Store Format Choice

Ming-Feng Hsieh

A dissertation submitted in partial fulﬁllment of the requirements for the degree of

Doctor of Philosophy (Agricultural and Applied Economics)

at the UNIVERSITY OF WISCONSIN-MADISON 2012

Date of ﬁnal oral examination: 12/15/11

The dissertation is approved by the following members of the Final Oral Committee: Kyle W. Stiegert, Professor, Agricultural and Applied Economics Jean-Paul Chavas, Professor, Agricultural and Applied Economics Brian Gould, Associate Professor, Agricultural and Applied Economics Paul D. Mitchell, Associate Professor, Agricultural and Applied Economics Timothy J. Richards, Professor, Agribusiness and Resource Management, Arizona State University c Copyright by Ming-Feng Hsieh 2012 All Rights Reserved i

Dedication

To my parents, Chung-Ho Hsieh and Hsiu-Mei Yeh, for their persistent support and encouragement.

To my wife, Dr. Hui-chen Wang, whose passion and knowledge in economics provide a constant source of inspiration.

To my son, Benjamin W. Hsieh, for his love. ii Acknowledgements

The writing of this dissertation has been a long but fruitful journey. I am grateful to many people who have made this work happen. First, I am deeply indebted to my advisor, Dr. Kyle W. Stiegert, for his thoughtful critique and helpful support throughout every stage of my dissertation. He has strongly inspired and influenced me with his great personality, wisdom, and dedication to research. Through his guidance in the dissertation process, I have developed proficiency to conduct rigorous research. I wish to thank my dissertation committee for their careful review. Dr. Jean-Paul Chavas, Dr. Brian Gould, Dr. Paul D. Mitchell and Dr. Timothy J. Richards have provided thoughtful and valuable guidance for my work. Their comments have greatly improved the quality of my dissertation. My graduate studies would not have been the same without the financial support from Food System Research Group and the department of Agricultural and Applied Economics at the University of Wisconsin-Madison. The assistance and experience from numerous other faculty members have also proved valuable for my academic development. While I cannot list all of the names with their contributions here, I am thankful. Finally, and most importantly, I would like to thank my wife Hui-chen. Her support, encouragement, patience and unwavering love were undeniably the bedrock upon which the past important years of my life have been built. Her passion and knowledge in economics have also been a constant source of inspiration and wisdom throughout my graduate study and the writing of this dissertation. I also thank my parents, brothers and sisters for their faith in me and persistent encouragement and support. Also, I thank my son, Benjamin, for his love and enormous energy and joy that he has brought. iii

Contents

List of Tables v

List of Figures vi

Abstract vii

1 Introduction 1 1.1 Organic Market and Food Retailing ...... 1 1.2 Consumer Behavior and Retailer Strategies ...... 4 1.3 Literature Review ...... 6 1.4 Objectives and Organization ...... 10

2 Theoretical Framework 12 2.1 The Basic Model ...... 12 2.2 SFCs in WTP Space ...... 17 2.3 Retailer Strategies and Market Shares ...... 28 2.4 The Role of Income in SFCs ...... 37 2.5 Concluding Remarks ...... 41

3 Empirical Analysis 45 3.1 Methodology and Model Speciﬁcation ...... 45 3.2 The Data ...... 54 3.3 Results and Discussion ...... 61 iv 3.4 Concluding Remarks ...... 68

4 Conclusion 70 4.1 Summary and Implications ...... 72 4.2 Limitations and Directions for Future Studies ...... 73 v

List of Tables

1.1 Organic Price Premium by Store Format and Shopper Type of Actual Pur- chase for Milk and Eggs, 2005-2008 ...... 3 1.2 Determinants of Frequent Organic Shoppers ...... 5

H M L 1 1 1 2.1 Shopper Preference and SFC for Product 2 (q1 > q1 > q1 & H > M > L ) 27 q1 q1 q1 2.2 [EXAMPLE 2.4] Parameter Values for Numerical Simulation ...... 33 2.3 Top & Bottom 10 Gini Coeﬃcients (Income Inequality) for U.S. States, 2010 40

3.1 The Base Basket for Format-specific Price Index ...... 51 3.2 The Consumer Profile, 2005-2008 ...... 57 3.3 The Retailer Profile by Store Formats, 2005-2008 ...... 59 3.4 Descriptive Statistics (Mean) of Variables for SFC Estimation, 2005-2008 . . 60 3.5 MLE Parameter Estimates of Mixed Multinomial Logit Model for SFC . . . 62 vi

List of Figures

1.1 The U.S. Organic Food Market, 1997-2009 (Organic Trade Association, 2007, 2009) ...... 2

2.1 Optimal SFC for product 2 among the format L’s shoppers for product 1 in

L M L M the benchmark example, where v2 > v2 and q2 < q2 ...... 21 2.2 Optimal SFC for product 2 among the format L’s shoppers for product 1 in

M L M L the reversed example, where v2 > v2 and q2 < q2 ...... 22

M L L M 2.3 Dominant Set of SFCs for Product 2 Purchase when q2 > q2 and v2 > v2 . 23

L M H 2.4 Optimal SFC for Product 1 for type Q shoppers, where v1 > v1 > v1 and

H M L q2 > q2 > q2 ...... 25

L 2.5 Eﬀect of Quality Change on Market Shares for various q1 in EXAMPLE 2.4 34

L 2.6 Eﬀect of Price Change on Market Shares for various v1 in EXAMPLE 2.4 . . 34 2.7 Market Shares and Proﬁt Maximizing Quality (top) and Value (bottom) . . 36 2.8 Market Shares and Income Distributions with Various Income Inequality . . 42

3.1 Structure and Description of A.C. Nielsen Homescan Data Set ...... 55 3.2 Price Eﬀect on SFC Probability by Income Groups ...... 67 vii

Abstract

In this dissertation, I develop a theory of consumer store format choice (SFC) that provides a framework to investigate how retailer strategies impact consumer shopping behavior. Moti- vated by the recent developments in food retailing and organic markets, I use the availability of organic products as a quality measure of a store format to analyze empirically the role of quality in value assessment and price sensitivity when consumers make SFC decisions. The theory chapter characterizes the rationales of consumer SFC. Within this framework, I explore consumers’ optimal SFCs resulted from the matching between a consumer’s willingness to pay (WTP) and retailer strategies. Market shares and the impacts of quality, price and shopping cost on the shares are analyzed and summarized in a set of remarks. In particular, the model suggests two main results. First, value, defined as quality-to-price ratio, is a key to consumer SFC and price sensitivity in SFC vary among consumers and among formats. Second, the shoppers of general formats, i.e. value-oriented stores and supermarkets, have diversified preferences while the high-end specialty format shoppers have a high WTP and quality perception. In addition, with incorporating income in the model, I show how difference in income inequality can contribute to market share patterns. This is important given the widening of wage differences between the wealthy and the poor in the U.S. I then specify an econometric model based on the theory and use actual purchase data to examine the predictions. The model predictions were addressed by the estimation of random parameters with a mixed multinomial logit model and the results of marginal effect of price provide supporting evidence for these implications. In this empirical investigation, viii I use consumer income to capture the variation in consumer willingness to pay and quality perception. The results verify that price sensitivity varies among consumers of different income levels. The model provides evidence that the higher income consumers are less price sensitive compared to low income consumers. In sum, this dissertation contributes to the literature by constructing a simple but rich theoretical framework for store format choice analysis. The empirical results provide supporting evidence on model predictions and establish useful insights for farmers and retailers in their marketing and developing decisions on organic agriculture. 1

Chapter 1: Introduction

1.1 Organic Market and Food Retailing

The organic food market has been one of the fastest growing segments in recent years. Aggregate organic food sales in the U.S. have maintained a 15-20% annual growth rate over the past decade. The report by Organic Trade Association (2009) indicates that the US sales of organic foods totaled nearly $23 billion in 2008, which marks a 15.8% increase compared to sales in 2007 and is over 6 times of the sales in 1997. The organic penetration rates, deﬁned as organic food as a percent of total U.S. food sales, have increased from 0.97% in 1997 to 3.59% in 2009 (see ﬁgure 1). According to The Hartman Group (2008), over two-third of U.S. consumers buy organic products at least occasionally and about 28 percent of these organic consumers are weekly organic users. Figure 1 shows that the traditional supermarkets and value-oriented retailers have become more important outlets where consumers shop for organic food as their combined market share for organic food have increased from 30% to 46% over the past decade. On the other hand, sales of organic foods through natural food chains, such as Whole Foods Market and Wild Oats, and other independent natural food stores peaked at 68% of total organic sales in 1995. By 2005, the market share of natural food channels had however dropped to 47% of sales. Consumers choose to purchase organic foods for a variety of reasons. Some of the commonly cited perceptions among consumers are that a) organic foods are grown without pesticides or other toxic chemicals and so they are healthier for them and their families, b) 2

Figure 1.1: The U.S. Organic Food Market, 1997-2009 (Organic Trade Association, 2007, 2009) organic farming relies on more sustainable natural biological systems, which are better for the environment, c) practices and standards have evolved in the U.S. to improve the treat- ment of organically raised livestock. However, cropping and livestock systems used in organic farming tend to have higher costs per unit of output than in conventional farming. When these costs are successfully passed downstream, it ultimately means higher retail prices for those products that use the organic label. The price of organic food is typically 30-40%, and sometimes over 100%, more than conventional (non-organic) alternatives. The hefty price premium of organic food has been one of the major reasons for consumers to choose conventional over organic foods (Kavilanz, 2008). Wal-Mart in 2006 launched an aggressive “going green and organic” campaign that would greatly increase the number of organic products they oﬀered with a price target of only 10% above the price for conventional counterparts. This market expansion and low pricing strategy has not only enhanced competition among food retailers in the United States but also encouraged consumers to rethink whether to buy 3

Table 1.1: Organic Price Premium by Store Format and Shopper Type of Actual Purchase for Milk and Eggs, 2005-2008

Store Format value-oriented supermarkets high-end Shopper Type Organic Conv Organic Conv Organic Conv 2005-06 18.81% 31.67% 38.07% 50.59% 56.46% 70.69% Milk 2007-08 19.92% 24.80% 30.03% 39.52% 51.51% 59.49% 2005-06 275% 282% 265% 286% 290% 175% Eggs 2007-08 106% 118% 132% 138% 233% 139% organic foods and to reevaluate their choice of store format(s) to purchase food. Table 1.1 summarizes the average price premium of organic versus conventional product for two of the most frequently purchased products, milk and eggs, by store format and shopper type based on the actual purchase of each transaction. The data show that price premium varies among stores of diﬀerent formats and between organic and conventional shoppers. First, organic price premiums are at minimum in value-oriented stores, while high-end stores feature much higher organic price premiums. In addition, consumers who purchased organic products in general face the lower organic price premium compared to those who purchased conventional alternative at the outlets of the same store format, except for the case of eggs at high-end stores. Conﬁrming the marketing activities led by Wal-Mart along with others in 2006, we observe sizably diminishing organic price premiums for all outlets over the two periods. For example, in the case of eggs, the organic price premiums for the organic shoppers at the value-oriented stores dropped from 275% to 106%, which is less than half of the former. The only exception is the case of milk purchased by the organic shoppers in the value-oriented stores, the price premium was 18.81% in 2005-06 and 19.92% in 2007-06. It likely indicates that organic price premium for milk may have reached these low-end retailers’ pricing constraint bounded by a certain level of markup above the high production costs of organics. 4

1.2 Consumer Behavior and Retailer Strategies

Patterns of buying behavior are typically based on a person’s needs and wants. The recogni- tion of needs is usually the first stage of consumer’s buying process, followed by the stages of searching for information, quality perception and value assessment, before the purchase decision is made. Need perception, i.e. how consumers perceive their needs accompanied with their wants, is believed to be one of the key determinants leading to the choice of consumer purchases. In this research, I focus on the role of quality perception and value assessment in consumers’ choosing where to shop for their food purchases. In particular, the linkages between store format choice and organic food consumption are the center of the study. First, let us examine the makeup of frequent organic shoppers. Table 1.2 presents how frequent organic shoppers would look like in terms of income, preference for discounts and individual demographic characteristics.1 The results show that frequent organic shoppers are likely to be consisting of high-income, discount lovers, small household, college educated, family with preschool children, but not families with school-age children and the elderly households. Second, increasing marketing efforts or introducing more organic products can be considered as a quality improvement for the retailer since organic products are considered as high-quality gourmet products. On one hand, this quality improvement for the store will likely enhance consumers’ perception on store image and so attract more visits to the store. It may increase store’s overall revenues or profits because of the increased sales from organics and other products. On the other hand, the higher prices due to higher supply costs of organics may offset or overturn the gain from improved quality if the package of higher quality with higher price does not appeal to the majority of consumers in their value assessment.

1A household is considered as a frequent organic shopper or not, based on whether the percentage of organic consumption is more than 10% of the total spending. 5

Table 1.2: Determinants of Frequent Organic Shoppers

2005-06 2007-08 constant 0.0309** (0.0034) 0.0897** (0.0035) income 0.0015** (0.0003) 0.0009** (0.0003) % discount -0.0140** (0.0020) -0.0174** (0.0021) household size -0.0013 (0.0011) -0.0124** (0.0008) less educated -0.0039 (0.0025) -0.0192** (0.0024) single 0.0217** (0.0021) -0.0148** (0.0022) preschool children 0.0408** (0.0047) 0.0701** (0.0041) school-age children -0.0200** (0.0026) -0.0287** (0.0027) elderly -0.0164** (0.0018) -0.0254** (0.0020) Note: Robust standard errors are in parentheses. *, ** denote statistical signiﬁcance at 5% and 1%.

