Store Brands and Retail Grocery Competition in Breakfast Cereals

Store Brands and Retail Grocery Competition in Breakfast Cereals Rong Luo The Pennsylvania State University April 15, 2013 Abstract This paper empirically analyzes the impacts of store brands on grocery retailers and consumers in the market for breakfast cereals. On the supply side, store brands help a retailer to avoid direct competition with other retailers and change the set of retailer s products. On the demand side, introducing store brands changes the national brands prices and consumers choice set. We analyze the effects via demand estimation for a single grocery store chain Dominick s at Chicago in 1997 and counterfactual exercises. The estimation results show that consumers unobserved utility of buying at a competing retailer is higher for consumers that value national brands, and is lower for ones that value Dominick s store brands. This is consistent with the claim that store brands help a retailer to avoid competition. The counterfactual calculations show that the profit loss from removing store brands is higher if the retailer has more competitors, with a median loss of 4.33% of profits from cereals. Existence of store brands increases national brands prices and consumer welfare increases slightly when store brands are removed. Keywords: store brands, national brands, competition, nested logit model, random coefficient model The author is grateful to Mark Roberts, Paul Grieco, Peter Newberry for advising my research process and helpful suggestions. All errors are the authors own responsibility. Department of Economics, Pennsylvania State University, University Park, PA 16802. Comments are welcome. 1

1 Introduction This paper studies grocery retailer s profits on store brands of cereal when the degree of competition among retailers varies. Competition among retailers is direct in the sense that they sell goods produced by the same manufacturers. Introduction of store brands helps a retailer to avoid direct competition because store brands are sold exclusively by the corresponding retailer. We emphasize the following substitution pattern in this paper. For a retailer A selling both national brands and store brands of cereal, entry of competing retail stores has different effects on the demand for national brands and store brands. Given that competing retailers sell the same set of national brands, A s national brands consumers may switch to competing stores since they can buy exactly the same products there. But A s store brands consumers can t find the exact same products in competing stores. That is to say, competing retailers sell products that are better substitutes for A s national brands than A s store brands. This substitution pattern justifies a retailer s decision of introducing store brands. As more retailers enter to share the market, a retailer may use store brands to lock-in some consumers when its market share decreases with competition. As a result, we expect store brands to be more important in markets where the degree of competition is high. In this paper, we empirically analyze the relationship between a retailer s profit gain from introducing store brands and its competition with other retailers. There are several reasons why the analysis of store brands demand and the relationship between store brands profit and retailers competition is important. First, the demand side welfare effect of store brands can t be neglected. Hong and Li (2013) documents that the market share of private label grocery products (mainly store brands) is about 20% in the United States in 2012, 35% in Europe and 50% in Britain in 2011. So store brands are more and more important in retail markets. So are the welfare effects. Second, measuring the profit effect of introducing store brands is critical to retailers. Introducing store brands is a strategy for retailers to avoid competition and increase profit. So it is important to predict whether and by how much store brands will increase profits. 1 literature has not studied store brands and competition very well. So we believe it is 1 This importance is not limited to retail markets. For most multiproduct firms, measuring the profit of introducing a new product that is a substitute for existing products has significant effect on market structure. For example, companies in industries like electronics and automobile frequently introduce new products, the effects of which can be studied by similar methods as developed in this paper. The 2

important to pay more attention to this topic. This paper uses a structural model of demand and data on ready-to-eat cereal products to analyze store brands profit effect and its relationship with competition among retailers. In the demand side model, to allow flexible substitution patterns across products and heterogeneity in consumer choices, we use the random coefficient demand model as used by Nevo (2000). One difference between the demand model in this paper and Nevo(2000) lies in the mean utility of outside option. Since data for only one chain retailer (Dominick s) stores is available, we define the outside option as not buying from Dominick s (including buying from other retailers and not buying ready-to-eat cereal). So in order to capture the competition effect, we assume the mean utility of outside option depends linearly on number of competitors. If the coefficient is positive, then additional competitors will increase the mean utility of outside option, and lower the market share of Dominick s. The supply model here is simple. We assume that each Dominick s store chooses the optimal prices for all brands. Cohen and Cotterill (2011) justifies this assumption by comparing different pricing models of milk products and finds that retailers maximize profits taken the wholesale prices as given. Given the demand estimates, each store knows the cross- and own- price elasticities of substitution of all products and solves its profit maximization problem. Given the observed prices are in equilibrium the cost of each product can be calculated from retailer s first-order conditions. The costs will be used in counterfactual analysis. The demand estimation results imply that consumers unobserved utilities of national brands and the outside option are positively correlated, with correlation coefficient 0.99. And the unobserved utilities of store brands and the outside option are negatively correlated, with correlation coefficient.44. These correlations explain the substitution patterns across national brands, store brands, and the outside option described earlier. The first counterfactual experiment calculates the effect of removing store brands. The results show that the median profit loss of removing Dominick s store brands readyto-eat cereals is $108 per store every week, which is 4.33% of profits in the ready-to-eat cereal category. When the number of competing retailers increases by 1%, the profit loss increases by 0.16%. If two Dominick s stores S1 and S2 have 100 and 101 competitors respectively, then the percentage profit loss of S2 is 0.15% higher than S1. Therefore, this counterfactual analysis implies that retailer s profit decreases from removing store 3

