Multiproduct retailing and consumer shopping behavior: The role of shopping costs

Multiproduct retailing and consumer shopping behavior: The role of shopping costs

Multiproduct retailing and consumer shopping behavior: The role of shopping costs Journal Pre-proof Multiproduct retailing and consumer shopping beh...

1MB Sizes 0 Downloads 68 Views

Multiproduct retailing and consumer shopping behavior: The role of shopping costs

Journal Pre-proof

Multiproduct retailing and consumer shopping behavior: The role of shopping costs Jorge Florez-Acosta, Daniel Herrera-Araujo PII: DOI: Reference:

S0167-7187(19)30088-8 https://doi.org/10.1016/j.ijindorg.2019.102560 INDOR 102560

To appear in:

International Journal of Industrial Organization

Received date: Revised date: Accepted date:

10 August 2018 14 September 2019 11 November 2019

Please cite this article as: Jorge Florez-Acosta, Daniel Herrera-Araujo, Multiproduct retailing and consumer shopping behavior: The role of shopping costs, International Journal of Industrial Organization (2019), doi: https://doi.org/10.1016/j.ijindorg.2019.102560

This is a PDF file of an article that has undergone enhancements after acceptance, such as the addition of a cover page and metadata, and formatting for readability, but it is not yet the definitive version of record. This version will undergo additional copyediting, typesetting and review before it is published in its final form, but we are providing this version to give early visibility of the article. Please note that, during the production process, errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain. © 2019 Elsevier B.V. All rights reserved.

Highlights • We estimate consumer shopping costs to be 104.7 euro cents per store visited, on average. • The omission of shopping costs yields biased predictions of substitution patterns and firm markups and marginal costs. • When a supermarket delists a product, it losses customers to rivals and its weekly revenue decreases by 5%. • The delisting supermarket reduces the price of substitutes and rises that of complements keeping the total price of baskets constant. • When a supermarket gives an aggresive price discount, it rises the price of complements in order to partially offset losses. • However, the extent to which it is able to compensate the loss in revenues depends heavily on shopping costs. • The rise in prices is up to 10% when shopping costs are present, whereas it is up to % when shopping costs are absent.

1

Multiproduct retailing and consumer shopping behavior: The role of shopping costs Jorge Florez-Acosta and Daniel Herrera-Araujo∗ This version: September 2019

Abstract We empirically examine the role of shopping costs in consumer shopping behavior in a context of competing differentiated supermarkets that supply similar product lines. We develop and estimate a model of demand in which consumers can purchase multiple products from multiple stores in the same week, and incur transaction costs of dealing with supermarkets. We show that a similar model without shopping costs predicts a larger proportion of multistop shoppers and overestimates own-price elasticities and product markups. Further, we use our model along with a model of competition between supermarkets to study two practices that are commonly used by supermarkets: product delisting and loss-leader pricing. We show that the presence of shopping costs makes product delisting less profitable whereas it makes loss-leader pricing more profitable compared to a context in which consumers do not incur shopping costs. Keywords: Supermarket competition, market power, multistop shopping, shopping costs, product delisting, loss-leader pricing. JEL Codes: D12, L13, L81. This research did not receive any specific grant from funding agencies in the public, commercial, or not-for-profit sectors. Declaration of interest: this paper was partially supported by the French Institut National de la Recherche Agronomique (INRA) in the following ways: first, Florez benefited from a PhD dissertation grant for three years between 2011 and 2014 while completing his PhD at the Toulouse School of Economics, and Herrera benefited from a PhD scolarship during the course of his PhD at the Toulouse School of Economics; and second, the permission granted by INRA to access the data base used in this paper.

Acknowledgments: This paper is a revised version of a joint chapter from each author’s 2015 PhD dissertation. Special thanks to Bruno Jullien, Pierre Dubois, James Hammitt and Patrick Rey for detailed comments and discussions. We also benefited from conversations with François Salanié, Andrew Rhodes, Michelle Sovinsky, Martin Peitz, Rachel Griffith, Andrea Pozzi, Helena Perrone, Juan Esteban Carranza, Tiago Pires, Alessandro Iaria and Laura Lasio, along with participants at various seminars and conferences. We are grateful to the Institut National de la Recherche Agronomique, INRA, for providing us with access to their data. All errors are our own.



Florez: Department of Economics, Universidad del Rosario, Calle 12C 4-59, Bogotá, Colombia (e-mail: [email protected]); Herrera: Université Paris-Dauphine, LEDA (CGEMP), Place du Maréchal de Lattre de Tassigny, (e-mail: [email protected]).

2

1

Introduction

The modern grocery retail industry is dominated by a small number of powerful large-scale supermarket chains1 that attempt to entice customers to favor one-stop shopping through aggressive non-price strategies, such as: first, the proliferation of superstores with vast floor areas (20,000+ sq. meters) that offer a large product range (200,000+ different brands);2 second, a joint location with suppliers offering parallel services (e.g., shopping malls, beauty salons, restaurants, car wash facilities, gas stations, and playgrounds for children); third, the promotion of private labels (PLs), whereby supermarkets are less dependent on branded products (the socalled national brands, NBs) and induce consumer loyalty; and lastly, the increasing emphasis on strategies that induce consumer retention and reinforce store loyalty, such as loyalty programs. In such a context, it becomes important to understand the role that consumers play in the way in which a set of differentiated supermarkets offering similar product lines interact. There are a number of key features that determine the effectiveness of supermarket price and non-price strategies and the exercise of market power. In particular, whether customers prefer concentrating their purchases within a single supermarket or sourcing multiple store chains in the same period, the transaction and opportunity costs associated with shopping activities (so-called shopping costs), and how these costs and shopping patterns shape the way in which consumers substitute across products and supermarkets. Precisely, a number of theoretical papers on multiproduct retailing shows that allowing customers to incur heterogeneous shopping costs in the analysis of multiproduct demand and supply may dramatically change policy conclusions (see Klemperer (1992); Klemperer and Padilla (1997); Armstrong and Vickers (2010); Chen and Rey (2012, 2019)). The contribution of this article is to study the role of shopping costs in explaining the consumer choice of multiple products and multiple shopping locations, and in the measurement of supermarkets’ market power. To this end, we develop a multiple-discrete choice model in the context of competition between supermarkets that offer the same product line to the same customers. Consumers can purchase baskets of products from either a single store (one-stop shopping) or multiple stores (multistop shopping) during a given period. Our key modeling strategy is to explicitly account for this observed heterogeneity by introducing consumer transaction costs related to shopping. Following Klemperer (1992), we comprehensively define shopping costs as all of the consumer’s real or perceived costs of using a supplier.3 These may include transportation costs and opportunity costs related to time spent parking, selecting products in 1 For instance, in 13 member countries of the European Union (including France, the UK, Germany, Sweden, Finland, and the Netherlands), the five leading retailers in each country have a combined market share of more than 60% (European Commission, 2014). 2 In Europe, for instance, between 2000 and 2011 the number of hypermarket outlets increased by 72% and the sales area increased by 46%, while the corresponding figures for supermarket outlets were 10% and 26%, respectively (European Commission, 2014). 3 Klemperer (1992) distinguishes among consumer costs in the following way: “(. . . ) a consumer’s total costs include purchase cost and utility losses from substituting products with less-preferred characteristics for the preferred product(s) not actually purchased [transport costs of the standard models à la Hotelling] (. . . ) Consumers also face shopping costs that are increasing in the number of suppliers used” p. 742.

3

the store, and waiting in line at the checkout; they may also account for the taste for shopping (Chen and Rey, 2012). Our general empirical strategy is to estimate basket-level demand using standard techniques from the discrete-choice literature, along with simulated methods. We specify the utility of each product as a function of observed and unobserved product and store characteristics, as well as parameters to be estimated. On every shopping occasion, each consumer faces idiosyncratic shopping costs that increase with the number of supermarkets visited. Each consumer weighs up the extra benefits of dealing with an additional store against the additional costs involved. If the benefits exceed the costs, the individual will visit an additional supermarket. Otherwise, she will make all her purchases at a single location. The total utility of a basket of products is the sum of the product-specific utilities minus the shopping costs. To consistently estimate the parameters of the model, we encounter this challenge: shopping costs vary across individuals and are unobserved (by the econometrician). We deal with this by decomposing shopping costs into two components: first, a mean shopping cost, which is common to all consumers; and second, an idiosyncratic deviation from the mean cost, which depends on both observed demographic characteristics and a random shock, which is known to consumers and assumed to follow a known parametric distribution. This shock captures all individual (unobserved) characteristics that may cause individual costs to differ from the average shopping cost. Once the parameters of the model are estimated, we perform two exercises that allow us to assess the relative importance of explicitly accounting for shopping costs in predicting reasonable substitution patterns in a multistop shopping environment. In the first exercise, we take our estimated model and simulate a scenario in which shopping costs fall to zero; that is, we assume that consumers no longer incur positive shopping costs. In the second exercise, we estimate an alternative specification that does not include any shopping costs, and compare the results with those obtained from our preferred specification with shopping costs. Furthermore, we apply our models of demand and supply to empirically examine some of the implications of two practices that are commonly used by supermarkets: product delisting and loss-leader pricing. The first case arises when a supermarket removes a particular product from its shelves while rivals keep supplying it. Although it can happen purely for commercial reasons (Davies, 1994), it can also be used by the supermarket in a strategic way to exploit manufacturers. In any case, delisting of products can entail losses for the delisting supermarket if a high proportion of its customers face high shopping costs.4 We simulate a large price increase for one of the products sold by a given supermarket so that it becomes prohibitively costly for consumers, while the same product continues being supplied by competing supermarkets at observed prices. The second case, which involves setting the price of certain products at or below its retail marginal cost (Lal and Matutes, 1994), often comes as a result of aggressive temporary price discounts. Even though this practice has been banned in a number of countries 4

For instance, in 2009 in Belgium, a request for a price rise by Unilever triggered the delisting of 300 of Unilever’s products by Delhaize, one of the largest supermarket chains in that country. Both parties ended up being harmed: Delhaize lost 31% of its customers to rivals and among those who remained, 47% substituted other brands for Unilever’s products. See http://in.reuters.com/article/delhaize-unilever-idINLG51937220090216.

4

because it can be predatory, recent literature shows that in the presence of shopping costs lossleading strategies and cross-subsidies are not predatory (see Chen and Rey (2012, 2019)).5 We simulate a situation in which one supermarket sets the price of one product at its marginal cost which turns out to be a discount of nearly 56% on the observed price. In both applications, we measure the net effect of the respective price change on demand and supply by allowing rival supermarkets to adjust their prices to a new equilibrium. For the sake of documenting the role of shopping costs, we perform two counterfactual experiments: one in which shopping costs are present and another in which we set shopping costs to zero for all consumers. Perhaps the most significant limitation of our approach is the dimensionality problem that arises when estimating demand for both baskets of products and multiple shopping locations. In our data set, we observe that a household purchases baskets containing 24 different products from two separate stores each week, on average, while some households can purchase up to 275 different products from up to nine separate grocery stores in the same week. We deal with the potential dimensionality problem as follows. First, we restrict our focus to three categories of products that are staple food items, among the most frequently purchased, and usually subject to unit demand. In our case, these categories are yogurt, biscuits, and refrigerated desserts. Second, we aggregate brands to the category level so that we end up with a reduced set of composite products. To flexibly evaluate the effects of product delisting and loss-leading, we allow for two alternatives of yogurt, namely, the leading national brand (NB) in France in 2005, and a composite yogurt “brand” that includes all of the remaining alternatives. Therefore, consumers have a set of four products from which they can choose at most three: one of the two alternatives of yogurt, biscuits, and desserts.6 Finally, we focus on a reduced set of three supermarket chains that were the leading grocery retailers in France in 2005 based on market share. The remaining supermarkets in our data set, along with the no purchase of the included goods option, are left as part of the outside good.7 We obtain several interesting results. First, from descriptive regressions, we find a significant relationship between the number of supermarkets visited by a household in a week and household characteristics that are a proxy for the opportunity cost of time. Second, our structural model allows us to retrieve consumer total shopping costs, which we estimate to be 104.7 euro cents per supermarket visited, on average. This cost includes a fixed shopping cost of 96.1 euro cents per supermarket visited and a transport cost of 8.6 euro cents per trip to a supermarket located at the average distance from consumer location. Third, when we unilaterally set shopping costs to zero, our simulations indicate that all consumers would visit at least one store every week with positive probability and, in particular, the probability of doing multistop shopping is similar to that of 5 A recent example happened in France in 2018, where one of the leading supermarket chains in the country, Intermarché, offered Nutella spread jars at 70% discount. Intermarché was fined for pricing some of its products below cost. 6 The choice of yogurt as the category offering two alternatives is arbitrary. 7 Of course, our data set indicates that consumers often purchase/visit more than the included products/stores. We take this into account in our empirical strategy. Following Gentzkow (2007), we assume that every basket includes a maximization over the included as well as the excluded (both observed and unobserved) products. See Section 3.3 for a detailed discussion.