Indeed, researches and marketing reports have suggested that price is probably the most important factor aﬀecting consumer’s purchase decisions. Understanding consumer’s responses to pricing strategy, i.e. price sensitivity, is therefore crucial to retailer’s strategy making. In sum, the evaluation of pricing strategy should be assessed together with the impacts from quality perception in decision making of retailers. Finally, understanding consumer behavior has direct impacts for marketing strategy too, i.e. for making better marketing campaigns. For example, by understanding that frequent organic shoppers are more receptive to the use of discounts when they make food purchases, retailers learn to schedule sales or issue coupons for organic foods. By understanding that new products are initially adopted by and more meaningful to a small group of consumers, we learn that 1) companies that introduce new products must be well targeted and positioned to this speciﬁc market segment and 2) it is important to maintain the relationships with the existing customers, who may prefer the old products to the new at least in the initial stage. 6

1.3 Literature Review

This dissertation concerning consumer behavior in choosing where to shop falls in line of several research fields, including studies on retail competition, product differentiation, product line strategies, store choice, price and store format choice. There are two main streams of literature that are related to retail competition: the literature on spatial competition (horizontal product differentiation) and the one on quality/variety competition (vertical product differentiation), along with price competition. In models of spatial competition, customers choose where to shop among a number of retail locations based on transportation costs and prices offered by the retailers (e.g. Hotelling, 1929; Lancaster, 1975; Salop, 1979; Economides, 1989, 1993). Unlike models of spatial competition, models of quality competition analyze the role of quality setting in competition among multi-product firms. In the research concerning product line strategy, there are competing views on market segmentation and so-called full-line strategies. For example, Brander and Eaton (1984) model the competition between multi-product duopolists, suggesting that producers of substitutable products are likely to monopolize on a particular market segment in order to deter entry. Champsaur and Rochet (1989) show that the duopolists would want to avoid head-to-head competition by specializing in a range of products with quality levels different from the other’s product offerings. On the other side of arguments, Gilbert and Matutes (1993) argue that firms may adopt full-line strategy rather than segmentation when allowing consumers to have idiosyncratic preferences for firms. In contrast to Brander and Eaton (1984)’s model, Doraszelski and Draganska (2006) allow firms to choose both the number and the type of products offered. They show that offering a targeted product may increase (decrease) some customers’ utility due to increased fit (misfit). In addition, the intensity of competition and the fixed cost of offering an additional product are also the key determinants for firms’ market 7 segmentation strategies. Anderson and de Palma (1992, 2006) study market performance of multi-product oligopoly firms using a nested logit framework. They argue, in equilibrium, greater heterogeneity among retailers leads to less variety, while greater heterogeneity among products within each store promotes variety. Hamilton and Richards (2006) combine a discrete (spatial) store-choice model and a within-store (quality/variety) product-choice model. They find that the range of product varieties may rise or fall in response to the restructuring of costs, but welfare unambiguously declines. Product variety is supplied less than social optimal resource allocation in both oligopolistic and monopolistically competitive equilibria. Ellickson (2006) examines the role of endogenous fixed costs in determining the equilibrium structure of the retail food industry, where he decomposed the retail food industry into two relatively distinct sub-markets: supermarkets and grocery stores. Although multiple retail formats, supermarkets and grocery stores, are explicitly analyzed, the study has been restricted to competition within the same retail format and there are no cross competition between two retail formats being discussed. In the existing literature, studies concerning store format choice have concentrated on empirical examinations rather than on theoretical modeling and analysis. Those empirical studies provide empirical evidences and useful insights for theoretical modeling. For example, Fox, Montgomery, and Lodish (2004) empirically examine competition between retail formats and explore how retailers’ assortment, pricing, promotional policies, and customers’ demographics affect shopping behavior. They find that levels of assortment and promotion are more important determinants than price on consumer expenditures. They also find cross-shopping behavior among retail formats and that visits to mass merchandisers do not substitute for trips to the grocery stores. Carpenter and Moore (2006) provided empirical analysis identifying demographic groups who frequent specific formats and examining store attributes, like price competitiveness, product selection, and atmosphere, as drivers of format choices. Fox, Postrel, and McLaughlin (2007) find that grocery retailers generally 8 benefit from agglomerating with discount stores while Wal-Mart Discount stores suffer from agglomerating with grocery stores. Substituting supercenters for discount stores causes these agglomeration effects to disappear because cross shopping with grocery stores is reduced. However, there are only few studies concerning competition and store choice among the retail formats ( i.e. the store format choice). Messinger and Narasimhan (1997) develop a model of retail formats based on consumers’ economizing on shopping time to explain the growth of one-stop shopping. Using the U.S. aggregate annual data for a 26-year period, their empirical analysis suggests that retail scale economies were not the drives of one-stop shopping, but the improvements in transportation and inventory-holding technologies were particularly important to the growth of supermarkets and the recent rising of one-stop shopping retail format. Their model and empirical analysis relying on simplifying assumptions and imperfect aggregate data have inevitably left several important and yet unanswered questions. Extending from Messinger and Narasimhan (1997), Bhatnagar and Ratchford (2004) study competition between supermarkets and convenience stores, and between food warehouses and supermarkets. Their model utilizes both consumers’ utility maximizing model and a model of retail firms’ profit maximizing in a competitive setting with free entry. They then construct several hypotheses and test them with data from a survey for consumers’ format choice. While their proposed hypotheses are intuitive, these hypotheses however disconnect with and lack support from their theoretical model. From a different angle of focus, several studies have looked into the issues concerning the role of retail pricing formats on consumers’ shopping decisions. The common results are that consumers would tend to visit HiLo stores more frequently and buy smaller baskets there, while they would buy larger baskets and are most weekenders at Every Day Low Pricing (EDLP) stores (e.g. Lal and Rao, 1997; Bell, Ho, and Tang, 1998; Bell and Lattin, 1998; Ho, Tang, and Bell, 1998). From another aspect, Ellickson and Misra (2007) model how supermarket chains make a choice of optimal pricing strategy, selecting among three 9 options: everyday low pricing, promotional pricing (HiLo), and a hybrid (combination of the former two) strategy. While they provided strong empirical evidence suggesting that supermarket chains are likely to coordinate (adopt the same pricing format) in their pricing strategies, they were however unable to pin down the exact source of the complementarities by a formal model. Finally, another line of empirical studies have examined the competitive effects of entry by a retailer with different format on the existing retailers. For instance, Basker (2005); Basker and Noel (2007) find that in response to Wal-Mart’s entry to grocery market, the smaller- scale grocery stores reduced about 10% of their prices, while the response of the big three supermarket chains (Albertsons, Safeway, and Kroger) is less than half that size. Franklin (2001) finds Wal-Mart supercenters’ entry has little impact on food seller concentration in major metropolitan areas between 1993 and 1999. There is no correlation found between entry and city size. Multiple linear regression analysis however indicates that Wal-Mart’s market shares are highest in low income and smaller metro areas. Stiegert and Sharkey (2007) find that both the market share of supercenters’ food sales and the marginal impact of supercenters’ entry did not have a significant impact on food prices in the metropolitan statistical areas analyzed. They also find that changes in market concentration were significantly and positively related to price changes. They argue that supermarket consolidation led to higher prices and any merger-related cost gains during this period were not passed on to consumers. Singh, Hansen, and Blattberg (2004) examines the impact of Wal-Mart’s entry on household purchase behavior based on a unique frequent shopper database. The empirical results show that the incumbent store lost 17% of revenue following Wal-Mart’s entry. These losses were mainly due to fewer store visits rather than the impact on basket sizes. They also find that households that respond to Wal-Mart are likely to be large basket consumers, especially have an infant and pets in the family, and are more likely to be weekend shoppers. 10

1.4 Objectives and Organization

In the existing literature, researchers have focused on a) the choice of stores within a certain format or b) empirical analysis of the store format choice. In this dissertation, I develop a novel theoretical framework that characterizes the rationalizes consumer SFC and to examine how retailers’ pricing and quality positioning affect consumer SFCs. I then examine the predictions of the theory using A.C. Nielsen Homescan purchase data of conventional and organic foods from different store formats. The results provide an improved understanding of consumer shopping behavior and SFC decision making in responding to retailers’ strategies. Throughout the study, I pay attention to the role of income differences, retailer pricing and marketing strategies, and heterogenous preferences have on the choice of store formats. With enhanced understanding of consumer demand and decision making, this study is expected to provide useful insights for organic farmers and food retailers in their marketing and development decisions. My theoretical model utilizes the concept of perceived utility and distinguishes the impacts of retailers’ quality positioning into two distinct groups: one on consumer’s fixed utility reflecting store image effect and the other on consumer’s variable utility affecting consumer’s evaluation of products and prices. Within this framework, I consider preference heterogeneity among consumers and examine the matching between retailers’ pricing and quality positioning and consumers’ preference characteristics in determining their SFC. I further discuss how those supermarket strategies shape consumer SFCs and retailer market shares and the impact of preference distribution on the results with the assistance of numerical methods. Finally, the theoretical results and implications are then examined in an empirical application with standard and mixed logit approaches using a unique set of household actual purchase data. The remainder of the dissertation is organized as follows. Chapter 2 presents a theoretical 11 framework modeling consumer SFC with a focus on consumer preference heterogeneity and food retailers’ pricing and quality positioning strategies. Chapter 3 utilizes consumer actual purchase data to examine the model implications on the SFC suggested from the theory. Chapter 4 concludes the dissertation by discussing the linkages between the theory and the data, the implications on the marketing and agricultural practices from the findings, and the future developments beyond this research. 12

Chapter 2: Theoretical Framework

In this chapter, I model consumer shopping behavior in choosing where to shop among the three differential store formats. I start from a simple model, where consumers maximize perceived utility consisting of a fixed and a variable component based on retailer prices and quality settings. I depict how consumers’ store format choice outcomes and associated market shares of retailers would look like in the space of willingness to pay. I focus on selected examples to demonstrate the effects on market shares due to a change in quality or a change in price. I then incorporate income and intercategory connection through the channels of consumer’s willingness to pay to the basic model. Numerical simulations and analysis are utilized to gain further understanding of this system of decision making. Finally, I conclude the chapter with the remarks on model implications for empirical applications.

2.1 The Basic Model

Three major retail formats in the U.S. food retailing sector are considered in this model: 1) Low-end value-oriented retailers (L), such as supercenters and price clubs, representing an inexpensive nontraditional shopping format characterized by low-pricing, broad product assortment, and low service; 2) Middle format - supermarkets (M): a format represented by traditional supermarkets and grocery stores, generally featuring promotional (HiLo) pricing, broad assortment in food categories and some service; 3) High-end specialty stores (H), such as natural food retail chains, providing consumers with high-priced upscale product offerings. The stores sell two quality-differentiated products, labeled 1 and 2, representing foods and 13 nonfood groceries respectively. The formats L and M sell both products, while format H sells only product 1. The retailers differ in quality positioning as well as price positioning. The model is built on a perceived utility framework, in which consumers vary in their willingness to pay (WTP) for product quality. In specific, consumer i’s deterministic utility consists of fixed and variable components:

fi fi fi TUi = FUi(q ) + VUi(q , xi ), (2.1)

where consumer i’s ﬁxed shopping utility at store format fi ∈ {L, M, H} is inﬂuenced by the

perceived store image, which is a positive function of the product qualities qfi ; her variable

fi fi shopping utility is a concave function of both quality (q ) and quantity purchased (xi ). To facilitate the analysis, I assume the ﬁxed utility to be linearly relative to product quality (qfi ) and the variable utility to be a quasilinear functional form, which is concave to qualities (q) and quantities demanded (x) for goods 1 & 2. That is,

fi f1i f2i FUi(q ) ≡ α1iq1 + α2iq2 , 2 fi fi X fji fji 1 fji fji 2 and VUi(q , xi ) ≡ x0i + θjiqj xji − 2 (qj xji ) , j=1

where x0 is the quantity demanded for the outside goods, α represents the ﬁxed WTP indicating the evaluation for store image aﬀected by product quality and θ represents the variable WTP for product quality. Consumers are heterogeneous in WTP, i.e. (α, θ). The utility maximization problem can be thus expressed as follows:

2 f1i f2i X fji fji 1 fji fji 2 Max α1iq1 + α2iq2 + x0i + θjiqj xji − (qj xji ) (2.2) x,f 2 j=1 14

2 X fji fji s.t. x0i + pj xji + c(f1i, f2i) = mi, j=1

fji x0i ≥ 0, xji ≥ 0, ∀j = 1, 2.

where c(f1i, f2i) denotes the total shopping cost for the purchase of product 1 & 2. Let cfji represent the shopping cost for consumer i to shop for product j at format f. Then

f1i f2i f1i f2i c(f1i, f2i) = c = c if f1i = f2i: one-stop shopping; c(f1i, f2i) = c + c if f1i 6= f2i, i.e. Additional shopping cost incurred when consumer purchases product 1 & 2 from two diﬀerent store formats. All parameters (αji, θji) are assumed positive for the concavity of the utility. Solving this utility maximization problem, we have the Marshallian demands derived from the ﬁrst-order conditions:

f f ∗ 1i 1i f1i θ1iq1 − p1 1 1 x = = θ1i − , (2.3) 1i f1i 2 f1i f1i (q1 ) q1 v1 f f ∗ 2i 2i f2i θ2iq2 − p2 1 1 x = = θ2i − , (2.4) 2i f2i 2 f2i f2i (q2 ) q2 v2 ∗ ∗ ∗ f1i f1i f2i f2i x0i = mi − c(f1i, f2i) − p1 x1i − p2 x2i , (2.5)

where I deﬁne “value” (v) as the ratio of quality to price, i.e.

qfji vfji ≡ j . (2.6) j fji pj

That is, the product’s value increases with its quality while decreasing with its price. Directly from the demand functions of (2.3) and (2.4), a type θ consumer only purchases the product when her reservation price (pr = θq), a product of her variable WTP (θ) and 15 the quality of the product (q), is greater than the purchase price p for the product. That is,

∗ fji fji fji xji > 0 ⇔ θjiqj > pj ∀j = 1, 2. (2.7) 1 or θji > fji vj

In other words, consumers who choose not to purchase at a certain store for the good must feel the price-quality combination oﬀered by the store is not worth buying. Furthermore, these demand functions suggest that consumers’ purchase decision reveals their reservation price and variable WTP: Consumers with higher reservation price will purchase more compared to those with lower reservation prices. In addition, the product with a superior value due to lower price, better quality or both, would be appealing to a wider range of consumers in terms of their variable WTP. Finally, these demand functions indicate that the own price eﬀect on demand for good 1 & 2 is:

∗ fji ∂xji 1 f = − 2 < 0. ∀j = 1, 2, (2.8) ∂p ji fji j qj

Thus, the size of own-price effect decreases with product quality. It is worth noting that there is neither an income effect nor a cross-price effect on demand in the present model setup, assuming the separability of consumption for the two goods. The following remark summarizes these results implied by the demand functions.

REMARK 2.1: A consumer only purchases when her variable WTP is greater than the inverse of value. Each consumer’s demand reveals her reservation price and variable WTP. Consumers with greater variable WTP have more quantity demanded. The format with greater value of product appeals to a wider range of consumers in terms of variable WTP. Finally, the price eﬀect on demand is more sensitive for products with lower quality. 16 Substituting (2.3) - (2.5) to the direct utility function as in (2.2), we can derive the

f1i f2i f1i f2i indirect utility function Vi(p1 , p2 , q1 , q2 , c, mi) as:

2 2 f1i f2i 1 1 1 1 Vi = α1iq + α2iq + mi − c(f1i, f2i) + θ1i − + θ2i − . (2.9) 1 2 f1i f2i 2 v1 2 v2

It suggests that the greater is consumer’s variable WTP above the ratio of price to quality, the greater is her utility. From the indirect utility function, we learn that an improvement of quality oﬀered by a retailer has two impacts:

1. It enhances consumer’s evaluation for the store and so induces more store visits;

2. It increases consumer’s perception of value and so leads consumers to purchase more while prices remain the same.

Here, as below, I formally deﬁne these two eﬀects on consumer’s SFC caused by a change in quality.

DEFINITION 2.1: The store image effect of a change in quality is defined as the change in consumer’s SFC outcomes, such as the change in market share, due to solely the change in consumer’s fixed utility caused by the change in quality.

DEFINITION 2.2: The value eﬀect of a change in quality is deﬁned as the change in consumer’s SFC outcomes, such as the change in market share, due to the change in consumer’s perception of value induced by the change in quality.

In addition, we learn from the setup of utility function (2.9) that there are three key factors aﬀecting consumers’ SFC decision, namely 1) quality, 2) value, 3) shopping cost. So, I further deﬁne the followings.