brands and the profit loss increases significantly as competition increases. In addition, the welfare of consumers increases after removing store brands. To clearly quantify the effects of competition, the second experiment assumes that each Dominick s store has 5 more competing retailers. The results show that the median profit loss is $502 per store every week, which is 20.21% of the ready-to-eat cereal profit. More importantly, Dominick s reduces prices of all brands to keep its consumers. On the consumer side, welfare increases when competition increases. The rest of the paper proceeds as follows. Section 2 describes related literature about store brands effects, demand estimation, and retailer-manufacturer relationship. In section 3, we introduce the data I used and some motivation facts on the substitution patterns across brands. The demand model and supply model are presented in section 4. Section 5 discusses the identification of parameters and estimation strategy of the demand model. The effects of two counterfactual cases, removing store brands and increasing competitors, are calculated in Section 6. The final section concludes the paper. 2 Literature There are several papers analyzing the effects of introducing store brands, but none has studied the horizontal competition effect of introducing store brands, which is also very important in addition to the vertical effects. Chintagunta, Bonfrer and Song(2002) analyzes the demand side price elasticity changes and supply side retailer-manufacturer relationships. Their test can t reject the hypothesis that there is no change in price e- lasticities after introducing store brands. On the supply side, they compare the markups before and after introducing store brands, and find that manufacturers taking a softer stand after store brands are introduced. Steiner (2004) theoretically argues that store brands are competitive weapons and consumer welfare is maximized when store brands and national brands compete vigorously. Cohen and Cotterill (2011) assesses the impact of retailer store brand products on manufacturer brand prices, profitability and consumer welfare. They calculate manufactural margins implied by different supply models between manufacturers and retailer and regress the implied costs of the models on input prices to choose the best fit supply model. They find that national brands prices are lower and consumer welfare is higher with store brands. In addition to analyzing prices and welfare effects of introducing store brands, our paper also emphasizes the relationship between 4

introducing store brands and competition among retailers, which hasn t been studied in the literature. Hong and Li (2013) analyzes the cost pass-though of national brands and private label products and finds that the pass-though rate of private labels is 40% higher than national brands. While we are also studying pricing of store brands, the purpose of our paper is quite different. The demand model and estimation strategy in this paper follow very popular specification in the literature, like BLP(1994), Nevo(2000), Petrin(2002), and so on. But there is one important distinction with the literature. To incorporate the effects of local competition, we allow the mean utility of outside option to depend on number of competitors. In contrast to being normalized to zero, the utility of the outside option in this paper is expected to increase with competition of retailers. If there are more retailers within a geographic area, consumers utility of not buying from Dominick s increases. So Dominick s consumers switch to the outside option as competition increases. Though including competition in the outside option utility is not an ideal way to model retailers competition, it is reasonable given we only have data for Dominick s stores and number of competing retailers. Analyzing effects of introducing store brands is also related to the vertical integration literature. Without store brands, retailers buy products from a manufacturer at wholesale prices. Theoretically, store brands eliminate the double marginalization problem of national brands, lower the national brands prices, and raise consumer welfare. However, Hastings(2004) analyzes the gasoline market in South California and finds that vertical integration significantly increases prices. The reason is that the vertical integration in her paper reduces competition among stations instead of increasing competition because large refining companies now control competing independent stations. In our paper, introducing low cost store brands makes the retailer to differentiate consumers, so the high quality national brands prices increase. Villas-Boas(2007) compares margins implied by different supply models between manufacturer and retailer to select the model most consistent with data. While comparing different supply models is important, our paper assumes reasonable pricing behavior of the retailer and focuses on relationship between vertical integrated store brands and retailer competition. 5

3 Data and Facts 3.1 Data and Ready-to-eat Cereal Market We analyze the ready-to-eat cereal products in Dominick s chain stores in this paper. Dominick s is a chain grocery store in Chicago area. The data used in this paper has three parts. The first part is store level weekly scanner data, which records manufacturer, price, quantity sold, markup and size of every product sold in each Dominick s store every week. This data also provides demographic characteristics for the market of each store. The second part is data on retailer competition. We find the number of food retailing stores at zip code level from 1997 U.S. Economic Census report, and number of population at zip code level from Illinois state census. The last part is the product characteristics data, which we collected manually from Mrbreakfirst s website. We consider the most important 20 brands sold in every Dominick s store. A brand here corresponds to a manufacturer-shape pair. There are four shape types: ring, flake, shredded and other. For example, all Cheerios by General Mills are combined into one brand. One reason to aggregate products in this way is to balance choice sets for consumers of different Dominick s stores. The product characteristics contain information on the grain type dummy, fruit dummy and nut dummy of each brand. Since products are combined into brands, we use weighted averages for brand characteristics. A market in this paper is defined as a store in a week. For example, the same store at two different weeks are treated as two different markets. In 1997, Dominick s has 86 stores, and almost all stores are located in areas with different zip codes. Eighteen stores are located in the downtown Chicago area, where the number of competitors and population data are the same and large for all stores. We drop observations of all stores in downtown Chicago area which leaves us data for 68 stores. To match the data of competition and population, we only use Dominick s data in 1997, which has 28 week observations for all stores. To avoid consistent prices in consecutive weeks, we only use data for the first week of each month during 1997. So the number of markets decreases to 288. To further restrict the set of products to be the same for every market, we select stores that carry all the 20 brands defined above. In the end, there are 285 markets. The outside option is defined as not buying from a Dominick s store. For each consumer in a market, the outside option includes two possibilities, buying from a retailer 6

