The Economics of the Internet and E-Commerce (Advances in Applied Microeconomics, Vol. 11)

Citation preview

THE ECONOMICS OF THE INTERNET AND E-COMMERCE

ADVANCES IN APPLIED MICROECONOMICS Series Editor: Michael R. Baye

ADVANCES IN APPLIED MICROECONOMICS

VOLUME 11

THE ECONOMICS OF THE INTERNET AND E-COMMERCE EDITED BY

MICHAEL R. BAYE Bert Elwert Professor of Business Economics & Public Policy, Kelley School of Business, Indiana University, USA

2002

JAI An Imprint of Elsevier Science Amsterdam – Boston – London – New York – Oxford – Paris San Diego – San Francisco – Singapore – Sydney – Tokyo

ELSEVIER SCIENCE Ltd The Boulevard, Langford Lane Kidlington, Oxford OX5 1GB, UK © 2002 Elsevier Science Ltd. All rights reserved. This work is protected under copyright by Elsevier Science, and the following terms and conditions apply to its use: Photocopying Single photocopies of single chapters may be made for personal use as allowed by national copyright laws. Permission of the Publisher and payment of a fee is required for all other photocopying, including multiple or systematic copying, copying for advertising or promotional purposes, resale, and all forms of document delivery. Special rates are available for educational institutions that wish to make photocopies for non-profit educational classroom use. Permissions may be sought directly from Elsevier Science Global Rights Department, PO Box 800, Oxford OX5 1DX, UK; phone: ( + 44) 1865 843830, fax: ( + 44) 1865 853333, e-mail: [email protected]. You may also contact Global Rights directly through Elsevier’s home page (http://www.elsevier.com), by selecting ‘Obtaining Permissions’. In the USA, users may clear permissions and make payments through the Copyright Clearance Center, Inc., 222 Rosewood Drive, Danvers, MA 01923, USA; phone: ( + 1) (978) 7508400, fax: ( + 1) (978) 7504744, and in the UK through the Copyright Licensing Agency Rapid Clearance Service (CLARCS), 90 Tottenham Court Road, London W1P 0LP, UK; phone: ( + 44) 207 631 5555; fax: ( + 44) 207 631 5500. Other countries may have a local reprographic rights agency for payments. Derivative Works Tables of contents may be reproduced for internal circulation, but permission of Elsevier Science is required for external resale or distribution of such material. Permission of the Publisher is required for all other derivative works, including compilations and translations. Electronic Storage or Usage Permission of the Publisher is required to store or use electronically any material contained in this work, including any chapter or part of a chapter. Except as outlined above, no part of this work may be reproduced, stored in a retrieval system or transmitted in any form or by any means, electronic, mechanical, photocopying, recording or otherwise, without prior written permission of the Publisher. Address permissions requests to: Elsevier Science Global Rights Department, at the mail, fax and e-mail addresses noted above. Notice No responsibility is assumed by the Publisher for any injury and/or damage to persons or property as a matter of products liability, negligence or otherwise, or from any use or operation of any methods, products, instructions or ideas contained in the material herein. Because of rapid advances in the medical sciences, in particular, independent verification of diagnoses and drug dosages should be made. First edition 2002 Library of Congress Cataloging in Publication Data A catalog record from the Library of Congress has been applied for. British Library Cataloguing in Publication Data A catalogue record from the British Library has been applied for. ISBN: 0-7623-0971-7 ∞ The paper used in this publication meets the requirements of ANSI/NISO Z39.48-1992 (Permanence of 䊊 Paper). Printed in The Netherlands.

CONTENTS LIST OF CONTRIBUTORS

vii

PREFACE

ix

THE IMPACT OF THE INTERNET ON HORIZONTAL AND VERTICAL COMPETITION: MARKET EFFICIENCY AND VALUE CHAIN RECONFIGURATION Anita Elberse, Patrick Barwise and Kathy Hammond

1

PRICE COMPETITION BETWEEN PURE PLAY VERSUS BRICKS-AND-CLICKS E-TAILERS: ANALYTICAL MODEL AND EMPIRICAL ANAYLSIS Xing Pan, Venkatesh Shankar and Brian T. Ratchford

29

PRICE DISPERSION THEN AND NOW: EVIDENCE FROM RETAIL AND E-TAIL MARKETS Patrick Scholten and S. Adam Smith

63

BUSINESS-TO-BUSINESS E-COMMERCE: VALUE CREATION, VALUE CAPTURE AND VALUATION Luis Garicano and Steven N. Kaplan

89

TRUST AMONG STRANGERS IN INTERNET TRANSACTIONS: EMPIRICAL ANALYSIS OF eBAY’S REPUTATION SYSTEM Paul Resnick and Richard Zeckhauser

127

TRANSACTION INNOVATION AND THE ROLE OF THE FIRM Daniel F. Spulber

159

COMBINATORIAL AUCTIONS IN THE INFORMATION AGE: AN EXPERIMENTAL STUDY John Morgan

191

v

vi

ANALYZING WEBSITE CHOICE USING CLICKSTREAM DATA Avi Goldfarb

209

CONSUMER ACQUISITION OF PRODUCT INFORMATION AND SUBSEQUENT PURCHASE CHANNEL DECISIONS Michael R. Ward and Michelle Morganosky

231

AN ECONOMIC ANALYSIS OF MULTIPLE INTERNET QoS CHANNELS Dale O. Stahl, Rui Dai and Andrew B. Whinston

257

LIST OF CONTRIBUTORS Patrick Barwise

Department of Marketing, London Business School, London, U.K.

Rui Dai

MSIS Department, University of Texas, Austin, TX, USA

Anita Elberse

Department of Marketing, London Business School, London, U.K.

Luis Garicano

Graduate School of Business, University of Chicago, Chicago, IL, USA

Avi Goldfarb

Department of Marketing, Joseph L. Rotman School of Management, University of Toronto, Toronto, Canada

Kathy Hammond

Department of Marketing, London Business School, London, U.K.

Steven N. Kaplan

Graduate School of Business, University of Chicago, Chicago, IL, USA

John Morgan

Haas School of Business, University of California at Berkeley, Berkeley, CA, USA

Michelle Morganosky

Department of Agricultural and Consumer Economics, University of Illinois, Urbana, IL, USA

Xing Pan

Department of Marketing, Robert H. Smith School of Business, University of Maryland, College Park, MD, USA

Brian T. Ratchford

Department of Marketing, Robert H. Smith School of Business, University of Maryland, College Park, MD, USA vii

viii

LIST OF CONTRIBUTORS

Paul Resnick

School of Information, University of Michigan, Ann Arbor, MI, USA

Patrick Scholten

Department of Business Economics & Public Policy, Kelley School of Business, Indiana University, IN, USA

Venkatesh Shankar

Department of Marketing and Entrepreneurship, Robert H. Smith School of Business, University of Maryland, College Park, MD, USA

S. Adam Smith

Newell Rubbermaid, Fayetteville, AR, USA

Daniel F. Spulber

Kellogg School of Management, Northwestern University, Evanston, IL, USA

Dale O. Stahl

Department of Economics, University of Texas, Austin, TX, USA

Michael R. Ward

Department of Agricultural and Consumer Economics, University of Illinois, Urbana, IL and Department of Economics, University of Texas at Arlington, USA

Andrew B. Whinston

MSIS Department, University of Texas, Austin, TX, USA

Richard Zeckhauser

John F. Kennedy School of Government, Harvard University, Cambridge, MA, USA

PREFACE The Internet has revolutionized the way consumers and firms interact in the marketplace, and it has dramatically changed the information enjoyed by market participants at various points in the value chain. This timely volume on the Internet and e-commerce provides academics and practitioners with cutting-edge research on the “glue” that holds the new economy together. The first six chapters of The Economics of the Internet and E-Commerce examine four broad issues: the role of the Internet in fostering competition, its impact on price dispersion and on business-to-business transactions, and the importance of reputation and trust in the new economy. The last four chapters examine the impact of the Internet on the organization of firms, the efficiency of auctions in the Internet age, how consumers choose websites and acquire product information, and the growing problem of congestion on the Internet. I am especially proud of the interdisciplinary scope of this volume and the international diversity of its elite group of contributors. The authors of these ten chapters are experts in diverse fields (ranging from economics, marketing and management to information sciences), and are affiliated with some of the leading academic institutions and businesses in the U.S., Canada, and Europe. Volume 11 is part of Elsevier’s Advances in Applied Microeconomics series – an annual research volume that disseminates frontier research well in advance of journals and other outlets. For additional information about e-commerce and Internet-related projects, or for more information about this series and planned future volumes, please visit the editor’s site at http://www. nash-equilibrium.com. Michael R. Baye [email protected] Bloomington, IN July 2002

ix

This Page Intentionally Left Blank

THE IMPACT OF THE INTERNET ON HORIZONTAL AND VERTICAL COMPETITION: MARKET EFFICIENCY AND VALUE CHAIN RECONFIGURATION1 Anita Elberse, Patrick Barwise and Kathy Hammond ABSTRACT In this chapter, we review research on the Internet’s impact on ‘horizontal’ and ‘vertical’ competition. First, focusing on horizontal competition, we examine theory and empirical evidence on the extent to which the Internet increases market efficiency, as well as possible underlying explanations for the observed empirical patterns. Second, turning to vertical competition among market players within the value chain, we analyze the extent to which the Internet leads to ‘disintermediation’, ‘reintermediation’, or other forms of value chain reconfiguration. We find little support for early predictions that the Internet will have a dramatic impact on horizontal and vertical competition.

INTRODUCTION Since the Internet took off in the mid-1990s, researchers in areas such as marketing, strategy, communications and economics have examined its impact The Economics of the Internet and E-Commerce, Volume 11, pages 1–27. Copyright © 2002 by Elsevier Science Ltd. All rights of reproduction in any form reserved. ISBN: 0-7623-0971-7

1

2

ANITA ELBERSE, PATRICK BARWISE AND KATHY HAMMOND

on markets. They have provided important insights into the extent to which the Internet changes the fundamental principles that govern markets and the relationship between market players. In this chapter, we set out to provide an overview of key findings to-date in two of the most active and advanced areas of the emerging field of ‘Internet economics’. First, we focus on ‘horizontal’ competition within markets. Here, we examine evidence on whether the Internet will, as is often claimed, lead to ‘frictionless markets’ – markets in which empowered customers, increasingly supported by intelligent agents, trusted intermediaries and third parties, shop around with minimal effort, playing one supplier off against another and relentlessly driving down prices. That is, in economic terms, we look at whether the Internet, by reducing transaction and search costs, decreases the ability of sellers to extract monopolistic profits and increases the ability of markets to optimally allocate resources (e.g. Bakos, 1997, 1998). After discussing theoretical notions and the likelihood of frictionless markets, we move to empirical research on the Internet’s impact on average market prices, price dispersion, price elasticity, and menu costs. We continue with underlying explanations for the observed empirical patterns. Here, we pay particular attention to research on so-called ‘shopbots’ or price search engines. We conclude with research on bundling, versioning, and auctions – mechanisms that sellers can employ to counter the potential erosion of their margins. Second, we turn to ‘vertical’ competition among players within the value chain. In this section, we examine another much debated issue – whether the Internet leads to disintermediation, i.e. threatens intermediaries between sellers and buyers. Those who argue in favor of this viewpoint generally point to the Internet’s potential to restructure and redistribute profits among players along the value chain. We review theoretical and empirical research in favor of this perspective as well as opposing claims of ‘reintermediation’, touching on such concepts as ‘cybermediation’ and ‘hypermediation’. This debate is strongly related to that about the Internet’s potential to increase market efficiency – some commentators have argued that disintermediation is an inherent feature of frictionless electronic markets. Throughout this chapter, we draw on refereed articles, books, and book chapters in marketing, economics, management or business studies, communications, as well as related disciplines. Although our focus is on academic research, we also discuss relevant practitioner-oriented studies (e.g. as published in Harvard Business Review). Importantly, we do not limit ourselves to published research – given the dynamic nature of the topic, we specifically seek to include working papers and manuscripts under review.