5

doing one-stop shopping. Once shopping costs are accounted for, the predicted probabilities of both one- and multistop shopping are lower, and consumers are less likely to visit a supermarket on a week-to-week basis. Fourth, when we compare the substitution patterns predicted by an alternative specification without shopping costs with those of our model, we find that shopping costs reinforce the complementarities between product categories that emerge when customers are able to purchase baskets of products. Such complementarities can be thought of as the “economies of scope” of buying related products from a single supermarket, as discussed by Klemperer (1992). In fact, the cross-price elasticities predicted by our model are larger than those of the alternative specification, suggesting that shopping costs make those economies of scope more valuable to consumers. Further, from the delisting application we find that the supermarket delisting the product strategically decreases the price of substitutes to try to encourage intra-store substitution. Conversely, we find that the price of complement products increases and that this increase offsets the decrease in the price of the substitute product, which suggests that an optimal strategy for the supermarket delisting the product is to keep the overall value of the basket constant and retain one-stop shoppers. Yet, the high shopping costs faced by one-stop shoppers implies that the delisting supermarket losses some demand to rivals and experiences a decrease in its revenues. Finally, from the aggressive discounting application, we find that the supermarket granting the discount increases the price of complement products up to a 10% and that the extent to which it is able to compensate the lower revenues on the loss leader depends on shopping costs. When shopping costs are not present, the supermarket giving the aggressive discount still increases the price of complements but up to 2%. The aggressive price discount generates a considerable increase in the demand for the loss leader whereas the demand for all other products of the supermarket giving the discount decreases. Rival supermarkets also experience a decrease in the demand for all of their products and the share of the outside good also decreases. Our results suggest that the loss-leading strategy stimulates some multistop shopping and a market expansion effect for the loss leader in addition to the inter-store substitution effect. This paper relates to a growing body of empirical literature that models consumer choice problems explicitly accounting for opportunity costs associated with shopping activities. This literature has focused on two types of costs, namely, search costs (e.g., Hortaçsu and Syverson (2014), Hong and Shum (2006), Koulayiev (2014), Moraga-Gonzalez, Sandor and Wildenbeest (2013), Kim and Bronnenberg (2010), Wildenbeest (2011), De los Santos, Hortaçsu and Wildenbeest (2012), Honka (2014), and Dubois and Perrone (2015)), and switching costs (e.g., Shy (2002), Viard (2007), and Honka (2014)). Less attention has been paid to shopping costs. Brief (1967) models consumer shopping patterns in a Hotelling framework, and estimates transportation costs to account for consumer shopping costs. Aguiar and Hurst (2007) evaluate how households substitute time for money by optimally combining shopping activities with home production. Customers incur a time cost of shopping that is explicitly accounted for. This paper also relates to literature that has developed demand models for multiple products. Hendel (1999) develops a multiple-discrete choice model to explain how firms choose multiple 6

alternative brands of personal computers. Further, Dubé (2004) applies Hendel’s model to the case of carbonated soft drinks. Gentzkow (2007) develops a flexible framework in which similar products can be either substitutes or complements. None of these studies incorporate consumer transaction costs into the choice problem. Wildenbeest (2011) sets out a search cost model in which consumers want to purchase a basket of products in a single stop and care about the total price of the basket. Finally, in analyzing multiproduct and multistore choice with shopping costs, our paper is closely related to that of Thomassen et al. (2017), who study pricing by grocery stores in the context of competition between specialized stores and supermarkets. To this end, they develop a model of demand in which consumers make discrete–continuous choices over multiple categories. Consumers can purchase from up to two stores in each period and incur a choicespecific fixed cost of shopping. Our approach, which we have developed contemporaneously and independently, differs from theirs in several important ways. First, our primary focus is on the role of shopping costs in predicting consumer substitution and shopping patterns. Second, we use our model to explore the effects of product delisting and loss-leader pricing, and though we add structure to the supply side in order to present it in a more realistic way, our main focus is on the demand side effects of such practices. Third, our empirical analysis is oriented by a model that is in line with the theoretical literature on multiproduct retailing with shopping costs. In our setting, the number of stores visited by a consumer is endogenously determined by a stopping rule involving the extra utility and extra costs involved in visiting an additional store. This enables us to empirically identify the distribution of shopping costs. Last but not least, while we are interested in analyzing the competition between supermarket chains of similar size and characteristics that supply the same product range to customers, they focus on competition between small specialized stores and large supermarkets. The rest of the paper is organized as follows. Section 2 outlines our structural model of multiproduct demand and consumer shopping behavior in the presence of shopping costs, as well as a supply model of supermarket oligopolistic competition. Section 3 describes the data and presents a preliminary analysis of consumer shopping behavior, our empirical strategy and discusses identification. Section 4 reports the estimation results and describes the role of shopping costs in predicting substitution patterns and supermarket markups and marginal costs by comparing the market equilibrium with shopping costs with the counterfactual in which shopping costs are set to zero. We then present a robustness analysis in which shopping patterns are defined in a different way, which gives rise to an alternative way of quantifying shopping costs. Section 5 presents and discusses the results of the applications on product delisting and loss-leader pricing. Finally, Section 6 concludes and discusses possible directions for further research.

7

2

The model

2.1

Consumer choice model

Consider a market in which there are I consumers indexed by i = 1, . . . , I with idiosyncratic valuations of grocery products indexed by k = 1, . . . , K. Suppose there are three store chains in the market indexed by r ∈ {A, B, C} that supply the same products to all consumers.8 Customer

i purchasing product k from store r in period t derives a net utility of v ikrt , which is a function of the price of the product and other characteristics.9 Consumers have unit demand for each product class and can purchase one, two or three products in the same period. Let B be the set of all exclusive and exhaustive baskets. Baskets with multiple products may be purchased from a single store (one-stop shopping) or from mul-

tiple stores (multistop shopping). A consumer favors multistop shopping if her shopping costs are sufficiently small, otherwise she will optimally make her purchases from a single store. In the formulation of the model, we distinguish between two important components of the total cost of shopping: first, a fixed shopping cost, denoted by si , which we assume is timeinvariant and independent of store characteristics (e.g., store size, facilities, location). It captures consumer i’s marginal (dis)utility from visiting an additional supermarket in a given shopping period. Given that it is constant across stores and periods, we interpret this cost as a pure measure of the individual’s taste for shopping activities. From now on, we will refer to this fixed component as “shopping costs”. Further, we assume that si is randomly drawn from a continuous distribution function G(·) and positive density g(·) everywhere. And second, transport costs, which we capture by including the distance to stores as an additive term in the utility function of a basket of products. Finally, we assume that consumers are well informed regarding prices and product characteristics. Therefore, consumers do not need to engage in a costly search to gather information about prices and product quality. Consumer i is supposed to exhibit optimal shopping behavior. This implies that she makes an optimal choice involving two elements: whether to be a one- or multistop shopper, and which stores to visit for each of the products she wants to buy. Roughly speaking, the choice set of consumer i will be restricted by the number of stores she can visit given her shopping costs, so that her choice will consist of selecting the mix of products and stores that maximize the overall value of the desired basket. In line with this, a three-stop shopper who can visit all stores and wants the three products will select the best product–store combination from the alternatives existing in the market within each category. A two-stop shopper will select the mix of two stores maximizing the utility of the desired basket from all possible product–store combinations. Her final basket will consist of the best of the two alternatives in each product category. Finally, a one-stop shopper will choose the store offering the largest overall value for the whole basket of 8 Assuming that all consumers have access to the same product range helps us to reduce dimensionality issues in the estimation of the model. An extension of the model would be to relax this assumption and allow for heterogeneous choice sets. 9 For now, we do not specify a functional form for the product-level utility, as it is not necessary for setting out the model. We will assume a parametric specification at the empirical implementation stage in Section 3.

8

products. Formally, let Dir for all r ∈ {A, B, C} denote the distance traveled by consumer i from his

household to store r0 s location, and τ denote a parameter that captures the consumer’s valuation of the physical and perceived costs of traveling that distance. We define the overall (i.e., basket level) utility net of transport costs of a shopper who is able to visit only one of the three stores in the market as follows:10 1 vit

= max

(K X

k=1

v ikAt − τ DiA ,

K X

k=1

v ikBt − τ DiB ,

K X

k=1

)

v ikCt − τ DiC .

(1)

Similarly, the overall net utility of a two-stop shopper is given by:

2 vit

= max

(K X

k=1 K X

k=1 K X k=1

max{v ikAt , v ikBt } − τ (DiA + DiB ) ,

max{v ikAt , v ikCt } − τ (DiA + DiC ),

(2)

)

max{v ikBt , v ikCt } − τ (DiB + DiC ) .

Finally, the overall net utility of a consumer who is able to visit all three stores is given by: 3 vit =

K X

k=1

max {v ikAt , v ikBt , v ikCt } −

X

τ Dir .

(3)

r∈{A,B,C}

Note that the expressions in (1), (2), and (3) are particular cases of a more general utility function, in which—conditional on shopping costs—an n-stop shopper selects the subset of stores that maximizes the overall utility of her desired basket. For a one-stop shopper, these subsets are singletons, for a two-stop shopper they contain two elements, and for a three-stop shopper each subset of stores contains precisely the number of stores in the market, which is why she does not need to maximize over subsets of supermarkets. We provide a general expression of the utility and choices of an n-stop shopper in Appendix A. 1 − s > 0 such that all consumers will visit at least one supermarket in each Suppose vit i

period. To determine the number of stops to be made, consumer i weighs the extra utility of

undertaking n-stop shopping with the extra costs involved, taking into account the fact that 2 ≡ v 2 − v 1 and the total cost of shopping increases with the number of stores visited. Let δit it it

3 ≡ v 3 − v 2 be the incremental utilities that consumer i derives from visiting, respectively, two δit it it

stores rather than one and three stores rather than two. A consumer may also consider visiting 3 −v 1 = δ 2 +δ 3 . either one or three stores, in which case her incremental utility will be given by vit it it it

Consumer i will optimally decide to undertake three-stop shopping only if the net utility

derived from visiting three stores is greater than that from either one- or two-stop shopping. 10

Note that the utilities below depend on the vector of all prices of products sold by the three stores in the market, which we denote by pt . However, we omit this for the sake of simplicity in the presentation of the model.

9

Formally: 3 2 1 vit − 3si > max{vit − 2si , vit − si }.

Rearranging, the optimal stopping rule for a three-stop shopper is given by: si 6 min

(

2 3 δit δit ,

3 + δit 2

)

(4)

.

Similarly, consumer i will optimally decide to undertake two-stop shopping if and only if: 2 1 3 vit − 2si > max{vit − si , vit − 3si }.

Hence, consumer i will undertake two-stop shopping as long as: 3 2 δit < si 6 δit .

(5)

Finally, consumer i will optimally decide to undertake one-stop shopping if and only if: 1 2 3 vit − si > max{vit − 2si , vit − 3si },

from which we can derive the optimal stopping rule for a one-stop shopper as follows: si > max

(

2 2 δit δit ,

3 + δit 2

)

.

(6)

In general, the optimal stopping rule for consumer i indicates that she will choose the mix of stores that maximizes her utility conditional on the extra shopping cost being at most the extra utility obtained from visiting additional stores. Note that equations (4), (5), and (6) imply that: 3 δit <

2 + δ3 δit 2 it < δit . 2

(7)

Therefore, the highest possible shopping costs for any consumer able to undertake multistop shopping at either two or three stores, respectively, in equilibrium are given by the following critical cutoff points: 2 s2it = δit ,

for two-stop shopping, and

3 s3it = δit ,

for three-stop shopping.

(8)

Note that these cutoff points depend on the period of purchase. The subscript t was added because it depends on utilities that may vary with time. The derived cutoffs for the distribution of shopping costs in (8) indicate that for given shopping costs, consumers only care about the marginal utility of visiting an additional store in making their final decision on how many stores to visit. Moreover, one-, two- and three-stop shopping patterns arise in equilibrium and will be

10

defined over the entire support of G(·) (see Figure 1). s3it

0

s2it

1 vit

s Three-stop

Two-stop

shoppers

One-stop

shoppers

shoppers

Figure 1: One-, two-, and three-stop shopping

2.2

Supply

Recall that we assume that all supermarkets supply the same product line to consumers consisting of k = 1, . . . , K products. Given that ex ante homogeneous products are ex post differentiated because they are sold by different supermarkets, let J = K ∗ 3 denote the total number of prod-

ucts existing in the market (K products sold by each of the three supermarkets in the market)

and Lr be the specific product line supplied by supermarket r. The profits of supermarket r are given by: Πrt =

X

k∈Lr

(pkrt − mckrt )Mkt qkrt (pt ),

where Mkt is the size of the market for product k in market t and qkrt is the market share of product k sold by supermarket r as defined by equation (14). We suppose that supermarkets compete in prices. Assuming that there exist a Nash-Bertrand equilibrium in pure strategies, and that the price vector that supports it has strictly positive entries, the price pkrt of some product k sold by r must satisfy the following first-order condition (FOC) derived from r’s profit maximization problem: qkrt (pt ) +

X

j∈Lr

(pjrt − mcjrt )

∂qjrt (pt ) = 0, for all r ∈ {A, B, C}. ∂pkrt

The FOCs yield a system of J equations that implies price-cost margins for each product in supermarket r’s product line. In order to write this system in matrix notation, let St be a J × J ∂q

t matrix containing demand responses to changes in retail prices, with entry S(j, k) = − ∂pjr krt 0

for j, k = 1, . . . , J and r0 ∈ {A, B, C}. Further, let Ωr be a matrix of dimension J × J with

jth entry equal to 1 if products j, k are in r’s product line and zero otherwise. The system of equations writes as: qt − (Ωr St )(pt − mct ) = 0, where qt , pt and mct are J-dimensional vectors of market shares, prices and marginal costs. Solving for the margins, we obtain: pt − mct = (Ωr St )−1 qt . 11

3

Empirical implementation

3.1 3.1.1

Data and preliminary descriptive evidence Overview of the data

Data on household purchases were obtained from the Kantar Worldpanel database. This is homescan data relating to grocery purchases made by a representative sample of 10,729 randomly selected households in France during 2005. These data were collected by household members using scanning devices. The data set contains information on 352 grocery product categories from approximately 90 grocery stores, including supermarket chains, hard discounters, and specialized stores. An entry in the data set records the purchase of a specific product from a given store on a particular date. Further, the data set includes information on household characteristics. In our data set, households are observed to purchase up to 275 different products from up to nine separate grocery retailers during a given shopping period. On average, a household purchases baskets containing 24 products from two stores each week. However, in order to overcome the dimensionality problem arising from such large baskets, we focus on a reduced set of three product categories and three supermarket chains that represent the most frequently purchased products and supermarkets most frequently visited by French households. In particular, we include yogurt, biscuits and refrigerated desserts, given that they meet several criteria that make our empirical exercise consistent with our structural model.11 First, they are staples, as most French households are heavy consumers of products from these categories and are not stored for long periods, so stockpiling is not a first order concern. Second, these categories are not close substitutes, which ensures that we can observe sufficient variation in shopping patterns, as consumers may tend to concentrate their purchases from the same category in a particular store but may wish to diversify purchases from other categories across supermarkets. Third, the expenditure on these three categories only corresponds to 8.5% of all grocery product expenditures in our sample. Finally, customers tend to consume one serving of a product from these categories at a time, which makes it convenient for a demand model that relies on a unit demand assumption (see Table 1). We supplement the homescan data with information on supermarket characteristics from the Atlas LSA 2005. This includes information by store format and type (regular and hard-discount stores) on aspects including the store’s location, sales area, number of checkouts, and number of parking bays. We merge both data sets using the name of the retailer and its zip code, the zip code of the consumer’s residence and the floor area of the outlet. We follow Dubois and Jódar-Rosell (2010) and compute distances between a household residence and a store by using zip codes. Using this distance, we then determine the outlet of each supermarket chain that is closest to the consumer’s dwelling and include only one outlet per retailer (in case there are 11

Our results are robust to the set of products selected. In Section 4.3, we present the results of the estimation of our demand model with an alternative set of product categories. Moreover, we discuss the potential biases from having a limited set of products in the subsection 3.3.