other than Dominick s and not buying ready-to-eat cereal at all. To calculate the market shares of the outside option, we assume each person s potential demand for ready-to-eat cereal is 1 pound per week 2. This leads the outside option s market share belongs to [33%, 99%], with median 91%. Dominick s stores sell national brands and store brand cereals. The main national brands are from manufacturers Kelloggs, General Mills, Post, Quaker, and Raltson. Store brands products have lower costs. The average Dominick s store brands cost is $1.58 per pound. But lowest national brands average cost is $2.35 per pound for Post and the highest is $3.35 for Ralston. Store brands prices are also lower. The average price for store brands is $2.60 per pound. But the lowest national brands average price is $2.97 for Post, and the highest is $4.23 for Ralston. While store brands have low prices, their market shares are quite low, 5.54% on average. So though store brands are available at all Dominick s stores, they are not good substitutes for national brands. In next section, we show some facts about the substitution pattern across brands and its relationship with retailer competition. 3.2 Facts about Substitution and Competition In this section, we first introduce summary statistics of the data, then reduced form regressions are used to show retailer competition s different effects on national brands and store brands. Table 1 shows mean prices, markups and market shares of all brands across markets. Market shares in the table are shares within each Dominick s store. Dominick s store brands have low prices (as low as $2.24 per pound), high markups and low market shares. Ralston brands have high prices (as high as $4.75 per pound) and low market shares. General Mill s ring brand and Kellogg s flakes have the highest market shares, with 11.61% and 14.04% respectively. Market shares do not change monotonously with prices. So besides prices, product characteristics also play an important role in consumer demand. 2 Granola and Breakfast Facts in Year 2006 stated that Americans consume 10 pounds per person per year, which means that actual consumption per person was 0.2 pound per week and potential demand should be higher than that. Assuming potential demand lower than 1 lb/week causes unreasonably negative market shares of the outside option in some markets. 7

Table 1: Product Summary Statistics Retail Price($/lb) Wholesale Cost Markup Within Store shares (%) Dominick s ring 2.58 1.55 39% 1.29 Dominick s flake 2.24 1.36 39% 2.93 Dominick s shred 2.94 1.90 39% 0.60 Dominick s puff 2.70 1.56 39% 0.54 General Mills ring 3.36 2.68 23% 11.61 General Mills flake 3.28 2.73 18% 9.48 General Mills shred 3.50 2.90 19% 2.47 General Mills puff 3.87 2.92 19% 4.37 Quaker flake 3.16 2.45 21% 1.85 Quaker shred 2.49 1.94 21% 4.11 Quaker puff 2.91 2.13 27% 4.77 Kellogg s ring 3.44 2.60 21% 4.77 Kellogg s flake 2.93 2.37 19% 14.04 Kellogg s shred 2.75 2.07 22% 10.28 Kellogg s puff 3.55 2.92 14% 7.83 Post flake 2.63 2.02 21% 8.07 Post shred 2.90 2.28 21% 4.09 Post puff 3.17 2.44 21% 3.69 Ralston shred 3.84 2.70 27% 2.43 Ralston puff 4.75 3.82 23% 0.77 The ready-to-eat cereal market is highly concentrated. The five biggest national manufacturers are: Kelloggs, General Mills, Post, Quaker and Raltson. According to the sales data of all Dominick s stores in 1996, the total market share (in terms of quantity sold) of the five companies is 94.91%. The decomposed shares are: 36.92% for Kelloggs, 27.93% for General Mills, 15.85% for Post, 10.73% for Quaker and 3.20% for Ralston. The market demographics we use are population, number of food retailers, logarithm of median income, age distribution statistics, and percentage of people owning a car in every market. Table 2 summarizes the market characteristics. The mean number of food retailers within a zip code area is 17.56, and the mean outside option share is 86.40%. Next, we use simple regressions to see the different effects of retailer competition on store brands and national brands. Let j = 0, 1,..., J denote different brands of cereal products, i = 1, 2,..., S denote consumers, and t = 1, 2,..., M denote markets. The first regression is as following: m jt = b 1 R t + b 2 Inc t + ϵ 1jt 8

Table 2: Market Demographic Statistics mean standard deviation Log Median Income 10.73 0.22 Population 36464 28419 Age< 9 Percentage 0.14 0.02 No car Percentage 0.05 0.05 # Competitors 17.56 11.18 Outside share 86.40% 12.18% where m jt is the markup of product j in market t, R t is the number of supermarkets in market t and Inc t be logarithm of median income in market t. m jt is defined as the ratio of gross profit per pound over price per pound, which is observed from data. For each product, gross profit is equal to revenue from sales of the product minus the wholesale cost of the product. Other than wholesale cost, there is distribution cost for each product, which will be calculated later. For example, m jt =.3 means that for every dollar revenue, $.30 is gross profit, and $.70 is wholesale cost. Table 3: Markups and Competition ˆb 1 ˆb2 Dominick s 0.0047 0.7021 GeneralMills 0.0045 0.7503 Quaker 0.0050 0.9814 Kellogg s 0.0028 1.0157 Post 0.0090 1.0373 Ralston 0.0102 0.4394 As table 3 shows, markups of most national brands decrease significantly with competition, while store brands markups don t decrease with competition. This fact implies that competing retailers affect the profitability of national brands more than that of Dominick s store brands. This is reasonable given that all competitors can sell national brands, but not store brands of Dominick s. Second, we regress the market (quantity) share of different brands within each store on number of competitors. log q jt q t = b 3 logr t + b 4 log p jt + ϵ 2jt 9