The Impact of the Internet on Horizontal and Vertical Competition

3

THE INTERNET’S IMPACT ON ‘HORIZONTAL COMPETITION’ Internet Markets: Less Friction? The ideas underlying the hypothesis that Internet markets exhibit less friction can be traced back decades. For example, Stigler (1961) attributed price dispersion to incomplete information, which implies that market friction should reduce when information is more readily available to consumers and/or when consumer search is less costly. In an early study of the effects of information technologies on market structures, Malone, Yates and Benjamin (1987) argued that electronic markets are likely to have less ‘friction’ than their offline counterparts, because they have lower ‘coordination’ or transaction costs. In a later article, Wigand and Benjamin (1995) gave a more comprehensive framework for understanding the effects of electronic commerce, including lower search costs and increased access to a broad selection of lower-priced goods for consumers, as well as reduced profit margins for retailers. Also focusing on the impact of reducing search costs, Bakos (1997) modeled their role in electronic markets with differentiated product offerings. He explored the implications for the incentives of buyers, sellers, and independent intermediaries to invest in such marketplaces. Importantly, and in line with Wigand and Benjamin’s (1995) main arguments, Bakos’s analysis provides formal support for the proposition that electronic marketplaces promote price competition and reduce the market power of sellers. More recently, Sinha (2000) argued that the ease of collecting and comparing information on the Web regarding product prices, features, and quality means that costs are becoming increasingly transparent. This, according to Sinha, will impair sellers’ ability to obtain high margins, turning most products and services into commodities. He suggested that the Internet “encourages highly rational shopping”, eroding the ‘risk premium’ that sellers have been able to extract from wary buyers (Sinha, 2000, p. 47). It also demands, he argued, that companies with different prices in different countries re-examine their price structure. Assessing possible solutions, Sinha indicated that one response for firms seeking to counter this threat is ‘smart’ pricing, through strategies like versioning or mechanisms like auctions. However, he noted that such ‘smart’ pricing may be extremely risky in the long term, as it may create perceptions of unfairness among consumers who are now able to share price information easily. Instead, Sinha favored another solution – a combination of product quality, innovation, and bundling.

4

ANITA ELBERSE, PATRICK BARWISE AND KATHY HAMMOND

Sinha’s succinct but wide-ranging article touches on key issues regarding the Internet’s impact on horizontal competition. Below, we first discuss empirical evidence to-date on dimensions of Internet market efficiency – price levels, price dispersion, price elasticity, and menu costs. We then discuss possible underlying explanations for these empirical patterns, both in general and as suggested by research on shopbots. We conclude the section with literature on bundling and versioning, and auctions. Internet Market Efficiency: Price Levels, Dispersion and Elasticity, and Menu Costs Smith, Bailey and Brynjolfsson (2000) reviewed early empirical evidence on the Internet’s impact on market efficiency. They found that Internet markets are more efficient than conventional markets with respect to average price levels, menu costs, and price elasticity. However, despite the presence of conditions to foster efficiency, they also found substantial and persistent price dispersion. They suggested that this may be partly explained by heterogeneity in retailerspecific factors such as trust and awareness (i.e. brand equity). In addition, Internet markets are still at an early stage and may change more radically with the development of cross-channel sales strategies, intermediaries and shopbots, improved supply chain management, and new information markets. In a recent book chapter aptly titled ‘The Great Experiment’, Clay, Krishnan and Smith (2001) updated this review by incorporating more recent academic studies on the Internet’s impact on retail prices. They also examined how firms set prices, and how consumers respond to prices on the Internet. As Smith, Bailey and Brynjolfsson (2000), and Clay, Krishnan and Smith (2001) also point out, signs of reduced friction are present in several studies. For example, Scott Morton, Zettelmeyer and Silva-Risso (1999) found that car buyers who used the Autobytel.com online referral service paid, on average, 2% less than customers who bought offline. A large share of these savings appears to be due to the bargaining power of the referral service and the lower cost of serving an online consumer. Similarly, in a study on life insurance, Brown and Goolsbee (2002) found that prices for term life policies fell 8% to 15% between 1995 and 1997. Although other factors contributed to lower prices, they showed that the rising use of shopbots, specialized web sites that facilitate comparison shopping, explained up to half of the total decline. Interestingly, they found that the introduction of comparison search sites initially increased price dispersion; it decreased only as use of these sites became more widespread.

The Impact of the Internet on Horizontal and Vertical Competition

5

Overall, however, Internet markets are not nearly as efficient as theories would predict. For instance, using data collected over two months in 1997 for Internet-only retailers and a matched set that primarily uses conventional channels, Bailey (1998a, b) found that prices for books, music CDs, and computer software were no lower on the Internet than in conventional channels. However, he did find evidence for a lower variance of prices for Internet retailers, as well as a higher frequency of price changes, compared with physical retailers. Earlier, Lee (1998) found that prices for used cars were on average higher on the Internet than in conventional channels – although it should be noted that his study has important shortcomings (see Clay et al., 1999). Erevelles, Rolland and Srinivasan (2001) explored the pricing behavior of Internet versus traditional firms for the vitamin industry. Their findings showed that, again contrary to theoretical notions, the average price/unit of vitamins is significantly higher for Internet retailers than for traditional retailers, even for private label brands. Their results also indicated that price dispersion (variation in prices) is significantly higher for Internet retailers than for traditional retailers, again contrary to what one would expect from theory. Clemons, Hann and Hitt (2002) found empirical evidence for wide price dispersion among online travel agents, with ticket prices varying by up to 28% for the same customer request, and up to 18% after accounting for ticket quality and route differences using a hedonic price model. They concluded that online travel agents engage in both horizontal product differentiation and price discrimination in addition to simply having a degree of random inefficiency. In a comprehensive study of price and non-price competition in the online book industry, Clay et al. (1999) collected prices of over 100 titles sold by 13 online and two physical bookstores. Controlling for book characteristics, they found that average prices in online and physical bookstores were the same. They found significant price dispersion among online bookstores, providing indirect evidence of perceived product or brand differentiation, enabling Amazon in particular to charge a “substantial premium, (. . .) even relative to barnesandnoble.com and Borders.com” (Clay et al., 1999, p. 1). In a larger follow-up study, using data collected over a period of 6 months, covering nearly 400 books, and 32 online bookstores, Clay, Krishnan and Wolff (2001) found no change in either prices or price dispersion over the sample period – again contrary to what search cost theories would predict. Furthermore, while average prices of advertised items or items that are purchased repeatedly are lower than those of unadvertised or infrequently purchased items (in line with theories about search costs), the same does not hold for price dispersion. According to the authors, this may be because stores

6

ANITA ELBERSE, PATRICK BARWISE AND KATHY HAMMOND

have succeeded in differentiating themselves (e.g. by carrying particular books) even though they are selling a commodity product. Examining yet another dimension of market efficiency, Clay and Tay (2001) investigate the cross-country differentials in online textbook pricing. They use data for nearly 100 textbooks, nine major online bookstores, and 4 countries. Clay and Tay discover that textbooks are substantially cheaper in foreign online bookstores, as compared with those in the U.S., with prices varying by as much as 42% for a single book using standard shipping, and 23% using expedited shipping. Interestingly, differences were significant even across Amazon branches. A thorough and influential study by Brynjolfsson and Smith (2000a) analyzed the prices of books and CDs on 41 Internet and conventional retail outlets. They found that prices on the Internet averaged 9% to 16% less than in conventional outlets (depending on whether taxes, shipping and shopping costs are included in the price), and that Internet retailers were willing to make smaller price changes than conventional retailers. However, they also found substantial price dispersion among Internet retailers, although weighting the prices by a proxy for market share reduced this dispersion. Based on their results, in line with Clay, Krishnan and Wolff (2001), they concluded that, “while there is lower friction in many dimensions of Internet competition, branding, awareness and trust remain important sources of heterogeneity among Internet retailers” (Brynjolfsson & Smith, 2000a, p. 563). Directly building on suggestions by Smith et al. (2000), Pan, Ratchford and Shankar (2001), using a range of price dispersion measures covering over 6,000 price quotes for nearly 600 products in a variety of product categories from over 100 online retailers, found that even after controlling for retailer characteristics, online price dispersion is large. Market characteristics are associated with a large portion of this price dispersion: it increases with consumer involvement or average price level of items, although at a decreasing rate, and decreases with the number of competitors, also at a diminishing rate. Their regression model accounts for over 92% of the variance in price dispersion. Degeratu, Rangaswamy and Wu (1999), investigating online versus traditional grocery stores, found that price sensitivity was higher online in the sense that online promotions were stronger signals of price discounts. However, the combined effect of price and promotion on consumer choice was found to be weaker online than offline. Using data that cover 4 million daily prices observed over an 8-month time period for 1,000 of the best-selling consumer electronic products found on a price comparison site, Shopper.com, Baye, Morgan and Scholten (2001) also find little support for the notion that prices on the Internet are converging to

The Impact of the Internet on Horizontal and Vertical Competition

7

‘the law of one price’. The same result emerges in a follow-up study, based on monthly price data for 36 products obtained over an 18-month period, again from Shopper.com (Baye, Morgan & Scholten, 2002). Price dispersion is remarkably persistent over the sample period, even after controlling for firm heterogeneity (Baye, Morgan & Scholten, 2002). They further find that price dispersion for a given product – measured by the gap between the two lowest prices listed – varies systematically with the number of firms listing price quotes for that product (Baye, Morgan & Scholten, 2001), which in turn is related to its life cycle stage (Baye, Morgan & Scholten, 2002). We provide a summary of key empirical findings regarding the four dimensions of Internet market efficiency in Table 1, which is partly adapted from Smith et al. (2000). Why Has the Internet’s Impact on Pricing Been Limited? Lal and Sarvary (1999) present a theoretical model which might help explain these unexpected empirical findings. Their model distinguishes between ‘digital’ product attributes (which can be communicated online at low cost) and ‘non-digital’ attributes (for which physical inspection of the product is needed). It assumes that consumers are faced with a choice of two brands but are familiar with the non-digital attributes of only the brand bought on the last purchase occasion. Based on this assumption, Lal and Sarvary showed that when: (1) the proportion of Internet users is high enough; (2) non-digital attributes are relevant but not overwhelming; (3) consumers have a more favorable prior about the brand they currently own; and (4) the purchase situation can be characterized by ‘destination shopping’ (i.e. the fixed cost of a shopping trip is higher than the cost of visiting an additional store), the use of the Internet can not only lead to higher prices but also discourage consumers from engaging in search. Their explanation is that, under these conditions, an online consumer who wishes to do so can avoid visiting any stores at all and therefore also avoid comparing the non-digital attributes of competing brands. A further insight is that physical stores may have a growing role for product demonstration and online customer acquisition. The underlying theory is based on Nelson’s (1970, 1974) distinction between search goods (whose quality can be judged by inspection) and experience goods (whose quality can be judged only through usage).

