 57, 307–333.

32

Online Appendix A

The French Grocery Retail Sector: Local Market Structure

An important feature of the French grocery retailing sector is the high degree of concentration observed at the local level. To illustrate this point, it is crucial to properly define the relevant market of a store. Basically, a store’s catchment area is delimited by the maximal distance a consumer is willing to travel. According to previous surveys on shopping patterns, consumers travel, on average, between 10 and 20 minutes (drive-time), depending on the store format, to reach a large grocery store. Based on these figures, the European Commission (EC) and the UK Competition Commission usually delimit the boundaries of a store’s catchment area to a distance corresponding to 20-minute driving time (see, e.g., decisions in cases No. IV\M.1085 Promod`es/Catteau or No. COMP/M.1221 Rewe/Meinl), and the CA retains two relevant markets (one accounts for all of the stores located within a 15-minute driving time, whereas the other focuses exclusively on hypermarkets located within a 30-minute drive). In line with the former definition, I compute some statistics that illustrate to what extent the degree of concentration at the local level differs from that observed at the national level. To this end, I use the Panorama TradeDimensions database, which tracks detailed information on the market structure of the French grocery retail sector (e.g., entry, exit and rebranding). Because I only know the ZIP codes of the stores, I assume that the stores are positioned at the center of their respective city. I choose a radius of 10 km to delimit the boundaries of a store’s catchment area. In addition, I limit the scope of my analysis to the 500 largest French cities to concentrate on the most representative markets. Therefore, a relevant market consists of one of these 500 cities and its neighboring cities located within a 10 km radius. For the first quarter of 2000, Table 7 reports the number of retail chains and retail groups per market, the market share of the leader, and a measure of concentration by computing the Herfindahl-Hirschman Index (HHI) based on selling areas (again considering retail chain and retail group). In the top panel, I display the results by the quartile classes of the population distribution of the cities. Looking at the HHI, I find that the level of market concentration is higher locally than nationally. For the 500 local markets, an average HHI of 2310.94 (at the retail group level) is computed far exceeding the HHI computed at the national level (i.e., HHI=1214). As shown, the difference is due to an incomplete geographical coverage of the retail groups (8.83 retail groups per market are observed, on average, from a total of 12 groups). More troublesome, however, is that a significant number of local markets are highly concentrated according to a standard interpretation.30 For instance, the HHI 30 Formally, the HHI is defined as the sum of squares of all of the market shares in the market. According to the 2004 EU Merger Guidelines, a HHI over 2000 indicates a highly concentrated market.

33

Table 7: Market structure for the 500 largest French cities Population distribution of cities [0, Q1] [Q1, Q2] [Q2, Q3] [Q3, Q4]

13.97 15.63 15.54 16.65

Retail chain Mkt Sh. 1 HHI>2000 (in %) (in %) 29.11 32.80 27.17 23.20 26.92 19.20 26.21 15.20

Total

Nb 15.45

Mkt Sh. 1 27.35

Nb

HHI 1590.42

8.42 8.80 8.89 9.22

Retail group Mkt Sh. 1 HHI>2000 (in %) (in %) 35.54 65.60 34.75 53.60 36.19 66.40 35.82 62.40

Nb 8.83

Mkt Sh. 1 35.57

Nb

HHI 2310.94

Notes: Descriptive statistics are reported for the first quarter of 2000 and are based on the 1999 census population. The database surveyed all of the hypermarkets, supermarkets, hard discount stores and convenience stores (with selling areas over 400 m2 ). In total, I count 30 retail chains that belong to 12 retail groups. The average of the statistics is reported. For the first quarter of 2000, the computation of the HHI based on national market shares is equal to 747 (retail chain level) and 1214 (retail group level). Source: Author’s calculations based on Panorama TradeDimensions database.

computed at the retail chain level stresses that 15.20% of the cities in the fourth quartile have an HHI above 2000, while the percentage increases to 62.40% at the retail group level. Finally, it is worth noting that the level of market concentration differs significantly according to city size. Beyond market size considerations, these discrepancies may be explained by the planning system on entry in effect prior to 1996 (i.e., the Royer Act), which imposed a higher sales area threshold for stores opening in cities of less than 40,000 people.

B

Data

Three data sources are used in this study. The most important is a survey that records the shopping patterns of households living in the French administrative aire urbaine of Montpellier (Montpellier AU) for the year 2000. Montpellier AU is one of the most important metropolitan areas in France, ranked 15th in terms of population size in 2000 (459,916 people). Its market structure closely reflects the patterns reported in Table 7. The urban area covers an urban center where the retail group leader has approximately 30% of the market share and peripheral areas where half of the large grocery stores compete with, at most, 4 rivals within a radius of 5 km. Compared with the figures of Table 7, the level of concentration observed in Montpellier city is somewhat below the average of the most populated French cities (HHI=1934.50 vs. HHI=2310.94), but it still concentrated. The survey, named PACHA, was conducted jointly by Montpellier’s chamber of commerce and the department of economics at the University Montpellier I. It provides detailed information on all of the stores visited by a household for a total of 49 food and non-food product categories. Overall, 1,654 households participated in the survey. The survey follows a quota sampling methodology to create a representative sample of the geographical and socio-economic group (age and income) composition of the population. For each product category, a household was asked to list all the 34

Table 8: Visiting a large grocery store by category (in %) Category Fruits & vegetables Meat Cooked meat Cheese

Mean 0.4730 0.6145 0.5760 0.7939

(S.D.) (0.4994) (0.4869) (0.4943) (0.4046)

Category Other dairy product Grocery item Alcoholic drink Soft drink

Mean 0.8827 0.8492 0.7033 0.8908

(S.D.) (0.3219) (0.3580) (0.4569) (0.3120)