where q jt is the quantity sold of brand j in market t, q t is total quantity sold in market t, and p jt is price of brand j in market t. Table 4: Quantity Shares and Competition ˆb 3 ˆb4 Dominick s.0363 0.2644 GeneralMills.0194 2.3521 Quaker.0210 2.7443 Kellogg s.0032 2.4333 Post.0103 2.7393 Ralston.0265 2.8952 Table 4 shows that the quantity shares of store brands increase significantly with number of competitors. This is consistent with the reason about competition stated above. When number of competitors increases, previous Dominick s customers of national brands may switch to other stores, because these stores also sell national brands. But Dominick s store brands consumers do not switch to competing stores because store brands are unique to Dominick s. Last, we regress the gross profit share of different brands on number of competitors. log π jt π t = b 5 logr t + ϵ 3jt where π jt is gross profit from brand j in market t, and π t is profit from market t of all brands. So π jt /π t is the profit share of brand j in market t. Table 5 shows the simple regression results. As we can see, the gross profit share of store brands is low relative to national brands, but increases significantly with number of competitors, while national brands profit shares decrease except for Post brands. Again, these regression results are consistent with the competition effects stated above. In summary, the facts shown above imply that when the number of competitors increases, national brands are affected in the opposite direction from Dominick s store brands. 10

Table 5: Gross profit shares and Competition Gross Profit Share ˆb5 Dominick s 6.62% (3.45%).0435 GeneralMills 31.82% (22.24%).0133 Quaker 13.64% (16.07%).0277 Kellogg s 28.58% (19.41%).0045 Post 11.49% (6.29%).0108 Ralston 7.85% (13.36%).0240 National brands markups and profit shares decrease with competition, while store brands profit share and quantity share increase with competition. This is due to the fact that store brands are sold only in Dominick s, while national brands are sold by all competing retailers. So retailers can use store brands to lock-in consumers and gain profit as more competitors enter the market. This is the question we are trying to answer in this paper -the relationship between store brands profit and retailer competition. We describe the demand and supply model in the next section. 4 Model In this section, we model a consumer s discrete choice over different brands and a retailer s profit maximization problem. For the demand model, we use random coefficient model, which is used in BLP(1995) and Nevo(2000). In the random coefficient model, consumer s utility contains two parts, the first is the mean utility of all consumers in the market, and the second part is the individual deviation from the mean, which depends on consumer characteristics. The outside good is defined as not buying from the 20 brands in Dominick s. So if a consumer buys cereal from other retailers or does not buy cereal at all, we interpret that he or she chooses the outside option. Defined in such a way, the outside good needs to capture the competition effect of other local retailers. Mean utility of the outside option is a linear function of number of competitors and local median income. The more competitive the market is, the higher utility the outside good provides. For the supply side, we assume that the retailer has monopoly power to choose prices for both store brands and national brands. The retailer chooses prices to maximize profit, given consumer s utility functions. For each price vector of both national brands and store brands, there is a corresponding market share vector of all brands. 11

4.1 Demand Model In the random coefficient model, consumers have heterogenous preferences over prices and brands. These heterogenous coefficients depend on individual demographics and random realization of unobserved shock term. The unobserved shock terms in the random coefficients of brand dummies can be correlated across different brands. If the correlation between national brands and the outside option is high, then when a consumer has high realization of random coefficient on national brands, he is also likely to have a high shock on the outside option coefficient. This feature of random coefficient model fits the facts we observed from data. If national brands and the outside option are highly positively correlated, then when there are more competitors so the mean utility of outside option increases, consumers of Dominick s national brands are likely to switch to the outside option. Let j = 0, 1,..., J denote different brands of cereal products, i = 1, 2,..., S denote consumers, and t = 1, 2,..., M denote markets. Products are differentiated in their characteristics. j = 0 denotes the outside option. Let x j be the vector of product characteristics of product j, which is same across markets. Let ξ jt be unobserved product characteristics of product j in market t and ϵ ijt be unobserved individual utility of product j in market t. We assume consumers have correlated heterogenous preferences on different brands. If we use brand dummies, we will need to estimate a 20 by 20 dimension covariance matrix. So instead of brand dummies, we use group dummies. There are three product groups: store brands, national brands and outside option. Let d j be the vector of dummies for the three groups. For example, if j is a national brand, then d j = [0, 1, 0], and if j is outside option, then d j = [0, 0, 1]. Hence, consumer i s utility of product j in market t is : u ijt = x j β + α i p jt + γ i d j + ξ jt + ϵ ijt, for j = 1, 2,..., J (1) β is the coefficients on product characteristics, which is the same across consumers. In BLP(1995), the product characteristic coefficients are individual specific for the automobile industry. Consumers with different demographic characteristics may value auto characteristics very differently. However, for ready-to-eat cereal, the product characteristics are grain type, fruits and nuts dummies, whose values do not seem to vary across 12