8

Table 1. Author(s)

Product Category

Data (Sample Period)

Findings

Prices for a matched set of 125 books, 108 CDs, and 104 software titles sold through 52 conventional and Internet outlets (March 1997).

Average prices are higher online than offline (varying from 3% for software to 13% for CDs).

Brown & Goolsbee (2002)

Life insurance

Prices for individual life insurance policies, for a sample of 30,000 policies per year, issued by 46 participating companies (1992–1997).

Substantial drop in prices (approximately 20% lower in 1996 and 1997, compared with 1992–1994, of which 8% to 15% is attributed to the rise of Internet).

Brynjolfsson & Smith (2000a)

Books, CDs

Prices for matched set of 20 books and 20 CDs sold through conventional and Internet outlets (1998–1999).

Average prices are lower online than offline (varying from 9% to 16%).

Clay, Krishnan, Wolff & Fernandes (1999)

Books

Prices for 107 books sold by 13 online and 2 physical bookstores (April 1997).

Controlling for book characteristics, prices online and offline are the same.

Clay, Krishnan & Wolff (2001)

Books

Prices for 399 books, 32 online bookstores (1999–2000).

No change in prices over sample period; Prices are lower for advertised vs. unadvertised and lower for frequently vs. infrequently purchased items.

Erevelles, Rolland & Srinivasan (2001)

Pharmaceuticals

Prices for multi-vitamins in 4 segments and across 5 retail formats.

Prices are higher online than offline, even for private label brands.

ANITA ELBERSE, PATRICK BARWISE AND KATHY HAMMOND

Average Price Levels Bailey (1998a); Books, CDs, Bailey (1998b) computer software

Empirical Evidence on Relative Efficiency of Internet Markets.

Continued.

Product Category

Data (Sample Period)

Findings

Lee (1998)

Cars

Prices for used cars sold in electronic and conventional auction markets (1986–1995).

Prices are higher online than offline; Online prices increase over time.

Scott Morton, Zettelmeyer & Silva-Risso (1999)

Cars

Prices for 360,000 purchased cars, 10,300 (3%) of which were processed online, via Autobytel.com (1999–2000).

Prices are lower online than offline (a 2% difference).

Bailey (1998a; b)

Books, CDs, computer software

Prices for a matched set of 125 books, 108 CDs, and 104 software titles sold through 52 conventional and Internet outlets (March 1997).

Price dispersion is not lower online than offline.

Baye, Morgan & Scholten (2001)

Consumer electronics

Four million daily price observations for the 1,000 most popular products, and for different sellers, listed at Shopper.com (8 months, 2000–2001)

Price dispersion is substantial over the sample period; Price dispersion varies with market structure: the higher the number of firms listing prices, the lower the price dispersion (measured by the price gap between the two lowest prices).

Baye, Morgan & Scholten (2002)

Consumer electronics

Monthly prices for 36 products, charged by on average 20 firms, listed at Shopper.com (18 months, 1999–2001).

Price dispersion is substantial over the sample period, even after controlling for firm heterogeneity; As products age, the number of firms listing prices and the average range in prices decline.

Brown & Goolsbee (2002)

Insurance

Prices for individual life insurance policies, for a sample of 30,000 policies per year, issued by 46 participating companies (1992–1997).

Price dispersion initially rises, then falls as the share of people researching prices online increases.

Price Dispersion

9

Author(s)

The Impact of the Internet on Horizontal and Vertical Competition

Table 1.

10

Table 1.

Continued.

Product Category

Data (Sample Period)

Findings

Brynjolfsson & Smith (2000a)

Books, CDs

Prices for matched set of 20 books and 20 CDs sold through conventional and Internet outlets (1998–1999)

Price dispersion is substantial online (average price differences range from 25% for books to 33% for CDs); Weighted by market share, price dispersion is lower online than offline.

Clay, Krishnan, Wolff & Fernandes (1999)

Books

Prices for 107 books sold by 13 online and 2 physical bookstores (April 1997).

Price dispersion is substantial online (ranging from 5% to 11% among 3 largest sellers).

Clay, Krishnan & Wolff (2001)

Books

Prices for 399 books, 32 online bookstores (1999–2000).

No change in price dispersion over sample period; No difference in price dispersion between advertised and unadvertised or frequently and infrequently purchased items.

Clay & Tay (2001)

Books

Prices for 95 textbooks at 9 online bookstores in the U.S., Canada, the U.K., and Germany (early 2001).

Cross-country price dispersion is substantial (prices are lower in foreign markets than in the U.S., up to 23% to 42% for a single book, depending on shipping).

Clemons, Hann & Hitt (2002)

Travel

Prices quoted by online travel agents for airline tickets (1997).

Price dispersion is substantial online (average price differences of up to 20%)

Erevelles, Rolland & Srinivasan (2001)

Pharmaceuticals

Prices for multi-vitamins in 4 segments and across 5 retail formats.

Price dispersion is higher online than offline.

ANITA ELBERSE, PATRICK BARWISE AND KATHY HAMMOND

Author(s)

Continued.

Author(s)

Product Category

Data (Sample Period)

Findings

Pan, Ratchford & Shankar (2001)

Books, CDs, DVDs, computer software, computers, PDAs, consumer electronics

A total of 6,700 price quotes for 580 products from 105 online retailers in a variety of product categories (November 2000).

Price dispersion is substantial online, even after controlling for online retailer characteristics (average price differences vary between 26% for laptops and 51% for CDs).

Degeratu, Rangaswamy & Wu (1998)

Groceries

Shopping behavior for groceries sold online (for 300 Peapod customers) and in conventional outlets (IRI scanner data) (1996–1997).

Sensitivity to the combined effect of price and promotion is lower online than offline.

Ellison & Ellison (2001)

Computer parts

Prices for computer parts (memory modules and motherboards) listed on price search engine PriceWatch for one online retailer (mid-2000).

Demand elasticity is high, particularly for lowquality products; Cross-elasticities of demand for high-quality products with respect to the price of low-quality products are large and negative.

Bailey (1998a); Bailey (1998b)

Books, CDs, computer software

Prices for matched set of 125 books, 108 CDs, and 104 software titles sold through 52 conventional and Internet outlets (1997).

Menu costs are lower online than offline (prices are changed more than twice as often online).

Brynjolfsson & Smith (2000a)

Books, CDS

Prices for matched set of 20 books and 20 CDs sold through conventional and Internet outlets (1998–1999).

Menu costs are lower online than offline (online adjustments are up to 100 times smaller than those offline).

Price Elasticity

Menu Costs

The Impact of the Internet on Horizontal and Vertical Competition

Table 1.

11

12

ANITA ELBERSE, PATRICK BARWISE AND KATHY HAMMOND

In addition to Nelson’s (1970) initial distinction, Darby and Karni (1973) proposed a third category, namely ‘credence’ goods, whose quality cannot be determined reliably even after usage. A classic example is wine. Online wine sales have been researched by Lynch and Ariely (2000) who found that first, lowering the cost of search for quality information reduced price sensitivity, and second, price sensitivity for goods common to two online stores increased when cross-store comparison was made easy. However, easy cross-store comparison had no effect on price sensitivity for unique goods. Third, making information environments more transparent by lowering all three search costs (for price information, for quality information within a given store, and for comparisons across the two stores) produced welfare gains for consumers. The implications are that retailers should aim to make information environments maximally transparent but try to avoid price competition by carrying more unique or differentiated merchandise – conclusions that are in line with empirical findings such as those of Brynjolfsson and Smith (2000), Clay et al. (1999), and Clay, Krishnan and Wolff (2001). Using survey data from the hospitality industry, Shankar, Rangaswamy and Pusateri (1999) studied two aspects of consumers’ price sensitivity: the weight a consumer attaches to price relative to other attributes, and the consumer’s propensity to undertake a search for better prices. They found that the Internet increases the latter but not the former. Furthermore, specific considerations in the design of web sites (e.g. the provision of in-depth price and non-price information) can help dampen price sensitivity. Also, they found that the Internet accentuates the effects of brand loyalty by decreasing price search – again in line with the empirical results by Brynjolfsson and Smith (2000), Clay et al. (1999), Clay, Krishnan and Wolff (2001), and others. Also relevant in this context is a theoretical paper by Chen and Hitt (2001) that bridges some of the economics and marketing literature. Chen and Hitt integrated two key aspects of consumer search in Internet markets – brand awareness (awareness of competing retailers) and brand sensitivity (the willingness to pay a premium to buy from a leading retailer) – and linked those to price dispersion. Their theoretical results, derived for a duopoly, suggest that when consumers are not indifferent to brands, or not all consumers consider all brands, aggressive (Bertrand) price competition is not optimal, and other pricing strategies deserve recommendation. Also focusing on pricing strategies, Smith (2001) examined how the unique characteristics of consumer search in Internet markets impact market structure and firm-level strategy. His analytical model emphasizes the interdependence in pricing strategies of well-known retailers, which allows them to cooperate to set high prices. He found his model

The Impact of the Internet on Horizontal and Vertical Competition

13

to be consistent with data collected in late 1999 for 24 book retailers for over 23,000 book titles. Carlton and Chevalier (2001) provided a new angle on price dispersion on the Internet by examining manufacturers’ decisions about whether and how to offer their products for sale online. Focusing on three product categories (fragrances, DVD players, and refrigerators), they found that manufacturers who limit distribution in the physical world also use various mechanisms to limit distribution online, particularly the sale of their products by online retailers who sell goods at deep discounts. Furthermore, they showed that manufacturers who distribute their goods directly through manufacturer websites tend to charge very high prices for the products, consistent with the hypothesis that manufacturers internalize free rider issues. Shopbots Price search engines, often referred to as ‘shopbots’ (short for ‘shopping robots’), play a key role in Internet market structures and pricing. Shopbots are sites that allow consumers to make ‘one-click’ price comparisons for product offerings from multiple producers and/or retailers, thereby substantially reducing buyer search costs for product and price information (Brynjolffson & Smith, 2000b). Such sites have been the topic of several studies, which together provide key insights into possible reasons underlying observed market inefficiencies. One straightforward rationale follows from research by Montgomery et al. (2001), who observed that the vast majority of Internet consumers do not use shopbots at all. Partly attributing that to the poor design of shopbots – which often require time-consuming searches that provide redundant or dominated alternatives – Montgomery and colleagues studied how such designs could be improved. Using a utility modeling approach, illustrated by means of data collected from online bookstores over a six-month period, they showed how shopbots can search more intelligently and increase consumer utility by selectively presenting results. Baye and Morgan (2001) provide a number of interesting insights into the role of shopbots (or, more generally, Internet ‘gatekeepers’), the markets in which they operate, and the markets they serve. Specifically, they examined how a gatekeeper’s decisions regarding the fees it charges to firms (that advertise prices) and consumers (who access the list of advertised prices) impact – and are impacted by – the competitiveness of the product market it serves. To this end, they introduced a market for price information where internal information flows are costless, but a gatekeeper sets ‘entrance’ fees for