12

Table 1: Characteristics of the selected categories

Position among 352 productsa Serving size (in grams)b Mean price (euro cents/serving) Mean consumptionc Average number of days between purchases

Yogurt

Biscuits

Desserts

2 125 26.27 9 8

3 30 9.77 12 10

6 80 45.42 8 10

Source: Kantar Worldpanel database 2005. Authors’ calculations. Notes: a Positions of the selected products in a ranking of the 352 products we observe, based on the number of purchases in 2005. b Servings are defined according to the most frequently purchased serving of each product. c Average number of servings purchased per household-week.

several stores meeting this criterion); further, we keep in the consumer’s choice set stores located within 20 kilometers of the household location. 3.1.2

Choice set

Our demand model allows shoppers to buy several different products in the same week, and assumes that shoppers are making a series of multiple-discrete decisions regarding which products to buy as part of a desired basket of products from a set of mutually exclusive and exhaustive alternatives. This choice set includes baskets of either one product or multiple products that can be purchased from either one store or several stores. Concerning products, we allow for two mutually exclusive alternatives in the yogurt category, one being the leading NB in France in 2005 and the other being a composite “brand” that consists of the remaining varieties (both other NBs and PLs) available in the market. This will be useful to capture consumers’ responses when the price of a particular product changes considerably, which is one of our goals of our applications section. As for the other two categories, in each case we treat the purchase of all brands as if they were purchases of a single general brand. Therefore, consumers face a set of four products from which they can choose at most three: one of the two yogurt alternatives, biscuits and desserts.12 A caveat related to our empirical definition of products is that by aggregating individual brands into composite products (categories), we are forcing consumers to be more price sensitive between supermarkets but less price sensitive within a supermarket, as long as we are reducing, and for some products eliminating, the set of substitutes available within the supermarket. Consider, for example, an increase in the price of either biscuits or desserts at, say, Supermarket 1, which by construction do not have any intra-store substitute. If the price of all other products remain constant, the total price of the baskets containing that product increases, which may 12

The choice of yogurt as the category with two alternatives is arbitrary. Our results are robust to the selection of the product category containing two options.

13

cause that one-stop shoppers switch supermarkets in order to find their desired basket at a cheaper price; and, on the other hand, that multistop shoppers look for that specific product at rival supermarkets. As a consequence, Supermarket 1 experiences a larger lose in demand than in a situation in which consumers have substitutes available at their preferred store. Implicitly, this implies that shopping costs act as a more important incentive to switch supermarkets than it should be. In practice, the fact that consumers are, on average, more price sensitive at the store level would be captured by the price coefficient (it would be higher) which will make the estimate of shopping costs lower in monetary terms. Therefore, we are estimating a lower bound of the true shopping cost. Regarding stores, we restrict our attention to the three leading supermarket chains in France based on national market share in 2005. These chains are present throughout the country and account for nearly 60% of groceries sales made by store chains in France (excluding hard discounters) in 2005. The remaining stores observed in our data are included in the outside option along with the no purchase of the included goods option (the interpretation of the outside good in this context is discussed below). Thus, we have four products that are available at three stores of similar size. This is consistent with our modeling framework of oligopolistic competition with differentiated product lines, where customers can visit multiple stores in the same shopping period to increase variety. A concern in the empirical implementation of our model is that being a discrete choice model, it does not allow customers to split the demand of products in two competing supermarkets. However, real life consumers can be observed to purchase different brands of the same product categories in different supermarkets within the same week. In our data set, the instances in which we observe customers splitting demands correspond to 4% of the total number of observations. We deal with this by eliminating those instances meaning that we restrict attention to those cases in which customers concentrate purchases of products in the same category in a particular supermarket within a week. In this context, a basket is a collection of product–store items, and given that there are four products, three stores and baskets that can consist of one, two or three items, we end up with a choice set of 112 mutually exclusive alternatives. When a household is observed to have made no purchases within the choice set in a given shopping period, we interpret it as having opted for the outside option. 3.1.3

The sample of households: composition and descriptive statistics

The selection of products and supermarkets discussed above imply an implicit selection of all of the households who registered purchases of the included products in at least one of the included supermarket chains. From that set, we randomly sample 518 households, which represent about 5% of the total number of households in the original data set.13 13

We give summary statistics on household characteristics and describe some shopping patterns of the full set of households in the original database in our Online Appendix.

14

Table 2 gives summary statistics for demographic characteristics of French households observed in our sample. The average household consists of three members, among whom there is a household head (the person primarily in charge of the household’s grocery shopping). On average, the age of the household head is 48 years old and the household income is 2,374 e per month. About half of the households in our sample reported having internet access at home, which may partially explain why internet purchases are not important in our data. As for storage capacity, 78% of households have storage rooms at home and 69% an independent freezer in addition to a refrigerator. Further, about 36% of households reported producing vegetables at home which, along with the fact that 27% of the households are located outside urban areas, can be a reason for the observed low frequency of shopping of some households. Table 2: Summary statistics for household characteristics Variable

Mean

Median

Sd

Min

Max

Demographics Household size Income (e/month) Expenditure on groceries (e/month) Expenditure on the included categories (e/month) Baby (=1 if yes) Household head’s age Lives in urban areas (=1 if yes) Car (=1 if yes) Internet access at home (=1 if yes)

3.08 2,374 261 27.05 0.20 48.24 0.73 1.54 0.55

3 2,100 245 23.24 0 45 1 1 1

1.37 1,162 131 18.00 0.40 14.28 0.44 0.80 0.50

1 150 0.51 0.25 0 22 0 0 0

7 7,000 948 128 1 91 1 6 1

Storage capacity Independent freezer Freezer capacity > 150L Storage room at home Home production

0.69 0.58 0.78 0.36

1 1 1 0

0.46 0.49 0.41 0.48

0 0 0 0

1 1 1 1

Source: Kantar Worldpanel database 2005. Authors’ calculations.

Table 3 displays details of consumer shopping patterns. On average, households tend to favor multistop shopping. The average French household visits more than one grocery store in a week most of the times over the year and tends to make between one and two trips per week to the same store. The average number of days between shopping visits is nearly six days. The preferred store type remains the regular supermarket over the hard discount store: only a 20.6% of weekly visits to grocery stores are to the latter. Larger store formats are preferred by consumers: on average, the two most frequently visited store formats are supermarkets and hypermarkets with 55.2% and 40.8% share on total visits per week, respectively. Smaller store formats (which supply a reduced product range generally at higher prices) receive the lowest number of visitors per week with 3.5%. Although these stores have the advantage of being within walking distance of a household’s location—as opposed to hypermarkets that are located 15

far from city centers—the preference for larger store formats can be explained by several factors, such as bulk shopping, lower prices, more intensive sales and promotional activities, and a larger product range. Table 3: Summary statistics for household shopping patterns Variable

Mean

Median

Sd

Min

Max

No. trips to same grocery store/week No. separate grocery stores visited/week Average number of days between visits Visits to Hard discounters (% of total/week)

1.40 1.65 5.48 20.64

1 1 4 0

0.78 0.81 6.07 40.47

1 1 1 0

7 6 131 100

Visits by format (% of total/week) Hypermarket Supermarket Convenience

40.76 55.16 3.45

31.89 59.78 0

34.63 34.59 11.35

0 0 0

100 100 89.47

Source: Kantar Worldpanel database 2005. Authors’ calculations.

3.1.4

Shopping period and shopper types

Consistent with this evidence, we define a shopping period as a week in which a household is recorded as making grocery purchases. During this period, a household that concentrates its purchases at a single supermarket chain can either make one visit to a store in the chain, make several visits to the same store or can visit several stores in the same chain. Regardless, we define this household as a one-stop shopper as long as it is observed to deal only with a single supermarket chain during the week. Conversely, when we see the household purchasing products from stores in competing chains during the week, we define it as a multistop shopper. 3.1.5

Preliminary evidence on the nature of multistop shopping

Scanner data on supermarket purchases by households is particularly well suited to the investigation of multistop shopping behavior. Supermarket chains with similar characteristics (size, variety of store formats, products and brands carried, and location) often offer similar product ranges to consumers who could opt for concentrating purchases with their preferred supermarket chain. Yet, we observe multistop shopping as a widespread pattern among households in our data. The right panel of Figure 2 shows the distribution of households by the average number of supermarkets visited within a week. Of the total number of households we observe, approximately 47% visit more than one supermarket per week, on average. Theory suggests that this observed heterogeneity in shopping patterns can be rationalized by the fact that for some consumers it is more costly to visit multiple supermarkets (see, for example, Klemperer (1992), Klemperer and Padilla (1997), Armstrong and Vickers (2010), and Chen and Rey (2012,

16

2019)).On the other hand, the left panel of Figure 2 shows the average share of weekly expenditure made by a household on the supermarket it spends most of its weekly budget. Interestingly, all multistop shoppers spend more than 50% of their budget on a particular supermarket, suggesting that there is, on average, a shopping location which is more important than the others.

Figure 2: Distribution of households and share of expenditure in the by average number of stores visited per week in 2005 Notes: The figure on the left shows the percentage of households by the number of shopping stops (i.e., different supermarket chains visited) they make in a week. The figure on the right shows the average share of weekly expenditure made by a household on the supermarket it spends most of its weekly budget, by the number of shopping stops made by that household in a week. Source: Kantar Worldpanel database 2005. Authors’ calculations.

To empirically check whether time constraints and other household characteristics determine the number of shopping stops to be made, we regress the number of separate supermarkets visited in a week on a set of household characteristics that proxy for households’ time constraints. We expect to find that more time constrained consumers (i.e., with a higher opportunity cost of time) tend to visit fewer supermarkets than less time constrained ones. We add some controls for household storage capacity (housing type, presence of a storage room and/or independent freezer, and the size of the largest freezer) to help rationalize—at least in part—the frequency of shopping. Further, we include supermarket, region and week dummies in all regressions. Table 4 shows the results. Of the coefficients which are statistically significant, the coefficients for the average distance between a household’s location and a store, those of the store format dummies (supermarket and hypermarket), that of car holdings and the one of the dummy for living in an apartment (as opposed to a house) are consistent with the geographic configuration of the French grocery store sector. In France, as a result of zoning regulations that limit store size depending on the zone of the city, larger store formats must be located farther away from the city centers. Accordingly, stores can be categorized into three formats depending on size and location: hypermarkets are the stores with the largest sales areas (> 2500 m2 ) and product range, are often only reachable by car and are located at a considerable distance from rival stores. Supermarkets are mediumsized stores (with sales areas between 400 m2 and 2500 m2 ) with a fairly varied assortment of 17

Table 4: Number of stores visited per week OLS

Poisson

Variable

Coefficient

Std. errors

Coefficient

Std. errors

Average distance to stores Hypermarket (=1 if yes) Supermarket (=1 if yes) Hard discount (=1 if yes) Household head’s age Log of Income Household size Car (=1 if yes) Lives in urban areas (=1 if yes) Lives in an apartment (=1 if yes) Lives in a farm (=1 if yes) Baby (=1 if yes) Internet access at home (=1 if yes) Home production (=1 if yes) Constant

0.041 -0.263 -0.193 0.089 0.005 0.013 0.032 -0.038 0.031 0.101 0.037 0.049 -0.015 0.084 0.316

0.01 0.085 0.072 0.115 0.002 0.042 0.020 0.104 0.052 0.055 0.151 0.058 0.043 0.047 0.232

0.066 -0.237 -0.154 0.098 0.005 0.012 0.027 -0.034 -0.007 0.100 0.037 0.047 -0.018 0.096 ——

0.012 0.076 0.059 0.109 0.002 0.043 0.020 0.100 0.051 0.054 0.158 0.059 0.043 0.047 ——

R2 Observations

0.132 453,725

—— 453,725

Source: Kantar Worldpanel database 2005. Authors’ calculations. Notes: The dependent variable is the number of separate supermarkets’ stores visited by household members within a week. Standard errors are clustered by household. The “average distance to stores” is computed as the average across distances between a household’s location and each of the stores visited by its members in a week. “Home production” indicates whether a household grows vegetables at home or not. All specifications include controls for household storage capacity and store, region and week fixed effects. The Poisson coefficients are the marginal effects of a Poisson model.

products, and are generally situated closer to the city center than hypermarkets. Smaller stores (with sales areas of less than 400 m2 ), are predominantly located in downtown areas and are easily reachable, but often offer a limited range of products (mostly staples), making them more suitable for top-up trips. In line with this, a positive coefficient for average distance suggests that households tend to make a combination of visits to bigger stores in order to make bulk shopping and to smaller stores in order to make top up trips. The negative coefficients of the store format dummies (supermarket and hypermarket) suggest that those customers who go to bigger stores tend to make fewer stops compared to those who visit convenience stores. This is consistent with economic intuition: given that hypermarkets and supermarkets are larger than convenience stores and carry a larger product range, consumers who source them need to make fewer shopping trips than those patronizing convenience stores because the larger stores make bulk shopping possible. In terms of housing type, our results suggest that people living in apartments—which are more likely to be in or closer to downtown areas—tend to visit a larger number of stores than those who live in houses. An alternative interpretation that is consistent with this result is

18

related to household storage capacity. Households with a lower storage capacity (those living in apartments) need to visit shops more often, and are thus more likely to be multistop shoppers. Finally, older people tend to make more shopping stops, presumably because they have more time available or a lower opportunity cost of time, which is consistent with previous evidence on the use of time (see Aguiar and Hurst (2007)). Further, we explore the relationship between the expenditure a household makes on a basket of products in a given period and the number of different supermarkets visited by that household in a week.14 We report the results in Table 5. Our estimates suggest that multistop shopping allows households to obtain the products they look for at lower prices, on average. Visiting an additional supermarket does not have a monotone impact on the expenditure. While visiting two different supermarkets in the same week decreases the expenditure by nearly 6%, visiting three different supermarkets decreases it by more than 30% (= 1 − exp(−0.438)). Therefore,

there is scope for savings when visiting multiple supermarkets.15

Table 5: Expenditure savings from multi-stop shopping Variable

Coefficient

Distance

-0.000131 [-0.0002, 0.00002]

Dummy for two supermarkets visited in a week (=1 if yes)

-0.066 [-0.0853, -0.0471]

Dummy for three supermarkets visited in a week (=1 if yes)

Time fixed effects Product fixed effects Retailer fixed effects

-0.438 [-0.645, -0.231] Yes Yes Yes

Observations R-squared

149,828 0.867

Source: Kantar Worldpanel database 2005. Authors’ calculations. Notes: The dependent variable is the log of expenditure in a basket that consists of one serving of each of the products included in our sample (i.e., yogurt, biscuits and refrigerated desserts). Standard errors are clustered by household. The “average distance to stores” is computed as the average across distances between a household’s location and each of the stores visited by its members in a week. The model includes a constant term.