individuals. So we assume β to be the same across all consumers. This reasonable assumption also decreases the number of parameters to be estimated. α i is consumer specific. For example, consumers with different income levels have different disutility on price. γ i is a vector of group dummy coefficients, and depend on consumer characteristics.. To allow competition to affect Dominick s profit, we assume that competition affects consumers mean utility of the outside good. Intuitively, consumers utility of consuming the outside good increases when there are more competitors. So we assume the utility of outside option in market t to be: u i0t = η R R t + η I Inc t + ϵ i0t, for j = 0 (2) η R and η I are coefficients on number of competing retailers and logarithm of median income. If η R is positive, then as competition increases, the market share of Dominick s will decrease. We assume that ϵ ijt has Gumbel distribution (type-1 extreme value distribution). α i γ is γ in γ i0 α γ s = + Π γ n 0 Inc i Inc 2 i + D(age < 9) D(car) ν ip ν is ν in ν i0 (3) [α; γ s ; γ n ; 0] is the mean coefficients on group dummies without consumer demographics. The mean of random coefficient of the outside dummy is normalized to be 0. Matrix Π has all coefficients on consumer characteristics. Each individual is characterized by D i. We use income, age dummy and car dummy for consumer characteristics. Let F Dt ( ) be the distribution function of consumer characteristics in market t. The unobserved part of the coefficients is [ν ip ; ν is ; ν in ; ν io ], which is distributed normal with mean [0; 0; 0; 0] and covariance matrix Σ. Let F ν be the joint distribution function of vector ν. We assume the price coefficient shock ν ip is independent with the ν is, ν in and ν io. Then the covariance matrix is: σp 2 0 0 0 0 σs 2 σ sn σ so Σ = 0 σ ns σn 2 σ no 0 σ os σ os σo 2 The covariance matrix can be written as the product of an lower-triangular matrix and 13

its transpose: Σ = L L, in which L = l p 0 0 0 0 l s 0 0 0 l ns l n 0 0 l os l on l o This decomposition is used to facilitate estimation. The elements in matrix L do not have economic meaning as those in Σ. We can rewrite consumer s utility in the following form: u ijt = δ jt + µ ijt + ϵ ijt (4) where δ jt = x j β + αp jt + γd j + ξ jt µ ijt = [p jt d j] (ΠD i + ν i ) δ jt is the mean utility of brand j in market t. µ ijt is individual deviation from mean utility. Given the distribution of ϵ ijt, we can calculate each individual s probabilities of choosing all brands. For consumer i, he chooses product j if u ijt u ikt, for any k = 0, 1,..., J. Define set of consumer draws such that brand j gives maximum utility to them. A jt = (D i, ν i ) u ijt >= u ikt, for all k = 0, 1,..., j Then market share of brand j is s jt = df Dt df ν (5) A jt The random coefficient model allows a very flexible substitution pattern. Conditional on product characteristics and prices, consumers demand elasticity is high between brands that have high correlation in ν. So if σ no is high, then Dominick s national brands consumers are likely to switch to outside option when outside option s mean utility increases. 14

4.2 Supply Model In this paper, we assume each Dominick s store chooses optimal prices for all products it carries. This assumption is different from papers that examine the pricing strategy of firms. In Nevo(2000), three competition models between manufacturers (single product firms, multi-product firms, and collusion between firms) are compared to match predicted markups. In Villas-Boas(2007), three vertical contract models between manufacturers (linear pricing model, vertically integrated model and nonlinear pricing model) are compared to match estimated cost and margins. Unlike Nevo(2000), we don t need to analyze competition between manufacturers, because the retailer Dominick s sells all products by different manufacturers. Unlike Villas-Boas(2007), the focus of this paper is not a comparison of the effects of different vertical models, but rather the downstream retailer s store brands effect and competition with other retailers. So it is reasonable to assume that Dominick s stores are free to choose optimal prices. We don t estimate the supply side game because we don t observe prices of other retailers than Dominick s in every market. And we don t compare different contracts between manufacturers and retailers, because the vertical relationship is not the focus of this paper and the assumption that retailers take wholesale prices as given is justified by Cohen and Cotterill (2011). For each product, there is a wholesale cost c jt and a distribution cost w jt. Wholesale cost is the price paid to manufacturers. Distribution cost includes the wage paid to store managers and other employees. Let M t be the market size of market t. Then each store s problem is to maximize total profit from all brands every week. max p t J (p jt c jt ω jt ) s jt (p jt )M jt (6) j=1 The first order conditions of the retailer s maximization problem is: s jt + J k=1 (p kt c kt ω kt ) s kt(p t ) p jt = 0, for j = 1, 2,..., J (7) Let Ω be the matrix of price elasticities. Its (i, j)th element is s it(p t) p jt. Then the vector of 15

distribution costs can be calculated from the FOCs. ω t = p t c t + Ω 1 s t (8) where the variables with subscript t are vectors of all brands in market t. This supply side model is quite standard except for the distribution cost. The distribution cost here is important to rationalize the prices of each store. In Nevo(2000) and Villas-Boas(2007), because cost data is not available, price-cost-margins are calculated using demand side estimates and firm s first order conditions. However, in this paper, both retail prices and wholesale prices are observed and the demand side will be estimated separately. If we also calculate the price-cost-margins and costs as these papers did, the results may be inconsistent with observed wholesale costs. One explanation for this inconsistence is that the wholesale prices are not true costs. First, there may be measurement error in the cost data. Second, in addition to the wholesale costs, there are other costs to be able to sell the products. So introducing the distribution costs justifies the observed prices given the observed wholesale costs. The next section discusses the estimation strategy of the demand model, and the identification of the parameters. 5 Estimation 5.1 Instruments and Estimation The estimation strategy follows Nevo(2000). We use generalized methods of moments estimation. The moment conditions are based on the orthogonality between unobserved product utility and instruments. In the random coefficient demand model, the parameters to be estimated are θ = (β, α, η, γ, Π, Σ). θ has a linear part θ 1 and a nonlinear part θ 2. θ 1 = (β, α, η, γ) and θ 2 = (Π, Σ). The parameters in θ 1 enter the utility function linearly, while parameters in θ 2 interacts with distributions of observed or unobserved terms in utility function. The moment conditions are the orthogonality between unobserved product utility and instruments: E(ξ jt z jt ) = 0. The instruments used here are lagged prices and lagged wholesale costs. Prices are correlated over time, but current demand shocks are indepen- 16