14

ANITA ELBERSE, PATRICK BARWISE AND KATHY HAMMOND

firms and consumers. They found that gatekeeper profits are maximized when access fees are sufficiently low that all consumers subscribe, advertising fees exceed socially optimal levels (thus inducing partial firm participation, and price dispersion in the product market), and advertised prices are lower than unadvertised prices. Intriguingly, the product market exhibits price dispersion even though (in equilibrium) all consumers purchase from a firm offering the lowest price. Focusing specifically on the effects of Internet-enabled price search on the elasticity of demand, an insightful study by Ellison and Ellison (2001) explored whether improved search on the Internet creates a real ‘Bertrand paradox’ with prices so low that firms can not cover their fixed costs. They viewed the effect of technological progress on search costs as a balance-of-power problem – a particularly useful perspective in this context. Their empirical analysis showed that price search on the Internet could make demand extremely elastic – so elastic that firms could never break even – if it were not for the power of the Internet that retailers are harnessing to thwart search engines and shopbots. Also, contrary to conventional wisdom, cross-price elasticities of demand for a higher quality product with respect to the price of a low-quality product were found to be large and negative, which provides strong empirical evidence of the effectiveness of loss-leader/bait-and-switch strategies. The study revealed that the most popular obfuscation strategy is to intentionally create an inferior quality good that can be offered at a very low price. That is, prices have advertising value. Opting out of search engines altogether is not the preferred strategy for firms due to a free rider problem (see Iyer & Pazgal, 2001, for a theoretical and empirical treatment of motivations of retailers that elect to join – or stay out of – shopbots or, in their terminology, ‘Internet shopping agents’). Smith and Brynjolffson (2001) empirically analyzed consumer behavior at Internet shopbots. Using panel data obtained from EvenBetter.com, a shopbot for books, analyzed by means of logit modeling techniques, they found evidence for limited price sensitivity. While price is an important determinant of customer choice, even among shopbot consumers – likely to be among the most price sensitive online consumers – branded retailers and previously visited retailers hold significant price advantages in head-to-head price comparisons. That is, shopbot consumers respond very strongly to well-known, heavily branded retailers – a particularly interesting finding given that the analysis focuses on homogenous products. However, customers are sensitive to how the total price is allocated among components (item price, shipping costs, and tax) and to the ordinal ranking of retailer offerings with respect to price. Furthermore, consumers use brand as a proxy for a retailer’s credibility

The Impact of the Internet on Horizontal and Vertical Competition

15

regarding non-contractible aspects of the bundle such as shipping time (also see a companion paper by the same authors, Brynjolfsson & Smith, 2000b). Finally, also focusing on shopbots, Chen and Sudhir (2001) provided an analytical argument for why, in contrast to conventional wisdom, reducing consumer search costs may lead to decreasing competition and increasing prices. Their argument rests on the intuition that as search costs are minimal, then a lack of search on the side of the consumer indicates strong loyalty, which in turn can be used by retailers as a proxy to segment the market into loyal and price sensitive segments. Aided by the low costs of personalized firmto-consumer communications means (e.g. e-mail), firms can improve the effectiveness of targeted pricing. This is a variant of ‘smart’ pricing which, as we already noted, could bring large long-term risks (Sinha, 2000). Bundling, Versioning, and Auctions In relation to the Internet, it is often said that ‘information wants to be free’. Here, ‘free’ can mean both liberated and priced at zero. On the Internet, the marginal cost of providing information to a customer is usually zero, so any pricing model for an information product based on equating marginal cost to marginal revenue would eventually lead to the information being given away. This raises the question of how a firm can make money from content creation or packaging. Although this is not a new issue – it is one which has always been faced by broadcasting, publishing and other media industries – the power and ubiquity of digital technology are increasing the scale of the problem. Arthur (1996) argued that new knowledge-based industries are characterized by increasing returns to scale, i.e. that if a product gains a dominant market share its advantage is magnified by increasing returns. In this world, ‘success accrues to the successful’ and ‘market share begets market share’. In contrast, traditional resource processing industries are characterized by diminishing returns. Arthur compared and contrasted these two interrelated worlds of business and offered advice to managers in knowledge-based markets. One important source of increasing returns in information and communication industries is network externalities – whereby the value of a product to one user depends on how many other users there are (e.g. Rohlfs, 1974; Katz & Shapiro, 1994). As mentioned, bundling and versioning are strategies that firms can employ to counter the possible erosion of margins that goes hand in hand with lower search costs (e.g. Sinha, 2000). Bakos and Brynjolffson (1999), leading researchers on bundling and the Internet, modeled optimal bundling strategies for a multi-product monopolist information supplier (i.e. with zero marginal

16

ANITA ELBERSE, PATRICK BARWISE AND KATHY HAMMOND

cost). They suggested that bundling large numbers of unrelated information goods might be surprisingly profitable because the law of large numbers makes it easier to predict consumers’ valuations for a bundle of goods than for the individual goods sold separately. They modeled the bundling of complements and substitutes, bundling in the presence of budget constraints, and the scope for offering a menu of different bundles if the market is highly segmented. Predictions from their analysis appear to be consistent with empirical observations of the markets for online content, cable TV programming, and music. In later work, Bakos and Brynjolfsson (2000a, b) extended their bundling model to a range of different settings. They argued that bundling can create ‘economies of aggregation’ for information goods, even in the absence of network externalities or economies of scale or scope. They drew four implications: (1) when competing upstream, for content, larger bundlers can outbid smaller ones; (2) when competing downstream, for consumers, bundling can discourage entry even when the prospective entrant has a superior cost structure or quality; (3) conversely, bundling by the new entrant can allow profitable entry; and (4) because a bundler can potentially capture a large share of profits in a new market, bundlers may have higher incentives to innovate than single-product firms. Shapiro and Varian (1998a, b) noted that the fixed costs of information products tend to be dominated by sunk costs – costs that are not recoverable if production is halted. They suggested that information providers therefore need strategies both to differentiate their products and to price them in such a way that the price varies between buyers, reflecting the sometimes markedly different value that the different segments place on the same (or almost the same) information product. The solution they proposed is ‘versioning’, i.e. offering the information in different versions targeted at different types of customer. Examples include release windows for movies and books. Shapiro and Varian described seven dimensions of versioning: convenience, comprehensiveness, manipulation, community, annoyance, speed, and support. Some of these issues were also explored by Adar and Huberman (2000), who focused on the possibility of exploiting the different but regular patterns of surfing demonstrated by different Internet users, by implementing ‘temporal discrimination’ through dynamically configuring sites and versioning information services. Finally, as also indicated by Sinha (2000), firms can use auctions to reduce the threat of eroding premiums. Chui and Zwick (1999) and Klein and O’Keefe (1999) are other sources for a preliminary review of relevant literature and frameworks for analyzing online auctions. Chui and Zwick explored the scope and scale of online auctions, as well as the range of business models, covering

The Impact of the Internet on Horizontal and Vertical Competition

17

business-to-consumer, business-to-business, and consumer-to-consumer auctions. Klein and O’Keefe (1999) described Teletrade.com, an example of a telephone-based auction which now also uses the Web. Also relevant, DeKoning et al. (1999) explored consumer motivations in using consumer-toconsumer online auctions. In a key contribution in this area, Lucking-Reiley (2000a) presented what he called ‘an economist’s guide’ to online auctions, including a brief history, and the results of a survey of 142 auction sites that were online in the Fall of 1998. His paper summarized the various business models they used, what goods they offered for sale, and what kinds of auction mechanism they employed. Lucking-Reiley argued that established auction theory from economics could be used to improve Internet auctions. He also presented detailed data on the competition between the incumbent (eBay) and the two well-funded entrants into the business-to-consumer online auction arena (Yahoo! and Amazon) in 1999. Among other things, he showed that the different auctioneers’ fee structures had measurable incentive effects on sellers’ choices and transaction outcomes. Building on this review, Lucking-Reiley (1999) tested different auction formats using field experiments in which collectible trading cards were auctioned. In addition, Lucking-Reiley et al. (2000b) presented an exploratory analysis of the determinants of prices in online auctions for collectible coins at eBay. Three findings stand out. First, a seller’s feedback ratings, reported by other eBay users, have a measurable effect on his auction prices. This is particularly true for negative feedback ratings. Second, minimum bids and reserve prices tend to have a positive effect on the final auction price. However, this finding does not take into account that these instruments also decrease the probability of the auction resulting in an actual sale. Third, when a seller chooses to have his auction last for a longer period of days, this significantly increases the average auction price. Regarding the first result, it is worthwhile to note that an empirical analysis by Houser and Wooders (2001) offered further support for the significance of the relationship between the seller’s reputation and prices.

THE INTERNET’S IMPACT ON ‘VERTICAL COMPETITION’ Extending the above ideas, some commentators have argued that as frictionfree electronic marketplaces lower the costs of transactions and facilitate search, it will become easier to match buyers and sellers. As a result, the role of intermediaries may be severely reduced or even eliminated, a process known

18

ANITA ELBERSE, PATRICK BARWISE AND KATHY HAMMOND

as ‘disintermediation’ (e.g. Gates, 1995; Gellman, 1996; also see Bakos, 1998). Opponents of this view, although sometimes convinced that certain types of intermediaries may disappear, believe that electronic markets will compensate for this by promoting the growth of new types of intermediaries (Bakos, 1998). Below, we discuss key theoretical and empirical research in this area, starting with work favoring the ‘disintermediation’ view, and then moving to studies that deal with ‘reintermediation’. The issue of the role of intermediaries in buyer-seller relationships has been a recurring theme in research and writing about the Internet, with much of the earlier work suggesting disintermediation and later papers generally arguing for some form of reintermediation. Disintermediation The ability of easily accessible electronic information to increase the efficiency of markets was an early topic addressed by academics in marketing and economics. Bakos (1991) used economic theory to develop models which showed that, where product quality and price information are easily available (as in electronic markets), search costs are reduced and benefits for buyers are increased which, in turn, can reduce sellers’ profits. Following Bakos’s work, Benjamin and Wigand (1995) suggested that the so-called ‘national information infrastructure’ (of which they believed the Internet was only a part) would cause a restructuring and redistribution of profits among stakeholders along the value chain, threatening all intermediaries between the manufacturer and consumer. These ideas were more fully developed in another paper (Wigand & Benjamin, 1995) which argued that the Internet holds great potential for efficiency gains along the whole industry value chain, primarily because of transaction cost savings (see also Rayport & Sviokla, 1995). Wigand and Benjamin, who foresaw a range of effects that include disintermediation, reduced profit margins, consumer access to a broad selection of lower-priced goods, various opportunities to restrict consumers’ access to the vast amount of available information, as well as potential commerce possibilities, voiced a need for public policy “to mitigate risks associated with market access and value chain reconfiguration”. An essential component of the evolution of the future world of electronic commerce, they suggested, is the ‘market choice box’ – the consumer’s interface between the increasing number of electronic devices inside and outside the home and the information superhighway (Wigand & Benjamin, 1995; Barwise, 2001). Shaffer and Zettelmeyer (1999) also sketched a fairly pessimistic scenario for intermediaries. In their view, manufacturers have traditionally had to rely on