DEFINITION 2.3: Store format SFC1 has quality advantage (QA) over store format SFC2 if qSFC1 − qSFC2 > 0. 17 DEFINITION 2.4: Store format SFC1 has value advantage (VA) over store format SFC2 if vSFC1 − vSFC2 > 0.

DEFINITION 2.5: SFC combination 1 has shopping cost saving advantage (CA) over SFC combination 2 if c(SF Cs1) − c(SF Cs2) < 0.

I will discuss how these three key factors aﬀect SFCs among consumers in the following sections.

2.2 SFCs in WTP Space

In this section, I depict consumers’ SFCs in the space of ﬁxed and variable WTP, where I identify the optimal SFC combinations of product 1 & 2 corresponding to various groups of consumers with heterogeneous preference. Under the basic model setup, consumer’s SFC decision making is indeed a discrete-choice utility maximization problem, in which the consumer compare and choose the one yielding the maximal utility of (2.9) among the choice

sets (f1i, f2i) ∈ {LL, LM, ML, MM, HL, HM} (big baskets) and {L0,M0,H0, 0L, 0M, 00} (small baskets), where the ﬁrst coordinate denotes the store format chosen for product 1 purchase and the second coordinate denotes the store format chosen for product 2 purchase,

f1if2i and 0 denotes the “non-buying” option. To simplify the notations, I will use Vi to denote

f1i f2i f1i f2i the consumer i’s indirect utility given by the specific set of SFC: Vi(p1 , p2 , q1 , q2 , c, mi). Because of taste difference, consumers are likely to choose different SFCs if they are not of the same type. In other words, facing a certain pair of SFC combinations, one group of consumers would prefer one choice while another group of consumers prefer the other choice due to the difference in their WTP parameters (α, θ). And beside these two groups, there may also be a group of consumers who feel indifferent between the two SFCs. For those indifferent consumers, their WTP parameters would follow a certain relation, so called the 18 indifference line in the space of WTP.

DEFINITION 2.6: The indifference line (IL) consists of the fixed and variable WTP combinations (α, θ) of the consumers who feel indifferent between a specific pair of

SF Cs1 SF Cs2 SFCs, i.e. Vi = Vi , ∀i ∈ IL.

Let α be in the horizontal axis and θ be in the vertical axis of the wtp space. This indifference line (if existed) will separate consumers into two groups in addition to the indifference group: The consumers with larger α, i.e. on the right of the indifference line, would prefer the format with QA, while the consumers with larger θ, i.e. on the upper of the indifference line, would prefer the format with VA. Clearly, the group in favor of the format with QA will be the same as the one in favor of the format with VA if the slope of the indifference line is negative. Otherwise, the QA and VA supporters will be in two separated groups in WTP space. Additionally, the condition (2.7) suggests that consumer i only make purchases for product j at the store format fji if and only if her variable WTP is greater than the inverse of the

fji format’s value, i.e. θji > 1/vj . So, I deﬁne the reservation line in WTP space identifying those whose reservation price is equal to the purchase price. That is,

DEFINITION 2.7: The reservation line (RL) consists of the ﬁxed and variable WTP combinations (α, θ) of the consumers whose reservation price is equal to the price of

SFC the goods at SFC, or θji = 1/vji , ∀i ∈ RL.

Thus, anyone whose variable WTP is below the reservation line of a speciﬁc store format would not make a purchase from that format. Instead, those consumers will look for a store format that provides better value so that the price is well below their reservation price. Based on consumer’s variable WTP and the associated value of reservation line, we further divide the consumers into various groups. 19 DEFINITION 2.8: The group of consumers whose variable WTPs are greater than the inverse of the lowest value provided by the retailers is referred as the high evaluation group of consumers. The group of consumers whose variable WTPs are smaller than the inverse of the highest value provided by the retailers are called the low evaluation group of consumers. Those whose variable WTPs are between reservation lines are called the middle evaluation group of consumers.1

SFC for Product 2: Two Format Case

Following the setup, consumers choose where to shop for product 2 between the two store formats: L and M. Within the group of consumers who chose to purchase product 1 from the format f1i ∈ {L, M, H, 0}, their SFC decision would depend on the utilities obtained for purchase of product 2 at L or M, given whichever format that the consumers have chosen for product 1 purchase.2 Consumers can also choose not to purchase any of product 2 if all product oﬀerings - price and quality combinations - available in the market are less preferred than the non-buying option. That is, based on their WTPs (α2i, θ2i), consumers will choose the SFC 2 among {L,M,0 } whichever gives them the higher utility for any given store format choice of product 1 purchase: f1i ∈ {L, M, H, 0}. To demonstrate the decision making of consumer’s SFC, I will ﬁrst discuss the following example.

EXAMPLE 2.1: Consider the selected SFCs with format L as the SFC for product 1:

(f1i, f2i) ∈ {LL, LM, L0}, where the format L (value-oriented retailers) has VA for

L M product 2 while format M (supermarkets) has QA for product 2, i.e. v2 > v2 and

M L q2 > q2 . 1There are n − 1 middle evaluation groups of consumers when there are n formats of retailers in the market. 2This can actually be a simultaneous or sequential decision making process, either of which starting ﬁrst will lead to the same conclusions. 20

LL In this example, consumer i would feel indiﬀerent between the two if and only if (Vi =

LM Vi ). That is,

2 2 L M 1 1 1 α2i(q2 − q2 ) + θ2i − L − θ2i − M − (c(L, L) − c(L, M)) = 0. (2.10) 2 v2 v2

Or alternatively,

M M L LL=LM 1 1 1 c q2 − q2 θ2 (α2i): θ2i = 2 M + L + + α2i . (2.11) v2 v2 1 1 1 1 M − L M − L v2 v2 v2 v2

Equation (2.11) specifies the indifference line in the preference space of (α, θ), i.e. these consumers who feel indifferent between LL and LM commonly have fixed and variable WTP

M L 1 1 combinations in a linear relation with a slope of q2 − q2 / M − L . It is straightforward v2 v2 to see that the slope of indiﬀerence line would be positive if there is not a single store format

M L M L having both quality and value advantage over the other format, i.e. q2 − q2 v2 − v2 < 0. In other words, the indiﬀerence line would be downward-sloping if one store format has both advantages. Since the low-end format (value-oriented retailers) has VA while the middle format (su-

L M M L M L permarkets) has QA, i.e. v2 > v2 ⇔ 1/v2 > 1/v2 and q2 > q2 , we have 1) the IL is M L upward-sloping and 2) RL2 is above RL2 . Figure ?? demonstrates the consumers’ optimal SFC decisions with respect to their WTPs. As shown in the ﬁgure, the consumers whose

L M variable WTPs (θ2i) are between RL2 and RL2 (the middle evaluation group) will purchase L product 2 at format L, and those whose variable WTP is below RL2 (the low evaluation group) will not purchase product 2, as neither format charges more than what they are will- ing to pay. In addition, those whose WTP combination is above the indiﬀerence line prefer LL, while those whose WTP combination is below ILLL=LM are in favor of LM. 21

Figure 2.1: Optimal SFC for product 2 among the format L’s shoppers for product 1 in the L M L M benchmark example, where v2 > v2 and q2 < q2

Next, let us consider an opposite example, where the format L has QA and format M has VA for product 2 instead.

EXAMPLE 2.2: Consider the selected SFCs with format L as the SFC for product 1:

(f1i, f2i) ∈ {LL, LM, L0}, where format L (value-oriented retailers) has QA for product

L M L M 2 while format M (supermarkets) has VA for product 2, v2 < v2 and q2 > q2 .

Here, we have the same indiﬀerence line between SFCs LL and LM (ILLL=LM ) as (2.11) in EXAMPLE 2.1. However, those whose WTPs fall in the upper/left side of ILLL=LM prefer LM because format M has VA in this example. In addition, the reservation lines here are in the reverse order compared to the previous example. Therefore, the middle evaluation

L M group of consumers have variable WTPs that are below 1/v2 and above 1/v2 , prefer LM to LL. And for this group of consumers, it is no longer true that everyone in this group would prefer LM to L0 because there involves additional shopping cost: cM in the SFC option LM compared to the option L0. That is, depending on their WTPs, some of the middle group 22

Figure 2.2: Optimal SFC for product 2 among the format L’s shoppers for product 1 in the M L M L reversed example, where v2 > v2 and q2 < q2 would choose format M for product 2 if the utility that they obtain from shopping is greater than that cost. Let us deﬁne the indiﬀerence line between these two SFC options:

q M LM=L0 1 M M c θ2 (α2i): θ2i = M + 2(c − α2iq2 ) for α2i < L . (2.12) v2 q2

LM=L0 Thus, those whose (α2i, θ2i) are in the right/upper side of indiﬀerence curve IL prefer

M LM to L0. Finally, consumers whose θ2i < 1/v2 , i.e. the low evaluation group, would not purchase product 2 from either M or L as the market prices are grater than their willingness to pay for the qualities provided. Figure 2.2 presents the optimal SFC decisions for product 2 in EXAMPLE 2.2.

The Dominant Set of SFC for Product 2

Returning back to the case of EXAMPLE 2.1, we can apply the same approach to analyze the cases where consumers pre-selected format M, H, or non-buying (0) for product 1 purchase. 23

M L L M Figure 2.3: Dominant Set of SFCs for Product 2 Purchase when q2 > q2 and v2 > v2

It is easy to show that the indiﬀerence lines for the pairs of SFCs with L or M for product

M L 1 1 2 purchase are with the same slope in all cases, i.e. (q2 − q2 )/( M − L ). Following the v2 v2 assumption that format L has VA and format M has QA in product 2 market, we have

ML=MM HL=HM LL=LM these indifference lines to be in the order of θ2 (α2i) > θ2 (α2i) > θ2 (α2i) for any given α2i. Figure 2.3 shows the dominant sets of SFCs for product 2 purchase for six identified shopper types of consumers. For example, the type V (Q) shoppers would choose format L (M) for their product 2 purchase no matter what SFC they selected for product 1 purchase. As to the groups of [2] VC and [3] QC shoppers, their WTPs are not in much difference; thus, the difference in shopping costs (CA) between the SFCs would play a key role in addition to the effects of quality and value, and so we are more likely to observe one-stop shopping behaviors in this middle range. Finally, non-buying option would be more appealing to the consumers whose WTPs, especially the variable WTP, are small, as the value offered by either format is not good enough. For the case of EXAMPLE 2.2, we will have similar figure with L and M reversed. 24 Remark 2.2 summarizes the results.

REMARK 2.2: The dominant sets of SFCs for product 2 purchase can be identiﬁed according to the formats’ relative advantages in quality and value settings. In particular,

M L L M 1. When format M has QA (q2 > q2 ) and format L has VA (v2 > v2 ), the dominant set of SFCs for product 2 purchase tends to be the quality-oriented format (M) for those whose fixed and variable WTPs are both large, and the value-oriented format (L) for those who has either low fixed WTP or low variable WTP. Non-buying is the dominant option for those whose variable WTP is lower than the inverse of value offered by the format with VA.

M L L M 2. When format L has QA (q2 < q2 ) and format M has VA (v2 < v2 ), the dominant set of SFCs for product 2 purchase tends to be the quality-oriented format (L) for those whose fixed and variable WTPs are both large, and the value-oriented format (M) for those who has either low fixed WTP or low variable WTP. Non-buying is the dominant option for those whose variable WTP is lower than the inverse of value offered by the format with VA.

SFC for Product 1: Three Format Case

As stated in the previous subsection, we have identiﬁed a dominant set of SFCs taking into account consumer’s preference for product 2 purchase. That is, for any speciﬁc consumer, she will make her SFC for product 1 based on the dominant set that her WTPs for 2 direct to. For example, the so-called type Q shoppers (Group [4]) would choose format M for their product 2 purchase no matter which SFC 1 they choose. In other words, the optimal

SFC combination will consist of f2i = M with one of 4 options for her SFC for product 1

f1i ∈ {L, M, H, 0} depending on her preference or WTPs about product 1 (α1i, θ1i). Next, 25

L M H Figure 2.4: Optimal SFC for Product 1 for type Q shoppers, where v1 > v1 > v1 and H M L q2 > q2 > q2

I will focus on this selected example to illustrate the typical outcomes for the three format cases.

EXAMPLE 2.3: Consider the selected SFCs with M as the SFC for product 2: (f1i, f2i) ∈ {LM, MM, HM, 0M}, where the low-end format (value-oriented retailers) has VA while the high-end format (specialty stores) has QA and three formats’ values and

L M H H M L quality settings are in the following order: v1 > v1 > v1 and q2 > q2 > q2 . In

LM=MM LM=HM addition, the indiﬀerence lines are in the order of θ1 (α1i) > θ1 (α1i) for all

H α1i and θ1i > 1/v1 .

Figure 2.4 depicts the optimal SFC for product 1 for this example. The last condition speciﬁed in the example ensures a positive market share for format M in the group of con-

H sumers who have variable WTP greater than the reservation line RL1 . In this case, the indiﬀerence line between LM and HM, i.e. ILLM=HM is not at work, as the other two indifference lines ILLM=MM and ILMM=HM dominate the outcomes. As also shown in the ﬁgure, 26

M H for those whose variable WTPs are above 1/v1 and below 1/v1 , they would choose between formats L and M for product 1 purchase, as format H’s value is not up to their satisfaction.

M Similarly, consumers with variable WTPs below 1/v1 will decide either shop at format L or not buying based on whether format L’s product 1’s price and quality combination is good enough to cover the additional shopping cost for a trip to format L’s store. The indiﬀerence lines will be in the stated order if the shopping costs for format L and format H combined are relatively large. Otherwise, their order will be reversed and the SFC

H - MM will be dominated by LM or HM for those whose variable WTP is greater than 1/v1 .

Shoppers’ WTP Patterns in Optimal SFCs

From the two-format cases to the three-format cases, we have observed that the optimal

SFCs in the space of WTP are all looked like the symbol angle (6 ), where the format with QA (but not VA) is the optimal SFC for those whose WTP combinations fall in the triangle area between the two sides of the angle: high-high WTP combinations, while format with VA is preferred by those whose WTP combinations are on the two rays of the angle: one-low WTP combinations. These patterns of SFCs suggest that lower-end format’s (format L’s or M’s) shoppers are more diversiﬁed in terms of WTP, as those who choose the same SFC can have very diﬀerent WTP combinations. For example, in the group of consumers who

¯ L choose LM in example 2.3, their θ1i can be close to the upper bound, θ1 or as low as 1/v1 . On the contrary, the high-end format’s shoppers and those who opt to not buying are more likely to have similar preference types, especially in variable WTP (θs). In particular, format H’s shoppers have high ﬁxed and variable WTPs, while the non-buying shoppers have low variable WTP. Here I further use numerical computation to demonstrate how each format’s shopper preference types distribute. Consider the case where retailers’ quality and value settings are: 27

H M L 1 1 1 Table 2.1: Shopper Preference and SFC for Product 2 (q1 > q1 > q1 & H > M > L ) q1 q1 q1

f2i α1i θ1i mean std dev mean std dev L 0.35 0.0857 0.49 0.0959 M 0.58 0.0652 0.61 0.0419 H 0.74 0.0271 0.97 0.0006 0 0.53 0.0831 0.18 0.0073

L M H L M H L M q1 = .025, q1 = .05, q1 = .1, 1/v1 = .2, 1/v1 = .25, 1/v1 = .4, q2 = .08, q2 = .1,

L M 1/v2 = .125, 1/v2 = .14, and both fixed and variable WTPs are uniformly distributed in the range of 0 and 1. From the simulation results shown in Table 2.1, we learn that low- end format’s shoppers are with low fixed WTP and mid variable WTP on average, while high-end shoppers have both high fixed and variable WTPs. It also clearly shows that the standard deviations of format L’s WTPs are the biggest, while the standard deviations of format H’s WTPs are the lowest. Even the middle format’s shoppers have relatively larger standard deviations compared to the high-end shoppers. I also observe that those who opt to not buy product 2 have low variable WTP with little variation. Remark 2.3 summarizes these results.