dent of lagged prices. Lagged costs are also correlated of current prices, but independent with current demand shocks. ( ) The GMM objective function is to minimize ξz 1 ZZ Zξ over θ. To calculate the objective function for a given parameter θ, we use data and the model s equations to calculate ξ. First, we simulate S number of consumers in each market t according to the distributions of demographics. Specifically, for a given θ 2 and mean utility δ, we can calculate each consumer s probability of choosing each product. Then we take the averages to be the predicted market shares of each product given θ 2 and δ. Next, we match the predicted market shares and observed market shares to solve for mean utilities δ, for the given θ 2. After calculating the mean utilities, we estimate the parameter θ 1, which enter linearly in δ. The residuals of this regression is the value of ξ for the given θ 2. So we can find the value of objective function for each given parameter θ 2. Then we can search over θ 2 to find the estimates that minimizes the objective function. 5.2 Identification The identification of the demand model is quite clear. (β, α) are identified by products that differ in their characteristics and prices. Suppose that two products are exactly the same except that one has fruit and the other one does not. The difference in their market shares imply the value of β fruit. If β fruit = 0, then they will have the same market share. As β fruit increases, the difference between the products increases. The identification for α can be explained in a similar way. γ n and γ s are identified by the mean utility of store brands and national brands, controlling for the product characteristics, number of competitors and log of median income. The nonlinear parameter is θ 2 = (Π, Σ). Π is the matrix of coefficients on consumer demographics, and it is identified by the market shares of markets with different demographics. Suppose two markets have the same prices for all products, same age distribution and same percentage of people having a car. Then the difference between the two markets product shares identifies the coefficients on income. If the coefficients on income are zeros, then the two markets will have the same market shares. But if the coefficients on the interaction of income and price is positive(less elastic), then the market with higher level of income will have higher market shares of the more expensive products. Similarly, if the coefficients on the interaction of income and national brand group dummy is posi- 17

tive, then market with higher level of income will have higher market shares of the same national brands. The identification of all parameters in Π can be explained in this way. The diagonal terms of Σ are identified by consumers loyalty to different groups. The higher the diagonal terms are, the more loyal the consumers are to these groups. For example, if σ n is high, then consumers are very heterogenous in their evaluation of national brands and this will lead consumers to substitute into similar products when price changes. These consumers are not very likely to switch to store brands or the outside option when national brands prices increase. If σ n is low, then unobserved utility is not an important part of consumers utility of national brands. In this case, when national brands prices increase, the consumers are likely to switch to store brands or the outside option. The extent to which national brands market share changes when national brands prices change identifies σ n. The identification of σ s and σ o can be explained similarly. In BLP(1995), the variances are estimated while the covariance terms are assumed to be zero. However, the covariance terms are critical in our paper. We want to emphasize the substitution patterns that are determined by correlation of peoples preferences across group dummies, which can t be a common product characteristic for brands in different groups. The off-diagonal terms determine the substitution patterns across groups, and they are identified by the asymmetric substitution patterns of store brands and national brands. Suppose σ no = 0, and purchases of national brands and store brands of Dominick s decrease when the number of competing retailers increases. If σ no > 0, then national brands consumers are likely to prefer the outside option and the market shares of national brands will decrease more than they do when σ no = 0. Similarly, if σ so < 0, then store brands consumers are not likely to prefer the outside option. Then market shares of store brands will decrease less than the case when σ so = 0. We can use similar arguments when the mean utilities of store brands and national brands change, and these arguments strengthen the identification of the correlations. 6 Estimation Results This section shows the estimation results of the random coefficient model. To compare results, we also estimate the model with simple Logit model with instruments. The simple Logit model can be estimated using a linear IV estimation. In the random coefficient model, we set the interaction terms between price and squared income, between national 18

brands group and squared income, between outside option and squared income, between price and age, and between national brands group and age to be zero. Table 6: Demand Estimation Results Logit-IV Random Coefficients β wheat -0.78 (0.08) -0.75 (0.23) β oat -0.48 (0.10) -0.48 (0.19) β bran 1.20 (0.20) 1.95 (0.56) β corn/rice 0.70 (0.05) 0.71 (0.13) β fruit -0.25 (0.06) -0.25 (0.17) β nut 2.59 (0.15) 2.60 (0.30) α -1.17 (0.09) -2.19 (0.12) γ store -6.11 (0.64) -1.77 (0.56) γ national -3.8 (0.68) 1.31 (0.48) η R 0.06 (0.00) 0.05 (0.00) η Inc -0.26 (0.06) -0.22 (0.00) l p -0.01 (0.11) l n 2.66 (2.00) l s -3.66 (1.06) l o 2.61 (1.58) l no 7.44 (0.80) l so -0.06 (2.86) l ns -0.21 (4.36) The demand estimation results of the two models imply significant coefficients of price. The Logit model underestimates the price coefficient because consumers who are less price elastic are not differentiated with more elastic consumers. The Logit model doesn t allow consumers characteristics to affect price coefficient. So high income consumers who have small price coefficients are mixed with lower income consumers who have larger price coefficients. As a result, price coefficient in Logit model is underestimated. The coefficient on the number of competitors in outside option s mean utility is also significant and positive. So as the number of competitors increases, the mean utility of outside option increases. In both models, the mean utility of outside option increases as local median income rises. The estimation results show the estimates of elements in the lower-triangular matrix L. All elements but the last two are significant. We set the last two entries to be zero 19