The Impact of the Internet on Horizontal and Vertical Competition

19

retailers to communicate product and category information that is either too technical or too idiosyncratic to be communicated effectively via mass media. The emergence of the Internet as a marketing communications channel now makes it possible for manufacturers (and third parties) also to provide such information. Shaffer and Zettelmeyer show how this may lead to channel conflict. Specifically, manufacturers gain and retailers lose from information that makes a retailer’s product offerings less substitutable. Rayport and Sviokla (1994) described how physical interactions in the marketplace were being replaced by virtual ‘marketspace’ transactions. They argued that the conventional value proposition was being disaggregated and that its three basic elements – content (the firm’s offering), context (how the content is offered), and infrastructure – could be managed in new and different ways. Building on these ideas, and perhaps warming to some form of reintermediation, Rayport and Sviokla (1995) suggested ways of managing and exploiting this new virtual value chain. Weiber and Kollman (1998) also evaluated the significance of virtual value chains and concluded that information, in its own right, would become a factor of competition in future markets. Reintermediation In contrast to these views, Sarkar, Butler and Steinfield (1998) argued against the idea that intermediaries are likely to disappear. Drawing on channel evolution literature and transaction cost economics, they proposed instead that virtual channel systems and new ‘cybermediaries’ would emerge. Jin and Robey (1999) who focused on business-to-consumers cybermediaries such as Amazon, Virtual Vineyards, and 1–800-FLOWERS, proposed six theoretical perspectives on cybermediation: transaction cost economics, consumer-choice theory, retailing as an institution, retailing as social exchange, retailers as bridges in social networks, and retailers as creators of knowledge. They concluded that a multi-theoretical approach (in contrast to transaction cost theory alone) shows both that the disintermediation hypothesis was overstated and that cybermediaries can exist for many reasons – very much in line with Sarkar, Butler and Steinfield (1998). More recently, Klein and Selz (2000) argued that cybermediaries will play an increasing role in electronic markets. Focusing on the automotive industry, they examined the roles and functions of such players and their impact on established distribution and sales channels. In a theoretical analysis rooted in economics, Bhargava, Choudhary and Krishnan (2000) explored the aggregation benefits that consumers derive from having access to multiple providers

20

ANITA ELBERSE, PATRICK BARWISE AND KATHY HAMMOND

through an intermediary. They concluded that when consumers are heterogeneous and differ in their willingness to pay for intermediation, the intermediary can offer two or more service levels at different price levels. Importantly, the propositions by Sarkar, Butler and Steinfield (1998), Jin and Robey (1999), Klein and Selz (2000) and Bhargava et al. (2000) are consistent with empirical findings. For example, in their own study, Klein and Selz (2000) identified two new types of cybermediaries: service brokers and information brokers. More generally, Bailey and Bakos (1997), working on the premise that markets do not necessarily become disintermediated as they become facilitated by information technology, explored thirteen case studies of firms participating in electronic commerce. They found evidence of various emerging roles for online intermediaries, including aggregating, matching sellers and buyers, providing trust, and supplying interorganizational market information. The authors discussed two specific examples to illustrate an unsuccessful strategy for electronic intermediation (BargainFinder) as well as a more successful one (Firefly). Support for these views also comes from research in business-to-business contexts. For example, Chircu and Kauffman (2000) described a framework whereby a traditional intermediary is able to continue to compete by combining web technology with its existing specialist assets (a ‘bricks and clicks’ strategy). An empirical analysis of the corporate travel industry showed that traditional travel firms have been able to avoid disintermediation and retain a highly profitable central role in this market (also see Ramsdell, 2000; Kaplan & Sawhney, 2000, for intermediation issues in business-to-business settings). In a short Harvard Business Review perspectives article, Carr (2000) took the above arguments a step further by arguing that, far from the widely predicted disintermediation, the Internet is in fact leading to ‘hypermediation’. That is, transactions over the Web, even very small ones, routinely involve many intermediaries – not only wholesalers and retailers, but also content providers, affiliate sites, search engines, portals, internet service providers, software makers and many others. He suggested that it is these largely unnoticed intermediaries who stand to gain most of the profits from electronic commerce. More generally, Christensen and Tedlow (2000) categorized the Internet as a ‘disruptive’ technology which enables innovative retailers to create new business models that significantly change the economics of the industry. They put this in a historical context by relating the Internet to three previous disruptive technologies in retailing: the department store, the mail order catalog, and the discount department store. They proposed that “. . . the essential mission of retailing has always had four elements: getting the right

The Impact of the Internet on Horizontal and Vertical Competition

21

product in the right place at the right price at the right time and the Internet has great potential for improving performance on various combinations of the first three of these. For information products and services, the Internet can also perform outstandingly on the fourth, time, dimension, but for physical products it does not. When shoppers need products immediately, they will head for their cars, not their computers” (Christensen & Tedlow, 2000, p. 42). They further argued that the Internet is unsuited for products which require ‘touch and feel’, not to mention ‘taste and smell’. Based on their analysis of the three earlier disruptive technologies, Christensen and Tedlow noted that one pattern has been that generalist stores and catalogs dominate at the outset of the disruption but are then supplanted by specialists. A second pattern has been that the disruptive retailers initially sold easy-to-sell branded mass-market products and then moved up-scale with higher-margin, more complex products. They suggested that it is too soon to say whether the first of these two patterns will recur on the Internet – more likely, the pattern will vary between categories – but that there is some evidence that the second pattern is recurring – and probably much faster than with the previous disruptive technologies, since the Internet enables firms to swiftly achieve high market reach combined with high richness of content and range (also see Evans & Wurster, 1999). If this analysis is correct, we will see consumer e-commerce growth for complex, high-ticket items such as durables (but excluding those needing ‘touch and feel’) as well as for simple branded products such as books and music.

CONCLUSIONS AND IMPLICATIONS We can safely conclude that early predictions of dramatic change – both in terms of ‘horizontal’ and ‘vertical’ competition – have proved overstated. Popular statements made in the mid-1990s that the Internet would lead to frictionless, commoditized markets where low transaction and search costs would drive down prices, reduce retailers’ margins to zero, and eliminate the role of intermediaries (e.g. Gates, 1995, Negroponte, 1995; Wigand & Benjamin, 1995), have so far proved wide of the mark. Empirical evidence to-date on the Internet’s potential to improve the efficiency of markets is mixed. While there are some indications that the Internet has increased competition and reduced average prices for some product categories, the impact on price levels, price dispersion, and price elasticity has been small (e.g. Brynjolfsson & Smith, 2000; Clay et al., 1999; Clay, Krishnan & Wolff, 2001). Furthermore, if the Internet indeed leads to value chain reconfiguration, then reintermediation – particularly the emergence

22

ANITA ELBERSE, PATRICK BARWISE AND KATHY HAMMOND

of so-called ‘cybermediaries’ – appears a much more likely scenario than disintermediation (e.g. Sarkar, Butler & Steinfield, 1998). Several studies provided insights into the reasons for why electronic markets have not (yet) attained the level of efficiency that was predicted by some commentators. Most importantly, brands continue to play a key role (e.g. Brynjolfsson & Smith, 2000a; Clay, Krishnan & Wolff, 2001; Lal & Sarvary, 1999; Shankar, Rangaswamy & Pusateri, 1999). The underlying reasons for why brands exist have not gone away (e.g. Barwise, 1997) – they are not dead. In fact, trusted brands may be even more important in a world of information overloaded, money-rich and time-poor consumers, particularly for the many product categories where product quality still cannot be reliably judged online (e.g. Barwise, 1997; Brynjolfsson & Smith, 2000a; Dayal et al., 2000; Lal & Sarvary, 1999). Brands are the primary tool by which manufacturers and retailers can differentiate their products from those of their competitors. Related explanations for the limited efficiency of online markets follow from insights into the role of shopbots. It is clear that price search engines are still used by only a small fraction of online users (e.g. Montgomery et al., 1999), so it comes as no surprise that their influence on market structure and dynamics has so far been limited. However, even if consumers and sellers were to adopt shopbots on a much larger scale, it is by no means guaranteed that price dispersion will decrease (e.g. Baye & Morgan, 2001). Objectives of shopbots and retailers (as well as other market players) are inherently different. As a result, retailers will seek to thwart shopbots and effectively use the Internet to prevent the erosion of their margins (Ellison & Ellison, 2001). This ‘battle for power’ is not likely to go away. Concerns about brand reputation should also limit firms’ enthusiasm for aggressive ‘smart’ pricing – a euphemism for charging more to less pricesensitive customers (Sinha, 2000). However, strategies such as bundling and versioning have been applied on a wide scale, particularly for information products (e.g. Bakos & Brynjolfsson, 2000a, b; Shapiro & Varian, 1998a, b). One area where the impact of the Internet is perhaps most felt is auctions, especially in business-to-business contexts, but also in business-to-consumer contexts where eBay is leading the pack. Again, however, brands play a key role in such environments – for sellers it is crucial to establish a solid reputation (e.g. Lucking-Reiley, 2000b). In that regard, rules of the ‘old’ economy remain valid. It may be that the limited impact of the Internet on horizontal and vertical competition is still the quiet before the storm. As intelligent agent software becomes more powerful and more widely used, we may see more pressure on prices and on brand loyalty – probably not so much on the choice of brands

The Impact of the Internet on Horizontal and Vertical Competition

23

itself, but rather on the choice of channels through which to buy the brand. Also, as other pricing models gain wider acceptance, such as ‘name-your-ownpricing’ (e.g. Simon & Schumann, 2001), market structures may undergo further changes. In any event, the Internet’s impact on market competition promises to be a deserving research topic for years to come.

NOTE 1. This chapter is partly based on two sections of ‘Marketing and the Internet: A Research Review’ by Barwise, Elberse and Hammond (2002), which considers the Internet’s impact on marketing in a wide range of areas. The full review is published as chapter in the ‘Handbook of Marketing’ (Weitz & Wensley, 2002). A – regularly updated – version is also available at http://www.marketingandtheinternet.com. We are indebted to numerous colleagues across the world who have provided useful comments on our research review since it was first made available.