REMARK 2.3: The general formats’ (L and M) shoppers are with diversiﬁed WTPs, while the high-end specialty format’s (H) shoppers are with high ﬁxed and high variable WTPs and the “small-basket” shoppers have low variable WTP for quality.

Before concluding this section, it is worth noting that EXAMPLE 2.1 and EXAMPLE 2.3 are in a “consistent” intercategory relationship, i.e. low-end format has VA and high-end format(s) has QA for both products. It is also possible to observe the case where one format has VA in one product while having QA in the other product, i.e. in a “mixed” intercategory relationship, like EXAMPLE 2.2 and EXAMPLE 2.3. When formats’ advantages are mixed between the two products, we would very likely observe that a SFC combination like 28 (L,H) may be appealing to a group of consumers who have similar tastes (WTPs), as it is actually a high-high quality combination. I will discuss the role of intercategory connection in consumer’s SFC decision making further in the later sections.

2.3 Retailer Strategies and Market Shares

In this section, I will derive the market shares for each format of retailers based on the analysis in the preceding section. First, I use a simple case - EXAMPLE 2.1 to illustrate how the changes in retailer’s quality, price and shopping cost affect market shares. I then employ numerical methods to see how market shares and the effectiveness of retailer strategies for various quality and price settings of retailers. To simplify calculations, let us assume that both fixed and variable WTPs are distributed uniformly with lower bounds at 0 and finite upper bounds: αji ∼ uniform[0, α¯j] and θji ∼ ¯ uniform[0, θj]. Then we can directly derive the market shares of product 2 for each store format in the case of EXAMPLE 2.1 by calculating the areas of SFCs in Figure 2.1. The resulted conditional market shares are:3

h i ¯ 1 ¯ 1 1 1 1 M θ2 − M θ2 − L M − L + c v2 v2 2 v2 v2 1 1 s(L|f1i = L) = M L + M − L α¯2. (2.13) (q2 − q2 ) v2 v2 h i θ¯ − 1 θ¯ − 1 1 1 − 1 + cM 1 2 vM 2 vL 2 vM vL ¯ 2 2 2 2 s(M|f1i = L) = θ2 − M α¯2 − M L . (2.14) v2 (q2 − q2 ) α¯2 s(0|f1i = L) = L (2.15) v2 3Those market shares are the shares out of those who choose low-end format (L) for their product 1 purchase. Therefore, they are conditional percentages. The total unconditional market share for each format requires taking the SFC results for both products together into account. 29

The Eﬀects of Quality Change

L Now, let’s observe how a change in q2 aﬀects the conditional market share of format L,

L s(L|f1i = L). In speciﬁc, we can also derive and decompose ∂s(L|f1i = L)/∂q2 as:

L ∂s(L|f1i = L) ∂s(L|f1i = L) ∂s(L|f1i = L) ∂v2 L = L + L L (2.16) ∂q ∂q L ∂v ∂q 2 2 dv2 =0 2 2

∂s(L|f1i = L) ∂s(M|f1i = L) L = − L ∂q L ∂q L 2 dv2 =0 2 dv2 =0

h i ¯ 1 ¯ 1 1 1 1 M θ2 − vM θ2 − vL 2 vM − vL + c = 2 2 2 2 , M L 2 (q2 − q2 ) h i ¯ 1 1 ¯ 1 1 1 1 M L θ2 − M (θ2 − L ) + ( M − L ) + c ∂s(M|f1i = L) ∂v2 v2 2 v2 2 v2 v2 − L L = L L M L , ∂v2 ∂q2 v2 q2 (q2 − q2 ) L ∂s(0|f1i = L) ∂s(0|f1i = L) ∂v2 α¯2 − L = − L L = L L . ∂q2 ∂v2 ∂q2 v2 q2

L L Here, the first term in (2.16): ∂s(L|f1i = L)/∂q2 L = − ∂s(M|f1i = L)/∂q2 L , rep- dv2 =0 dv2 =0 resents the store image effect: the effect on market shares due to the change in fixed evaluation of store image caused by the change in quality, while the latter terms: (∂s(L|f1i = 30

L L L L L L L L)/∂v2 )(∂v2 /∂q2 ) consisting of (∂s(M|f1i = L)/∂v2 )(∂v2 /∂q2 ) and ∂s(0|f1i = L)/∂q2 , represents the value eﬀect: the eﬀect on market shares due to the change in value caused by the change in quality. Similarly, as shown in (2.18) and (2.19), we can also decompose

L the eﬀect of change in format L’s quality q2 on format M’s conditional market share: 1)

L L store image effect: ∂s(M|f1i = L)/∂q2 = ∂s(M|f1i = L)/∂q2 L and 2) value effect: dv2 =0 L L L ∂(s(M|f1i = L)/∂v2 )(∂v2 /∂q2 ). Directly from these equations, we learn that the effects of quality change tend to be greater when 1) format L’s VA is larger, or 2) format M’s QA smaller. In particular, the difference in both store image and value effects are approaching infinity when the difference in quality settings between the two formats (L and M) is close to 0. In sum, we have the following.

REMARK 2.4: The effects of a change in quality on market shares become greater, as the difference in quality between two formats becomes small. In such a case, the store image effect is especially greater than the value effect due to a change in quality.

The Eﬀect of Price Change

L Similarly, we can examine how a change in p2 aﬀects the market share of format L, s(L|f1i =

L L). In speciﬁc, we can also derive and decompose ∂s(L|f1i = L)/∂p2 in the following ways:

L ∂s(L|f1i = L) ∂s(L|f1i = L) ∂v2 L = L L (2.20) ∂p2 ∂v2 ∂p2 L L ∂s(M|f1i = L) ∂v2 ∂s(0|f1i = L) ∂v2 = − L L − L L ∂v2 ∂p2 ∂v2 ∂p2 where

h i ¯ 1 1 ¯ 1 1 1 1 M L θ2 − M (θ2 − L ) + ( M − L ) + c ∂s(M|f1i = L) ∂v2 v2 2 v2 2 v2 v2 L L = L M L , ∂v2 ∂p2 q2 (q2 − q2 ) L ∂s(0|f1i = L) ∂v2 α¯2 L L = L . ∂v2 ∂p2 q2 31

L Therefore, a reduction in format L’s price of product 2 (4p2 < 0) would increase the market share of format L at the expense of format M’s and induce new customers who did not participate in product 2 market. It is worth noting that the price effect shown in (2.20) is dependent of the inverse of quality. The price effect for the format with QA would be smaller than the one without. The key reason behind this is that the price effect occurs through the impact on the value, i.e. the ratio of price to quality. In the case where format M has better

M L M L quality: q2 > q2 , the equal change in prices between formats ∆p2 = ∆p2 would lead to a M L p2 p2 smaller amount of change in price-per-quality ratio for format M than L, i.e. ∆ M < ∆ L . q2 q2 As a result, the impact of price changes in higher-quality format on SFC would be smaller than that of the same changes in lower-quality format on SFC. In other words, the price sensitivity would be lower for the SFC with respect to the higher-quality format. In addition, we can re-express the value eﬀect due to quality change in terms of the price eﬀect here, i.e.:

L L ∂s(M|f1i = L) ∂v2 1 ∂s(M|f1i = L) ∂v2 L L = − L L L , (2.21) ∂v2 ∂q2 v2 ∂v2 ∂p2 L L ∂s(M|f1i = L) ∂v2 ∂s(M|f1i = L) ∂v2 or L L = − L L . (2.22) ∂v2 ∂ln(q2 ) ∂v2 ∂ln(p2 )

Equation (2.22) implies that the value effect due to 1% increase in quality is equal to the price effect due to 1% reduction in price. I summarize the effect of price change as the following.

REMARK 2.5: From example 2.1, we learn that the price effect on consumer’s SFCs reflected in the retailer’s market share varies among the formats with various quality settings. The price effect, or so-called price sensitivity, is smaller for higher-quality format shoppers. Interestingly, an 1% change of price reduction yields the same amount of effect on market share as the value effect induced by an 1% change in quality improvement. 32 The Effect of Shopping Cost Change The effect of a change in shopping cost (cM in this case) on the conditional market share of format L in EXAMPLE 2.1 is as:

¯ 1 1 1 θ2 − M M − L ∂s(L|f1i = L) ∂s(M|f1i = L) v2 v2 v2 M = − M = M L (2.23) ∂c ∂c q2 − q2

It is straightforward to see that the effect of shopping cost change is greater when format M’s value is smaller (i.e. away from the upper bound) or the slope of the indifference line becomes smaller (i.e. quality difference decreases or value difference increases between the two formats). Intuitively, the shopping cost matters in this case only to the choice between LL (one-stop shopping) and LM (cross-shopping). Therefore, the bigger area above the

M reservation line RL2 , referred as the group, the greater impact from the difference in the shopping cost between the two. In addition, shopping costs affect consumer’s purchase power; therefore, it would affect consumer’s SFCs and retailer’s market shares through the change in quantity demanded, and so the influence in WTP space is vertically rather than horizontally. Thus, the effect of shopping cost change is greater when the slope of indifference line is smaller, i.e. when the two formats’ value difference is greater or their quality difference is smaller. In sum, I have the following remark.

REMARK 2.6: The effect of shopping cost change on retailer’s market share is greater when there is a large high evaluation group of consumers. The shopping cost advantage is more helpful for the one-stop shopping format to gain market shares when the value difference is greater or the quality difference is smaller between the one-stop and cross shopping options. 33

Table 2.2: [EXAMPLE 2.4] Parameter Values for Numerical Simulation

Parameter Values Cases Product 1 Product 2 L M H 1 1 1 L M 1 1 q1 q1 q1 L M H q2 q2 L M v1 v1 v1 v2 v2 Base 0.02 0.05 0.10 0.2 0.25 0.50 0.075 0.10 0.13 0.14

L L Various q1 and Fixed p1 = 0.01 L 4q1 (0.01,0.10) 0.05 0.10 (0.10,0.50) 0.25 0.50 0.075 0.10 0.13 0.14

L L Fixed q1 = 0.02 and Various p1 L 4p1 0.02 0.05 0.10 (0.10,0.50) 0.25 0.50 0.075 0.10 0.13 0.14

Market Share and Proﬁt Maximization

Following the preceding comparative analysis, I will now turn to a numerical example (EX- AMPLE 2.4) to see how quality change and price change affect retailer’s market shares. Table 2.2 lists the parameter values adopted in this example. In this example, format L has VA and format M has QA in product 2 market. I then use various quality of format L in the first case to see how change in quality affect consumer’s SFC decisions and so retailer’s market shares. As shown in Figure 2.5, the effect of a change

L L in q1 on format L’s market share in product 1 market, i.e. ∂S1(L)/∂q1 , is in a reverse U

L M shape, maximized around the point where v1 = v1 . The eﬀects on other formats’ market

L shares due to a unit increase of q1 are all negative, as format L’s quality improvement attracts consumers to switch to format L.

L We observe similar pattern of results for the eﬀects of 4p1 on market shares, as shown

L in Figure 2.6. The price eﬀect of format L on its own market share: ∂S1(L)/∂p1 , is in a U

L M shape, minimized around the point where v1 = v1 as well. This similarity between price and quality eﬀects is indeed coherent with the results summarized in REMARK 2.5. I propose the following simple example to see how the retailer’s proﬁt maximizing value 34

L Figure 2.5: Eﬀect of Quality Change on Market Shares for various q1 in EXAMPLE 2.4

L Figure 2.6: Eﬀect of Price Change on Market Shares for various v1 in EXAMPLE 2.4 35 would behave. In particular, I assume that retailer’s proﬁt (π) is a function of market share (S) and price margin (p − C), i.e.

π = S(v)(p − C(q)) , (2.24) where market share S(v) is an increasing function of value (v) and production cost C(q) is an increasing function of quality (q). Then at maximization, the optimal quality setting must satisfy the following:

∂π ∂S(v) ∂C(q) = 0 = (p − C(q)) − S(v) , (2.25) ∂q ∂q ∂q ∂S(v) S(v) ∂C(q) ⇐⇒ = . (2.26) ∂q p − C(q) ∂q

Similarly, the optimal price setting must satisfy:

∂π ∂S(v) = 0 = (p − C(q)) − S(v), (2.27) ∂p ∂p ∂S(v) S(v) ⇐⇒ = . (2.28) ∂p p − C(q)

Therefore, we can use (2.26) and (2.28) together with the results from Figures 2.5 and 2.6 to identify the profit maximizing quality and profit maximizing value for each format. Since production cost C(q) is an increasing function of quality (q) and market share is increasing with quality too, I expect the right hand side of (2.26), referred as marginal cost curve, is an increasing function of quality. Thus, possible profit maximizing quality can be identified as in Figure 2.7. As shown in the figure, there are multiple quality levels that satisfy the first order con-

L∗ dition. However, only the quality level q1 is the profit maximizing quality. To see that, let’s examine the excess marginal benefit, i.e. the difference between marginal benefit and 36

Figure 2.7: Market Shares and Proﬁt Maximizing Quality (top) and Value (bottom) marginal cost, before and after the intersected points E0 and E∗. It is straightforward to see that excess marginal beneﬁt is negative before but positive after the point E0 ; thus, the

L0 ∗ value q1 is actually the proﬁt minimizing value. Alternatively, at point E , excess marginal

L∗ beneﬁt is decreasing around the intersected point; thus, the value q1 is indeed the proﬁt

L∗ maximizing value. In other words, we have the optimal q1 falling somewhere below format M’s quality position in this example. In the event that format L has an improvement in its production leading to a reduction in costs, the proﬁt maximizing quality will be smaller

L∗ reflecting the gain in cost saving. Similarly, profit maximizing value 1/v1 can be identified

∗ L∗ M L∗ M as in E of Figure 2.7, where 1/v1 is below 1/v1 , i.e. v1 > v1 . In sum, I conclude the 37 following.

REMARK 2.8: The proﬁt maximizing quality for each format can be identiﬁed by the

∂S(v) S(v) ∂C(q) marginal benefit per quality ( ∂q ) and marginal cost per quality ( p−C(q) ∂q ) curves. We can also identify the profit maximizing value by identifying the marginal benefit per

∂S(v) S(v) price ( ∂p ) and marginal cost per price ( p−C(q) ) curves. Finally, the proﬁt maximizing quality and value together determine the proﬁt maximizing price accordingly.