14

We thank an anonymous referee for suggesting this point to us. In addition, we performed a regression of the number of stops made by a household on the log of the potential savings that the household could have made by visiting the lowest priced stores for each product. The potential savings are computed by comparing the expenditure of the basket actually purchased by the household with the same basket consisting of the lowest priced products available on the market for that household. We find further evidence of the negative relationship between multistop shopping and weekly expenditure, which indicates that the number of supermarkets visited increases with the potential savings. We report the results in Table B.1 of the Online Appendix. 15

19

3.2 3.2.1

Empirical specification of consumer demand Product-level demand

We empirically specify product-level utility as a function of observed and unobserved product and store characteristics, and time fixed effects. We allow consumer heterogeneity to enter the model through the price coefficient, which is a function of observed and unobserved household characteristics. Formally, let the utility of consumer i from purchasing product k from store r at time t be given by: v ikrt = −αi pkrt + xkr β + ξkr + φt ,

(9)

where pkrt is the price of product k at store r, xkr is a vector of observed product-store characteristics, ξkr captures the mean valuation of unobserved product-store characteristics, which we capture by including product-store dummy variables, and φt are time fixed-effects. Finally, αi is an individual-specific coefficient that captures the marginal utility of price and β is a vector of parameters common to all households. We model the distribution of the price coefficients as a function of observed and unobserved demographic characteristics as follows: αi = α + di π α + σ α νi ,

νi ∼ N (0, 1),

(10)

where α captures the mean (across consumers) marginal utility of the price of product k sold by r at t, di is a vector of observed household characteristics, π α is a vector of coefficients that measure how the taste for price change with household characteristics, νi is a random variable that captures unobserved household attributes that influence consumer choices and σ α is a scaling parameter. Further, we assume that individual shopping costs are a parametric function of a fixed shopping cost which is common across all consumers, ς, and can be thought of as the average disutility that every consumer bears as a result of the need to engage in shopping, and an idiosyncratic deviation from this mean that consists of observed, di , as well as unobserved household characteristics, ηi , which rationalize the observed heterogeneity in shopping patterns across individuals. This yields: si = ς + di π s + σ s ηi ,

ηi ∼ N (0, 1),

(11)

where π s is a vector of coefficients measuring the change in costs with household characteristics and σ s is a scaling parameter. In line with our modeling framework, we empirically define the utility that a n-stop shopper (n ∈ {1, 2, 3}) derives from purchasing a basket b ∈ B as: uibt =

X

kr∈b

v ikrt − τ Dir − si nt + εibt ,

n − si nt + εibt , = vibt

20

(12)

n is the overall utility of basket b net of transport costs as defined by equations (1) where vibt

through (3) above, nt is the number of stores visited in period t, si is the individual shopping cost, and εibt is an idiosyncratic basket-level shock to utility. Note that equation (12), along with equations (9) through (11), fully specify the utilities of one- and multistop shoppers as a function of price, product characteristics, distance to stores, and individual shopping costs. Thus, our utility accounts for both the vertical and horizontal dimensions of consumers’ valuations of products. The vertical component is captured by product-store characteristics, while the horizontal component is captured by distance, which varies across store formats and zip codes, and shopping costs. Finally, we normalize the utility net of shopping costs of the basket containing only excluded products (which we denote as basket b = O) to zero. Thus, it is modeled as a function of an individual random shock to utility, uiOt = εiOt . A consumer who wishes to buy a basket b at time t faces a choice set B of mutually exclusive

and exhaustive alternatives consisting of combinations of products and stores. The basket she chooses is such that she obtains the highest possible utility net of shopping costs. This is, for all b0 ∈ B, consumer i chooses basket b at time t if: uibt > uib0 t , ∀ b0 6= b. Let θ = (α, β 0 , π α , π s , σ α , σ s , ς, τ )0 be a vector containing the parameters to be estimated. 0

0

We assume that the random shocks to utility, εibt , are distributed i.i.d. type I extreme value. Integrating over εibt yields the closed-form choice probability of basket b, at time t, as a function of the characteristics of products and supermarkets: qibt (X, pt , di , νi , ηi ; θ) =

1+

n − ns ) exp(vibt i , m b0 ∈B exp(vib0 t − msi )

P

(13)

where X is the matrix of characteristics of all of the products, pt is the vector of prices of all of the products and n and m correspond to the number of supermarkets visited to purchase baskets b and b0 , respectively. 3.2.2

Aggregate demand and demand elasticities

The market level demand for product k supplied by supermarket r (product kr) at time t is obtained by aggregating individual level demands for all of the baskets containing product kr. Let the subset of all baskets containing kr be denoted by Bkr , the individual demand for product

kr is:

qikrt (X, pt , di , νi , ηi ; θ) =

X

qilt .

l∈Bkr

The market level demand for product kr at time t is obtained by aggregating individual level

21

demands for all of the baskets containing product kr as follows: qkrt (X, pt , di , νi , ηi ; θ) =

Z

qikrt dF (d)dF (ν)dF (η).

(14)

Further, let us define the individual level demand for the pair of products kr and jr0 is given by: qikr(jr0 )t (X, pt , di , νi , ηi ; θ) =

X

qilt .

l∈Bkr(jr0 )

The own- and cross-price elasticities of demand defined by equation (14), for all r0 ∈

{A, B, C}, are given by:   

3.3

∂qjr0 t pkrt = ∂pkrt qjr0 t  

pkrt qjr0 t

R

krt − pqkrt

R



αi qikrt 1 − qikrt dF (d)dF (ν)dF (η), if j = k, r = r0 , 

αi qikrt qijr0 t − qikr(jr0 )t dF (d)dF (ν)dF (η), if j 6= k, r 6= r0 .

The outside option

(15)

In our context, the interpretation of the outside option differs from that of a standard discretechoice model of demand for a single product. As pointed out by Gentzkow (2007), in a model that allows consumers to choose multiple products simultaneously, every choice involves a maximization over all excluded alternatives, unlike the case of a standard multinomial model, where only the utility from good ‘zero’ is implicitly maximized over all excluded products. We follow Gentzkow’s approach and restrict the choice set to all baskets that result from the combination of up to three stores and up to three products from the four alternatives available in order to keep our problem involving multiproduct and multistore choices tractable. The remaining observed purchases are treated as part of an outside good. Accordingly, we define our outside good as all of the observed purchases of both included and excluded products at excluded supermarkets and excluded products at included stores, as well as all of the unobserved purchases and visits to unobserved grocery stores. We discuss the implications of this definition for our estimates in what follows. To fix ideas, consider a household that registers purchases of biscuits from one of the included supermarkets in a given week. If this is the only grocery shopping activity observed for that household in that week, we would interpret that this household was a one-stop shopper that has purchased a basket of one product from a single shopping location, and conclude—according to our structural model—that its overall utility net of shopping costs is larger than that it would have derived had through any other shopping pattern. However, it may be the case that this household has purchased yogurt and/or some excluded products at excluded stores, or that it has also purchased some excluded products from the included stores. In any case, our interpretation is that this household is better off choosing that basket involving a single stop at our included supermarkets, and possibly an outside option, rather than choosing a basket obtained from any combination of additional products and multiple supermarkets among the ones we included in our analysis. 22

In line with this, we must include a caveat related to the interpretation of the shopping patterns that we observe in our sample. If a consumer is observed to have made purchases at two of the three included stores, we interpret this to mean that she was able to make two visits in addition to the visits she may have made to any excluded stores. In this sense, the number of supermarket visits by a household that possibly also purchased products at excluded stores is interpreted in the context of this paper as the number of additional visits the household has made to the included supermarkets. We check the robustness of our estimates to the definition of the outside option by restricting our sample to the consumers who were observed to make purchases only from the three supermarkets included in our main analysis, which implies that we are not excluding any observed stops. In this case, we define the outside good as all of the purchases made from one of the three included supermarkets. Our results are robust to this alternative definition. See Section 4.3 for details and results.

3.4

Identification

Our general identification strategy follows the standard literature that estimates the demand of differentiated products. In particular, it follows Nevo (2001) and exploits the panel structure of the data to identify the parameters of the demand model independent of the supply side. Given that the data is at the transaction level, we further exploit the rich observed variation across individuals in key variables, such as prices, the number of visited supermarkets and the baskets chosen by individuals for the identification of the demand parameters. There are, however, two challenges to identification. The first is related to the potential endogeneity of prices, which is common to the literature of demand estimation and stems from the potential correlation of prices with unobserved product characteristics or shocks to demand, which are common to all consumers in a local market. The second challenge concerns the identification of the fixed component of individual shopping costs separately from both other components of shopping costs that vary with time and across supermarkets and other fixed components of the utility, such as the mean valuation of unobserved product and supermarket characteristics and aggregate shocks to demand. In what follows, we discuss the way in which we deal with each of this challenges in turn. A common concern in the identification of the price parameters is that firms often adjust prices in response to changes in local market preferences for product characteristics that are unobserved to the econometrician (Berry, Levinsohn and Pakes, 1995; Nevo, 2001). To correct for unmeasured product characteristics, we include product-supermarket dummy variables. Further, to control for aggregate shocks to product-supermarket demand that may affect preferences over time, we include time dummy variables. Endogeneity would still be a concern if prices were correlated with the basket-level idiosyncratic shock to demand, εibt . This would arise, for example, in situations in which supermarkets can set price schedules that affect differently one- and multistop shoppers (e.g., the total price of a basket of three products from the same supermarket may be lower than that resulting from purchasing the same basket of products

23

from different supermarkets). Given that our controls may not fully account for all of the sources of variation in prices, we estimate our demand model by including a control function for prices. In a first stage, we use Hausman instruments (Hausman (1997)), which we define as the price of the same product at the same supermarket in other local markets (i.e., geographic regions) of France, excluding the region in which the price to be instrumented was observed. The identifying assumption is, therefore, that after controlling for market-level aggregate shocks and unobserved product and supermarket characteristics, the prices at other local markets contain information of the prices observed in a particular region, but are not correlated with the basketlevel idiosyncratic shocks to demand of that region. Turning to the identification of the the fixed shopping costs, our key strategy for identifying the fixed component of the individual shopping costs is to exploit the observed, both cross-section and time-series, variation in individual shopping patterns; this is, the number of supermarkets visited by the household members each week. However, a concern with this variable is related to the fact that it may be capturing multiple factors that simultaneously determine consumer shopping patterns, such as other components of shopping costs that vary with supermarket characteristics and time, consumer preferences for unobserved product and supermarket characteristics, and market-level demand shocks. To control for unobserved factors that affect the choice of products and supermarkets and, thus, shopping patterns, we include product-supermarket dummies, which capture, among other things, supermarket-level factors that may make more or less difficult to purchase the products that a customer is looking for (e.g., the average store size, the organization of the products in the store and shelves, the average number of checkouts), and time dummy variables, which capture local market shocks that may make more or less costly to do shopping activities. Further, in order to account for the fact that some households may be more or less time constrained than others due to idiosyncratic factors, we allow household observed and unobserved demographics to enter the model as interactions with the number of supermarkets visited. Further, to account for transport costs (e.g., driving to the supermarket, parking the car and the opportunity cost of time associated to these activities), we include the distance to supermarkets from the household’s location. Due to the fact that we do not observe the exact address of the households in our sample, but their zip code only, we follow Dubois and Jódar-Rosell (2010) and assume that the distance between household location and nearby stores is the same for all households with the same zip code. The inclusion of distances to stores is also useful to capture part of the horizontal dimension of consumers’ preferences for product characteristics. The key identifying assumption is, therefore, that after controlling for all of these factors, along with the unit demand assumption, the observed differential variation in household-level shopping patterns allows us to isolate the parameters of the component of consumer shopping costs that is constant across supermarkets and time periods, which is consistent with our modeling framework. A threat to identification would be if supermarkets offered specific promotions and rebates to individual consumers to compensate their shopping costs. However, there is no evidence, to the best of our knowledge, of supermarkets offering either individual-specific pro24

motional schedules or personalized pricing to customers. Therefore, after controlling for both individual and aggregate unobserved factors, the number of supermarkets visited by a consumer is unlikely to be correlated with the basket-level individual specific shock to demand. Finally, equation (8) shows that we can identify critical cutoff points of the distribution of shopping costs if we are able to both observe the optimal shopping patterns of one- and multistop shoppers and identify the parameters of the product-specific utilities involved in the computation of the nth cutoff point. For each individual, we need to be able to compute both the utility of her actual choice and the utility she would have derived had she chosen any other basket. This exercise involves a challenge as long as we do not observe neither the true set of alternatives available for each consumer at the moment of a purchase nor the prices of those alternatives. We deal with this challenge as follows: first, we assume that all the products included in our analysis were available for all consumers at each shopping occasion. Therefore, we supplement the observed transaction made by each household in each week with the remaining alternatives. And second, we replace the unobserved price of each product with the average price of that product in the same local market. If we do not observe the purchase of that product in a given market, then we replace the price of that product by the average price of the same product in other markets in the same period.

3.5

Estimation

To estimate the parameters of our model, we use the data set described in Section 3.1. Given that we allow for random coefficients of price and shopping costs, our choice probabilities do not have a closed-form solution. Thus, we use simulated methods to estimate them. In our data, we observe consumers choosing a basket of products (which may be the outside option) at each period t during a number of T periods. Let H index the set of all possible sequences of choices our data takes; that is, all basket sequences at all choice occasions during our period of observation. The probability of observing consumer i making a sequence of choices h ∈ H is given by:

Pih (X, di , νi , ηi ; θ) =

T Y

qibt (X, pt , di , νi , ηi ; θ).

(16)

t=1

Let h be the vector of observed choices of each consumer, a natural way to estimate θ would be to maximize the log-likelihood function: L(X, h; θ) =

X i

ln

Z

Pih (X, di , νi , ηi ; θ)dF (ν)dF (η).

(17)

However, the integral over unobserved household characteristics, νi , ηi , in equation (17) does not have a closed-form solution. Therefore, we use the Simulated Maximum Likelihood (SML) method (see Lerman and Manski (1981), and Pakes and Pollard (1989)) in order to overcome this problem. As the SML method requires the number of simulation draws, S, to approach infinity with

p

S/I = O(1), we use 100 draws in our simulations. The SML estimator is given

25

by: θˆSM L = arg max θ

4

(

X i

#)

"

1X Pih (X, di , νis , ηis ; θ) ln S s

.