Notes: The number of observations is 1611. Source: Author’s calculations.

stores visited during the year, and next to order them according to their frequency of visit. To simplify the survey administration, the frequency of visit was discretized into 5 classes: 100%, 75%, 50%, 25%, and a marginal response ε%. The total of the shares reported for a product category must equal 100%. For example, a household may have declared visiting “Store A” 50% of time, “Store B” 25%, “Store C” 25%, and “Store D” ε% for a given product category. The household was also able to declare visiting a specialized store, a farmers market, to home make the product, or to not consume this product at all. The data reveal heterogeneous shopping patterns. For instance, some households visit a multitude of shopping places to do their food shopping (up to 11 places), whereas others concentrate their purchases in one or two places. Regarding the supermarket channel, the attractiveness of large grocery stores differs significantly depending on the product category. In Table 8, I report the mean probability of visiting a large grocery store by product category. For instance, the probability of choosing this retail channel is twice as large for soft drinks than for fruits and vegetables. However, it must be noted that important heterogeneity around theses means are observed. Finally, aggregating the data, I find that a household visits, on average, 1.78 store (S.D.=0.61). Departing from the information on household store choices, I determine the primary shopping destination of a household for the 8 food product categories retained (i.e., fruits and vegetables, meat, cooked meat, cheese, other dairy products, grocery items, alcoholic drinks, and soft drinks). To give greater importance to the choices that lead to more spending, I weight each choice by the mean expenditure share of that category. To that end, I use the TNS Worldpanel database, which contains detailed information on households spending by product. I select the data for the year 2000 and for the same d´epartment (French administrative division) as this study and classify each product into one of the eight product categories. Table 9 reports the weights computed. The survey also contains information on household characteristics, such as the age and the socio-economic status of the household head and the location of residence of the household. Some summary statistics on the household characteristics used in this study are presented in Panel A of Table 10. The second source of data is store characteristics obtained from the Atlas de la distribution, a national survey on French grocery stores. This database brings together information on store characteristics, such as store size, number of employees, number of parking spaces and the store location. Additional information on store characteristics (like the presence of a mall or a gas station) were filled in by in situ surveys. Panel B in Table 10 presents the summary statistics of store characteristics 35

Table 9: Expenditure share by product category (in the total food spending) Category % in food spending Fruits & vegetables 6.87 Meat 7.77 Cooked meat 6.59 Cheese 8.86 Other dairy product 12.17 Grocery item 27.53 Alcoholic drink 6.93 Soft drink 5.60 Other foodstuffs 17.68 Source: Author’s calculations based on TNS Worldpanel survey for the year 2000 and the d´ epartement of H´ erault.

used in the study. Finally, I supplement the information at the store level by collecting data on population and average household income at the commune level from the Census (French National Institute of Statistics, INSEE). The data were collected at the same time as the French grocery retail sector was experiencing a major event that significantly changed its market structure: the merger between the second-largest retail group (Carrefour) and the fifth-largest group (Promod`es). After a long inquiry by the competition authorities, the merger was approved on May 3, 2000. The merging firms kept almost all their existing store locations, but rebranded two of the pre-existing retail chains (Continent hypermarkets and Stoc supermarkets). The rebranding operations began on June 2000 for the hypermarkets and August 2000 for the supermarkets. It has been shown by Allain et al. (2013) that the merger led to substantial price increases in local markets where the merging group was able to internalize the competitive effect. Fortunately, the data used in this study have been only marginally affected by this event. The household survey was conducted mainly before summer 2000, and consequently, the store choices reported over the last 12 months are not contaminated by the merger. Further, the rebranding operations concerned only a handful of stores in Montpellier AU. Only 4 Stoc supermarkets were rebranded into Champion supermarkets. The operations began at the end of summer 2000 and lasted a couple of weeks. Together, these 4 stores represent approximately 3% of the total market share. Considering the timing of the rebranding operations and the fact that none of these stores compete directly with a former Champion supermarket or a Carrefour hypermarket, I assume that the merger has not modified the demand and the pricing decisions of the stores.

36

Table 10: Summary statistics Panel A: Store data Variable Units Hypermarket Binary Supermarket Binary Hard discount Binary Convenience store Binary Gas station Binary Mall Binary Surface m2 # cash registers Nb of cash reg./100m2 2 # employees/m Nb of employees/m2 HHI HHI/1000 Market income e/household Market population Sum of inhabitants Population density Nb of inhabitants/104 m2

Panel B: Household data Variable Units Age group 1 Binary Age group 2 Binary Age group 3 Binary Age group 4 Binary Age group 5 Binary Age group 6 Binary Credit card Binary House Binary Montpellier city Binary Rural town Binary Single household Binary Work Binary # cars Nb of cars/household log(Size-adjusted log(e/person) household income)

Mean 0.19 0.40 0.32 0.08 0.50 0.26 2167.14 0.66 0.03 2.62 14671.63 113592.70 13.11

S.D. 0.40 0.50 0.47 0.27 0.50 0.44 2503.56 0.20 0.02 2.32 3037.13 109437.60 11.69

Min 0 0 0 0 0 0 450.00 0.33 0.01 0.63 10654.51 3713 1.25

Max 1 1 1 1 1 1 11799.94 1.25 0.06 10 23469.04 251270 40.00

Obs. 62 62 62 62 62 62 62 62 62 62 62 62 62

Mean 0.1192 0.1136 0.1930 0.2272 0.1540 0.1930 0.8231 0.5587 0.5444 0.2806 0.2259 0.5307 1.3563 7.06