2.4 The Role of Income in SFCs

In the basic model, consumer’s utility is structured by a quasilinear functional form, which restricts income effect by nature of the setup. With no income effect, the present simplified model fails to explain how an increase in income would have affected consumer’s consumption decision and how SFC decisions vary among income groups. There is a surprisingly simple solution to this problem. That is, we can use the channel of WTPs to incorporate income in the model and to construct the intercategory connection in explaining consumer’s SFC decision making.

Incorporating Income in the Model

In literature, for example Baumgartner et al. (2011), consumer survey studies have suggested that consumer’s variable WTP varies with income, and the richer households tend to have higher variable WTP for product quality. Therefore, it is natural to hypothesize that consumer’s variable WTP (θ) follows a linear relation with her income level, i.e.

0 m θji = θji + θj mi, ∀j = 1, 2, (2.29) 38

0 where the constant parameter (θji) is assumed to the heterogeneous unobservable part of

m consumer’s variable WTP and the marginal parameter (θj ) represents the marginal propen- sity to pay for quality (MPP), which quantiﬁes the increase in consumer’s willingness to pay occurs with an increase in (disposable) income. The MPPs are expected to be positive, as higher income households are more likely to have larger WTP. By substituting (2.29) into (2.9), the resulted form of the indirect utility function thus becomes:

2 !2 f1i f2i X 1 0 m 1 Vi = α1iq1 + α2iq2 + mi − c(f1i, f2i) + θji + θj mi − (2.30) 2 fji j=1 vj

f1i f2i = α1iq1 + α2iq2 + mi − c(f1i, f2i) (2.31)   2 0 m θ + θ m 2 X 1 0 2 0 m 1 m2 2 ji j i fji 1 fji +  θji + θjiθj mi + θj mi − f pj + 2 pj  2 2 q ji fji j=1 j 2qj

Expressing the squared terms consisting of income and value in equation (2.30) to the terms in equation (2.31), we learn that consumer’s utility is dependent of 1) quality, 2) income, 3) shopping cost, 4) price (or value), 5) WTP as well as 6) the interaction terms:

0 m fji fji θji + θj mi /qj ∗ pj . Those interaction terms representing the price eﬀect imply the following.

REMARK 2.9: Consumer’s SFC decision is subject to the maximization of perceived utility from the format choices. Price eﬀect or price sensitivity in SFC will vary among consumers due to heterogeneous unobservables and diﬀerences in income.

With the setting of (2.29), we can also derive Marshallian demand functions as:

0 f1i m f1i f1i ∗ θ q + θ m q − p xf1i = 1i 1 1 i 1 1 , (2.32) 1i f1i 2 (q1 ) 39

0 f2i m f2i f2i ∗ θ q + θ m q − p xf2i = 2i 2 2 i 2 2 . (2.33) 2i f2i 2 (q2 ) ∗ ∗ ∗ f1i f1i f2i f2i x0i = mi − c(f1i, f2i) − p1 x1i − p2 x2i

It is straightforward to see that the income eﬀect on demand for product 1 & 2 is:

∗ ∂xfji θm ji = j > 0 for j = 1, 2. (2.34) ∂m fji i qj

This veriﬁes the common property of demand, i.e. consumer’s demand increases with income,

m fji and the income eﬀect would be greater with a larger MPP (θj ) or a less quality (qj ). For the outside good, the income eﬀect on demand is:

∗ 2 m ∂x X θj 0i = 1 − . (2.35) ∂m fji i j=1 vj

Thus, the income eﬀect on demand for the outside good decreases with the MPPs but increases with the values of good 1 & 2.

Market Shares and Income Distribution

Following the setup of (2.29), consumer’s variable WTP is increasing with income. Suppose

0 that the heterogeneous unobservable term of consumer’s variable WTP: θji is relatively small compared to the part involving income. I expect that the distribution of variable WTP follows the same / similar pattern of income distribution. Therefore, consumer’s SFC decisions and the resulted retailer’s market shares are closely connected to the distribution of income as their connection to variable WTP. To show how market shares vary with various patterns of income distribution, I next compare the resulted market shares corresponding to two income distributions with diﬀerence inequality (high- and low- inequality). 40

Table 2.3: Top & Bottom 10 Gini Coeﬃcients (Income Inequality) for U.S. States, 2010

Top 10 States Bottom 10 States Rank State Gini Rank State Gini 1 New York 0.499 1 Utah 0.419 2 Connecticut 0.486 2 Alaska 0.422 3 Louisiana 0.475 3 Wyoming 0.423 3 Massachusetts 0.475 4 New Hampshire 0.425 5 Florida 0.474 5 Iowa 0.427 6 Alabama 0.472 6 Wisconsin 0.430 7 California 0.471 7 Nebraska 0.432 8 Texas 0.469 8 Hawaii 0.433 9 Georgia 0.468 8 Idaho 0.433 9 Mississippi 0.468 8 North Dakota 0.433 The information was tabulated from 2010 data from the American Community Survey conducted by the US Census Bureau

According to data from American Community Survey (ACS) conducted by the US Census Bureau, U.S. household income inequality has grown 18% from 1967 to 2010. There are sizable variations among the states in the U.S. as well. Table 2.3 lists the states with the top and bottom 10 Gini coefficients in 2010.4 As shown in the table, New York state has the biggest Gini coefficient, which is 19% more than Utah state’s. Here, I use a log-normal distribution (lnN(µ, σ)) to construct possible income distributions with various levels of inequality. In particular, I use lnN(ln(.5), 1.1) representing high income inequality (Gini coefficient = .5) and lnN(ln(.6136), 1.35) representing low income inequality (Gini coefficient = .42) to observe how the difference in income distributions affects the market shares for each format. The computation of market shares for each distribution is conducted using the selected parameter values from the numerical example introduced in the preceding section: EXAM-

4The Gini coeﬃcient summarizes income inequality in a single number and is one of the most commonly used measures of income inequality. It uses a scale from 0 to 1, where 0 represents perfect equality and 1 represents perfect inequality. 41

L M H PLE 2.4. In the selected case, format L has VA (1/v1 = .2, 1/v1 = .25, 1/v1 = .5) while

L M H format H has QA (q1 = .02, q1 = .05, q1 = .10) in product 1 market. For product 2 market,

L M L M format L has VA (1/v2 = .13, 1/v2 = .14) and format M has QA (q2 = .075, q2 = .10). The simulation results, shown in the bottom diagram of Figure 2.8, indicate that the middle format (M) has 1.29% less market share, while the rest three format options gains market shares as income inequality increases. Among the three, the high-end format’s market share increases from .05% to 06%,˙ which is actually a 10.75% increase. The low-end format (L) has marginally more of market share, which is about 1.77% due to the increase in income inequality. From this exercise, we learn the following.

REMARK 2.10: In the market where income inequality is greater, it is expected that there are more high-end or low-end shoppers but less middle format’s shoppers. This result is consistent with the observation of the rising of value-oriented and high-end specialty formats in the recent years.

2.5 Concluding Remarks

In this chapter, I construct a perceived utility framework characterizing the rationales of SFC decision making of heterogeneous consumers among three quality and price differentiated formats. Consumers make two SFCs: one for the targeted product category (available at all three formats) and the other for the remaining items on a shopping list (available only from the general formats: L and M). Shoppers face a fixed cost for each shopping trip; hence, additional cost would occur if they visit different store formats for the two choices. Consumers are considered to be heterogeneous in both fixed and variable WTPs for product quality. With the model setup, a retailer’s quality setting affects consumers’ overall evaluation for the store (store image effect) as well as their purchase decisions based on 42

Figure 2.8: Market Shares and Income Distributions with Various Income Inequality 43 the retailer’s value and consumers’ variable WTPs. This two-dimensional impact of quality setting on consumer SFCs plays a key role in the findings of this study. With this model, I have identified how consumer’s SFCs behave in WTP space. I have, specifically, depicted the matching between consumers’ WTPs and their optimal SFC combinations. I have also analyzed how retailer’s pricing and quality settings would affect the market shares for each format of retailers, and how one may utilize these results from market shares to identify profit maximizing quality and value. Finally, with incorporating income in the model, I have shown that difference in income inequality may have contributed to the outlook of market shares in the present markets. While consumer’s SFC involves a complex system of decision making and the present model provides rich implications and results, I will focus on the role of “price sensitivity” in SFC for my empirical applications in next chapter. In particular, the following remarks the findings regarding to this emphasis.

1. Price sensitivity is dependent of income, quality and heterogeneous unobservables. As demonstrated in section 2.4, the price sensitivity is dependent of the ratio of

0 m fji 0 θji + θj mi /qj , that includes a heterogeneous unobservable term (θji), income (mi)

fji and product quality qj . Therefore, price sensitivity is reﬂected by the eﬀects of these three factors on consumer’s SFC decisions.

2. Price sensitivity in SFC varies among consumers and among formats. Directly following the previous remarks, variation in household income and variation in consumer unobservable preference heterogeneity direct price sensitivity to vary among consumers. And due to diﬀerential quality settings and various shopper compositions, price sensitivity varies among formats.

3. Variation in price sensitivity is smaller among the high-end format’s shoppers than among the general formats’ shoppers. This is directly from REMARK 2.3, where I 44 observe that high-end specialty format’s shoppers are with high fixed and high variable WTPs while the other two formats’ shoppers are with diversified WTPs. Given the price sensitivity is closely tied with variable WTP, it is expected that variation in price sensitivity follows the similar pattern of variation in consumers’ WTPs among the shoppers of differential formats.

The established model suggest two important benefits from further understanding of price sensitivity. First, it helps to estimate the value effect due to quality change. As illustrated and described in REMARK 2.5, an 1% change of price reduction yields the same amount of effect on market share as the value effect induced by an 1% change in quality improvement. Therefore, the value effect of quality change can be obtained through the estimation of price sensitivity. Second, it helps to identify the profit maximizing value with the availability of supply- side / production cost data. As summarized in REMARK 2.8, we can identify the profit

∂S(v) maximizing value by identifying the marginal beneﬁt per price ( ∂p ), i.e. price sensitivity, S(v) and marginal cost per price ( p−C(q) ) curves. This marginal cost curve is actually empirically approachable as it only involves two well established factors, i.e. the market shares and the price margins in the market. This approach for identifying proﬁt maximizing values will help food retailers on their production and marketing decision making. 45

Chapter 3: Empirical Analysis

In this chapter, I investigate empirically the rationales of consumer SFC decision making among three differential store formats: low-end value-oriented retailers (L), the middle format: supermarkets (M) and high-end specialty stores (H). The estimation results are used to evaluate the impact on consumer’s SFCs due to price change, i.e. price effect or price sensitivity, and the effects of quality change caused by expanded marketing and new introduced lines of organic products. The remainder of this chapter is organized as follows. First, I construct an econometric model that quantifies the impacts of economic and socio-demographic factors on the probability of a household choosing a specific store format for food purchases. I then discuss the estimation approaches and the methods used for testing the predictions and hypotheses from the theory. The data and descriptive statistics are discussed next. I then present and discuss the regression results. Next, I discuss how the marginal effects of price vary among various groups of income. The chapter concludes with a summary of findings and remarks.

3.1 Methodology and Model Speciﬁcation

Consumer’s store format choice (SFC), i.e. choosing where to shop, is a discrete choice problem based on consumers’ utility maximization among available alternatives. That is, household h maximizes the following indirect utility associated with purchasing foods at a speciﬁc store format f on trip t 46

Vhft = Vbhft + εhft (3.1)

= Ahf αh + Xhftβh + εhft,

where Ahf αh is the ﬁxed component of deterministic utility, Xhftβh represents the variable component of deterministic utility which is dependent of the household’s shopping list for

the specific trip, and εhft is the error term being assumed with independence of irrelevant alternatives (IIA), i.e. the odds of preferring one choice over another do not depend on the presence or absence of other ”irrelevant” alternatives. This specification takes into account the repeated choices by each individual decision maker by treating the coefficients that enter utility as varying over households but being constant over choice situation for each household. Following the theoretical framework, the set of explanatory variables affecting the fixed component of utility (A) includes 1) income, 2) shopper types represented by a shopping preference for organic foods and a discount use rate in this study, 3) individual household characteristics including educational and marital status, household size and family compositions, 4) shopping costs and convenience reflected by store locations, average days between trips and network sizes of store formats, and 5) quality index represented by the percentage of organic products to total products sold in each format. The set of factors influential to the variable component of utility (X) contains 1) prices represented by price indexes and 2) shopping basket size and components. To capture the preference heterogeneity, I include an interaction term of income and price, i.e. income∗price, as well as random price coefficients to reflect the variation in price sensitivity among different income groups of households. It is worth noting that I do not interact prices with quality indexes to represent the value variable in the estimation due to the multi-linearity issue with such a variable and the main 47 term of price. However, as suggested by the theory, the differences in quality perception among consumers will cause differential price sensitivities in SFC. This calls for mixed logit models for estimation of random preference variation.

Mixed Logit Model with Repeated Choices

The SFC decision based on maximization of utility (3.1) ﬁts the use of a mixed logit model with repeated choices, which has been well established in the literature, such as Train (2003). The motivation for the mixed logit model arises from the limitations of the standard logit model. Particularly, the mixed logit model solves the shortcomings of the standard logit model by allowing for 1) random taste variation, 2) unrestricted substitution patterns, and 3) correlation in unobserved factors over time (Train, 2003). Therefore, the mixed logit models especially outscore the standard logit models for estimations using panel data, like in the present study. In addition, unlike the probit model which is limited to the normal distribution, the mixed logit can also utilize any distribution for the random coeﬃcients. For estimation, I follow the simulation approach of Train (2003) (page 151) and utilize the STATA command: MIXLOGIT by Hole (2004).1 The approach considers a sequence

of alternatives (store formats), one for each trip, f = f1, ..., fT . The probability the the household makes this sequence of choices is the product of logit formulas:

T " A α +X # Y e hfi h hfit Lhf (α, β) = (3.2) P eAhj αh+Xhjt t=1 j

assuming the εhft’s are independent over time. The unconditional probability is the integral of this product over all values of (α, β):

Z Z Phf = Lhf (α, β)f(β)dβ f(α)dα. (3.3)

1For the choice with multiple alternatives (greater than two) like the present one, the approach works equally well after creating duplicate records of data (see Hole (2004), pages 397-398). 48 The estimation of the probability for the sequence of choices consists of a great deal of computations and time due to the consideration for repeated choices, requiring T times as many draws compared to the case where coeﬃcients being constant over time for each household. Due to the computational issues with a fairly large size of analyzed sample facing in this study, I allow random parameters for only the quality and price variables to obtain converged estimation results. The speciﬁcation of the deterministic component of latent utility adopted for estimation is:

P P Vbhft = α1 incomeh + f α2f loyaltyhf + α3 (organic shopper)h + α4 %discounth + j α5j j P demographich + α6 (days between trips)ht + α7 distancehft + f α8f n(store)hft P P P + f α9fh qualityhft + f β1fh pricehft + f β2f incomeh * pricehft + β3 (total P k spending)hft + k β4k (% spending on product category)hft,

where α9fh and β1fh are the random parameters. The demographic variables include household size, education, marital status, children and age (elderly), and product category variables include dry grocery, fresh produce and organics.