REFERENCES Adar, E., & Huberman, B. A. (2000). The Economics of Surfing. Quarterly Journal of Electronic Commerce, 1(3), 203–214. Arthur, W. B. (1996). Increasing Returns and the New World of Business. Harvard Business Review (July–August), 100–109. Bailey, J., & Bakos, J. Y. (1997). An Exploratory Study of the Emerging Role of Electronic Intermediaries. International Journal of Electronic Commerce, 1 (3, Spring), 7–20. Bailey, J. P. (1998a). Intermediation and Electronic Markets: Aggregation and Pricing in Internet Commerce. Ph.D. Technology, Management and Policy, Massachusetts Institute of Technology, Cambridge, MA. Bailey, J. P. (1998b). Electronic Commerce: Prices and Consumer Issues for Three Products: Books, Compact Discs, and Software. Organisation for Economic Cooperation and Development, OCDE/GD, 98(4). Bakos, J. Y. (1991). A Strategic Analysis of Electronic Marketplaces. MIS Quarterly (September), 295–310. Bakos, J. Y. (1997). Reducing Buyer Search Costs: Implications for Electronic Marketplaces. Management Science, 43(12), 1676–1692. Bakos, J. Y. (1998). The Emerging Role of Emerging Marketplaces on the Internet. Communications of the ACM, 41(8), 35–42. Bakos, J. Y., & Brynjolfsson, E. (1999). Bundling Information Goods: Pricing, Profits and Efficiency. Management Science, 45(12), 1613–1630. Bakos, J. Y., & Brynjolfsson, E. (2000a). Aggregation and Disaggregation of Information Goods: Implications for Bundling, Site Licencing, and Micropayment Systems. In: D. Hurley, B. Kahin, & H. R. Varian (Eds), Internet Publishing and Beyond: The Economics of Digital Information and Intellectual Property. Cambridge MA: MIT Press. Bakos, J. Y., & Brynjolfsson, E. (2000b). Bundling and Competition on the Internet. Marketing Science, 19(1), 63–82. Barwise, P. (1997). Editorial: Brands in a Digital World. Journal of Brand Management, 4(4), 220–223.

24

ANITA ELBERSE, PATRICK BARWISE AND KATHY HAMMOND

Barwise, P. (2001). TV, PC or Mobile? Future Media for Consumer E-Commerce. Business Strategy Review, 12(1), 35–42. Barwise, P., Elberse, A., & Hammond, K. (2002). Marketing and the Internet: A Research Review. Working Paper, London Business School. [http://www.marketingandtheinternet.com] Baye, M., & Morgan, J. (2001). Information Gatekeepers on the Internet and the Competitiveness of Homogenous Product Markets. American Economic Review, 91(3), 454–474. Baye, M., Morgan, J., & Scholten, P. (2001). Price Dispersion in the Small and the Large: Evidence from an Internet Price Comparison Site. Working Paper, Indiana University, July 2001. Baye, M., Morgan, J., & Scholten, P. (2002). Persistent Price Dispersion in Online Markets. Working Paper, Indiana University, April 2002. Benjamin, R. I., & Wigand, R. T. (1995). Electronic Markets and Virtual Value Chains on the Information Superhighway. Sloan Management Review (Winter), 62–72. Bhargava, H., Choudhary, V., & Krishnan, R. (2000). Pricing and Product Design: Intermediary Strategies in an Electronic Market. Working Paper, MIT Ecommerce Forum, 1–20. Brown, J. R., & Goolsbee, A. (2002). Does the Internet Make Markets More Competitive? Evidence from the Life Insurance Industry. Journal of Political Economy, 110(3), 481–507. Brynjolfsson, E., & Smith, M. D. (2000a). Frictionless Commerce?: A Comparison of Internet and Conventional Retailers. Management Science, 46(4), 563–585. Brynjolfsson, E., & Smith, M. D.(2000b). The Great Equalizer?: Consumer Choice Behavior at Internet Shopbots. Mimeo, MIT Sloan School. Carlton, D. W., & Chevalier, J. (2001). Free Riding and Sales Strategies for the Internet. Working Paper, National Bureau of Economic Research. Carr, G. N. (2000). Hypermediation: Commerce as Clickstream. Harvard Business Review (January–February), 46–47. Chen, Y., & Sudhir, K. (2001). When Shopbots Meet Emails: Implications for Price Competition on the Internet. Review of Marketing Science, 1(3). Chircu, A. M., & Kauffman, R. J. (2000). Reintermediation Strategies in Business-to-Business Electronic Commerce. International Journal of Electronic Commerce, 4(4), 7–42. Christensen, C., & Tedlow, R. S. (2000). Patterns of Disruption in Retailing. Harvard Business Review (January–February), 42–45. Chui, K., & Zwick, R. (1999). Auctions On The Internet – A Preliminary Study. Working Paper, Department of Marketing, Hong Kong University of Science and Technology, 1–27. Clay, K., Krishnan, R., Wolff, E., & Fernandes, D. (1999). Retail Strategies on the Web: Price and Non-Price Competition in the On Line Book Industry. Working Paper, Carnegie Mellon University. Clay, K., Krishnan, R., & Smith, M. (2001). The Great Experiment: Pricing on the Internet. In: Lowry, Cherrington & Watson (Eds), The Handbook of Electronic Commerce in Business and Society (pp. 139–152). CRC Press. Clay, K., Krishnan, R., & Wolff, E. (2001). Prices and Price Dispersion on the Web: Evidence from the Online Book Industry. Working Paper, National Bureau of Economic Research, No. 8271. Clay, K., & Tay, C. H. (2001). Cross-Country Price Differentials in the Online Textbook Market. Working Paper, Carnegie Mellon University. May 2001. Clemons, E., Hann, I. H., & Hitt, L. (2002). Price Dispersion and Differentiation in Online Travel: An Empirical Investigation. Management Science, 48(4), 534–549.

The Impact of the Internet on Horizontal and Vertical Competition

25

Darby, M., & Karni, E. (1973). Free Competition and the Optimal Amount of Fraud. Journal of Law and Economics, 16, 67–88. Dayal, S., Landesberg, H., & Zeisser, M. (2000). Building Digital Brands. McKinsey Quarterly, 2, 42–51. Degeratu, A., Rangaswamy, A., & Wu, J. (2000). Consumer Choice Behavior in Online and Traditional Supermarkets: The Effects of Brand Name, Price, and Other Search Attributes. International Journal of Research in Marketing, 17(1), 55–78. DeKoning, V., Giles, T., Glufing, L., & Leigh, V. (1999). Online Auctions: Consumer Behavior in Local vs. Global Formats. Working Paper. Ellison, G., & Ellison, S. F. (2001). Search, Obfuscation, and Price Elasticities on the Internet. Working Paper, MIT. January 2001. Erevelles, S., Rolland, E., & Srinivasan, S. (2001). Are Prices Really Lower on the Internet?: An Analysis of the Vitamin Industry. Working Paper, University of California, Riverside. Evans, P., & Wurster, T. S. (1999). Getting Real About Virtual Commerce. Harvard Business Review (November–December), 85–94. Gates, B. (1995). The Road Ahead. Viking. Gellman, R. (1996). Distintermediation and the Internet. Government Information Quarterly, 13(1), 1–8. Houser, D., & Wooders, J. (2001). Reputation in Auctions: Theory and Evidence from eBay. Working Paper, University of Arizona. October 2001. Iyer, G., & Pazgal, A. (2001). Internet Shopping Agents: Virtual Co-Location and Competition. Working Paper, University of California at Berkeley & Washington University in St. Louis. May 2001. Jin, L., & Robey, D. (1999). Explaining Cybermediation: An Organisational Analysis of Electronic Retailing. International Journal of Electronic Commerce, 3 (4, Summer), 47–65. Kaplan, S., & Sawhney, M. (2000). E-Hubs: The New B2B Marketplaces. Harvard Business Review (May–June), 97–103. Katz, M., & Shapiro, C. (1994). Systems Competition and Network Effects. Journal of Economic Perspectives, 8(2), 93–115. Klein, S., & O’Keefe, R. M. (1999). The Impact of the Web on Auctions: Some Empirical Evidence and Theoretical Considerations. International Journal of Electronic Commerce, 3 (3, Spring), 7–20. Klein, S., & Selz, D. (2000). Cybermediation in Auto Distribution: Channel Dynamics and Conflicts. Journal of Computer-Mediated Communication, 5(3). Lal, R., & Sarvary, M. (1999). When and How Is the Internet Likely to Decrease Price Competition? Marketing Science, 18(4), 485–503. Lee, H. G. (1998). Do Electronic Markets Lower the Price of Goods? Communications of the ACM, 41 (1, January), 73–80. Lucking-Reiley, D. (1999). Using Field Experiments To Test Equivalence Between Auction Formats: Magic on the Internet. American Economic Review, 89(5), 1063–1080. Lucking-Reiley, D. (2000a). Auctions on the Internet: What’s Being Auctioned and How? Journal of Industrial Economics, 48 (3, September), 227–252. Lucking-Reiley, D., Bryan, D., Prasad, D., Naghi, D., & Reeves, D. (2000b). Pennies from eBay: The Determinants of Price in Online Auctions. Working Paper, VanderBilt University. Lynch, J. G. J., & Ariely, D. (2000). Wine Online: Search Costs Affect Competition on Price, Quality, and Distribution. Marketing Science, 19 (1, Winter), 83–103.