S.D. 0.3241 0.3174 0.3948 0.4191 0.3610 0.3948 0.3817 0.4967 0.4982 0.4494 0.4183 0.4992 0.8472 0.68

Min 0 0 0 0 0 0 0 0 0 0 0 0 0 5.54

Max 1 1 1 1 1 1 1 1 1 1 1 1 5 8.70

Obs. 1611 1611 1611 1611 1611 1611 1611 1611 1611 1611 1611 1611 1611 1611

Notes: S.D. corresponds to standard deviation. The Herfindahl-Hirschman Index (HHI ) variable is computed by assuming that the stores are positioned at the center of their respective cities. A store’s catchment area is composed of cities located within a radius of 10 km around the store’s city. The HHI is then computed based on selling surfaces at the retail group level. The Market income variable corresponds to the mean household income calculated over the set of cities that belong to a store’s catchment area. The Market population variable is computed as the sum of inhabitants living in cities that belong to a store’s catchment area. The Population density variable corresponds to the number of inhabitants of the commune per ten thousand square meters. There are 6 age groups: 20 to 24, 25 to 29, 30 to 39, 40 to 49, 50 to 59, and 60 and over. There are 5 household income groups: less than e765, e765 to e2265, e2265 to e4500, e4500 to e7500, and e7500 and over. The Size-adjusted household income variable is computed by first assigning a household to the mid-point of its income range. I then regress the log of the household’s income group on dummies for household size, the age of the household head, the number of cars within the household, the number of persons available to work and those working within the household, a dummy variable indicating whether the household lives in a house, and the sex of the head of household. The size-adjusted household income is calculated by subtracting its size fixed effect from the log of its income group and adding back the single household fixed effect. Source: Author’s calculations.

C

Results from the Bivariate Probit Models

The first stage of the demand model relies on the estimation of several Bivariate Probit Models (BPMs). For each product category, I estimate a binary discrete 37

choice model to determine the probability that a household will visit a large grocery store to buy this family of items. To control for a potential selection bias induced by household location, I specify a Bivariate Probit Model (BPM). Once the estimated parameters are obtained, I compute a household-specific store price index that is next plugged into the random coefficients logit model. By formulating a householdspecific store price index, I introduce an important source of heterogeneity among households that reflects the large variety of shopping patterns observed in the data at the product category level. Let us assume that a household h (h = 1, . . . , H) decides to buy a product in category c (c = 1, . . . , C) in a large grocery store. Following the discrete choice literature, I represent the purchase incidence of these product categories with an indicator variable ihc . The probability that a household h attends a large grocery store to buy a product in category c is then estimated through a Probit model of the form Pr (ihc ) = Φ (Xh ), where the explanatory variables correspond to a set of household characteristics describing their composition (Age group, Single household), their purchasing power (Size-adjusted household income, Credit Card, House), their location (Montpellier city), their availability to shop (Work), and their retail environment (# hypermarkets ≤ 10 km). However, the choice to reside in an urban center (i.e., Montpellier city) is not without consequences on the retail offer a household faces. Due to the zoning regulation in force, large grocery stores (e.g., hypermarkets and large supermarkets) are less likely to be available in inner cities where barriers to entry are the highest. Further, people living in inner cities are on average less inclined to travel to go shopping, which reduces their probability of visiting a large grocery store. To control for this selection mechanism, I augment the model with a selection equation that explains the household’s decision to live in Montpellier city. One way to consistently estimate this selection model is to consider a BPM. Let i∗hc denote the underlying latent variable associated with the purchase incidence of a product category c in a large grocery store and m∗hc the latent variable associated with the decision to live in Montpellier city. The model may be written as: i∗hc = Xh β1c + ε1c m∗hc = Zh β2c + ε2c with



ε1c ε2c



  ∼N

0 0

(12)

    1 ρc , ρc 1

The correlation between unobserved confounders of both equations is captured through the parameter ρc . If ρc 6= 0 it is necessary to take both equations’ factors into account to generate consistent estimates. For instance, the correlation of the error terms may result from unobservable household ability to travel long distances. The inability to drive (because of health conditions or simply due to the absence of a car) may motivate a household to settle in an urban center where corner shops are more numerous, but may also limit the ability to move to city outskirts where the largest grocery stores are located. 38

The results of the selection equation are presented in Table 11. Overall, we observe that the age of the household head and the number of cars per household significantly explain the decision to live in Montpellier city. This is not a surprising result because one-fifth of the population of Montpellier city consists of students. A young household without a car is thus a typical household of there. Looking at the decision to shop in the supermarket channel, I find important heterogeneity among the product categories (see Table 12). If living outside Montpellier city or facing a large number of hypermarkets are good predictors of patronizing a large grocery store, other features may explain this decision, depending on the product category. For instance, older households are less likely to visit a large grocery store for products usually available at farmers’ markets (e.g., fruits and vegetables, meat, cooked meat, cheese), as are single households. Surprisingly, the size-adjusted household income does not explain to any great extent the decision to shop in a large grocery store, except for fruits and vegetables.