Variables and Measures

The measurements of the variables, including shopper types, price index, basket size and composition, shopping cost and convenience, and demographics, are described as follows.

Shopper Types

To identify shopper types, I adopt the following three measures: format loyalty, discount shoppers, and organic shoppers. First, the format-specific loyalty for a household is represented by the percentage of trips that the household made to the format during the initialization period. This loyalty index reveals the shopper’s preference toward a specific format 49 due to probably the familiarity about the store layout, the general prices and assortments, and the convenience and quality of service, based on his/her past shopping experience. Second, I use household discount-use rate, defined as the ratio of the number of items purchased with coupons or at discounted prices to the total number of items in a shopping trip, to capture their preference between promotional pricing (HiLo) and everyday low pricing (EDLP). These rates are calculated from the household purchase information during the initialization period. It is expected that a household with a high discount-use rate would prefer the format in which stores/chains use HiLo pricing instead of EDLP. Third, I use the household purchase information during the initialization period to identify whether the householder is a frequent organic shopper or not, depending on whether the percentage of organic consumption is more than 10% of the total spending. As illustrated in the introduction chapter, organic shoppers are likely to consist of high-income, discount lovers, small household, college educated, family with preschool children, but not families with school-age children and the elderly households. It is expected that the likelihood for frequent organic shoppers to choose high-end format stores is relatively higher than the one for non-organic shoppers.

Shopping Cost and Convenience

As suggested in my theoretical model as well as in store choice literature, shopping cost and convenience are important to consumer decision making in SFC. In this research, I use “days between trips” and the average “distance” between the stores of certain format and the household’s home to measure the shopping costs. It is expected that the more days between trips reflects the higher shopping cost that the consumer may encounter for each individual trip. The longer distance between stores and household’s home, the higher shopping cost for the household. Regrading to the measure of “distance”, there is no detailed location information for both 50 stores and households except the zip code for store and the census track code for household in the data available to us. Thus, the distance measure between a specific pair of store and household is calculated by the distance between the center of the area with the zip code of the store and the center of the area with the census track code of the household. Although this is less than ideal and can potentially cause measurement error, it is likely to yield better outcomes with less bias compared to the case without a distance measure, where the estimation suffers from the omitted variable bias. To better represent the location impacts, I use the number of stores for each format that are located with 25 miles from the household’s home to capture the location convenience of the format, called the network effects. I expect that the more number of stores of the format is expected to have positive impacts on the SFC for the format.

Demographics

The demographics available to this study include household size, educational and marital status of the householders, and the family compositions - the ages and numbers of children and the ages of the householders. Selecting the major group of consumers (college educated, married, no young children and non-elderly) as the default group, I examine the impacts of demographic characteristics diﬀering from the default on the decision of SFC.

Quality Index

Measuring store format quality represents a challenge for this study due to the limitation of this sort of data. . Past researcher employed measures such as whether or not prepared food service is available, the number of product varieties, product categories, check-out lines etc, to proxy store quality. In this study, I adopt a new approach to measure store quality: the share of the store’s sales coming from organic food products. The organic label covers a wide range of products, it reﬂects rigorous supply management on the part of food processors, 51

Table 3.1: The Base Basket for Format-speciﬁc Price Index 1. Dairy-Milk-Refrigerated 2. Bakery - Fresh Bread 3. Cereal - Ready To Eat 4. Soft Drinks - Carbonated 5. Yogurt-Refrigerated 6. Fruit 7. Soup-Canned 8. Cookies 9. Vegetables 10. Eggs-Fresh 11. Precut Fresh Salad Mix 12. Candy-Chocolate 13. Fruit Drinks-Other Container 14. Water-Bottled 15. Beef 16. Snacks - Tortilla Chips 17. Fresh Carrots 18. Fresh Strawberries 19. Fresh Fruit-Remaining 20. Rice - Mixes 21. Fruit Juice - Apple 22. Fruit Juice-Remaining 23. Prepared Foods 24. Yogurt-Refrigerated-Shakes & Drinks 25. Frozen Fruits 26. Vegetable Juice And Drink Remaining 27. Meat Products-Imitation & Additives 28. Fish 29. Whipping Cream 30. Seafood-Shellﬁsh.

and farmers must meet strict guidelines regarding the production of organic commodities. As a result, organic food is perceived as an indicator of food quality leading to perceptions by consumers about the quality of the store’s product offerings (Bruns, Fjord, and Grunert, 2002). Consequentially, I hypothesize a store’s new introduction or increase in the supply rate of organic foods as an increase in the store’s quality. Specifically, a store format’s quality index in this study is represented by the total number of organic products sold (org). The log ratio of organic supply, defined as the log of ratio of numbers of organic products sold in

the two formats: ln(orgf /orgM ), is used to capture the impact of quality in my estimation.

Price Index

To generate the format-specific price index, I first select a comparable basket of items available for all three formats. Table 3.1 lists the base basket containing 30 most frequently purchased product categories: After the base basket is constructed, I then calculate the average household consumption pattern for the selected product categories in the basket from the initialization sample. Using these base quantities together with the format-specific cate- 52 gory price indexes, we can compute the cost of the base basket at each format. Specifically, the cost measure can be written as:

J X CIt = pt · xti , (3.4) f f,j h,j j=1

where pt is the price index of category j at format f in the estimation month t and xti is f,j h,j

the quantity that average household h purchases in category j in the initialization period ti. With the cost measures for the formats, I use the cost for purchasing the base basket at supermarkets (format M) in the ﬁrst period as the base cost to calculate the ratio of every cost measure to this base cost. This yields the format-speciﬁc price index:

t t 1 PIf = CIf /CIM . (3.5)

Basket Size and Composition

I use “total spending,” deﬁned by the total transaction amount recorded, to measure the shopping basket size for each individual shopping trip. In addition, I use the average shares of dry grocery, fresh produce and organic foods in the total consumption during the initialization period to measure how consumers’ consumption preference pattern aﬀects their choice of store format. Instead of using the basket composition of the current trip, this alternative measure can reduce the bias due to endogeneity.

Testable Hypotheses

Based on the model predictions from the theoretical framework built in Chapter 2, I will test for the following hypotheses in this empirical application.

H1. Price sensitivity varies among households of different types and among formats. 53 Consumers of different types would be attracted to the formats that offer a matching combination of price and quality positioning to their tastes. Thus, price sensitivity is expected to vary across store formats and among households of different types, likely reflected by income, preference for organic consumption and demographic characteristics. In addition, variation in consumers’ quality perception will have a systematic effect on SFC. High-income households will prefer high-end format retailers, while low-income households will prefer the other two formats. Similarly, frequent organic shoppers will prefer high-end format retailers, while conventional shoppers will prefer the general formats. Cor- responding to shoppers’ characteristics for each format, the variation in price sensitivity within a group of consumers who patronizing high-end specialty stores would be less than the one for those who shop at the other two formats. In other words, we have the following hypothesis:

H2. Variation in price sensitivity for the general formats’ shoppers is expected to be greater than the variation for the high-end format’s shoppers.

The above two hypotheses are testable directly from the parameter estimates of random coefficients of price in the mixed logit models. The hypothesis H1 would involve the tests on whether random coefficients are different from 0, while the hypothesis H2 relies on the comparison of standard deviations of price coefficients between choice equations. Finally, since consumer’s SFC is a choice decision concerning multiple products; hence, the inter-category connection, especially cross-price effects, will have impacts on SFC. As a result, we may observe certain format combinations that are appealing to a group of consumers but not others. Shopping costs revealed by the distance between retailer and consumer will help explain SFC. The households, who travel longer distance, shop less frequently, or have less access to a certain format, will likely bear higher travel costs; thus, they will prefer formats with broader assortments, reflecting one-stop shopping behavior. I will 54 examine whether these findings that have commonly been established in literature hold in our estimation as well.

3.2 The Data

I use a multi-outlet panel data set - Homescan by A.C. Nielsen for a non-coastal U.S. city that covers a 208-week period between December 26, 2004 (hereafter January, 2005) and December 27, 2008.2 Figure 3.1 provides a general presentation of this data set. The data panel includes a representative of households in 52 market areas and 9 remaining (rural) areas in the 48 continental states of the US. Panelists report their purchases by scanning either the Uniform Product Code (UPC) or a designated code for random weight products of all their purchases from grocery stores or other retail outlets. The data set contains purchase/sales information on eight product departments (dry goods, frozen, dairy, deli, meat, fresh produce, non-food, and alcoholic beverages) and over 600 product categories of food and non-food items sold in grocery stores or other retail outlets. The households report weekly purchase data including price, quantity, promotional information, and product characteristics. One of the product characteristics contained in the data is the identiﬁer for organic products. For UPC-coded products, organic products can be identiﬁed by the presence of the USDA organic seal or with organic-claim codes cre- ated by Nielsen. In addition, the data set contains store-level information including store location, store size (as represented by an estimate of annualized value of sales), store formats (as described below), and demographics of store users. For identifying store formats, the retail outlets included in the data set are categorized into eight groups: 1) grocery stores, 2) drug stores, 3) mass merchandisers, 4) supercenters, 5) clubs, 6) convenience stores, 7) health food stores, and 8) all other. The household demographic information is collected for

2I thank Dr. Ephraim Leibtag and Dr. Biing-Hwang Lin at Economic Research Service of the U.S. Department of Agriculture, and the A.C. Nielsen Company for supplying key data for this study. 55

Figure 3.1: Structure and Description of A.C. Nielsen Homescan Data Set each panelist, including household size, income, age, employment, education, marital status, race, type and location of residence, and selected household equipment characteristics (e.g. kitchen appliances, TV items, internet accessibility). These individual household demographic characteristics will be aggregated by store based on the purchase/sale information to generate the store-level demographics of store users. This is a useful data set to study the linkages of organic consumption and store format choice in the US grocery sector. It will allow for an analysis of the role of major retailers such as Wal-Mart and Whole Foods that traditionally do not cooperate with scanner data collection companies, since the data are recorded from the consumers’ end instead of by the retailers. The rich information of household purchasing pattern, especially with the organic 56 claim, allows me to study the present research that cannot be addressed using other forms of data. In the database available and used for my analysis, households are included only if they participated in at least 10 of the 12 months during the year; these households are referred to as the ”static” panel. Only a subset of this static panel, called the Fresh Foods panel, records purchases of random-weight foods without UPCs. During 2005-06, the full static panel consists of approximately 40,000 households per year, with 8,000 recording both UPC and random weight items and 32,000 recording only UPC items. In 2007, Nielsen replaced the Fresh Foods Panel with the Total Shopper View, which no longer contains details on the random weight items. Specifically, item characteristics and the quantity purchased are no longer recorded. However, the sample size increased significantly to 22,000 households for those households reporting both UPC and non-UPC items, and a total sample of 60,000 households recording just UPC purchases. Due to the inconsistency on the coverage of random weight items over the analyzed period, the proposed research will focus on the UPC items but include the non-UPC items in the estimation of total purchases/sales. Because of the data redesign, I separate the four-year period into two i.e. 2005-06 and 2007-08.3 In addition, the separation of the two also allow us to analyze how consumer behavior responds to the Wal-Mart and other retailers’ market expansion in organic market around the end of 2006. Within each set of two years, I use the first 26 weeks as the “initialization” period to identify shopper types and format-specific indexes to avoid potential endogeneity between organic consumption decision and SFCs. The remaining 78 weeks were used as the “estimation” sample. The estimation is based on every shopping trip of households with shopping duration being no longer than 30 days during the estimation period at the selected retail chains in the market, to ensure that each panelist was faithful

3organic information is not available for random-weight fresh produce for 2007-08 data set. This will likely lead to underestimate consumers’ organic penetration, as fresh produce has always been on the top of consumers’ shopping list for organics. 57

Table 3.2: The Consumer Proﬁle, 2005-2008 2005-06 2007-08 Number of households 710 942 Number of shopping trips 161.34 137.83 Average spending per trip 23.06 18.40 Organic/total frequency 1.20% 1.84% Organic/total spending 1.24% 1.93% Household size 2.36 2.40 Income ($0000s) 6.33 6.86 Some college educated 87.9% 88.1% Married 57.9% 58.0% Preschool children (age <6) 5.8% 9.8% School-age children (age 6-18) 21.2% 21.8% Elderly (age>65) 22.5% 22.3% in recording purchases and remained in the panel for the entire period. The resulting data set had 710 households with a total of 45,877 shopping trips in 2005-06 sample and 942 households with 48,469 trips in 2007-08 sample. The selected retail chains for my analysis include 2 value-oriented retail chains consisting of 29 (37) stores, 4 traditional supermarket chains featuring 172 (147) stores, and 1 high-end specialty supermarket chain with 6 (7) stores in my 2005-06 (2007-08) sample.

Consumer Proﬁle

Descriptive statistics of the consumer profile are provided in Table 3.2. The statistics show that there were significant reductions in shopping frequency and basket size over the two sample periods, which may indicate a greater reliance on food away from home during the latter period. my data may also pick up some impact from the economic downturn for the U.S., particularly in the latter half of 2008 when the housing related credit crisis began to pick up steam. In this trend of consumption reduction, organic food is however 58 relatively less affected as its share to total food consumption has increased from 1.20%/1.24% to 1.84%/1.93% in terms of frequency/spending (dollar amount). I observe no significant changes in household demographics, with an exception that the percentage of household with pre-school children (age<6) had increased from 5.8% (2005-06) to 9.8% (2007-08) on average.