when calculating the covariance matrix. The implied covariance matrix is: ˆΣ = 0.001 0 0 0 0 7.06 9.72 6.93 0 9.72 68.72 9.54 0 6.93 9.54 6.80 The variance of unobserved utility for store brands is 68.72. This high variance means that store brands consumers are very loyal to Dominick s store brands. These consumers are not likely to switch to national brands or the outside option when store brands prices change. The variances for national brands and the outside option are lower, meaning that consumers of national brands and the outside option are likely to switch to other groups when prices change. Let ρ denote the implied correlation between groups. In the random coefficient demand model, the estimates of the covariance matrix implies that correlation between random shocks on national brands and outside option is ρ no = 0.99. So consumers with a strong preference for national brands will also have a strong preference for the outside option. The estimated correlation between the random coefficients of store brand dummy and outside option dummy is ρ so = 0.44. Then if a consumer has high value of store brands, he is likely to have low value of the outside option. So Dominick s national brands and the outside option are close substitutes for each other. In addition, ρ ns = 0.44, which means that consumers preferences for Dominick s national brands and store brands are negatively correlated. Given these estimated results, retailers competition, which changes the utility of outside option, will substantially affect Dominick s national brands demand. The estimates of parameters in Π are shown in the table 7. The positive coefficient of interaction term between price and income level is positive, which means that the higher income the consumer has, the less dis-utility he gets from prices because we estimate negative price coefficients. The own- and cross- elasticities are shown in following table. elasticities are higher within the groups than across groups. We find that price 20

Table 7: Demand Estimation Results Random Coefficient Model(std) π price,inc 0.12 (0.07) π price,car 1-0.32 (0.10) π national,inc -4.09 (0.19) π national,car 1-0.21 (0.23) π store,inc 8.31 (0.07) π store,inc 2-1.35 (2.03) π store,age<9 6.01 (4.61) π store,car 1-2.22 (1.03) π out,inc -3.94 (2.44) π out,age<9 0.35 (1.41) π out,car 1-0.56 (2.63) 7 Counterfactual Comparison 7.1 Without Store Brands One question that we are interested in is whether store brands are more important in markets with more competitors. In order to answer this question, we consider the case in which there is no store brand. One important assumption made is that wholesale costs stay the same when store brands are removed. Though this assumption is not reasonable when analyzing the vertical relationships between retailer and manufacturers, it is not problematic to analyze the relationship between store brands and retailers competition. We compare the optimal profits with and without store brands, and analyze the relationship between the profit change and competition. We also compare the welfare effect of removing store brands. Without loss of generality, assume the first K brands are national brands. Now each store chooses optimal prices of all national brands to maximize profit, given the mean utility of outside option. The profit maximization problem of any store is now: max p t K (p kt c kt ω kt ) s kt (p kt )M kt The F.O.C.s with respect to prices of national brands are: j=1 p t = c t + ω t Ω 1 (p t ) s(p t ) (9) 21

Table 8: Demand Elasticities Own KL s flake Dominick s puff Dominick s ring -3.03 0.00 0.10 Dominick s flake -2.64 0.01 0.22 Dominick s shred -3.60 0.00 0.04 Dominick s puff -3.30 0.00-3.30 General Mills ring -4.13 0.08 0.06 General Mills flake -3.98 0.05 0.04 General Mills shred -4.31 0.01 0.01 General Mills puff -4.33 0.03 0.02 Quaker flake -3.75 0.02 0.00 Quaker shred -2.99 0.01 0.01 Quaker puff -3.53 0.04 0.03 Kellogg s ring -3.98 0.03 0.02 Kellogg s flake -3.51-3.51 0.05 Kellogg s shred -3.18 0.05 0.03 Kellogg s puff -4.11 0.05 0.04 Post flake -3.09 0.03 0.02 Post shred -3.49 0.02 0.01 Post puff -3.74 0.01 0.01 Ralston shred -4.34 0.02 0.01 Ralston puff -5.99 0.01 0.00 where the upper script denotes variables in counterfactural case. Ω is the matrix of partial derivatives, the (i, j)th element is s it(p t) p jt optimal prices solve the vector equations. Let Π t at any given price vector p t. So the new be the new profit for the Dominick s store in market t. So Π t = K (p kt c kt ω kt ) s kt (p kt)m kt j=1 For stores in different markets, the profit loss from not selling store brands will be different. Market sizes, consumer characteristic distributions, and competition degrees are all different for different markets. We are especially interested in the relationship between profit loss and local competition. Previous facts show that the profit share of store brands increases significantly with competition. So when store brands are removed, the profit loss in terms of profit percentage is expected to increase with competition. We find that after Dominick s removing store brands, its profits decrease in all 285 markets. The median prices of the 16 national brands decrease. For all national brands in all markets, prices of 65% of them decrease when store brands are removed, and 35% 22