Results

4.1

Empirical estimates

We report estimates of the utility parameters, distance and shopping costs according to two specifications in Table 6. Column two includes a control function for price. The two regressions include product, supermarket and time fixed-effects. Results are as expected: demands are downward sloping with statistically significant estimates for price in all of the specifications. The estimate of the mean valuation of distance is negative and statistically significant, which means that the mean valuation of a basket of products is lower the greater the distance to a store from a customer’s location. The estimate for the number of supermarkets visited in a week is negative and statistically significant, which we interpret as consumers facing, on average, a positive fixed cost of dealing with multiple stores. Moreover, the estimates for the standard deviations of the coefficients of price and number of visits are significant and larger than one in both cases. This indicates that unobserved household characteristics are important in order to explain observed heterogeneity in household choices and shopping patterns. Our preferred specification is the model with a control function for price (second column of Table 6). We use the results of this specification to obtain measures in euros of the average shopping and transport costs by dividing the estimated coefficient of the number of supermarkets visited in the first case, and distance, in the second case, by the estimate of the price coefficient. We report the results in Table 7. The average fixed shopping cost a consumer incurs when sourcing a supermarket is 96 euro cents. The transport cost per store, taking the average distance of 7.2 km between consumer’s dwelling and a store, is 8.6 euro cents per trip. Summing up, the total shopping costs (fixed shopping costs plus transport costs) incurred by the consumer in visiting a store is, on average, 104.7 euro cents. This cost is 2.87% of the mean total expenditure on groceries made by a household at one supermarket in a week. Similarly, we translate into euros our measures of the shopping cost cutoffs, computed as differences in utilities between one-, two- and three-stop shopping according to equation 8 of our model and calculate the predicted distribution of shoppers in each category. We report the results in Table 8. The threshold between zero and one stops is 1.8 euros. Customers with shopping costs beyond that threshold find it more costly to visit one store in a given week in relation to its benefits, which prevent them from grocery shopping every week. This corresponds to a proportion of shoppers of 61.8% as predicted by our model. One-stop shoppers are those whose shopping costs lie between −3.4 and 1.8 euros, and comprise 37.1% of all customers.

Two-stop shoppers are those whose shopping costs lie between −3.4 and −6.5 euros, and are 1%

of the total number of shoppers. Finally, we estimate the maximum shopping costs necessary to 26

Table 6: Demand estimates Covariate Mean coefficients Price (e/basket) No. of visited stores Distance (km) Price × Household head’s age Price × Car Price × Log of income Price × Household size No. of visited stores × Household head’s age No. of visited stores × Car No. of visited stores × Log of income No. of visited stores × Household size Standard deviations Price No. of visited stores

Observations

Uncorrected

With control function

-1.547 [-0.968, -2.002] -4.987 [-4.656, -5.320] -0.062 [-0.059, -0.066] -0.019 [-0.034, -0.003] -0.519 [-1.149, 0.090] -0.072 [-0.462, 0.325] 0.582 [0.367, 0.747] 0.033 [0.023, 0.044] 0.906 [0.444, 1.284] 0.364 [0.113, 0.658] -0.047 [-0.153 0.078]

-5.242 [-2.549, -7.381] -5.035 [-4.709, -5.353] -0.063 [-0.059, -0.067] -0.021 [-0.037, -0.004] -0.667 [-1345, -0.022] -0.032 [-0.407, 0.352] 0.642 [0.430, 0.824] 0.037 [0.028, 0.050] 0.907 [0.389, 1.318] 0.336 [0.106, 0.611] -0.101 [-0.207, 0.003]

2.139 [1.966, 2.322] 1.833 [1.704, 1.956]

2.202 [1.989, 2.349] 1.838 [1.712, 1.957]

1,483,104

1,483,104

Notes: The 95% confidence intervals are given in square brackets. All regressions include product, store and month fixed effects. A basket can consist of unit servings of one, two or three products: desserts (80 g), biscuits (30 g) and yogurts (125 g).

rationalize the lower proportion of three-stop shoppers (which is just 0.02%) as negative: -6.5 euros. Consistent with our model, we interpret a negative shopping cost as a measure of a strong taste for shopping. Elasticities We report the averages of the estimated own- and cross-price elasticities in Table D.1 of the Online Appendix. Own-price elasticities are negative and most are larger than −1. Cross-price

elasticities intra-store are positive for products in the same category (yogurts) and negative otherwise. We interpret this as complementarity between categories in the same store. Further, cross-price elasticities are positive for the same product across stores. When the price of a

27

Table 7: Mean fixed shopping cost, mean distance cost and total shopping cost at the supermarket level Euro cents

% of mean weekly expenditure

Mean shopping cost

96.06

2.63

Mean transport cost

8.64

0.24

104.70

2.87

Mean total shopping cost

Notes: The mean shopping cost is obtained by dividing the estimate of the number of visited stores by the the mean price coefficient. The mean transport cost is obtained by multiplying the transport cost per kilometer of 1.2 euro cents, times the average distance to a supermarket, 7.2 km. The percentages shown in the second column are computed with respect to the average weekly expenditure of a household on groceries per supermarket, which is 37 euros.

Table 8: Average implied shopping cost cutoffs (across periods and consumers, in euros) and distribution of shoppers (in percentages) Number of stops

Mean shopping cost cut-offs (Euros)

Distribution of shoppers (%)

Zero-stop shopping

——-

One-stop shopping

1.810 [0.547,2.803] -3.399 [-4.863, -1.851] -6.545 [-8.061,-4.914]

61.8 [59.5, 64.3] 37.1 [34.5, 39.3] 1.0 [0.7, 1.22] 0.02 [0.01, 0.03]

Two-stop shopping Three- stop shopping

Notes: Mean shopping cost cutoffs are computed using demand estimates and equation 8 of the model presented in Section 2. To translate estimates into euros, we divide the absolute value of each coefficient by the absolute value of the mean price coefficient. The distribution of shoppers is computed using the probabilities of engaging in one-, two- or three-stop shopping predicted by our structural model. The 95% confidence intervals are given in square brackets.

product rises in a given supermarket, the demand for products from different categories within that supermarket decreases, whereas the demand for all products in competing stores increases. This result is a consequence of two aspects that we observe in our data, namely, the large number of consumers who purchase products from the three included categories we consider in the same week, and the large proportion of one-stop shoppers. According to our framework, and due to the fact that these consumers find it very costly to make an additional stop to purchase only one product, they prefer substituting stores instead and buy the entire basket at a competing location where its total value is lower.

28

4.2

The role of shopping costs

Do shopping costs matter when it comes to explaining consumer shopping behavior? We answer this question with the help of two exercises. On the one hand, we take our estimated model and simulate a scenario in which shopping costs fall to zero for all consumers. On the other hand, we estimate an alternative specification that does not include any shopping costs, and compare the results with those obtained from our preferred specification with shopping costs. In the first exercise, we compare the predicted probabilities of being a zero-, one- or a multistop shopper obtained from our estimated model (baseline) with a counterfactual scenario in which all consumers bear zero fixed shopping costs (keeping transport costs unchanged). The results are reported in Table 9. In the absence of shopping costs, consumers choose the outside option with smaller probability. Shopping costs introduce frictions that deter consumers from purchasing the included products on a regular basis, which is consistent with the theory and shows the importance of accounting for such costs in a model of multistop shopping. Further, a scenario with zero shopping costs predicts a larger proportion of multistop shoppers; nearly 30% as opposed to 1% in our estimated model with shopping costs. Note that removing shopping costs need not translate into a situation in which all consumers are multistop shoppers. In fact, most shoppers optimally choose to visit a single store. This might be the result of transport costs along with unobserved idiosyncratic valuations of product-store combinations that yield a greater utility through concentrating purchases at a single store (and that are captured by the error term of the model). Table 9: Probability of visiting a given number of stores with and without shopping costs Number of stops 0 1 2 3

With shopping costs (%)

Zero shopping costs (%)

Percentage change

61.8 [59.5, 64.3] 37.1 [34.5, 39.3] 1.0 [0.7, 1.2] 0.02 [0.01, 0.03]

1.32 [1.0, 1.7] 69.53 [67.8, 71.6] 27.90 [25.8, 29.8] 1.25 [1.1, 1.4]

-98% 87% 2,179% 6,242%

Notes: This table shows the predicted probabilities of engaging in a given number of shopping stops within a week (which we interpret also as the distribution of shoppers by number of stops). The first column shows the results under our preferred specification with shopping costs. The second column shows analogous results obtained by simulating a counterfactual situation in which shopping costs are zero for all consumers. If the consumer visited more stores than the three included in our sample, the number of stops in this table should be interpreted as “additional” stops. The 95% confidence intervals are given in square brackets.

In the second exercise, we compare estimated substitution patterns and implied markups and marginal costs from two alternative specifications of the demand model, namely, one with 29

shopping costs (our base model) and another one without shopping costs. We use the semielasticities obtained under each model and compute the ratio between the two as an indicator of estimation bias (assuming that our base model is the correct specification). Table 10 shows the average ratios in three panels: the top panel (Panel A) presents the ratios of own- and cross-price semi-elasticities computed for each supermarket and averaged across supermarkets, the middle panel (Panel B) reports the ratios of inter-supermarket cross-price semi-elasticities averaged across supermarkets, and the bottom panel (Panel C ) reports the ratios between estimated semi-elasticities with respect to the outisde good.16 All of the entries on the main diagonal of Panel A are positive and greater than one, which means that the estimated own-price semi-elasticities under the no-shopping-costs scenario are, on average, biased upwards compared to those obtained when shopping costs are accounted for. Similarly, the cross-price semi-elasticities between substitutes in the category of yogurts are also biased upwards. Conversely, a model without shopping costs appears to predict lower cross-price semi-elasticities for products in different categories in the same supermarket. This suggests that estimates of intra-store substitution patterns between product categories are biased downwards when shopping costs are not accounted for in a model of multistop shopping. Alternatively, all of the entries of Panels B and C are positive and larger than one. This suggests that a model of multistop shopping which does not account for shopping costs tends to overestimate inter-store price semi-elasticities both for products within the same category and across categories (taking that the outside good as an additional category). Our interpretation of this result is that in the absence of shopping costs, consumers are more sensitive to a price change because they are able to substitute products across stores at no additional cost. Interestingly, the two specifications predict the same direction for consumer substitution patterns (i.e., all of the ratios in Table 10 are positive). When the price of a product (say, for example, biscuits) changes, the demand for the remaining products decreases. This suggests that the cross-price semi-elasticities are capturing the economies of scope that arise when concentrating purchases at a single supermarket. In this case, shopping costs make those economies of scope more valuable as it is more costly for consumers to substitute a single product and deal with multiple supermarkets than switch between stores. This is consistent with the theoretical findings of previous literature (see, in particular, Klemperer (1992)). Next, we report average prices of the observed equilibrium, as well as average implied markups and marginal costs, according to the two specifications in in Table 11. The markups predicted by our base model are lower, on average, when compared to those predicted by an alternative model without shopping costs. This is in part explained by the lower intra-store cross-category elasticities, and by the higher market shares that the model without shopping costs predicts. As a consequence, assuming that our preferred specification is the true model, the marginal costs predicted by a model without shopping costs are biased downwards. 16

We report mean estimated own- and cross-price elasticities for each specification in Tables D.1 and E.1 of the Online Appendix.

30

Table 10: Average ratios of own- and cross-price semi-elasticities estimated from two alternative models Product

Yogurt

Yogurt NB

Biscuits

Desserts

Panel A: within supermarket Yogurt Yogurt NB Biscuits Dessert

1.56 2.51 0.77 0.83

2.79 1.65 0.86 0.93

0.82 0.82 1.62 0.83

0.81 0.80 0.76 1.58

Panel B: across supermarkets Yogurt Yogurt NB Biscuits Dessert

2.91 2.85 1.66 1.90

3.07 3.00 1.74 1.99

1.85 1.81 3.31 1.96

1.81 1.77 1.64 2.89

Panel C: outside good

9.99

10.86

11.25

10.85

Notes: Ratios are computed by dividing the estimated semi-elasticity of the model without shopping costs with the corresponding semi-elasticity of the model with shopping costs. An entry in Panel A corresponds to the average across supermarkets of intra-store, own- and cross-price semi-elasticity ratios. An entry in Panel B corresponds to the average across supermarkets of inter-store cross-price semi-elasticity ratios. Finally, an entry in Panel C corresponds to ratio of the the semi-elasticities with respect to the outside good obtained from each model.

4.3

Robustness

In order to check the robustness of our results, we perform three exercises in which we reestimate our demand model under different conditions: first, we try a different specification in which we account for the number of trips to each store in addition to the number of supermarkets visited; second, we estimate our preferred specification with a different selection of products, namely, ham, cheese and pastry; and third, we change the definition of the outside good to check the implications on our demand estimates of our preferred definition. Overall, we find that our results are robust to the definition of shopping patterns, the selection of products and the definition of the outside good. 4.3.1

An alternative identification strategy for shopping costs

In the first exercise, we test our estimates of shopping costs to one potential concern. Our strategy for the identification of shopping costs does not consider the intensity of the relationship of a consumer with each supermarket in the same shopping period as multistop shopping. For example, if we observe that a consumer uses only one supermarket chain in a given week, we would interpret this as the consumer being a one-stop shopper, even if during that week we observe that the consumer has made multiple trips to the stores of that supermarket. A concern with this interpretation of shopping patterns is that additional trips to the same store may imply 31

Table 11: Average prices and implied marginal costs and markups according to two alternative models With shopping costs Product

Price (Euro cents)

Marginal cost (Euro cents)

Markup (%)

Without shopping costs Marginal cost (Euro cents)

Markup (%)

Yogurt Yogurt NB Biscuits Desserts

27.55 25.21 9.84 49.52

Supermarket 13.60 11.18 0.76 40.05

1 50.63 55.65 92.27 19.13

11.10 8.97 0.44 35.40

59.69 64.43 95.53 28.53

Yogurt Yogurt NB Biscuits Desserts

26.95 27.69 10.34 48.77

Supermarket 13.10 13.85 0.23 38.98

2 51.39 49.98 97.81 20.08

10.48 11.14 0.13 33.86

61.13 59.78 98.75 30.58

Yogurt Yogurt NB Biscuits Desserts

30.16 26.11 11.02 52.79

Supermarket 16.43 12.21 0.32 42.13

3 45.53 53.22 97.08 20.19

12.93 9.29 0.17 36.11

57.14 64.42 98.43 31.59

Notes: This table reports implied marginal costs and markups from two alternative models. The first column reports the average across products within the same category of the observed prices.

positive costs to the consumer, even if she is dealing with a single supermarket. If this is so, our estimates of shopping costs may be biased upwards because we are counting fewer stops per individual, on average. In order to check if our results are sensitive to the definition of shopping patterns, we re-estimate the demand model with a more flexible definition of multistop shopping which will count every trip to a store made by an individual as a stop, in addition to accounting for the number of supermarkets she is dealing with. Specifically, we include an additional variable to our main specification, that accounts for the number of shopping trips that each household makes to the same supermarket chain within a week, and a random coefficient to allow unobserved individual characteristics that explain observed heterogeneity in the intensity of shopping for each supermarket to enter the model. This implies that the definition of shopping patterns can change with respect to our main model. For example, in the alternative specification a shopper dealing with a single chain but that made a number of trips to the stores of that chain is now a multistop shopper. Similarly to our main case, if a household is observed to deal with multiple supermarket chains but made one trip to each of them is also called a multistop shopper. We report the results of the demand estimates of this new specification in Table B.1 of the Appendix. The results of this model are similar to our main specification: the coefficients of price and distance to stores are statistically significant, of the same sign and close in magnitude. We compare the estimates of shopping costs obtained from the two alternative definitions of 32

shopping patterns, that are reported in Table 12. The estimate of the cost of dealing with a supermarket is 63.2 euro cents which captures, among other things, the cost of one trip to that supermarket. We add to this cost the average cost of making and additional shopping trip to a supermarket, which is 20.1 euro cents and is the result of translating the estimate of the variable “number of shopping trips” of Table B.1 to euros and multiplying it with the number of additional trips made to a supermarket within a week, which in our sample is 0.4 additional trips (see Table 3). Overall, the mean shopping cost of dealing with each supermarket in this context is 83.3 euro cents, on average. This figure is close to the estimate we obtained with our preferred definition of multi-stop shopping. Our main results are, therefore, robust to the definition of shopping patterns. Table 12: Estimated mean shopping costs according to two alternative definitions of shopping patterns (in euro cents) Definition of shopping patterns Preferred

Alternative

Mean cost per additional trip to a store

——-

20.1

Mean cost per supermarket

96.0

63.2

Mean shopping cost

96.0

83.3

Notes: Mean shopping costs for each case are obtained, respectively, using results in column 2 of our main results (Table 6) and column 2 of the robustness results (Table B.1) and translating the respective figure into euros by dividing it by the estimate of the price coefficient of the respective specification. The average cost of making and additional shopping trip to a supermarket is the result of translating the estimate of the variable “number of shopping trips” of Table B.1 to euros and multiplying it with the number of additional trips made to a supermarket within a week, which in our sample is 0.4 additional trips.