26

ANITA ELBERSE, PATRICK BARWISE AND KATHY HAMMOND

Malone, T., Yates, J., & Benjamin, R. (1987). Electronic Markets and Electronic Hierarchies: Effects of Information Technology on Market Structure and Corporate Strategies. Communications of the ACM, 30(6), 484–497. Montgomery, A. L., Hoanager, K., Krishnan, R., & Clay, K. B. (2001). Designing a Better Shopbot. Working Paper, Carnegie Mellon University. Negroponte, N. (1995). Being Digital. Knopf. Nelson, P. (1970). Information and Consumer Behavior. Journal of Political Economy, 78(2), 311–329. Nelson, P. (1974). Advertising as Information. Journal of Political Economy, 81 (July/August), 729–754. Pan, X., Ratchford, B. T., & Shankar, V. (2001). Why Aren’t the Prices of the Same Item the Same at Me.com and You.com?: Drivers of Price Dispersion Among E-Tailers. Working Paper, Robert H. Smith School of Business, University of Maryland. November 2001. Ramsdell, G. (2000). The Real Business of B2B. McKinsey Quarterly, 174–183. Rayport, J. F., & Sviokla, J. J. (1994). Managing in the Marketspace. Harvard Business Review (November–December), 141–150. Rayport, J. F., & Sviokla, J. J. (1995). Exploiting the Virtual Value Chain. Harvard Business Review (November-December), 75–85. Rohlfs, J. (1974). A Theory of Interdependent Demand for a Communications Service. Bell Journal of Economics, 5(1), 16–37. Sarkar, M., Butler, B., & Steinfield, C. (1995). Intermediaries and Cybermediaries: A Continuing Role for Mediating Players in the Electronic Marketplace. Journal of Computer Mediated Communication, 1 (3, December). Sarkar, M., Butler, B., & Steinfield, C. (1998). Cybermediaries in Electronic Marketspace: Toward Theory Building. Journal of Business Research, 41, 215–221. Scott Morton, F., Zettelmeyer, F., & Silva-Risso, J. (1999). Internet Car Retailing. Working Paper, National Bureau of Economic Research. Shaffer, G., & Zettelmeyer, F. (1999). The Internet as a Medium for Marketing Communications: Channel Conflict Over the Provision of Information. Working Paper, MIT Ecommerce Forum. Shankar, V., Rangaswamy, A., & Pusateri, M. (1999). The Online Medium and Customer Price Sensitivity. Working Paper, University of Maryland & MIT Ecommerce Forum, 1999. Shapiro, C., & Varian, H. R. (1998a). Information Rules: A Strategic Guide to the Network Economy. Harvard Business School Press. Shapiro, C., & Varian, H. R. (1998b). Versioning: The Smart Way to Sell Information. Harvard Business Review (November/December), 106–114. Simon, H., & Schumann, H. (2001). Pricing Opportunities in the Digital Age. In: J. Wind & V. Mahajan (Eds), Digital Marketing: Global Strategies from the World’s Leading Experts (pp. 362–380). New York: John Wiley. Sinha, I. (2000). Cost Transparency: The Net’s Real Threat to Prices and Brands. Harvard Business Review (March/April), 43–50. Smith, M. D., Bailey, J., & Brynjolfsson, E. (2000). Understanding Digital Markets: Review and Assessment. In: E. Brynjolfsson & B. Kahin (Eds), Understanding the Digital Economy. MIT Press. Smith, M. (2001). The Law of One Price? Price Dispersion and Parallel Pricing in Internet Markets. Working Paper, Carnegie Mellon University. Smith, M., & Brynjolfsson, E. (2001). Customer Decision-Making at an Internet Shopbot. Working Paper, Carnegie Mellon University.

The Impact of the Internet on Horizontal and Vertical Competition

27

Stigler, G. J. (1961). The Economics of Information. The Journal of Political Economy, 69(3), 213–225. Weiber, R., & Kollmann, T. (1998). Competitive Advantages in Virtual Markets – Perspectives of “Information-Based Marketing” in Cyberspace. European Journal of Marketing, 32(7/8), 603–615. Weitz, B., & Wensley, R. (2002). The Handbook of Marketing. London: Sage. Wigand, R. T., & Benjamin, R. I. (1995). Electronic Commerce: Effects on Electronic Markets. Journal of Computer Mediated Communication, 1(3), 1–11.

This Page Intentionally Left Blank

PRICE COMPETITION BETWEEN PURE PLAY VERSUS BRICKS-AND-CLICKS E-TAILERS: ANALYTICAL MODEL AND EMPIRICAL ANALYSIS Xing Pan, Venkatesh Shankar and Brian T. Ratchford ABSTRACT In this paper, we first develop a game theoretic model of price competition between a pure play e-tailer and a bricks-and-clicks e-tailer. We show that in general, the pure play e-tailer has a lower equilibrium price. We then develop a simultaneous equation model of e-tailer price and traffic and estimate this model using data collected from 905 e-tailers across eight product categories. The empirical results show that after controlling for the effects of other variables, prices at pure play e-tailers are generally lower. E-tailers with high traffic do not always command higher prices. Etailers with high level of reliability, shopping convenience, and deep information, generally do not generate high web traffic and do not enjoy high prices. However, trust enhances e-tailer traffic and early online entry is associated with both high e-tailer traffic and high prices.

1. INTRODUCTION Pricing is a well-researched topic in economics and marketing (Rao, 1984) and pricing on the Internet has attracted a lot of theoretical and empirical research The Economics of the Internet and E-Commerce, Volume 11, pages 29–61. Copyright © 2002 by Elsevier Science Ltd. All rights of reproduction in any form reserved. ISBN: 0-7623-0971-7

29

30

XING PAN, VENKATESH SHANKAR AND BRIAN T. RATCHFORD

attention (e.g. Bakos, 1997; Brynjolffson et al., 2000; Clay et al., 1999; Clemens et al., 1998; Lal & Sarvary, 1999). In the economics literature, much research has focused on price dispersion in the offline environment (e.g. Cohen, 1998; Carlson & McAfee, 1983; Dahlby & West, 1986; Dana, 1999; Mitchell & Sorensen, 1986; Salop & Stiglitz, 1982; Sorensen, 2000; Stigler, 1961). It has been found that price dispersion in online environments, which are conducive to perfect competition, is persistent, even for homogeneous products (e.g. Baye & Morgan, 2001a, b, 2002; Morgan et al., 2001; Smith & Brynolfsson, 2000; Pan, Ratchford & Shankar, 2001, 2002; Ratchford, Pan & Shankar, 2002). Online price dispersion, defined as a non-degenerate distribution of prices (such as a positive range or standard deviation) of an item with the same measured characteristics across e-tailers of the item at a given point in time, has generally been found to be substantial by these studies. The general conclusion is that online price dispersion is no less than offline price dispersion (e.g. Smith & Brynjolffson, 2000) and that market characteristics explain a significant portion of online price dispersion (e.g. Pan, Ratchford & Shankar, 2001). Not much, however, is known about whether prices of pure play Internet e-tailers are lower or higher than prices of bricks-and-clicks retailers (retailers with bricks-and-mortar stores who also have an Internet store). Increasingly, many bricks-and-mortar stores are having online presence and turning themselves into bricks-and-clicks retailers (Zettelmeyer, 2000). Pricing differences between these types of retailers are important from both theoretical and empirical perspectives and have implications for retailer pricing strategies. Are there differences in the characteristics and strategies between these types of e-tailers that may suggest differences in their prices in a competitive context? Can bricks-and-clicks e-tailers leverage their bricks-and-mortar presence offering inspection, pick-up and return benefits to charge higher prices than pure play e-tailers? Do pure play e-tailers have to price lower than bricksand-clicks e-tailers to draw traffic to their web sites? Or can they afford to price as high or higher than bricks-and-clicks e-tailers based on better convenience or Web site features? These questions have not been adequately explored. There has been limited research on possible price differences between different types of e-tailers. There have been few theoretical papers in this regard. Druehl and Porteus (2001) analytically model price competition between an Internet firm and a bricks-and-mortar firm in a vertically differentiated product market and conclude that the Internet firm may not dominate the market even if it delivers the product at a lower cost. Their comparison is not between bricks-and-clicks and pure play Internet e-tailers, but between bricks-and-mortar and pure play Internet retailers – a situation that

Price Competition Between Pure Play Versus Bricks-and-Clicks e-Tailers

31

is not as common as the former in today’s environment. Balasubramanian (1998) analyzes a model of competition between a bricks-and-mortar store and a direct mail store. His results show that the direct marketer’s price is higher (lower) than that of the conventional retailer when the product is not well (well) adapted to the direct channel. His comparison, however, is also not between bricks-and-clicks and pure play Internet stores. The number of empirical studies on this issue has been sparse. Most studies on this topic have compared prices at online and traditional retailers. Bailey (1998) compared the prices of books, CDs and software sold online with those sold through traditional channels between 1996 and 1997, that is, during the early days of Internet, and found higher prices online for all the product categories. Clay et al. (1999) did not find any significant differences in the two channels for books. Brynjolfsson and Smith (2000) found that prices of CDs and books sold online are much lower than those sold through traditional channels. Brown and Goolsbee (2002) found decreasing price levels in the life insurance industry due to the impact of the Internet. Erevelles, Rolland and Srinivasan (2000) found higher levels prices of vitamins for Internet retailers than for traditional retailers. Morton, Zettelmeyer and Risso (2001) studied dealer pricing of automobiles in California and found that prices are lower online than offline although the difference was only 2%. Ancarani and Shankar (2002) found that the prices of books and CDs in Italy were lower at online retailers than those at offline retailers by about 4–6%. Taken together, these studies suggest that prices at online retailers are generally lower than those at offline retailers in the current environment. However, these studies did not examine price differences between pure play and bricks-and-clicks e-tailers. In a study comparing prices at pure play and bricks-and-clicks e-tailers, Tang and Xing (2001) found that prices of DVDs at pure play Internet retailers were significantly lower than prices at bricks-and-clicks e-tailers (an average of 14%). However, they focused on only one product category, did not control for the effects of possible drivers of prices, and did not offer a theoretical basis for their results. Based on a hedonic regression analysis, Pan, Ratchford and Shankar (2002) found that prices at pure play e-tailers relative to bricksand-clicks e-tailers were lower for CDs, DVDs, desktop computers, and laptop computers, the same for PDAs and electronics, and higher for books and computer software. Ancarani and Shankar (2002) found that prices of books in Italy are lower at pure play e-tailers than those at multichannel e-tailers by about 6%. Combined, these studies suggest that generally, the prices at pure play e-tailers are less than those at bricks-and-clicks e-tailers. However, these studies were not based on theoretical models. Our purpose in this paper is to

32

XING PAN, VENKATESH SHANKAR AND BRIAN T. RATCHFORD

investigate theoretically and empirically price competition between pure play and bricks-and-clicks e-tailers. We first develop a game theoretic model of price competition between a pure play e-tailer and a bricks-and-clicks e-tailer and derive their equilibrium prices. We show that in general, the price at the pure play e-tailer is lower than that at the bricks-and-clicks e-tailer. We then develop a simultaneous equation model of e-tailer price and traffic for empirical analysis, in which price and traffic are functions of e-tailer and market characteristics and the type of e-tailer (pure play or bricks-and-clicks e-tailer). We estimate this model using data collected from 905 e-tailers across eight product categories, including apparel, gifts and flowers, health and beauty, home and garden, sports equipment, computer hardware, consumer electronics, and office supplies. The empirical results support our analytical model in that after controlling for the effects of other variables, prices at pure play e-tailers are generally lower than prices at bricksand-clicks e-tailers. The rest of the paper is organized as follows. In the next section, we develop and discuss a game theoretic model of price competition between a pure play and a bricks-and-clicks e-tailer. In Section 3, we describe the data. In Section 4, we develop the empirical model and discuss its estimation. We present the results from empirical analysis in Section 5. In Section 6, we discuss the implications and outline avenues for future research and offer our conclusions.