D

Estimation Strategy and Identification

Estimation method. Let θBP M = {β, ρ} and θM XL = {α, λ, γ, φ, ϕ, ξ, Π, Γ} denote the set of demand parameters derived from the bivariate probit models and the mixed logit model, respectively. Whereas the bivariate probit models can be estimated by a standard maximum likelihood method, the store choice probabilities involve the computation of high-dimensional integrals and do not give closed-form solutions. I then estimate the demand parameters θM XL with a simulated maximum likelihood estimator (SML). Conditional on θM XL and I, the log-likelihood function to be maximized is: L Y ;θ

M XL

Jh XX 

 ,I = 1 log shj θM XL , I




h

(13)

j=0

where Y is the vector of store choices, and shj is the probability that household h chooses store j as its primary shopping destination conditional on the distribution parameters of λh and γh . For a mixed logit model, this probability is defined as: eVhj (θ

shj (λ, γ, Π, Γ) = PJh

j=0

M XL ,I

eVhj (θ

)

M XL ,I)




(14)

where Vhj corresponds to the indirect utility function net of the idiosyncratic term εhj . Denoting F ∗ the standard normal probability distribution function, the unconditional probability of household h’s choice is the integral of shj over all possible values of (λ, γ, Π, Γ): Z
M XL Phj θ , I = shj (λ, γ, Π, Γ) dF ∗ (λh ) dF ∗ (γh ) . (15)

39

40

Cooked meat -0.1725 (0.1514) -0.5199*** (0.1396) -0.5867*** (0.1430) -0.4081*** (0.1526) -0.6507*** (0.1433) 0.0191 (0.0620) 0.1220 (0.0925) -0.0157 (0.0792) 0.0971 (0.0915) -0.4919*** (0.0510) 0.9884** (0.3854) -0.1735 (0.1512) -0.5199*** (0.1395) -0.5863*** (0.1428) -0.4029*** (0.1524) -0.6482*** (0.1430) 0.0175 (0.0619) 0.1257 (0.0922) -0.0122 (0.0792) 0.0958 (0.0913) -0.4918*** (0.0509) 0.9942*** (0.3846)

Cheese

Other dairy product -0.1701 (0.1513) -0.5167*** (0.1395) -0.5838*** (0.1426) -0.4066*** (0.1524) -0.6513*** (0.1431) 0.0207 (0.0619) 0.1237 (0.0924) -0.0174 (0.0791) 0.0996 (0.0913) -0.4923*** (0.0508) 0.9750** (0.3846)

Grocery item -0.1722 (0.1513) -0.5195*** (0.1396) -0.5861*** (0.1428) -0.4093*** (0.1524) -0.6492*** (0.1431) 0.0170 (0.0619) 0.1252 (0.0923) -0.0142 (0.0793) 0.0990 (0.0915) -0.4908*** (0.0510) 0.9985*** (0.3846)

Notes: Standard errors in parentheses.*, **, *** indicate significance at the 10%, 5%, 1% level, respectively. Source: Author’s calculations.

Dependent variable: Montpellier city binary variable Variable Fruits & Meat vegetables Age group 2 -0.1716 -0.1712 (0.1515) (0.1516) Age group 3 -0.5186*** -0.5190*** (0.1397) (0.1398) Age group 4 -0.5879*** -0.5883*** (0.1430) (0.1432) Age group 5 -0.4070*** -0.4064*** (0.1526) (0.1527) Age group 6 -0.6468*** -0.6478*** (0.1432) (0.1432) Size-adjusted household income 0.0174 0.0172 (0.0621) (0.0620) Credit card 0.1236 0.1239 (0.0922) (0.0924) Work -0.0151 -0.0138 (0.0793) (0.0793) Single household 0.0992 0.1013 (0.0915) (0.0916) # cars -0.4909*** -0.4905*** (0.0510) (0.0510) Constant 0.9960*** 0.9963*** (0.3857) (0.3849) Alcoholic drink -0.1672 (0.1516) -0.5166*** (0.1398) -0.5886*** (0.1432) -0.4058*** (0.1526) -0.6453*** (0.1431) 0.0191 (0.0621) 0.1233 (0.0923) -0.0150 (0.0793) 0.0978 (0.0915) -0.4910*** (0.0510) 0.9831** (0.3854)

Table 11: Results from the Bivariate Probit Models (Montpellier city’s equation) Soft drink -0.1791 (0.1516) -0.5268*** (0.1399) -0.5934*** (0.1431) -0.4151*** (0.1528) -0.6620*** (0.1434) 0.0211 (0.0620) 0.1196 (0.0923) -0.0165 (0.0790) 0.0985 (0.0915) -0.4954*** (0.0504) 0.9896** (0.3844)

41

1.0318 (0.3097) -2,086.07 1611

0.4409 (0.5067) -2,035,87

2.1638 (0.1413) -2055.69

2.4441 (0.1180) -1791.05

3.6174 (0.0572) -1551.10

whether or not a household visits a large grocery store to Meat Cooked Cheese Other dairy meat product -0.8828*** -1.0564*** -0.7123** -0.6668** (0.2815) (0.2646) (0.2966) (0.3386) -0.1088 -0.1319 0.1233 0.0840 (0.1427) (0.1400) (0.1578) (0.1782) -0.1855 -0.3049** 0.1070 0.2504 (0.1394) (0.1355) (0.1561) (0.1892) -0.1748 -0.2963** 0.0367 -0.0340 (0.1495) (0.1450) (0.1653) (0.1880) -0.3623** -0.5022*** -0.2297 -0.0264 (0.1515) (0.1474) (0.1633) (0.1927) -0.4575*** -0.6460*** -0.2739* -0.3152* (0.1432) (0.1383) (0.1544) (0.1749) -0.0689 0.0073 -0.0150 0.0313 (0.0610) (0.0603) (0.0672) (0.0771) -0.0348 -0.0871 0.1084 0.1624 (0.0906) (0.0894) (0.0969) (0.1076) -0.0578 -0.0998 -0.0879 -0.1392 (0.0782) (0.0770) (0.0878) (0.1026) -0.0860 -0.1650* -0.2312** -0.2179* (0.0920) (0.0921) (0.1019) (0.1183) -0.0141 -0.0140 -0.0623 -0.1622 (0.0819) (0.0802) (0.0902) (0.1025) 0.1100*** 0.0986*** 0.0607*** 0.0387* (0.0170) (0.0167) (0.0186) (0.0207) 0.8925* 0.6115 1.0267** 1.1730** (0.4627) (0.4552) (0.4952) (0.5584) 0.1101 0.2468 0.2891 0.4307* (0.1666) (0.1699) (0.1859) (0.2291) 1.4710 (0.2252) -1663.83