4.3.2

Robustness to alternative selection of products and definition of the outside option

In the second exercise, we check the robustness of our structural estimates to changes in the composition of the basket of products and the subsample of households that purchase a particular set of products. In order to do that, we estimate our main model of demand with a different subset of frequently purchased products: ham, cheese and pastry. Following the same procedure for the estimation of our main model, we randomly sample a number of households that registered purchases of one or several of these products during 2006. We report the results of the demand estimation in Table B.2 of the Appendix. (We also offer descriptive statistics of the sample of households that registered purchases of those three products in Tables C.4, C.5 and C.6, and the distribution of shopping patterns and the share of expenditure in Figure C.2 of the Online Appendix C). 33

Finally, in the third exercise, we change the definition of the outside option to check the sensitivity of our results to the existence of additional stops not included in the analysis (i.e., assumed to cost zero to the consumers). We restrict our sample to the consumers who were observed to make purchases only from the three supermarkets included in our main analysis. Moreover, we do not allow for the no purchase option (which also captures the shopping visits made by the consumers to shopping locations that are not observed in the data). Therefore, we compute product level market shares for the included products so that they sum to one. As a consequence, the model is inelastic to generalized shocks. We report the results of the demand estimation in Table B.3 of the Appendix. The results of our main estimates of shopping costs with those obtained with the two alternative models are shown in Table 13. The estimates of the mean fixed shopping cost, the mean transport costs and the total shopping costs are similar. Table 13: Estimated mean shopping and distance costs according to our main model and two variations (in euro cents) Main model

Alternative set of products

Alternative definition of the outside option

Mean shopping cost

96.1

81.9

74.9

Mean transport cost

8.6

9.0

11.1

104.7

90.9

86.0

Mean total shopping cost

Notes: Mean shopping and distance costs for each case are obtained, respectively, using the results in column 2 of our main results (Table 6) and column 2 of the robustness results (Tables B.2 and B.3) and translating the respective figure into euros by dividing it by the estimate of the price coefficient of the respective specification. The mean total shopping cost is obtained as the sum of the shopping and distance costs.

5

Applications

In this Section, we apply our models of demand and supply to study two marketing strategies that are widespread and that can be straightforwardly captured with our framework: the effects of product delisting by a particular supermarket on consumer shopping behavior and the effects of aggressive price discounts that may be presented as price promotions but can be hiding reselling at or below cost.

5.1

Product delisting

We are interested in how consumers substitute products and stores when one of the products is no longer available at their preferred shopping location. To capture this demand-side effects in 34

a flexible and realistic way, we take into account that rival supermarkets may react to product delisting by adjusting their prices. Thus, we use our model of supermarket competition in order to capture the strategic reactions of rivals when a supermarket removes a product from its stores’ shelves. There is, however, an important limitation in our approach. In our model of supply, we implicitly assume that wholesale prices are exogenous and do not adjust when a product is delisted by one supermarket. In other words, we are assuming that upstream firms do not respond to product delisting by a downstream counterpart. Of course, this is a very restrictive and unrealistic assumption given that—as pointed out in the Introduction—delisting can be used strategically as a vertical restraint imposed by supermarkets on manufacturers in order to obtain better terms of trade. We, however, make this assumption for the sake of simplicity. A complete study of the effects of product delisting would consider how both rival stores and manufacturers optimally adjust prices in the same period. This is outside the scope of this paper and we leave it for future research. 5.1.1

Counterfactual simulations

Retail stores often use the threat of delisting either a product or a range of products so as to obtain better deals with manufacturers. Although the supermarket may benefit from this practice, it also entails some losses, as consumers may be tempted to switch to rival stores when a product they desire is unavailable at their usual store. For instance, in 2009 in Belgium, a request for a price rise by Unilever triggered the delisting of 300 of Unilever’s products by Delhaize, one of the largest supermarket chains in that country. Both parties ended up being negatively affected: Delhaize lost 31% of its customers to rivals, and among those who remained, 47% substituted other brands for Unilever’s products.17 In the UK, a similar dispute between Tesco, the largest supermarket chain in the country, and Premier Foods resulted in the delisting of a number of the supplier’s products in 2011, resulting in a 1% fall in sales and an 18% fall in the value of Premier Foods’s shares.18 Finally, in the United States, a dispute regarding prices between Costco and Coca-Cola in 2009 led Costco to delist all Coca-Cola products for some weeks. As a consequence, Costco’s shares went down 0.3%.19 If consumers find it costly to visit alternative supermarkets (e.g., because of strong feelings of loyalty, very large shopping costs or head-to-head competition between supermarkets), the delisting of a product will only harm the manufacturer, making the threat of delisting an effective bargaining strategy for retailers. However, when consumers find it optimal to undertake multistop shopping, the delisting of a product can also harm the delisting store: this occurs as a result of a reduction in demand from shoppers who either continue to undertake one-stop shopping at a competing store or who visit an additional store if their shopping costs are sufficiently low. 17

See http://in.reuters.com/article/delhaize-unilever-idINLG51937220090216. See http://www.telegraph.co.uk/finance/8684844/Premier-Foods-crumbled-by-Tesco-bust-up.html. 19 See http://www.reuters.com/article/cocacola-costco-idUSN1020190520091210. 18

35

To assess the effects of product delisting on consumer shopping behavior and supermarket competition, we perform a counterfactual experiment in which we simulate a large increase in the price of NB yogurt in supermarket 1 so that it becomes prohibitively expensive. We believe that this hypothetical rise in price will capture what happens when a product is removed (i.e., zero demand for that product), and is convenient from an empirical perspective as we do not have to erase observations from our data. We use our demand estimates and the supply model introduced above to compute prices and market shares of the resulting counterfactual equilibrium. 5.1.2

Results

Table 14 reports the observed prices by product category for each supermarket as well as their adjusted prices when supermarket 1 delists yogurt NB. We compare the model’s predictions prior to delisting and after the delisting of yogurt NB by supermarket 1 under two distinct counterfactual scenarios: one in which consumers bear shopping costs and a second one in which we assume that there are no shopping costs. For each situation, except for the observed case (column 1), we recompute equilibrium prices and market shares. Notice that with or without shopping costs, the pricing strategies adopted by supermarkets after the delisting of the yogurt NB is similar: supermarket 1 decreases the price of the composite yogurt while it increases the price of biscuits and desserts. As a response, rival supermarkets increase the price of all of the products they supply. This is explained by two factors: the higher proportion of one-stop shoppers in the two cases and the fact that even in the absence of (fixed) shopping costs consumers still bear transport costs, which make substitution between similar products across stores imperfect. In an attempt to retain those one-stop shoppers who no longer find their preferred alternative, Supermarket 1 decreases the price of the substitute to make it more attractive to them. Alternatively, the substitution costs faced by consumers, in addition to the fact that one-stop shoppers may care more about the overall value of the basket than product-wise prices, allows Supermarket 1 to increase the price of both biscuits and desserts. Moreover, given that Supermarket 1 losses some customers to rivals, they experience a rise in demand and, as a consequence, respond by increasing their prices for all products. The difference between the two counterfactual scenarios is in the magnitude of the responses rather than in their direction. In effect, the ability of Supermarket 1 to increase the price of biscuits and desserts is higher in the presence of shopping costs compared to the scenario with zero shopping costs, as they act as a barrier for consumers to source additional supermarkets. Therefore, shopping costs create inter-category complementarities between product categories that otherwise would be independent. Table 15 reports the overall value of baskets with two and three products under different scenarios. In line with the previous reasoning, the decrease in the price of the composite yogurt and the increase in the price of biscuits and desserts implies that the overall value of baskets containing a combination of composite yogurt and one or two of the complement products

36

Table 14: Prices and price changes when supermarket 1 delists yogurt NB Scenario /

Observed

Counterfactual 1: with shopping costs

Product

Price

Post delisting price

Yogurt Yogurt NB Biscuits Desserts

27.55 25.21 9.84 49.52

Yogurt Yogurt NB Biscuits Desserts Yogurt Yogurt NB Biscuits Desserts

Counterfactual 2: zero shopping costs

% change

Pre delisting price

Post delisting price

% change

27.25 —— 10.09 49.79

Supermarket 1 -1.09 —— 2.54 0.54

33.74 31.43 17.04 56.04

33.37 —— 17.22 56.25

-1.10 —— 1.06 0.38

26.95 27.69 10.34 48.77

26.97 27.71 10.35 48.77

Supermarket 2 0.07 0.06 0.02 0.01

33.70 34.44 18.20 56.14

33.81 34.55 18.24 56.17

0.31 0.31 0.24 0.05

30.16 26.11 11.02 52.79

30.19 26.14 11.02 52.80

Supermarket 3 0.09 0.11 0.07 0.03

38.33 34.33 20.47 61.59

38.47 34.47 20.54 61.63

0.35 0.39 0.31 0.07

Notes: The results in this Table were obtained under the assumption that Supermarket 1 delists yogurt NB while its rivals continue to stock the product. “Counterfactual 1” is a simulated scenario in which we assume that supermarket 1 sets the price of its Yogurt NB to its marginal costs, which implies a discount of 55.7% on the observed price. “Counterfactual 2” is a simulated scenario in which we set shopping costs to zero for all consumers. Marginal costs and prices are given in Euro cents.

remains constant after the delisting of the yogurt NB. Again, the higher proportion of one-stop shoppers and the inter-store substitution frictions enable Supermarket 1 to keep the overall value of baskets that include composite yogurt (almost) constant as long as one stop shoppers are weakly indifferent between baskets containing yogurt NB and those containing composite yogurt if the overall value of the baskets is the same. Similarly, Table 16 reports the market shares and their percentage changes by product for each supermarket when Supermarket 1 delists yogurt NB. In the presence of shopping costs, when Supermarket 1 decreases the price of its composite yogurt its demand raises by 3.3%. Nevertheless, the market shares of other product categories decrease by about 4% in both cases. This decrease in market shares is explained by two effects: first, the increase in prices due to the removal of the downward pricing pressure generated by the delisting of the yogurt NB; and second, consumer shopping costs, which makes it easier for shoppers to switch supermarkets. The combination of both effects makes the decrease in prices of the composite yogurt insufficient to compensate shoppers and retain them. Instead, consumers switch to rival supermarkets leading to an increase in market shares for all of the products offered by them. As expected, the increase in market shares is larger for yogurt NB than for the composite yogurt. Rivals’ biscuits and desserts also experience an increase in market shares. 37

Table 15: Overall value and value changes of baskets of two and three products at supermarket 1 Scenario /

Observed

Counterfactual 1: with shopping costs Post delisting value

% change

Counterfactual 2: zero shopping costs

Basket

value

Pre delisting value

{YN B , B, D} {Y, B, D}

84.6 86.9

Three-product baskets —— —— 87.1 0.2

{YN B , B} {Y, B} {YN B , D} {Y, D} {B, D}

35.0 37.4 74.7 77.1 59.4

Two-product baskets —— —— 37.3 -0.1 —— —— 77.0 0.0 59.9 0.9

Post delisting value

% change

104.5 106.8

—— 106.8

—— 0.0

48.5 50.8 87.5 89.8 73.1

—— 50.6 —— 89.6 73.5

—— -0.4 —— -0.2 0.5

Notes: The table is organized in three vertical panels. “Observed value” refers to the data used to estimate our preferred specification with shopping costs. “Counterfactual 1” is a simulated scenario in which we assume that Supermarket 1 delists its Yogurt NB. “Counterfactual 2” is a simulated scenario in which we set shopping costs to zero for all consumers. YN B : Yogurt NB, Y : Yogurt, B : Biscuits and D: Desserts.

In a context without shopping costs, the delisting of yogurt NB by supermarket 1 generates an increase in the demand for the composite yogurt of 5.5%, which is larger compared to the increase in demand when shopping costs are present. By contrast, the market shares of biscuits and desserts decrease by only 0.76% and 1%, respectively, which are lower than the corresponding numbers obtained with nonzero shopping costs. Precisely, the difference in the reactions in demand reflects the complementarities generated by shopping costs across seemingly unrelated categories. In line with our previous results, market shares for both yogurts and the other products increase for rival supermarkets. Table 17 reports the probabilities of visiting each store under the two scenarios. Results are consistent with theory and conventional wisdom. The delisting of yogurt NB by Supermarket 1 makes consumers be less likely to visit supermarket 1 in both scenarios: the probability of visiting supermarket 1 decreases by 4.9% with shopping costs whereas this probability decreases by 2.3% when consumers do not bear any shopping costs. Alternatively, the probability of visiting rival supermarkets increase for all other supermarkets. These results constitute further evidence on the role of shopping costs: a one-stop shopper who wants to purchase yogurt NB, which is no longer available at Supermarket 1, might prefer to switch to rival supermarkets even if Supermarket 1 is her preferred shopping location; similarly, a two-stop shopper may prefer a different mix of stores and products. This effect is softer in the absence of shopping costs, given that consumers can still purchase the yogurt NB by making an additional stop at a lower cost without leaving completely their preferred shopping location. Finally, we report in Table 18 weekly revenues for all supermarkets under the observed and 38

Table 16: Market shares and changes in market shares when supermarket 1 delists yogurt NB Scenario /

Observed

Counterfactual 1: with shopping costs

Product

share

Post delisting share

Yogurt Yogurt NB Biscuits Desserts

8.28 0.92 5.49 6.55

8.56 0.00 5.26 6.27

Yogurt Yogurt NB Biscuits Desserts

6.93 0.40 4.57 4.97

Yogurt Yogurt NB Biscuits Desserts Outside option

% change

Counterfactual 2: zero shopping costs Pre delisting share

Post delisting share

% change

Supermarket 1 3.32 -100 -4.19 -4.24

20.93 2.39 13.79 14.82

22.07 0 13.69 14.68

5.48 -100 -0.76 -1.00

7.00 0.41 4.61 5.02

Supermarket 2 1.01 1.02 0.94 1.01

20.00 1.17 13.61 12.90

20.25 1.18 13.69 12.98

1.23 1.25 0.60 0.65

9.02 0.52 6.33 7.24

9.12 0.53 6.40 7.33

Supermarket 3 1.15 1.13 1.05 1.14

21.35 1.27 16.18 16.38

21.64 1.29 16.26 16.47

1.35 1.32 0.50 0.58

62.13

62.47

2.84

2.91

2.51

0.54

Notes: The results in this Table were obtained under the assumption that Supermarket 1 delists yogurt NB while its rivals continue to stock the product. “Counterfactual 1” is a simulated scenario in which we assume that supermarket 1 sets the price of its Yogurt NB to its marginal costs, which implies a discount of 55.7% on the observed price. “Counterfactual 2” is a simulated scenario in which we set shopping costs to zero for all consumers. Market shares are given in percentages.