Retailer Proﬁle

Table 3.3 depicts the characteristic differences among the retailers of three store formats. Location or network wise, high-end specialty stores are much less accessible compared to the other two formats as shown in number of stores, share of trips, share of spending, as well as by the average travel distance from consumer’s home to the store. However, it is documented that these high-end specialty stores are the major outlets for organic food, as their organic shopping rates are by far higher than those of the other two formats. In the selected sample market, traditional supermarkets remain the most important outlets among the three formats, although increasing market shares of value-oriented stores are observed in the data. Regarding to pricing factors, I observe no significant price difference between value- oriented retailers and supermarket chains, but much higher prices at high-end specialty stores in both organic and non-organic alternatives. The data of discount use rates suggest that unlike the other two, traditional supermarkets promote promotional pricing. However, interestingly, I observed a much higher discount use rate applied to organic purchases at high-end stores than elsewhere. As to the coverage of product assortments, measured by the number of UPCs, value-oriented retailers have broadest coverage but supermarket chains offer more varieties per category on average. The high-end specialty stores carried a much higher percentage of organic products in terms of both broadness and variety, but with a much small scale of assortments in general. 59

Table 3.3: The Retailer Proﬁle by Store Formats, 2005-2008 2005-06 2007-08 value- super- high- value- super- high- oriented markets end oriented markets end Number of stores 29 172 6 37 147 7 Average travel distance (miles) 9.02 8.87 16.96 8.74 9.54 14.45 Share of trips 19.32% 79.46% 1.21% 21.47% 78.11% 0.43% % Organic in frequency 0.27% 0.78% 25.07% 0.80% 1.38% 35.07% Share of spending 18.49% 79.69% 1.81% 21.34% 78.01% 0.64% % Organic in spending 0.32% 0.96% 21.91% 1.02% 1.66% 29.99% Pricing & Discount Price index (selected basket) 0.968 1 1.505 0.919 0.929 1.373 Organic PI (selected basket) 0.977 1 1.357 1.046 1.039 1.449 % discount (overall) 12.81% 40.12% 11.69% 10.25% 35.99% 9.51% % discount (organics) 0.05% 0.29% 4.06% 0.08% 0.43% 3.42% Broadness (# UPCs) & Depth (# Brands) of Assortments Average broadness per store 2038 1505 659 1557 1517 201 % Organic in broadness 0.79% 2.28% 25.84% 1.35% 3.62% 31.84% Average variety per category 33.98 63.72 9.07 32.86 57.78 4.68 % Organic in variety 7.47% 8.35% 49.54% 8.91% 10.52% 61.03%

Descriptive Statistics

Table 3.4 lists the definitions and descriptive statistics (means) of the explanatory variables used in the estimation of SFC models for the markets of a non-coastal U.S. city in 2005- 2008. One of the notable differences between the two samples is that I observe a fast growing pattern of organic consumption. In particular, the percentage of organic food purchase in total spending of the trip increases from 1.24% to 1.93%, which is with over 50% of growth. Furthermore, the percentage of frequent organic shoppers also increases from 3.33% to 4.68%, that is a 40% increase. We also learn that the average transaction amount recorded for each shopping trip decreased by about 20% over the two periods, likely reflecting the economy downturn occurred in the later period. 60 6 & 0 otherwise 0.0611 0.1127 < Table 3.4: Descriptive Statistics (Mean) of Variables for SFC Estimation, 2005-2008 Variableprice V/Sprice H/Stotal spending Definition% price dry index grocery (value-oriented) / total% price transaction price fresh index amount index produce recorded (supermarkets) (high-end) for / the% % price shopping organics of index trip % dry (supermarkets) of grocery freshincome purchase produce in purchase total in spending total 0.9754 oforganic spending the of trip 1.0597 the trip %%discount 22.8651 of 18.1503 organic foods purchaseloyalty in V total spending 1.5390 of household theloyalty income S 1.5438 trip 0.0657 (in 0.5284 $1,000) 1 discount 0.0741 ifloyalty use 0.3508 H householder rate is during a the frequent initializationdays organic period between shopper % trips of & 0.0124 number trips 0 ofdistance that otherwise days household 0.0193 between format made two (value-oriented, to shopping supermarkets, then(stores) trips high-end) V during 0.0333 the initializationn(stores) period 0.0468 Sn(stores) H the distance number between ofhousehold 0.3421 consumer’s value-oriented size home stores and within 0.3033 store 25 numberless miles of educated from supermarket household stores number within number 2005-06 1.1312 ofsingle 25 of 2007-08 high-end 0.8050 miles persons stores from in 1.6564 within household the 0.7636 25preschool household miles 1 children from if 4.9563 household 5.7222 householder 1 is if 5.4318 not family 5.6494 6.3300 college has educated child(-ren) & under 6.8776 0.1806 0 aged otherwise 0.2245 0.2619 8.9900 0.2062 1 9.3263 if single householder & 0 otherwise 0.0144 0.1115 0.0119 0.1081 2.3583 2.3924 0.4248 0.4275 school-age children 1 if familyelderly has child(-ren) aged 6 18 & 0 otherwise 1 if householder is aged 65 and above & 0 0.2290 otherwise 0.2354 0.2403 0.2320 61 As to market shares, the data show that value-oriented format was gaining market shares, as its customer loyalty increases 24%. The market expansion of value-oriented format is also reflected in more number of stores over the two sample period. On the other hand, high-end specialty store format was losing its shares over time.

3.3 Results and Discussion

In this study, the store format choice consists of the choice of three alternatives, i.e. low-end value-oriented (L), middle format: supermarkets (M), and high-end specialty (H) formats. The alternative, middle format: supermarkets, is chosen as the base outcome for estimation in the multinomial logit model setting. The analyzed samples are weighted by sampling weights, and heteroskedasticity-robust standard errors are calculated and reported. To account for differential price sensitivities among households, I first allow for price coefficients to vary among households. In addition, I interact the ratios of price indexes with household income and the income squared to capture the variation in price sensitivity due to the differences in WTP or quality perception associated with household income.

Parameter Estimates

Table 3.5 reports the MLE parameter estimates for SFC from a mixed multinomial logit model. These parameter estimates represent the marginal utilities of the determinants, or equivalently the impacts of the determinants on log odds: ln(Phf /PhM ), where f = L (value- oriented) or H (high-end) is the SFC in selection with format M (supermarkets) as the base. Several key ﬁndings emerge from the regressions. First, the price sensitivity varies by the income level. The results from the parameter estimates of interaction term: income*price suggest that price sensitivity is greater for lower-income households, especially for those own- price terms. In addition, the parameter estimates from the price variables show statistically 62

Table 3.5: MLE Parameter Estimates of Mixed Multinomial Logit Model for SFC

2005-06 2007-08 value-oriented high-end value-oriented high-end Variable Coef. S.E. Coef. S.E. Coef. S.E. Coef. S.E. Random Parameters price L/M mean -4.238** 1.479 -7.779* 3.841 -1.396* 0.669 -0.142 2.615 standard deviation 0.863** 0.213 0.805 0.508 0.759** 0.112 -0.177 0.384 price H/M mean -1.136 0.802 -0.689 0.885 0.057 0.426 -0.817 2.329 standard deviation 0.949** 0.242 0.456** 0.111 0.589** 0.060 0.601 0.362 ln(org L/M) mean -0.123 0.093 1.372** 0.494 0.006 0.137 0.199 0.955 standard deviation 0.148 0.120 1.718** 0.378 0.800** 0.193 1.348 0.864 ln(org H/M) mean 0.320* 0.165 0.456 0.427 -0.323 0.210 2.124* 0.983 standard deviation 0.215** 0.066 0.062 0.086 0.190** 0.043 0.799** 0.267 income*price L/M 1.070** 0.244 0.770 1.174 0.073** 0.014 -0.347 0.610 income2*price L/M -0.017 0.024 -0.066 0.074 0.005 0.007 -0.005 0.027 income*price H/M 0.136 0.240 0.300** 0.058 0.047 0.092 0.120** 0.044 income2*price H/M 0.002 0.015 0.040 0.043 -0.005 0.004 -0.006 0.019 income -1.140** 0.243 0.234 0.640 -0.157* 0.075 0.507 0.339 % discount -3.427** 0.272 -5.786** 0.828 -3.204** 0.223 -5.665** 1.186 days between trips 0.006 0.009 -0.024 0.047 0.002 0.007 -0.073 0.054 distance -0.032** 0.010 -0.010 0.009 -0.028** 0.005 -0.024 0.024 n(store) L 1.318** 0.191 0.371 0.312 0.740** 0.131 0.475 0.308 n(store) M -0.127** 0.045 -0.157* 0.075 -0.213** 0.041 -0.348** 0.097 n(store) H 1.514** 0.313 3.137** 0.844 0.665** 0.219 5.135** 0.980 total spending 0.001 0.003 -0.001 0.008 0.002 0.003 -0.030 0.025 % dry grocery 1.924** 0.164 -0.520 0.597 0.604** 0.153 0.973 0.632 % fresh 0.367 0.377 0.827 0.909 0.263 0.252 2.058** 0.769 % organics -2.820 2.019 6.448** 1.086 -1.311 0.775 4.258* 1.862 loyalty L 6.291** 0.460 -0.554 1.081 5.736** 0.291 2.439* 1.238 loyalty H -1.370 1.448 15.379** 1.789 0.318 2.011 9.166* 3.625 organic shopper 0.099 0.580 0.701 0.479 0.086 0.310 2.056** 0.669 household size -0.100 0.118 -0.117 0.499 -0.150 0.081 -1.857 1.007 less educated 0.057 0.385 -0.076 0.769 0.319 0.211 -24.693** 1.298 single 0.089 0.231 -0.320 0.942 -0.114 0.195 -1.318 1.142 preschool children 0.043 0.371 0.722 1.147 -0.179 0.246 2.370 1.341 school-age children 0.539 0.335 0.702 1.209 0.001 0.257 3.264 1.982 elderly -0.506* 0.229 -0.241 0.410 -0.107 0.184 1.355 1.159 Note: Supermarkets (middle format) is the base outcome for all variables. Heteroskedasticity- robust errors are reported. *, ** denote statistical significance at 5% and 1%, respectively. 63 significant standard deviation, indicating price sensitivity varies among households. These results provide supporting evidence for hypothesis H1. Second, the estimated standard deviations for price coefficients are significantly different from zero for both equations and both sample periods. The sizes of standard deviations of price coefficients are uniformly larger in value-oriented equations than in high-end equations. It implies that variation in price sensitivity for the general formats’ shoppers (value-oriented and supermarkets) is greater than the variation for the high-end format’s shoppers. It supports the expectation in H2. Third, the results from the income coefficients suggest that high-income households are more likely to shop at supermarkets than value-oriented stores, and at high-end than supermarkets. However, the latter is not statistically significant. In addition, the coefficients of organic shopper type in both periods confirm the prediction that frequent organic shoppers prefer high-end format retailers. Fourth, the coefficients of quality indexes (log of the ratio of organic supply) represent the store image effects associated with consumer fixed utility component. The estimation result shows an interesting pattern of cross-format effects in the estimation of 2005-06 sample: an increase in quality ratio of value-oriented to supermarkets has positive effects on log odds of high-end formats versus supermarkets, and an increase in quality ratio of high-end to supermarkets also has positive effects on log odds of value-oriented versus supermarkets. In 2007-08, the own-format effect is significant in the high-end equation. Along with the observation that the difference between a value-oriented format’s organic supply and the other two formats’ was reduced over the two periods, the result suggests that own-format effect of a quality improvement arises only when quality difference between formats is small enough. In other words, less effective impact of quality may hinder quality improvement for a market with well distanced quality settings among formats. In addition, I observe statistically significant standard deviations in these random coefficients from both periods of 64 estimation. This implies that the store image effect varies among households in the estimated samples. Regarding to location and factors affecting shopping costs in SFC, the results confirm the impacts of shopping cost on SFC. In particular, the households with shopping patterns of more days between trips or longer distance tend to choose supermarkets over the value- oriented retailers or high-end specialty stores. This suggests that the households with higher shopping costs prefer formats featuring one-stop shopping environment - supermarkets, which feature broader assortments and more varieties in food categories. The results from the coefficient of number of stores in each format within the shopping range of households suggest a network effect on consumer’s shopping decision. As its store number increases, the format itself gains in the likelihood of consumer store visits. The cross-format coefficients indicate a complimentary relationship between value-oriented and high-end formats, as an increase in the number of stores of either format would have positive impacts on the probability for consumers to visit the other format. This implies that the low-high bundle of value- oriented and high-end formats appeals to households rather than other combinations, such as supermarkets and high-end formats together. Furthermore, a household’s SFC decision is dependent of her basket size and composition. The estimation results show no significant difference in the impact of total spending per trip on the choice among the alternatives. Confirming the common belief, I find that households with higher expense share in dry grocery are most likely to shop at value-oriented stores than supermarkets, while organic food consumption helps the decision to shop at the high- end stores over the other two. This is also consistent with the implications from coefficient estimates of organic dummy, i.e. organic shoppers are more likely to shop at high-end stores than supermarkets, while they prefer supermarkets to the value-oriented format. The results do not show a differential shopping pattern for households of different types, from the parameter estimates of the demographic variables. 65 Finally, the scale of price sensitivity reduced significantly in value-oriented equation for 2007-08 compared to 2005-06, while price sensitivity in high-end.equation increases over time. This is likely a result of increasing supply in organics due to Wal-Mart’s market expansion in the summer of 2006. The connection suggests that retailers’ increased supply on organic products have a positive impact on consumers’ quality perception and retailers’ store image.

Marginal Eﬀect of Price

A direct interpretation of the parameter estimates reported in Table 3.5 provides only limited substantive behavioral meaning about the effect of specific factor on consumer’s SFC outcomes. This arises when dependent variables appear as stand alone variables and as nonlinear transformed variable and when they are interacted with other dependent variables. In these instances, a standard approach is to compute the full marginal effect of a dependent variable’s impact on the model, and the statistical significance of a utility parameter does not imply the same significance for the marginal effect. Therefore, I further compute the marginal effect of price for various income groups of households. A marginal effect, defined as the derivative of the probability, is the influence a one unit change in an explanatory variable has on the probability of selecting a particular outcome. Under the present model setup, the effect of a change in price involves three terms: 1) main term (price), 2) income interaction term (income*price), and 3) income squared interaction term (income2*price). Therefore, the marginal effect of price would vary among income groups as a result of these effects combined. Furthermore, the price effect will vary among households, regardless of their income, since the mixed logit model allows random parameters for the main term of price. Let y denote the dependent variable, which is a dummy variable of SFC. The conditional 66 mean of the dependent variable is

2 E[y|p, m, X] = L(β1p + β2mp + β3m p + Xβ) ≡ L(u) (3.6)

2 where L is the logistic function and u denotes the index β1p+β2mp+β3m p+Xβ, p denotes price and m denotes income. Suppose that price and income (p, m) are both continuous. The marginal eﬀect of price is thus

∂L(u) = (β + β m + β m2)L0(u). (3.7) ∂p 1 2 3

2 0 Unlike linear models, the marginal effect consists of (β1 + β2m + β3m ) and L (u). Figure 3.2 presents the marginal effects for various income groups, by plotting the percentage change in predicted probabilities for a percentage change in prices for value-oriented and high-end equations for both periods of samples, where the horizontal axis is income in percentile and the vertical axis is the marginal effect of price.4 The solid lines represent the means of marginal effects of price calculated from individual household’s marginal effect, and the dotted lines represent the upper and lower bounds of confidence intervals (95% & 5%).5 As shown in the figures, those results from comparison of marginal effects of price among households and across formats support the hypotheses: H1. Price sensitivity varies among households of different types and among formats, and H2. Variation in price sensitivity for the general formats’ shoppers is expected to be greater than the variation for the high-end format’s shoppers. First, the marginal effects of price vary among income groups. Particularly, the magni-

4In estimation, one percentage change in price was applied to compute the difference in predicted probabilities between with and without the change in price for each individual. The difference in predicted probabilities was then divided by the predicted probability before change to be reported as the marginal effect due to a 1 % change in price. This measure is actually in a form of elasticity. 5The upper bound is valued at (mean + 1.05*sd) and the lower bound is valued at (mean - 0.96*sd) for our sample, of which size is over 1000. 67

(a) Value-oriented, 2005-06 (b) High-end, 2005-06

Figure 3.2: Price Effect on SFC Probability by Income Groups 68 tude of marginal effect decreases as income increases. For those rich households, the marginal effect of price is even greater than zero at the mean values and even the entire confidence interval for the top income groups. For value-oriented equation (versus supermarkets) in the sample of 2005-06, the marginal effects vary from -4.475 to 0.176, and I observe significant differences in standard deviation across income groups: the variation of marginal effect is much greater for the poor than for the rich. I observe an upward-sloped curve for high- end equation in 2005-06 as well, but the variation in marginal effect is maximized at the top income groups instead. Upward-sloped curves are observed for both value-oriented and high-end equations in 2007-08 sample too, although the variations in marginal effects are more uniformly sized across income groups. In sum, those results support hypothesis H1. The marginal effects of price differ across formats as well: the effects in value-oriented equation are about 2-3 times of those in high-end equations. This provides support for the latter part of hypothesis H1. Next, the variation in marginal effect for high-end shoppers is relatively smaller than that for value-oriented shoppers as well – The coefficient of variation (CV), defined as the ratio of the standard deviation to the mean, is 0.27 for high-end and 0.90 for value-oriented in 2005-06; 0.57 for high-end and 0.92 for value-oriented in 2007-08. It thus provides supporting evidence for hypothesis H2.