1.1821 (0.2767) -1919.67

8.6639 (0.0032) -1527.45

buy a given product category Grocery Alcoholic Soft item drink drink -0.7209** -1.2177*** -1.1891*** (0.3232) (0.2939) (0.2778) 0.0847 0.1166 0.2806 (0.1650) (0.1450) (0.1783) 0.2140 0.1223 0.2780 (0.1667) (0.1437) (0.1822) 0.0223 -0.0067 0.0181 (0.1732) (0.1534) (0.1799) 0.0694 -0.0407 -0.0511 (0.1797) (0.1563) (0.1817) -0.1538 -0.2890** -0.2736 (0.1663) (0.1442) (0.1668) 0.0549 0.1062* -0.0090 (0.0722) (0.0637) (0.0749) 0.1418 0.2544*** 0.1431 (0.1028) (0.0906) (0.1060) -0.1395 0.0379 -0.1446 (0.0951) (0.0823) (0.1002) -0.0691 -0.1208 -0.0594 (0.1078) (0.0946) (0.1104) -0.0713 -0.0406 -0.0285 (0.0979) (0.0849) (0.0977) 0.0678*** 0.0986*** 0.0605*** (0.0203) (0.0189) (0.0213) 0.5935 -0.2961 1.4418*** (0.5370) (0.4905) (0.5135) 0.2361 0.1946 0.6559*** (0.1958) (0.1806) (0.2251)

Notes: Standard errors in parentheses.*, **, *** indicate significance at the 10%, 5%, 1% level, respectively. The reference Age group is Age group 1. The variable # hypermarket ≤ 10 km corresponds to the number of hypermarkets within 10 km of the household’s residence. Source: Author’s calculations.

LR-test {ρ = 0} p-value Log-likelihood Observations

Dependent variable: binary variable indicating Variable Fruits & vegetables Montpellier city -0.6976** (0.2755) Age group 2 0.0699 (0.1373) Age group 3 -0.1859 (0.1339) Age group 4 -0.0976 (0.1443) Age group 5 -0.1600 (0.1473) Age group 6 -0.3795*** (0.1385) Size-adjusted household income -0.1167** (0.0594) Credit card -0.0073 (0.0884) Work -0.0888 (0.0761) Single household 0.0917 (0.0895) House 0.0143 (0.0803) # hypermarket ≤ 10 km 0.0745*** (0.0161) Constant 0.8372* (0.4539) ρ 0.1701 (0.1685)

Table 12: Results from the Bivariate Probit Models (Shopping’s equation)

The unconditional store choice probability of a household h is then approximated by numerical simulations and can be written as: R

SPhj θ

M XL

1X shj (λ, γ, Π, Γ) . ,I = R r=1


(16)

By proprieties, these simulated probabilities are unbiased, and their variance diminishes as the number of draws increases. Rather than using random draws, I follow recent advances in simulation methods and generate 100 Halton draws. Note that I keep the same set of draws for each iteration.31 Endogeneity issue. Using a simulated maximum likelihood estimator, I obtain valid estimates of the demand parameters expressed in Eq.(3) under a condition of exogeneity of the idiosyncratic term εhj with all the covariates. As stressed many times in the empirical IO literature, the estimates of differentiated products demand systems may suffer from an endogeneity bias depending on whether managers choose some product characteristics (store characteristics, here) based on attributes that are unobserved by the researcher (Berry, 1994; Berry et al., 1995). If so, the variables of concern are correlated with the error term, and a traditional endogeneity issue arises. Past research has emphasized the strong presumption of price endogeneity because of the difficulty of controlling for the product characteristics (especially product quality) that influence pricing decisions. Consequently, the central identification assumption of the mixed logit model requires that εhj be independent with price, and thus that no unobservable store components (such as quality) are correlated with observable store and household characteristics, and more particularly with price. One way to control for unobserved store attributes that may be capitalized in price is to introduce retail chain fixed-effects ξr , assuming that part of price decisions is driven by unobserved attributes expressed at the retail chain level (e.g., advertising, quality of products sold). Thus, by introducing retail chain fixed-effects I control for an endogeneity bias as long as the unobserved store attributes do not deviate from the mean of their respective retail chain. However, it is likely that other unobserved store-specific attributes influence the price positioning of the stores. Therefore, to soften the previous condition, I incorporate additional store-specific variables, denoted by κq in the household indirect utility function, to control for the demand and the competitive environment of a store. Specifically, I introduce two sets of interaction terms: the first one accounts for the number of rivals (decomposing by store format) a particular store faces within 1 km and 2.5 km radii, while the second set of variables captures the demographic features of the city where a particular store operates (i.e., Montpellier city and Rural town). By introducing these interaction terms, I expect to control for the main unobserved drivers of the 31

A consensus exists in the literature regarding the superiority of Halton draws over random draws (see Bhat, 2001; Train, 2003; Chiou and Walker, 2007; among others). For a given number of draws, Halton draws achieve greater efficiency and coverage than random draws because the observations of a Halton sequence are negatively correlated.