Table 17: Predicted probabilities of visiting stores 1, 2 and 3 under different scenarios when supermarket 1 delists yogurt NB Scenario / Supermarket

Observed

Counterfactual 1: with shopping costs

Counterfactual 2: zero shopping costs

probability

Post delisting probability

% change

Pre delisting probability

Post delisting probability

% change

Supermarket 1

13.68

13.02

-4.88

40.29

39.36

-2.31

Supermarket 2

10.71

10.83

1.08

36.68

37.02

0.92

Supermarket 3

15.14

15.32

1.23

43.96

44.33

0.86

Notes: The results in this Table were obtained under the assumption that Supermarket 1 delists yogurt NB while its rivals continue to stock the product. Probabilities are given in percentages.

the counterfactual scenarios. In order to compute revenues, we assume that the market size for each product is equivalent to the average number of servings consumed by an individual

39

in a week multiplied by the french population in 2005. According to our data, on average, an individual consumes three servings of yogurts, four servings of biscuits and three servings of desserts per week. Our results suggest that a supermarket loses revenues when it delists one product while competing rivals continue to stock the product. Consistent with our previous results, the delisting supermarket suffers more in a situation in which shopping costs are present as a result of two opposing effects: first, an extensive margin effect that is present whenever consumers decide to switch supermarkets due to high shopping costs; and second, an intensive margin effect that is present whenever consumers who previously purchased a single product now purchase two or more products due to a decrease in the price of baskets that include the composite yogurt. The fact that the loss in revenue is lower in the case without shopping costs implies that the intensive margin effect is stronger under this scenario than in the scenario where shopping costs are present. Table 18: Weekly revenues (in million euros) and changes in weekly revenues under different scenarios when supermarket 1 delists yogurt NB Scenario / Supermarket

Observed

Counterfactual 1: with shopping costs

Counterfactual 2: zero shopping costs

revenues

Post delisting revenues

% change

Pre delisting revenues

Post delisting revenues

% change

Supermarket 1

38.3

36.4

-4.97

112.3

109.3

-2.66

Supermarket 2

29.7

30.0

1.04

102.5

103.6

1.08

Supermarket 3

44.9

45.4

1.19

133.9

135.3

1.09

Notes: The results in this Table were obtained under the assumption that Supermarket 1 delists yogurt NB while its rivals continue to stock the product. Revenues were obtained under the assumption that the market size for each product was equivalent to the average number of servings consumed by an individual in a week multiplied by the french population in 2005. According to our data, on average, an individual uses to consume 3 servings of yogurt, 4 servings of biscuits and 3 servings of desserts in a week.

5.2

Agressive price discounts

Consider the case of a supermarket making an aggressive discount on one or more products to attract demand.20 For products with certain characteristics (e.g., staples, frequently purchased and that make part of the regular product assortment of discount stores), the price after the discount often is at or lies below the product’s unit cost. Anecdotal evidence shows that this practice, best known as loss-leader pricing, is widespread in multiproduct retailing markets even though it has been banned in several countries of the European Union, such as Germany, France, Ireland and Belgium, and in at least 25 states of the United States. Loss leading practices are usually banned because they are considered predatory, which is often supported by one-stop shopping models. However, a more general approach that considers 20

We thank an anonymous referee for suggesting this application to us.

40

multistop shopping and that accounts for shopping costs may lead to the opposite result. For instance, Chen and Rey (2012, 2019) show that in the presence of shopping costs, loss-leading strategies and cross-subsidies are not predatory, and the latter might even be welfare enhancing. We apply our models of demand and supply to this case in order to examine how, in the presence of shopping costs, both consumers and rival supermarkets react when one supermarket gives an aggressive price discount on one of its products. 5.2.1

Counterfactual simulations

A recent example of agressive price discounts happened in France in 2018, where one of the leading supermarket chains in the country, Intermarché, offered Nutella spread jars at 70% discount. After an investigation, the French competition authority found that such a low price was below Intermarché’s unit cost. It also found that the supermarket chain was also selling other products below cost. As a consequence, it imposed the maximum fine possible (about 375 thousand euros). This case motivated new regulations imposing tighter constraints on discounts that supermarkets can make on products.21 In order to assess the effects of aggressive discounting on consumer shopping behavior and supermarket competition, we simulate a situation in which Supermarket 1 sets the price of its yogurt NB at its marginal cost which turns out to be a discount of nearly 56% on the observed price. We use our structural models of demand and supply along with our estimated coefficients to compute counterfactual prices, market shares and consumer shopping patterns in this context. As in the delisting applications and for the sake of documenting the role of shopping costs, we perform two counterfactual experiments: one in which shopping costs are present and another in which we set shopping costs to zero for all consumers. 5.2.2

Results

We report the price responses of supermarkets to a discount in yogurt NB by Supermarket 1 in Table 19. Our results are consistent with anecdotal evidence and economic literature according to which loss leading is often used by multiproduct retailers in order to attract demand but compensate losses by setting higher prices on other products. The decrease in the price of yogurt NB triggers a decrease in the price of the substitute yogurt, and an increase in the price of biscuits and desserts. Rival supermarkets react by lowering the prices of their products in most cases. In this case, decreasing the price of yogurt NB may stimulate multistop shopping: customers with low enough shopping costs can find it optimal to purchase yogurt at Supermarket 1 and biscuits and desserts at rival supermarkets. When shopping costs are not present (“counterfactual 2” of Table 19) Supermarket 1 still increases the price of biscuits and desserts but in lower percentages. To understand why, one should look at the overall price of baskets, which we report in Table 20. When shopping costs are present, most consumers are one-stop shoppers which means that those looking for purchasing 21

See https://www.journaldeleconomie.fr/Nutella-a-prix-casse-Intermarche-convaincu-de-revente-a- pertea 5684.html.

41

Table 19: Prices and price changes when supermarket 1 offers an aggressive discount on yogurt NB Scenario / Product

Observed equilibrium Marginal cost

Price

Yogurt Yogurt NB Biscuits Desserts

13.6 11.2 0.8 40.1

27.5 25.2 9.8 49.5

Yogurt Yogurt NB Biscuits Desserts

13.1 13.9 0.2 39.0

Yogurt Yogurt NB Biscuits Desserts

16.4 12.2 0.3 42.1

Counterfactual 1: with shopping costs Post discount price

% change

Counterfactual 2: zero shopping costs Pre discount price

Post discount price

% change

Supermarket 1 26.8 -2.69 11.2 -55.65 10.8 9.58 50.7 2.32

33.7 31.4 17.0 56.0

32.7 11.2 17.4 56.6

-3.00 -64.43 2.03 0.97

27.0 27.7 10.3 48.8

Supermarket 2 26.9 -0.08 27.7 -0.08 10.3 -0.08 48.8 0.00

33.7 34.4 18.2 56.1

33.4 34.1 18.1 56.1

-0.87 -0.85 -0.74 -0.17

30.2 26.1 11.0 52.8

Supermarket 3 30.1 -0.10 26.1 -0.12 11.0 -0.13 52.8 0.00

38.3 34.3 20.5 61.6

37.9 33.9 20.3 61.4

-1.01 -1.13 -1.01 -0.24

Notes: The table is organized in three vertical panels. “Observed equilibrium” refers to the data used to estimate our preferred specification with shopping costs. The marginal cost column shows the figures we backed out using our demand estimates and the model of supply. “Counterfactual 1” is a simulated scenario in which we assume that supermarket 1 sets the price of its Yogurt NB equal to its marginal cost, which implies a discount of 55.7% on the observed price. “Counterfactual 2” is a simulated scenario in which we set shopping costs to zero for all consumers.

more than one product must concentrate purchases on a single supermarket. The aggressive discount on yogurt NB attracts one-stop shoppers. Those who prefer the yogurt NB to the alternative yogurt can now purchase their preferred basket at a cheaper total price. However, those who prefer the alternative yogurt to the yogurt NB (because of strong feelings of loyalty, for example) should pay a higher overall price for their preferred basket. Conversely, in the absence of shopping costs, Supermarket 1 decreases the overall value of all baskets in which any of the yogurt varieties is present. Therefore, shopping costs enables Supermarket 1 to exploit those shoppers who do not prefer the loss leader. The huge price discount generates a considerable increase in the market share of yogurt NB whereas there is a decrease in market shares in all other products, which is a consequence of a substitution effect in the case of the composite yogurt and the increase in prices in the case of biscuits and desserts. Similarly, the demand for all of the products of rival supermarkets decreases in spite of their efforts in decreasing their prices; the outside good also experiences a decrease in market share. This is the result of two effects: first, a business stealing effect, as long as former customers of supermarkets 2 and 3 and those not included in our analysis (but

42

Table 20: Overall value and value changes of baskets of two and three products at Supermarket 1 Scenario /

Observed

Counterfactual 1: with shopping costs Post discount value

% change

Counterfactual 2: zero shopping costs

Basket

value

Pre discount value

{YN B , B, D} {Y, B, D}

84.6 86.9

Three-product baskets 72.6 -14.1 88.3 1.6

{YN B , B} {YN B , D} {Y, B} {Y, D} {B, D}

35.0 74.7 37.4 77.1 59.4

Two-product baskets 22.0 -37.3 61.9 -17.2 37.6 0.5 77.5 0.5 61.5 3.5

Post discount value

% change

104.50 106.82

85.14 106.70

-18.53 -0.12

48.5 87.5 50.8 89.8 73.1

28.6 67.8 50.1 89.3 74.0

-41.1 -22.5 -1.3 -0.5 1.2

Notes: The table is organized in three vertical panels. “Observed value” refers to the data used to estimate our preferred specification with shopping costs. “Counterfactual 1” is a simulated scenario in which we assume that supermarket 1 sets the price of its Yogurt NB equal to its marginal cost, which implies a discount of 55.7% on the observed price. “Counterfactual 2” is a simulated scenario in which we set shopping costs to zero for all consumers. YN B : Yogurt NB, Y : Yogurt, B : Biscuits and D: Desserts.

captured by the outside good) now prefer to purchase (at least) yogurt NB from Supermarket 1; and second, a market expansion effect as the cheaper baskets that include the loss leader stimulates some zero-stop consumers to become shoppers. As expected, we observe similar patterns but with higher demand reactions in counterfactual 2 because consumers incur lower transaction costs of shopping.

6

Conclusions

In this article, we develop and estimate a model of multiproduct demand for groceries in which customers with heterogeneous shopping costs can choose between visiting one or multiple supermarkets in a given week in order to empirically examine the role of individual shopping costs in consumer shopping behavior and supermarket pricing. We obtain that the mean (across consumers) total shopping costs are 104.7 euro cents per supermarket visited. This cost includes a fixed shopping cost of 96.1 euro cents per supermarket visited and a transport cost of 8.6 euro cents per trip to a supermarket located at the average distance from consumer location. In order to assess the role of shopping costs in consumer choice and supermarket competition, we compare our results with those obtained from a situation in which shopping costs are zero, and confirm that accounting explicitly for the observed heterogeneity in shopping patterns through positive shopping costs is important from an empirical standpoint. For instance, when we unilaterally set shopping costs to zero, the model predicts that all consumers would visit at least one store every week with positive probability (which is not consistent with what is observed in 43

Table 21: Market shares and changes in market shares when supermarket 1 offers an aggressive discount on yogurt NB Scenario /

Observed

Counterfactual 1: with shopping costs

Counterfactual 2: zero shopping costs

Product

share

Post discount share

% change

Pre discount share

Post discount share

% change

Supermarket 1 Yogurt Yogurt NB Biscuits Desserts

8.28 0.92 5.49 6.55

8.17 1.77 5.44 6.49

-1.40 91.82 -0.97 -1.02

20.93 2.39 13.79 14.82

20.06 6.93 13.49 14.55

-4.15 189.82 -2.20 -1.86

Supermarket 2 Yogurt Yogurt NB Biscuits Desserts

6.93 0.40 4.57 4.97

6.89 0.40 4.54 4.95

-0.64 -0.64 -0.53 -0.41

20.00 1.17 13.61 12.90

19.40 1.13 13.43 12.78

-3.01 -3.05 -1.32 -0.91

Supermarket 3 Yogurt Yogurt NB Biscuits Desserts

9.02 0.52 6.33 7.24

8.96 0.52 6.30 7.21

-0.67 -0.71 -0.55 -0.43

21.35 1.27 16.18 16.38

20.69 1.23 16.01 16.28

-3.09 -3.25 -1.01 -0.62

Outside option

62.13

61.86

-0.45

2.84

2.60

-8.61

Notes: The results in this Table were obtained under the assumption that Supermarket 1 delists yogurt NB while its rivals continue to stock the product. “Counterfactual 1” is a simulated scenario in which we assume that supermarket 1 sets the price of its Yogurt NB to its marginal costs, which implies a discount of 55.7% on the observed price. “Counterfactual 2” is a simulated scenario in which we set shopping costs to zero for all consumers. Market shares are given in percentages.

our data); conversely, once shopping costs are accounted for, the predicted probabilities of both one- and multistop shopping are lower, and consumers are less likely to visit a supermarket on a week-to-week basis. Our results also suggest that a model of multiproduct choice that does not account for shopping costs tend to overestimate, on average, the implied own-price elasticities, market shares and markups and underestimate underestimates cross-price elasticities. We apply our models of demand and supply to empirically examine two practices that are commonly used by supermarkets: product delisting and loss-leader pricing. Even though we obtain interesting results, our analysis presents some limitations as we do not fully capture all of the potential effects at play. For example, in the delisting application we do not capture the pricing reactions of upstream firms as we do not model the vertical relations between manufacturers and supermarkets. As a result, we are not able to measure neither the full effects on supermarket pricing caused by delisting of products nor the role of product delisting as a vertical restraint. Precisely, our two applications motivate two avenues for future empirical research that can incorporate our demand framework. Understanding the full implications of these two practices 44

that are commonly used by supermarkets is important from a competition policy perspective. The first avenue would be an in depth analysis of product delisting and its implications for vertical bargaining and equilibrium pricing at both upstream and downstream levels when supermarkets use product delisting strategically to increase their bargaining power vis-à-vis manufacturers. A model that accounts for the vertical dimension of the supply of groceries along with our model of multiproduct and multistore choice with shopping costs is a natural step toward understanding an overall picture of the consequences of product delisting, even though it represents a complex and challenging task. A second avenue concerns below-cost pricing strategies by supermarkets, which is what raises concerns on competition authorities about loss-leader pricing being predatory. According to the OECD (2005), laws preventing resale below cost (RBC) and claiming to protect high-price, low-volume stores from large competitors who can afford to offer lower prices might be introducing unnecessary constraints. Evidence from countries without RBC laws shows that smaller competitors need not be pushed out of the market if they are not protected. Chen and Rey (2012, 2019) show that in the presence of shopping costs, loss-leading strategies and cross-subsidies are not predatory, and the latter might even be welfare enhancing. Empirical evidence that fully accounts for the supply and demand implications of eliminating RBC laws may help to clarify this issue and motivate policy changes.