3.4 Concluding Remarks

In this chapter, I use a unique household purchase panel data to investigate the determinants of consumer SFC decision making and examine the theoretical predictions from chapter 2. Several key findings emerge from the analysis of a non-coastal U.S. metropolitan area in 2005- 06 and 2007-08. First, I find that household’s price sensitivity in SFC decreases with income, and varies among households. The variation in price sensitivity is relatively noteworthy for the general formats’ shoppers compared to high-end shoppers. Second, organic food 69 demand and supply are both important factors influencing households in choosing the high- end specialty format. In particular, the estimated random parameters for organic supply ratio show statistically significant variation in store image effect among households and the effect due to own format’s quality improvement is significant only when quality settings are more comparable among formats. On the demand side, the results suggest that organic food consumption is an important reason leading consumers to choose a high-end format over the other two formats. Third, the results from the network effect of the number of stores in the area suggest a complementary relationship between value-oriented and high- end formats, a likely pair of inferior and superior good combination, while supermarkets format presents the features of an one-stop shopping outlet. In addition, I find supportive evidence that shopping costs measured by days between trips and the distance between store and household are important to consumers’ format choice between the specialty format and one-stop shopping formats, confirming the common findings in the literature. Finally, basket size and composition on the shopping list, format loyalty and shopper types are also found to be important to a household’s SFC decision making. In sum, my findings verify key predictions from the theory. First, the shoppers of general formats (value-oriented stores and supermarkets) have more diversified preferences than those that shop at high-end specialty formats. Moreover, high-end specialty format shoppers have higher WTP and quality perception compared to general format shoppers. Second, the value factor, defined as the ratio of quality to price, is central in the theoretical model in determining the consumer SFC decision. Since value is the deterministic factor, both price and quality effects are considered empirically. Even with a quality proxy like the organic supply rate used in this study, the estimated price effect may still contain partially the impact due to unobserved difference in quality perception among households. As a result, observed price sensitivity should vary among consumers and among formats. And indeed, the estimation results verify the prediction. 70

Chapter 4: Conclusion

Motivated by the emerging high-end food shopping formats the U.S. (e.g. Whole Foods Market), and the growing market shares of value formats world-wide (e.g. Wal-Mart), I study the question of store format choice. The theory of store format choice has been poorly developed in the extant literature . Without a strong theoretical foundation, empirical findings have been largely exploratory in nature. For example, Fox, Montgomery, and Lodish (2004) empirically examined competition between retail formats and explored how retailers’ assortment, pricing, promotional policies, and customers’ demographics affect shopping behavior. They found statistically insignificant price effects and so concluded that levels of assortment and promotion are more important determinants than price on consumer expenditures. However, value should be the key to consumer’s shopping decision as suggested by my theory. Therefore, price and quality effects should be considered together. Without allowing for variation in price parameters in their study, the mean estimates would likely suffer from bias due to unobserved taste heterogeneity, and so misrepresent the actual price effect and lead to incorrect arguments. My theory of store format choice is grounded in the neoclassical tradition of consumer economics (see for example Moescallell). I study the rationales of consumer SFC using a novel two part utility function. The fixed utility component allows me to introduce store image effect due to fixed evaluation of a store’s quality, The variable utility component introduces the traditional consumer demand factors, like price and shopping baskets, relevant to consumer purchase decisions. Together, the two-part utility structure provides a way to investigate consumer responses to retailer format choices of quality and pricing. The model is 71 used in a simulation that shows how retailer strategies impact consumer shopping behavior. The simulation aggregates the individual shopping decisions leading to predictions about market shares of each format . The simulation results show that the effect of quality change on retailer market share is in a U shape with respect to quality levels, while the effect of price change on retailer market share is in an inverse U shape with respect to the inverse of value. These marginal benefit curves together with the marginal cost curves derived from market share and price margin can be used to identify profit-maximizing value, quality and price. This result provides a useful approach for retailer to construct the optimal production and supply decisions. The theoretical model provides several standard hypotheses tests such as the impact of shopping cost advantage to promote one-stop shopping behavior. Additionally, two unique hypotheses tests provided a motivation for my empirical work. First, the price sensitivity varies among households and among formats. Second. variation in price sensitivity is smaller among the high-end format’s shoppers than among the general formats’ shoppers. In order to make evaluations of these hypotheses, data on store quality is essential. In the U.S. one seemingly important dimension in store formats is the development and promotion of organic food, which is widely perceived as providing quality benefits to consumers. I then specified an econometric model based on the theoretical framework and used actual purchase data to examine the theoretical predictions. The estimation results from a mixed multinomial logit model show supporting evidence for the theory and provide further understanding of consumer SFC and shopping behaviors in responding to retailers’ strategies. Next, I summarize the findings and remark on the limitations of the present study and the directions for future studies. 72

4.1 Summary and Implications

The theoretical framework was built on consumer’s maximization of perceived utility. Con- sumers are considered to be heterogeneous in both fixed and variable WTPs for product quality. With the model setup, a retailer’s quality setting affects consumers’ overall evaluation for the store (store image effect) as well as their purchase decisions based on the retailer’s value and consumers’ variable WTPs. This two-dimensional impact of quality setting on consumer SFCs plays a key role in directing the findings of this study. With this model, I have identified how consumer’s SFCs behave in WTP space. I have also analyzed how retailer’s pricing and quality settings would affect the market shares for each format of retailers, and how one may utilize these results from market shares to identify profit maximizing quality and value. Finally, with incorporating income in the model, I have shown that difference in income inequality may have contributed to the outlook of market shares in the present markets. The model suggests two main results. First, value, defined as quality-to-price ratio, is a key to consumer SFC and price sensitivity in SFC vary among consumers and among formats. Second, the shoppers of general formats (L: value-oriented and M: supermarkets) are with diversified preferences while the high-end specialty format H’s shoppers have high WTP and quality perception. Both model predictions were addressed by the estimation of random parameters with a mixed multinomial logit model and the results of marginal effect of price provide supporting evidence for these implications. In particular, ‘income’ was used in this empirical investigation to capture the variation in consumer willingness to pay and quality perception. The results show that price sensitivity varies among consumers of different income levels, and the richer is less price sensitive compared to the poorer. In addition, the theory and empirical applications also confirm a common finding in the literature. That is, shopping costs revealed by the distance between retailer and consumer, 73 shopping frequency and location accessibility to the formats, help explain SFC. In addition, I find a complimentary relationship between the low-end value-oriented and the high-end specialty formats, reflecting the importance of inter-category connection in consumer SFC. In sum, I have built a great foundation for SFC analysis and provided supporting evidence on model predictions from a unique actual purchase data set. I showed that 1) organic consumption and retailer strategies around this product are important to consumer SFC, 2) value is the key to consumer SFC decision making and it is essential to take differential quality perception into account when estimating price effect, and 3) income distribution and inequality has influential impacts on the structure of food retailing industry. These results provide useful insights for farmers and retailers in their marketing and developing decisions on organic agriculture.

4.2 Limitations and Directions for Future Studies

While this dissertation provides useful insights on consumer SFC decision making, there are several noteworthy limitations that are explored as part of this proposal. First, the model presented above focuses on the utility associated with a single SFC decision. In this context, the model will not capture SFC decisions for food buyers that exhibit more complex shopping patterns. For instance, consumers decide which store (format) to visit or not visit, as well as decide the set of product categories to purchase and what quantity to purchase for each store (format). Therefore, the format choice problem would likely contain both discrete (SFC and product category choice) and continuous (quantity) choice dimensions. Without the integration of several consumer decisions in a unified model of consumer utility maximization, it is difficult to justify an empirical approach that avoids omitted variable bias and associated endogeneity problems. Second, the above model presumes well-behaved functional forms leading to interior 74 solutions. However, an important feature of consumer choice behavior is the prevalence of corner solutions, wherein consumers are observed not to visit certain store formats or not to purchase any quantity of certain commodities. Analyzing demand decisions for general corner solutions are complex. In the literature, the Kuhn-Tucker (KT) approach suggested by for example Hanemann (1984); Vsquez Lavin (2007); Vsquez Lavin and Hanemann (2008) provides a promising avenue for this type of choice problem with corner solutions. Third, the relationship between income and WTP is important to explain SFC and purchase decisions. Utilizing a linear function to link WTP and income serves a good start point. However, further investigation, such as nonlinear or nonparametric approaches, on this issue is not only conceptually important, but also can provide enhanced understanding of consumer food purchase decision making and help resource allocation and efficiency in food retailing and farming. Finally, the model can be extended by considering more realistic cases in which shopping costs may vary based on consumers’ spatial location and opportunity cost of time, instead of the homogeneous shopping costs as assumed in this dissertation. Consumer heterogeneity in these contexts can be specified by either a discrete or continuous distribution. 75

Bibliography

Anderson, S.P., and A. de Palma. 2006. “Market Performance with Multiproduct Firms.” Journal of Industrial Economics 54:95–124.

—. 1992. “Multiproduct Firms: A Nested Logit Approach.” Journal of Industrial Economics 40:261–276.

Basker, E. 2005. “Selling a Cheaper Mousetrap: Wal-Mart’s Eﬀect on Retail Prices.” Journal of Urban Economics 58:203–229.

Basker, E., and M. Noel. 2007. “The Evolving Food Chain: Competitive Eﬀects of Wal- Marts Entry into the Supermarket Industry.” PhD dissertation, Department of Economics, University of Missouri, Working Papers.

Baumgartner, S., M.A. Drupp, J.M. Munz, J.N. Meya, and M.F. Quaas. 2011. “Income Distribution and Willingness to Pay for Ecosystem Services.”

Bell, D.R., T. Ho, and C.S. Tang. 1998. “Determining Where To Shop: Fixed and Variable Costs of Shopping.” Journal of Marketing Research 35:352–369.

Bell, D.R., and J.M. Lattin. 1998. “Shopping Behavior and Consumer Preference for Store Price Fromat: Why ”Large Basket” Shopping Prefer EDLP.” Marketing Science 17:66–88.

Bhatnagar, A., and B.T. Ratchford. 2004. “A Model of Retail Format Competition for Non- durable Goods.” International Journal of Research in Marketing 21:39–59. 76 Brander, J.A., and J. Eaton. 1984. “Product Line Rivalry.” American Economic Review 74:323–334.

Bruns, K., T.A. Fjord, and K.G. Grunert. 2002. “Consumers’ Food Choice and Quality Perception.” Working paper, the MAPP Centre, Aarhus School of Business, University of Aarhus, June 1, 2002.

Carpenter, J.M., and M. Moore. 2006. “Consumer Demographics, Store Attributes, and Retail Format Choice in the US Grocery Market.” International Journal of Retail and Distribution Management 34:434–452.

Champsaur, P., and J.C. Rochet. 1989. “Multiproduct Duopolists.” Econometrica 57:533– 557.

Doraszelski, U., and M. Draganska. 2006. “Market Segmentation Strategies of Multiproduct Firms.” Journal of Industrial Economics 54:125–149.

Economides, N. 1989. “Quality Variations and Maximal Variety Diﬀerentiation.” Regional Science and Urban Economics 19:21–29.

—. 1993. “Quality Variations in the Circular Model of Variety-Diﬀerentiated Products.” Regional Science and Urban Economics 23:235–257.

Ellickson, P., and S. Misra. 2007. “Supermarket Pricing Strategies.” PhD dissertation, Duke University, Department of Economics, Working Papers.

Ellickson, P.B. 2006. “Quality Competition in Retailing: A Structural Analysis.” Interna- tional Journal of Industrial Organization 24:521–540.

Fox, E.J., A.L. Montgomery, and L.M. Lodish. 2004. “Consumer Shopping and Spending across Retail Formats.” Journal of Business 77:S25–60. 77 Fox, E.J., S. Postrel, and A. McLaughlin. 2007. “The Impact of Retail Location on Retailer Revenues: An Empirical Investigation.”

Franklin, A.W. 2001. “The Impact of Wal-Mart Supercenters on Supermarket Concentration in U.S. Metropolitan Areas.” Agribusiness 17:105–114.

Gilbert, R.J., and C. Matutes. 1993. “Product Line Rivalry with Brand Diﬀerentiation.” Journal of Industrial Economics 41:223–240.

Hamilton, S.F., and T.J. Richards. 2006. “Variety Competition in Retail Markets.”

Hanemann, W.M. 1984. “Discrete/Continuous Models of Consumer Demand.” Econometrica 52:541–561.

Ho, T., C.S. Tang, and D.R. Bell. 1998. “Rational Shopping Behavior and the Option Value of Variable Pricing.” Management Science 44:145–160.

Hole, A.R. 2004. “Fitting Mixed Logit Models by Using Maximum Simulated Likelihood.” The Stata Journal 7:388–401.

Hotelling, H. 1929. “Stability in Competition.” The Economic Journal 37:41–57.

Kavilanz, P.B. 2008. “The High Price of Going ’Organic’.” CNNMoney.com, pp. 2.

Lal, R., and R. Rao. 1997. “Supermarket Competition: The Case of Every Day Low Pricing.” Marketing Science 16:60–80.

Lancaster, K. 1975. “Socially Optimal Product Diﬀerentiation.” American Economic Review 65:567–85.

Messinger, P.R., and C. Narasimhan. 1997. “A Model of Retail Formats Based on Consumers’ Economizing on Shopping Time.” Marketing Science 16:1–23. 78 Organic Trade Association. 2009. “The OTA 2009 Organic Industry Survey.” Working paper, Organic Trade Association, May 2009.

Salop, S. 1979. “Monopolistic Competition with Outside Goods.” Bell Journal of Economics 10:141–56.

Singh, V.P., K.T. Hansen, and R.C. Blattberg. 2004. “Impact of Wal-Mart Supercenter on a Traditional Supermarket: An Empirical Investigation.”

Stiegert, K.W., and T. Sharkey. 2007. “Food Pricing, Competition, and the Emerging Su- percenter Format.” Agribusiness 23:295–312.

The Hartman Group. 2008. “The Many Faces of Organic 2008.” Working paper, The Hart- man Group, Inc., Summer 2008.

Train, K. 2003. Discrete Choice Methods with Simulation. Canbridge: Cambridge University Press.

Vsquez Lavin, F.A. 2007. “Discrete/Continuous Choice Models and Consumer Heterogene- ity.” Doctor of philosophy, University of California, Berkeley.

Vsquez Lavin, F.A., and M. Hanemann. 2008. “Functional Forms in Discrete/Continuous Choice Models With General Corner Solution.” Working paper, Department of Agricul- tural and Resource Economics, UCB, University of California-Berkeley, December 30, 2008.