of national brands prices increase. The decrease in national brands prices implies that Dominick s chooses high prices for national brands after introducing store brands. With store brands, Dominick s increase national brands prices to differentiate national brands consumers from store brands consumers. Table 9: Median Prices of National Brands No Store Brands/$ Data/$ General Mills ring 3.51 (.26) 3.52 (.26) General Mills flake 3.32 (.09) 3.32 (.09) General Mills shred 3.57 (.12) 3.58 (.12) General Mills puff 3.74 (.41) 3.74 (.40) Quaker flake 3.12 (.14) 3.12 (.14) Quaker shred 2.46 (.11) 2.46 (.11) Quaker puff 2.94 (.28) 2.94 (.28) Kellogg s ring 3.35 (.24) 3.35 (.24) Kellogg s flake 3.07 (.34) 3.08 (.34) Kellogg s shred 2.68 (.19) 2.68 (.20) Kellogg s puff 3.62 (.46) 3.62 (.46) Post flake 2.61 (.19) 2.61 (.19) Post shred 2.88 (.11) 2.88 (.10) Post puff 3.21 (.32) 3.21 (.32) Ralston shred 3.98 (.79) 3.98 (.79) Ralston puff 5.11 (.31) 5.12 (.31) Table 10 shows the own elasticities of national brands with and without store brands. We find that consumers are less elastic when there is no store brand. This result is intuitive. First, consumers demand are more elastic when store brands are available because they have more alternatives in their choice set. Second, consumers are more elastic when prices are higher. The national brands prices are higher when store brands are sold, so the consumers elasticities are higher. The first two columns of table 11 shows the profit loss and market share changes of removing store brands. We find that the median profit loss is $108 per week per store, 23

Table 10: Demand Elasticities Without Store Brands Store Brands Dominick s ring -3.03 Dominick s flake -2.64 Dominick s shred -3.60 Dominick s puff -3.30 General Mills ring -4.06-4.13 General Mills flake -3.93-3.98 General Mills shred -4.24-4.31 General Mills puff -4.25-4.33 Quaker flake -3.71-3.75 Quaker shred -2.98-2.99 Quaker puff -3.50-3.53 Kellogg s ring -3.93-3.98 Kellogg s flake -3.47-3.51 Kellogg s shred -3.16-3.18 Kellogg s puff -4.05-4.11 Post flake -3.07-3.09 Post shred -3.45-3.49 Post puff -3.69-3.74 Ralston shred -4.25-4.34 Ralston puff -5.80-5.99 Table 11: Percentage of Profit Loss and Competition Profit Loss Outside Share Loss Elasticity Median 4.33% +0.24% ˆλ 1 0.16 St.d 7.07% 0.73% ˆλ 2 0.21 which corresponds to 4.33% of profits before removing store brands. And the total profit loss for all the 285 markets is $37853 every week. The $108 loss per week leads to a $5616 loss per store every year. Removing store brands increases market share of the outside option, with a median increase of 0.24%. The last column of table 11 shows the relationship between profit loss of removing store brands and retailers competition. To check this relationship, we regress log of profit losses of all markets on corresponding log of number of competitors and log market median income. Let π st be profit loss in market t after removing store brands. The results show that profit loss increases with competition significantly, with elasticity 0.16. That is to say, when the degree of competition increases by 1%, the profit loss increases by.16%. 24

Suppose a Dominick s store has 10 competitors, and its profit loss from removing store brands is $500 every week, then this loss will increase to $540 if it has 15 competitors. Profit loss decreases with the market level income, with elasticity 0.21. So when a Dominick s store is in a market with higher income level, the profit loss is smaller when store brands are removed. log( π st π st ) = λ 1 log(r ( t)) + λ 2 log(inc t ) We calculate consumers welfare change when store brands are removed. Given prices and characteristics of all alternatives, define consumer i s expected utility as E(u it ) = 20 k=0 s ikt (δ kt + µ ikt ), where 0 represents the outside option. We find that welfare increases when store brands are removed. The mean welfare increase is about $6662 per market every year. Two sources affect the welfare of consumers. On one hand, removing store brands reduces consumers welfare because there are less alternatives. On the other hand, removing store brands changes the optimal prices of national brands. These two effects lead to the relatively small change of consumer welfare. 7.2 Increasing Competition In this part, we want to find out the effects of increasing competition on Dominick s prices and profits. To calculate new optimal prices for all markets, we first check whether wholesale costs vary with number of retailers in a market. We use wholesale cost data to check whether wholesale prices change with competition. Using fixed effects for all brands, we found that coefficient in front of competition is 0, and not significant. This result implies that wholesale prices don t vary with number of retailers in a market. So we assume that wholesale prices in all markets remain the same when competition increases. Suppose each Dominick s number of competitors increases by 5, then mean utility of the outside option increases by 5 η R. Now consumers mean utility of the outside option increases and we expect Dominick s to decrease prices and lose profit. Given there are 5 more competitors, we solve the supply side s profit maximization problem to calculate the new optimal prices and find the profit loss of each store. The first two columns of table 12 are median prices after and before increasing competition. We see that prices of all brands decrease. The third column shows the median changes in markups of all brands. Markup for every brand is defined as the ratio of profit 25