42

price decision and therefore, the potential endogeneity issue.32 Identification. The demand parameters are identified through several sources of variation. First, each choice occasion differs from the others because of the heterogeneity observed in household characteristics. This heterogeneity makes it possible to identify the parameter α. Further, for a given choice occasion, a household faces a set of stores whose characteristics differ from one another. Hence, the average valuation of store characteristics identifies {φ, γ, ϕ} and, for the same reason, the unobserved characteristics of each retail chain ξr . Note that by specifying the fixed effects at the retail chain level rather than at the store level, I avoid the identification problem that may arise for the parameters of store characteristics, as the store’s dummy variable should be strongly correlated with the observed store characteristics. The main source of variation across choice occasions is derived from the heterogeneity in the spatial distribution of households’ residence and stores’ location, which induces different choice sets among households. As a result, different distributions of distance among households located in different cities (or IRIS ) can be observed. This finding enables me to identify parameters {λ, Π}. Finally, because a store price index is specific to a household and varies across the alternatives for a given choice set, there is sufficient variation to identify the parameters associated with pehj .

E

Selection Tests of Supply Models

Given that the pricing models defined in subsection 4.2 are non-nested, I infer the model that fits the data best by applying a non-nested testing procedure. I follow Rivers and Vuong (2002)’s procedure, which proceeds via pairwise comparisons of the competing models. For each competing model (denoted h), there is a corresponding pricing equation derived from the profit maximization problem and expressed as a function of the implied price-cost margin and cost shifters: phj = µhj + ωjh + Wjh λh +ηjh {z } | cost shifters

32

The data prevent me from adopting correction methods that address more directly the price endogeneity bias. One of these methods-known as the “fixed-effects” approach (Berry et al., 1995; Nevo, 2001)-addresses this issue by introducing alternative-specific fixed-effects that separate the market-specific valuation of unobserved attributes from the mean valuation of the alternatives, with the aim of applying traditional instrumental methods. However, to identify the parameters, I need sufficient variations in the data (i.e., a cross-section of markets or a longitudinal data set). As a result, I rule out this correction method because I can observe only one market for a given period of time in the data. Alternatively, I could refer to the “control function” approach developed by Petrin and Train (2010). Its principle consists of regressing in a first step the price variable on all exogenous factors. The residuals are then plugged into the indirect utility function along with the price variable. Because the price variable already includes an error term (as it corresponds to an estimated variable), I cannot use this two-stage error correction method. Therefore, I reject this approach in favor of the retained specification.

43

where µhj corresponds to the price-cost margin of store j under model h, while ωjh and Wjh are respectively the unobservable and observable costs shifters. By subtracting the stores’ price-cost margin from both sides of the equation, I can express the estimated marginal cost as follows: chj = ωjh + Wjh λh + ηjh The test principle consists in comparing the explanatory power (or, more precisely, the lack-of-fit criterion Qhn , see definition below) of the cost shifters on the marginal costs recovered from two competing models, h and h0 . Hence, the implied marginal costs of the preferred model are best explained by exogenous cost shifters. Identifying the cost parameters requires sufficient variation among the stores’ marginal cost, even though part of this variation may be independent from a rival’s behavior. In what follows, I assume that the marginal costs may be expressed as an exponential function of the cost shifters:
chj = exp ωrh + Wjh λh + ηjh where the unobserved cost variables ωrh are supposedly identical among the stores of the same retail group. The parameters of the cost equations are estimated using a nonlinear least squares (NLLS) estimator. Thus, the lack-of-fit criterion, Qhn , is defined as the objective function of the NLLS (i.e., the residual The Rivers-Vuong    sumof squares). √  0 0 n test statistic, defined as Tn = σcn Qhn θbnh − Qhn θbnh , is expressed as a function of the difference in the lack-of-fit criteria between the two competing models and an estimate of the sampling variance between the objectives (σ 2 ). The RiversVuong test statistic is asymptotically distributed according to a standard normal distribution. Based on Rivers and Vuong (2002), the null hypothesis is that the two competing models, h and h0 , are asymptotically equivalent when n h  h  h0 o h0 H0 : lim Qn θ − Qn θ =0 n→∞

 h
h where Qn θ is the expectation of the lack-of-fit criterion Qhn θh evaluated at the h

pseudo-true values of the parameters of model θ (similar to model h0 ). The first alternative hypothesis is that h is asymptotically better than h0 when n h  h  h0 o h0 H1 : lim Qn θ − Qn θ < 0. n→∞

Similarly, the second alternative hypothesis is that h0 is asymptotically better than h when n h  h  h0 o h0 H2 : lim Qn θ − Qn θ > 0. n→∞

The Rivers and Vuong (2002) test is a generalization of the Vuong (1989) test to a broad class of estimation methods (e.g., NLLS or GMM). Compared with other 44

45 4.23 (0.00)

(1) 1.7366*** (0.0549) 1.8141*** (0.0457) 1.6920*** (0.0251) 2.1324*** (0.0525) 0.0846** (0.0333) 0.0276 (0.0325) 0.0174 (1.1720) 0.0024** (0.0009) 4.04 (0.00)

(2) 1.7677*** (0.0738) 1.8909*** (0.0516) 1.6942*** (0.0263) 2.2049*** (0.0508) 0.0615 (0.0370) -0.0027 (0.0409) -1.5938 (1.2091) 0.0031*** (0.0010) 4.02 (0.00)

(3) 1.5156*** (0.1644) 1.7338*** (0.1208) 1.7026*** (0.0601) 2.1472*** (0.1060) 0.0743 (0.0642) 0.1593* (0.0884) -4.6815 (3.3053) 0.0035** (0.0016) 3.39 (0.00)

Pricing (4) 1.3854*** (0.1846) 1.6469*** (0.1453) 1.4642*** (0.1460) 2.0962*** (0.1212) 0.1351 (0.0807) 0.2018** (0.0853) -4.8219 (3.5061) 0.0050** (0.0021) 1.66 (0.14)

model (5) 1.6967*** (0.0771) 1.7756*** (0.0568) 1.6485*** (0.0421) 2.0793*** (0.0623) 0.1002** (0.0418) 0.0550 (0.0433) -1.4016 (1.1922) 0.0040*** (0.0011) 0.54 (0.82)