References Aguiar, Mark, and Erik Hurst. 2007. “Life-cycle prices and production.” The American Economic Review, 97(5): 1553–1559. Armstrong, Mark, and John Vickers. 2010. “Competitive non-linear pricing and bundling.” The Review of Economic Studies, , (77): 30–60. Berry, Steven, James Levinsohn, and Ariel Pakes. 1995. “Automobile prices in market equilibrium.” Econometrica, 63(4): 841–890. Brief, Richard. 1967. “Measuring Shopping Costs.” Journal of Industrial Economics, 15(3): 237–241. Chen, Zhijun, and Patrick Rey. 2012. “Loss leading as an exploitative practice.” The American Economic Review, 102(7): 3462–3482. Chen, Zhijun, and Patrick Rey. 2019. “Competitive Cross-Subsidization.” forthcoming RAND Journal of Economics. Davies, Gary. 1994. “The delisting of products by Retail Buyers.” Journal of Marketing Management, , (10): 473–493. De los Santos, Babur, Ali Hortaçsu, and Matthijs Wildenbeest. 2012. “Testing models of consumer search using data on web browsing and purchasing behavior.” The American Economic Review, 102(6): 2955–2980. 45

Dubois, Pierre, and Helena Perrone. 2015. “Price Dispersion and informational frictions: Evidence from supermarket purchases.” Dubois, Pierre, and Sandra Jódar-Rosell. 2010. “Price and Brand Competition between Differentiated Retailers: a Structural Econometric Model.” IDEI Working Paper Series, , (64): 42. Dubé, Jean-Pierre. 2004. “Multiple discreteness and product differentiation: demand for carbonated soft drinks.” Marketing Science, 23(1): 66–81. European ern

retail

Commission. on

choice

2014. and

“The innovation

economic in

the

http://ec.europa.eu/competition/publications/KD0214955ENN.pdf

impact EU

of food

(accessed

modsector.”

July

17,

2015). Gentzkow, Matthew. 2007. “Valuing new goods in a model with complementarity: online newspapers.” The American Economic Review, 97(3): 713–744. Hausman, Jerry. 1997. “Valuation of new goods under under perfect and imperfect competition.” In The Economics of New Goods. , ed. R. Gordon and T. Bersnahan, Chapter 5, 207–248. University of Chicago Press. Hendel, Igal. 1999. “Estimating multiple-discrete choice models: an application to computerization returns.” The Review of Economic Studies, , (66): 423–446. Hong, Han, and Matthew Shum. 2006. “Using price distributions to estimate search costs.” RAND Journal of Economics, 37(2): 257–275. Honka, Elizabeth. 2014. “Quantifying search and switching costs in the U.S. auto insurance industry.” RAND Journal of Economics, 45(4): 847–884. Hortaçsu, Ali, and Chad Syverson. 2014. “Product differentiation, search costs, and competition in the mutual fund industry: A case study of S&P 500 Index Funds.” The Quarterly Journal of Economics, 119(2): 403–456. Kim, Jun, Paulo Albuquerque, and Bart Bronnenberg. 2010. “Online demand under limited consumer search.” Marketing Science, 29(6): 1001–1023. Klemperer, Paul. 1992. “Equilibrium product lines: competing head-to-head may be less competitive.” The American Economic Review, 82(4): 740–755. Klemperer, Paul, and Jorge Padilla. 1997. “Do firms’ product lines include too many varieties?” RAND Journal of Economics, 28(3): 472–488. Koulayiev, Sergei. 2014. “Search for differentiated products: identification and estimation.” RAND Journal of Economics, 45(3): 553–575. 46

Lal, Rajiv, and Carmen Matutes. 1994. “Retail pricing and advertising strategies.” Journal of Business, 67(3): 345–370. Lerman, Steven, and Charles Manski. 1981. “On the Use of Simulated Frequencies to Approximate Choice Probabilities.” In Structural Analysis of Discrete Data with Econometric Applications. , ed. Charles Manski and Daniel McFadden, Chapter 7, 305–319. MIT Press. Moraga-Gonzalez, Jose Luis, Zsolt Sandor, and Matthijs Wildenbeest. 2013. “SemiNonparametric estimation of consumer search costs.” Journal of Applied Econometrics, 28(7): 1205–1223. Nevo, Aviv. 2001. “Measuring market power in the ready-to-eat cereal industry.” Econometrica, 69(2): 307–342. OECD. 2005. “Resale below cost.” Policy Roundtables, 231. Pakes, Ariel, and David Pollard. 1989. “Simulation and the asymptotics of optimization estimators.” Econometrica, 57(5): 1027–1057. Shy, Oz. 2002. “A quick-and-easy method for estimating switching costs.” International Journal of Industrial Organization, , (20): 71–87. Thomassen, Oyind, Howard Smith, Stephan Seiler, and Pasquale Schiraldi. 2017. “Multi-Category Competition and Market Power: A model of Supermarket Pricing.” American Economic Review, 107(8): 2308–2351. Viard, Brian. 2007. “Do switching costs make markets more or less competitive? The case of 800-number portability.” RAND Journal of Economics, 38(1): 146–163. Wildenbeest, Matthijs. 2011. “An empirical model of search with vertically differentiated products.” RAND Journal of Economics, 42(4): 729–757.

Appendix A

The utility function of a n-stop shopper

We can give a general expression for the optimal decision rule of a n-stop shopper, n ∈ N =

{1, · · · , Ri }, Ri 6 R, being R the total number of grocery stores in the market, as follows. Assume a n-stop shopper compares baskets of the desired products from all the possible com-

binations of n stores. Denote each of these combinations by j ∈ {1, · · · , Jin }, where according to combinatorics theory, the total number of combinations of R elements taken n at a time is given by Jin = Ri !/n!(Ri − n)! Consumer i will choose the mix j of n stores such that

47

K X

k=1

max{vikrt }r∈j >

K X

k=1

max{vikr0 t }r0 ∈l ∀ l = 1, · · · , Ji

For instance, in a context with R = 3 stores, a one-stop shopper n = 1 will pick the best combination of one store out of Ji1 = 3 possible {A},{B},{C}, and pick the best mix such that it yields the largest overall value of the desired bundle. Similarly, a two-stop shopper, n = 2, will compare all Ji2 = 3 possible combinations of two stores ({A,B},{B,C},{A,C}) and pick the best according to the rule above. For a three-stop shopper, n = 3, the number of combinations of three stores taken three at a time is Ji3 = 1, i.e., {A,B,C} which explains why she is not maximizing over several subsets of stores in equation (3).

B

Robustness checks

In order to check the robustness of our results, we perform three exercises in which we reestimate our demand model with different conditions: first, we try a different specification in which we account for the number of trips to each store in addition to the number of stores visited; second, we estimate our preferred specification with a different selection of products, namely, ham, cheese and pastry; and third, we change the definition of the outside good to check the implications on our demand estimates of our preferred definition. Overall, we find that our results are robust to the definition of shopping patterns, the selection of products and the definition of the outside good.

B.1

An alternative identification strategy for shopping costs

We include an additional variable to our main specification, that accounts for the number of shopping trips that each household makes to the same supermarket chain within a week, and a random coefficient to allow unobserved individual characteristics that explain observed heterogeneity in shopping patterns to enter the model. This implies that the definition of shopping patterns can change with respect to our main model. For example, a household that dealt with a single supermarket chain in a week, regardless of the number of shopping trips it made to the stores of that chain is a one-stop shopper in the context of our main model. However, in the alternative specification in which we account for the number of trips made, a shopper dealing with a single chain but that made a number of trips to the stores of that chain is now a multistop shopper. Similarly to our main case, if a household is observed to deal with multiple supermarket chains but made one trip to each of them is also called a multistop shopper. We report the results of the demand estimates of this new specification in Table B.1. The columns of this table are analogous to those of Table 6. The results of this model are similar to our main specification: the coefficients of price and distance to stores are statistically significant, of the same sign and close in magnitude.

48

Table B.1: Demand estimates using an alternative definition of shopping patterns Covariate Means Price (e/basket) No. of visited stores No. of trips to stores Distance (km) Price × Household head’s age Price × Car Price × Log of income Price × Household size No. of trips to stores × Household head’s age No. of trips to stores × Car No. of trips to stores × Log of income No. of trips to stores × Household size Standard deviations Price No. of visited stores No. of trips to stores Observations

Uncorrected

With control function

-1.763 [1.224 , 2.301] -4.161 [-3.792, -4.529] -4.175 [-4.968, -3.381] -0.0796 [-0.084, -0.074] -0.0246 [-0.045, -0.003] -0.861 [-1.682, -0.0397] -0.0402 [-0.559, 0.479] 0.621 [0.432, 0.809] 0.0227 [0.007, 0.0379] 0.784 [0.248, 1.319] 0.0693 [-0.289, 0.427] -0.0568 [-0.205, 0.092]

-5.840 [-8.448, -8.448] -3.690 [-4.029, -3.350] -1.858 [-2.029, -1.686] -0.0792 [-0.0741, -0.0842] -0.0412 [-0.056, -0.0263] -0.872 [-1.610, -0.133] -0.313 [-0.559, 0.479] 0.691 [0.502, 0.879] 0.00862 [-0.001, 0.0181] 1.146 [0.640, 1.651] 0.0460 [-0.232, 0.324] -0.176 [-0.290, -0.0619]

2.038 [1.757, 2.318] 1.557 [1.349, 1.764] -0.899 [-1.001, -0.796]

2.781 [2.506, 3.055] 1.062 [0.933, 1.190] -1.129 [-1.242, -1.015]

2,556,265

2,556,265

Notes: The 95% confidence intervals are given in square brakets. All regressions include product, store and time (months) fixed effects. A basket can consist of unit servings of one, two or three products: desserts (80 g), biscuits (30 g) and, yogurts (125 g).

B.2

Demand estimates with an alternative set of products

We reestimate our demand model with this alternative set of products. We report the results in Table B.2, which is analogous to Table 6 of the main text. Results are similar to those obtained with our selected basket. The coefficients for price, the number of visited stores and distance are of the same order of magnitude as those of our main results. As a consequence, the estimates of mean shopping cost, 81.9 euro cents, and mean transport cost, 9 euro cents, are close to their counterparts obtained from our main sample of 96 and 8.64 euro cents, respectively. Our results are robust to the selection of product categories.

49

Table B.2: Demand Estimates Covariate Means Price (e/basket) No. of visited stores Distance (km) Price × Household head’s age Price × Car Price × Log of income Price × Household size No. of visited stores × Household head’s age No. of visited stores × Car No. of visited stores × Log of income No. of visited stores × Household size Standard deviations Price No. of visited stores Observations

Uncorrected

With control function

-1.671 [-2.112, -1.231] -4.735 [-5.109, -4.361] -0.0693 [-0.074, -0.065] 0.00830 [-0.008, 0.025] 0.897 [0.254, 1.539] 0.863 [0.529, 1.198] -0.231 [-0.436, -0.027] -0.0112 [-0.023, 0.001] 0.285 [-0.357, 0.926] -0.179 [-0.535, 0.177] 0.380 [0.240, 0.521]

-5.603 [-8.409, -2.797] -4.589 [-4.939, -4.240] -0.0702 [-0.075, -0.066] 0.0242 [0.012, 0.037] 1.443 [0.843, 2.044] 0.676 [0.378, 0.975] -0.162 [-0.297, -0.028] -0.0247 [-0.035, -0.015] 2.315 [1.771, 2.858] -0.425 [-0.707, -0.144] 0.0919 [-0.017, 0.201]

1.572 [1.431, 1.713] 2.174 [1.993, 2.355]

1.621 [1.515, 1.727] 1.937 [1.808, 2.066]

1,647,296

1,647,296

Notes: The 95% confidence intervals are given in square brackets. All regressions include product, store and month fixed effects. A basket can consist of unit servings of one, two or three products: ham (42.5 g), cheese (80 g) and, pastry (40 g).

B.3

An alternative definition of the outside good

We estimate the model with a different definition of the outside good. The definition of the outside option restricts our sample to the consumers who were observed to make purchases only from the three supermarkets included in our main analysis, which implies that we are not excluding any observed stops. In this case, we define the outside good as all of the purchases made from one of the three included supermarkets. We report the results in Table B.3. The estimates for price, distance and shopping costs are of the same order of magnitude to those obtained with our main model.

50

Table B.3: Demand Estimates Covariate Means Price (e/basket)

With control function

No. of visited stores Distance (km) Price × Household head’s age Price × Car Price × Log of income Price × Household size No. of visited stores × Household head’s age No. of visited stores × Car No. of visited stores × Log of income No. of visited stores × Household size Standard deviations Price

-5.951 [-8.725, -3.124] -4.456 [-5.193, -3.648] -0.0919 [-0.089, -0.078] -0.0406 [-0.057, -0.020] -0.956 [-2.087, 0.104] -0.0456 [-0.617, 0.533] 0.628 [0.385, 0.835] 0.00803 [-0.026, 0.035] 0.228 [-1.393, 2.301] 0.344 [-0.616, 0 .533] -0.221 [-0.662, 0.1304] 1.944 [1.693, 2.178] 1.916 [1.151, 2.519]

No. of visited stores Observations

899,010

Notes: The 95% confidence intervals are given in square brackets. All regressions include product, store and month fixed effects. A basket can consist of unit servings of one, two or three products: desserts (80 g), biscuits (30 g) and, yogurts (125 g).

51