(6) 1.3145*** (0.3100) 1.4941*** (0.2250) 1.2928*** (0.3679) 1.8882*** (0.1441) 0.2951* (0.1712) 0.1876 (0.1298) -2.8163 (2.3676) 0.0088*** (0.0032) 0.49 (0.86)

(7) 1.2131*** (0.3709) 1.4166*** (0.2698) 1.2212*** (0.4485) 1.8371*** (0.1732) 0.3114 (0.1992) 0.2413 (0.1531) -3.1990 (2.9438) 0.0094** (0.0038)

2.41 (0.03)

(8) 3.6650*** (0.6664) 3.5795*** (0.4887) 2.8314*** (0.7036) 3.8639*** (0.3372) 0.3308 (0.3626) 0.0511 (0.3032) -10.9563 (7.3622) 0.0086 (0.0081)

Notes: Robust standard errors in parentheses.***, **, *: indicate significance at the 1%, 5%, 10% level respectively. The Gas station variable informs about the presence of a gas station at the store location as does the Mall variable. The # employees/m2 variable corresponds to the number of employees per square meter inside the store. The population density variable is equal to the number of inhabitants of the commune per ten thousand square meters. Retail group fixed-effects ωrh are not reported. The omitted retail group is Leclerc. Source: Author’s calculations.

 LR-test ωrh = 0 p-value

Population density

# employees/m2

Mall

Gas station

Convenience store

Hard discount

Supermarket

Cost variable Hypermarket

Dependent variable: Estimated marginal costs chj

Table 13: NLLS estimates of the cost equations

Table 14: Rivers and Vuong test statistics for pricing model selection H1 /H2 Model 1 Model 2 Model 3 Model 4 Model 5 Model 6 Model 7

Model 2 -2.73

Model 3 -3.76 -3.12

Notes: The test statistic is Tn =

√ n σ c n

Model 4 -5.08 -4.65 -3.16

Model 5 -2.70 -0.08 3.20 4.78

Model 6 -3.15 -2.98 -2.04 -1.49 -3.01

Model 7 -3.20 -3.07 -2.37 -1.95 -3.09 -0.83

Model 8 -5.70 -5.69 -5.69 -5.69 -5.69 -5.67 -5.70

    0  h0 θ bh bh → N (0, 1). A test statistic value below Qh n n θn − Qn

(above) the critical value of -1.64 (1.64) means that the row model (h) is better (worse) than the column model (h0 ) at the significance level of 5% (test statistics distributed standard normal) ; otherwise I cannot reject the null hypothesis that the competing models are asymptotically equivalent. Source: Author’s calculations.

non-nested model selection procedures, such as the Cox-type statistic developed by Bresnahan (1987), one advantage of this test is that it addresses a mis-specification of the competing models. However, by definition, this test is non-transitive and therefore may fail to determine a unique preferred model. NLLS estimates of the cost equation derived from the different pricing models are reported in Table 13. As expected, hard discount stores operate with the lowest distribution costs, followed by hypermarkets, supermarkets, and convenience stores. The results show that the stores’ marginal cost rises significantly with the presence of a gas station and a shopping mall. On the other hand, the number of employees per square meter (which may proxy for the level of service inside the store) does not influence the stores’ distribution cost. It also appears that the population density of the city where the store is located is positively correlated with the stores’ marginal cost, reflecting an inflationary effect on costs of the scarcity of retail space. Finally, the unobserved cost shifters captured through the retail group fixed-effects appear to be highly significant, as suggested by the likelihood ratio test. This suggests that the buying power of the central purchasing unit greatly influences the stores’ marginal cost. I select the preferred pricing model based on the results of the Rivers and Vuong tests. Table 14 displays the test statistics of the pairwise comparisons among the alternative models. A column (row) model is rejected in favor of a row (column) model if the test statistic value is below -1.64 (above 1.64) at the significance level of 5%. Using this decision rule, I find that the store-level pricing model (model 1) is never rejected in favor of the other models. Therefore, model 1 is the preferred model.

F

A Sample of Price Elasticities

46

47

1.6447

0.0115

0.0005

0.0000

0.0001

0.0542

Intermarche 7

Intermarche 1

Intermarche 8

Lidl 1

Outside option

0.0032

0.0045

0.0005

0.0025

0.0088

0.0061

0.0188

0.0291

0.0153

0.0891

-9.0701

0.0075

0.0322

Champion 1

0.0329

0.0743

0.0104

0.0592

0.0741

0.0473

0.1248

0.2066

0.0362

-8.6554

0.5485

0.0294

0.2147

Leader Price 3

0.0003

0.0293

0.3768

0.0167

0.0465

0.5174

0.0627

0.0814

-9.6726

0.1819

0.4739

1.4084

0.2846

Auchan 1

0.0204

0.0074

0.0017

0.1361

0.2139

0.0891

0.2271

-10.5675

0.0117

0.1497

0.1296

0.0098

0.0169

0.0122

0.0227

0.0095

1.3318

0.2727

0.0637

-9.2097

0.2113

0.0084

0.0841

0.0780

0.0068

0.0111

0.0119

0.0384

0.0781

0.7333

1.9804

-5.8807

1.7641

2.2980

1.9240

0.8831

0.7017

2.0205

0.1817

Intermarch´ e Intermarch´ e Carrefour 5 6 3

0.0434

0.0050

0.0020

0.6534

-5.4597

0.2118

0.8083

0.5900

0.0185

0.1481

0.1077

0.0174

0.0177

0.0582

0.0515

0.0193

-5.1945

0.4616

0.0554

2.7897

0.2652

0.0047

0.0836

0.0214

0.0043

0.0006

0.0301

3.4755

-5.3852

0.1812

0.0131

0.0554

0.1875

0.0315

0.9947

0.1382

0.0446

0.7118

0.0001

0.0466

-8.9481

0.1291

0.0180

0.0012

0.0010

0.0166

0.0050

0.0029

0.0366

0.0136

0.0017

0.0000

Intermarch´ e Intermarch´ e Intermarch´ e Lidl 1 7 1 8

Notes: Cell entries (i, j) give one hundred times the percentage change in the market share of i with a one percent change in the price of j. Source: Author’s calculations.

0.0007

0.0524

0.8162

0.0459

0.1324

0.1545

0.0126

0.2066

4.2630

0.4467

0.0214

0.0735

Auchan 1

Carrefour 3

0.2786

Leader Price 3

0.7010

Intermarche 6

0.2575

Champion 1

-6.5687

0.8556

0.0303

0.0730

Carrefour 2

Carrefour 2

Intermarche 5

-4.2127

Shopi 1

Shopi 1

Table 15: A sample of own- and cross-price elasticities from the mixed logit demand model

G

Estimates Obtained without the BPM Stage

This section reports the demand estimates obtained by fixing the category purchase probabilities in Eq.(2) to one. Put differently, I assume in this specification that households give equal importance to all of the product categories when deciding which store to visit. As expected, the price coefficient is somewhat different from the model estimated with a household-specific shopping basket, but it can be observed that the other coefficients are largely unresponsive to this change (see Table 17). On average, the recovered stores’ margin is 10.7% lower than those obtained with the two-stage model. This difference is explained by the fact that, by defining the households’ shopping basket more narrowly, one concentrates more on the product categories usually purchased in the grocery retail industry. As a result, household demand becomes less sensitive to price. However, this change does not modify the conclusion of the Rivers and Vuong tests. As shown in Table 16, store-level pricing (model 1) is still the preferred pricing model. Table 16: Rivers and Vuong test statistics (without first stage) H1 /H2 Model 1 Model 2 Model 3 Model 4 Model 5 Model 6 Model 7

Model 2 -2.42

Model 3 -3.74 -2.93

Notes: The test statistic is Tn =

√ n σ c n

Model 4 -4.85 -4.41 -3.05

Model 5 -3.88 -1.56 2.46 4.10

Model 6 -3.40 -3.24 -2.58 -2.05 -3.16

Model 7 -3.47 -3.35 -2.86 -2.45 -3.29 -0.92

Model 8 -5.48 -5.48 -5.48 -5.47 -5.48 -5.47 -5.48

    0  h0 θ bh bh Qh → N (0, 1). A test statistic value below n θn − Qn n

(above) the critical value of -1.64 (1.64) means that the row model (h) is better (worse) than the column model (h0 ) at the significance level of 5% (test statistics distributed standard normal) ; otherwise I cannot reject the null hypothesis that the competing models are asymptotically equivalent. Source: Author’s calculations.

48

Table 17: Results from the mixed logit model (without first-stage) Hypermarket

Interactions with store formats Supermarket Hard discount

# hypermarket ≤ 1 km

41.3297*** (10.5805) –

42.1707*** (9.9992) –

# supermarket ≤ 1 km

–

–

# hard discount ≤ 1 km

0.3993 (0.3533) –

-0.5353*** (0.1965) -0.3431 (0.3279) –

Variable Price Price × age group 2 Price × age group 3 Price × age group 4 Price × age group 5 Price × age group 6 Ln(distance) Mean S.D. Ln(distance) × # cars Ln(distance) × house Ln(distance) × rural town Cash registers Ln(surface) Mean S.D.

-1.4237*** (0.4073) 0.1508 (0.1792) 0.3873** (0.1620) 0.1699 (0.1701) 0.0775 (0.1809) 0.2526 (0.1708) -2.1547*** (0.1216) 0.8306*** (0.0890) 0.1217* (0.0635) -0.2332** (0.1135) -0.6942*** (0.1659) 1.8809*** (0.4539) 1.1387*** (0.3699) 2.4052*** (0.2757)

Constant

# hypermarket ≤ 2.5 km # supermarket ≤ 2.5 km

0.4269*** (0.1296) -0.2369*** (0.0693) -1.7898 (1.9683) -6.7482*** (1.9089) -0.2099 (1.6579) 0.0710 (1.1592)

# hard discount ≤ 2.5 km Single household Montpellier city Rural town Size-adjusted household income Observations Log-likelihood

Convenience store

-0.0134 (0.1021) -2.0468 (1.8148) -5.0880*** (1.7654) 1.2684 (1.5554) -0.2653 (1.0851)

40.1851*** (9.1963) 1.0703** (0.5047) -1.0774** (0.4218) –

47.8308*** (11.3554) –

0.5859*** (0.2247) -0.5490*** (0.1651) –

–

-2.5153 (1.7280) -4.1357** (1.6960) 2.5603 (1.5802) -0.4636 (1.0237)

-0.9253 (1.8423) -3.2290* (1.7491) –

– –

– –

-0.7922 (1.1721)

1,611 -3,414.64

Notes: Standard errors in parentheses.*, **, *** indicate significance at the 10%, 5%, 1% level, respectively. The store price index variable is divided by 10. The results with 100 Halton draws and retail chain fixed-effects. The omitted retail chain of hypermarkets is Leclerc. The omitted retail chain of supermarkets is Casino. The omitted retail chain of hard discounts is Norma. The omitted retail chain of convenience stores is March´ e U. Source: Author’s calculations.

49