2. ANALYTIC MODEL OF PRICE COMPETITION BETWEEN PURE PLAY AND BRICKS-AND-CLICKS E-TAILERS Our model is based on the Hotelling framework (1929). We first derive the equilibrium prices and profits of a bricks-and-mortar retailer and a pure play e-tailer. We then discuss the conditions under which the bricks-and-mortar retailer will launch an Internet store and compete against the pure play e-tailer through both the channels, that is, as a bricks-and-clicks retailer. We derive the equilibrium prices and show that the pure play e-tailer’s price is lower than that at the bricks-and-clicks e-tailer. 2.1. The Hotelling Framework In the Hotelling (1929) model, consumers are homogeneous in their preferences except for their locations in the “main street,” i.e. the market. Consumers have the same valuation for the product they purchase. However, to

Price Competition Between Pure Play Versus Bricks-and-Clicks e-Tailers

33

acquire the product, a consumer incurs a “transaction cost” that is heterogeneous across consumers and is determined by the distance from the consumer’s location to the store’s location. The Hotelling model has been well discussed in economic literature and the “main street” and “location” can represent a physical map or a perceptual preference map. For example, it may refer to how far the consumer lives away from a retail store, or how much the consumer likes the assortment or the convenience in buying the items at the store. A consumer whose “location” is exactly the same as the store’s location does not need to travel to the retail store and thus his/her “transaction cost” is zero. However, a consumer whose “location” is farther away from the store incurs a greater transportation cost to acquire the product. Figure 1 illustrates consumer preference in the Hotelling framework. In Fig. 1, the horizontal lines represent Hotelling’s “main street.” Consumers are uniformly distributed in this linear city,1 whose two ends are denoted by 0 and 1. There exist two stores (retailers) in this linear market, who are in “location” a and 1 ⫺ b and sell identical products but whose sales people dress uniforms with different patterns, e.g. with dashed line and dadot line respectively. Alternatively, they can be viewed as two retail stores that sell identical products but sell through different channels, a bricks-and-mortar store and an Internet store. Hereafter, we refer to Store 1 as a bricks-and-mortar store and Store 2 as a pure play Internet store (e.g. 800.com, buy.com, and TheNerds.net).2 The distance from Store 1’s location to the right end of the market is b, and the distance from Store 2’s location to the left end of the market is a. The distance between the two stores is then 1 ⫺ a ⫺ b. All the three distances are of course, non-negative. For a consumer whose “location” is at x, his/her relative preference toward the two stores is given by the relative

Fig. 1. The Hotelling Main Street.

34

XING PAN, VENKATESH SHANKAR AND BRIAN T. RATCHFORD

distances from his/her “location” to the stores’ “locations”, i.e. | x ⫺ a | / | x ⫺ (1 ⫺ b) |. For a consumer between 0 and a, Store 1 is closer than Store 2, so he/she strictly prefers Store 1 to Store 2. Similarly, for a consumer between 1 ⫺ b and 1, Store 2 is closer and is strictly preferred. Different consumers may prefer the Internet channel or the bricks-and-mortar channel for a variety of reasons. For example, some consumers prefer the bricks-and-mortar channel: (1) to see, touch, and feel the product before purchase; (2) to acquire the product instantly rather than wait for delivery; (3) to avoid security concerns on online transactions and give credit card numbers and personal information on the Internet; and (4) for lack of computer availability, Internet access, or computer related skills. In contrast, some other consumers may prefer the Internet channel for reasons such as: (1) flexible shopping time since the Internet is open 24 ⫻ 7; (2) ease of finding the product and comparing prices on the Internet; and (3) the lack of need for transportation to the bricks-and-mortar store. Based on a survey by Forrester Research of 90,000 U.S. households in December 1998, Goolsbee (2001) found that when prices are equal, 68% of computer shoppers would buy from local retailers while 32% of them would bypass the local stores and buy from the Internet. This suggests that the bricksand-mortar and the Internet channels are horizontally differentiated and can be appropriately modeled in a Hotelling framework. Notably, the consumer preference we have discussed is the preference toward a distribution channel rather than the preference toward particular retailers or stores. Although a particular retailer with appropriate marketing strategies can be more attractive than its competitors, consumer preference toward the channel through which the retailer sells is relatively much more stable.3 Therefore, our model focuses on a game in which price rather than location is the decision variable. 2.2. Modeling Assumptions We assume that consumers buy a product for which the consumers’ reservation value is high enough to guarantee that every consumer will buy one and just one unit of the product. Consumer x’s utility (in monetary value) from buying this product is given by: (1) Ux = Ix + Vx ⫺ FPx where Ix is the utility from income (or from buying other goods), Vx is the utility from the product, and FPx is the full price that consumer x pays, which

Price Competition Between Pure Play Versus Bricks-and-Clicks e-Tailers

35

is the sum of the price paid (Px) and the transaction cost (TCx) incurred by the consumer. We assume that the “transaction cost” that consumer x has to pay is a quadratic function of the distance between the consumer and the store, i.e. TCx = t(x ⫺ li)2

(2)

where x is the “location” of the consumer (assumed to be uniformly distributed along the linear market), li is the “location” of the store, and t is the unit transaction cost (t > 0). Thus, consumer x’s utility from buying the product is: Ux = Ix + Vx ⫺ P2 ⫺ t(x ⫺ a)2, if buying from the Internet channel

(3)

Ux = Ix + Vx ⫺ P1 ⫺ t[x ⫺ (1 ⫺ b)] , 2

if buying from the bricks-and-mortar channel

(4)

Consumers will buy from the store that offers a lower full price and thus a higher utility level. Furthermore, for simplicity, we assume that only one bricks-and-mortar retailer and one pure Internet retailer compete in this market. They both seek to maximize their own profits. We also assume that both the stores have the same marginal cost C.4 2.3. Bricks-and-Mortar Versus Pure Play Internet e-Tailers We first examine the competition between the bricks-and-mortar retailer and the pure play Internet retailer and then extend it to competition between the bricks-and-clicks retailer and the pure play Internet e-tailer. Consumers choose to buy from the store that offers a lower full price. The demand for the bricks-and-mortar store (Store 1) is given by Eq. (5) and the demand for the pure play Internet store (Store 2) is given by Eq. (6). P1 + t[x ⫺ (1 ⫺ b)]2 < P2 + t(x ⫺ a)2

(5)

P1 + t[x ⫺ (1 ⫺ b)] > P2 + t(x ⫺ a)

(6)

2

2

Solving Eqs (5) and (6) for x, we can obtain the demand expressions for the two stores as follows. D1 =

(P2 ⫺ P1) (1 ⫺ a + b) + 2 · t · (1 ⫺ a ⫺ b) 2

(7)

D2 =

(P1 ⫺ P2) (1 + a ⫺ b) + 2 · t · (1 ⫺ a ⫺ b) 2

(8)

36

XING PAN, VENKATESH SHANKAR AND BRIAN T. RATCHFORD

Thus, the profit functions of the two stores are: ␲1 = (P1 ⫺ C) ·

␲2 = (P2 ⫺ C) ·

冋 冋

(P2 ⫺ P1) (1 ⫺ a + b) + 2 · t · (1 ⫺ a ⫺ b) 2 (P1 ⫺ P2) (1 + a ⫺ b) + 2 · t · (1 ⫺ a ⫺ b) 2

册 册

(9)

(10)

In a Nash simultaneous decision game, the first order conditions are: 1 1 P1 = · t · (1 ⫺ a ⫺ b) · (1 ⫺ a + b) + · (P2 + C) 2 2

(11)

1 1 P2 = · t · (1 ⫺ a ⫺ b) · (1 + a ⫺ b) + · (P1 + C) 2 2

(12)

From these equations, we can solve the equilibrium prices of the two stores as follows. 1 P1具Opt ⫺ Nash典 = · t · (1 ⫺ a ⫺ b) · (3 ⫺ a + b) + C 3

(13)

1 P2具Opt ⫺ Nash典 = · t · (1 ⫺ a ⫺ b) · (3 + a ⫺ b) + C 3

(14)

The equilibrium price difference is: 2 P1具Opt ⫺ Nash典 ⫺ P2具Opt ⫺ Nash典 = · t · (1 ⫺ a ⫺ b) · (b ⫺ a) 3

(15)

Thus, it is not certain which type of retailer has a higher price. This is consistent with Druehl and Porteus (2001) who obtain a similar result when they model competition using a vertically differentiated model. If b > a, i.e. if more consumers prefer the bricks-and-mortar channel to the Internet channel for reasons discussed before, then the bricks-and-mortar store is able to charge a higher price and vice versa.5 We can also calculate the equilibrium profits of the two stores. If b > a, then the bricks-and-mortar store will enjoy higher profits than the pure play Internet store. ␲1具Opt ⫺ Nash典 =

1 · (3 ⫺ a + b)2 · (1 ⫺ a ⫺ b) · t 18

(16)

␲2具Opt ⫺ Nash典 =

1 · (3 + a ⫺ b)2 · (1 ⫺ a ⫺ b) · t 18

(17)

Price Competition Between Pure Play Versus Bricks-and-Clicks e-Tailers

37

Figure 2 illustrates the competition and Nash equilibrium with a numerical example (when t = 10, C = 1, a = 1/4, b = 1/3). The horizontal axis represents the linear market and the vertical axis represents price. The solid curve represents the full price for consumer x when buying from the pure play Internet store and the dotted curve represents the full price when buying from the bricks-andmortar store. Consumers between the left end of the market and the intersection of the two curves will buy from the pure play Internet store, while consumers between the intersection of the two curves and the right end of the market will buy from the bricks-and-mortar store. The dadot line is the bricks-and-mortar store’s price and is higher than that at the pure play Internet store (the dashed line). 2.4. Bricks-and-Clicks Versus Pure Play Internet e-Tailers We now analyze the situation when the bricks-and-mortar retailer (Store 1) will launch an Internet branch (denoted as Store 3) to compete against the pure play e-tailer (Store 2). For simplicity, we assume that the cost of online entry is zero for both the stores.6 Since Store 1’s Internet branch sells through exactly the same channel as Store 2, Store 2 and Store 3 are both at the “location” a. However, consumers

Fig. 2. Competition Between a Bricks-and-Mortar Store and a Pure Play Internet Store.

38

XING PAN, VENKATESH SHANKAR AND BRIAN T. RATCHFORD

may usually perceive the Internet branch of the bricks-and-mortar store to be superior to the pure play Internet store. This is because the store name and the presence of the physical bricks-and-mortar store are likely to increase consumers’ familiarity with its Internet branch and leverage consumer loyalty. Further, the physical presence of the physical bricks-and-mortar store is likely to increase consumer trust toward its Internet branch even though the consumer may have never experienced the bricks-and-mortar store before, because one of the main concerns of online shopping is the separation of payment and delivery and thus the uncertainty of order fulfillment (Brynjolfsson & Smith, 2000; Pan, Ratchford & Shankar, 2002). In addition, the presence of the bricks-and-mortar store may be associated with better customer support. For example, consumers can buy products from Staples.com and return at any Staples’ local store. Further, the presence of the bricks-and-mortar store may increase consumer convenience. For example, consumers can order from Circuitcity.com, check any local store inventory online, and choose to pick up the product at any Circuit City local store with the product in stock. George Barr, Circuit City’s director of Internet merchandising, reported that about 50% of their online shoppers choose to pick up products at their local stores (Messmer, 2001). Given these, we conclude that the consumer transaction cost of buying from the bricks-and-mortar retailer’s Internet branch (Store 3) is possibly lower than that at the pure play e-tailer (Store 2). Specifically, we assume that the unit transaction cost at Store 3 is lower than at Store 2, i.e. t3 = k · t, where 0 < k ≤ 1. When k = 1, i.e. when consumers actually perceive the two type of Internet stores are indifferent, Store 2 and Store 3 will play a Bertrand game and both will have to price at the marginal cost to generate positive sales (Economides, 1993). As a result, both Store 2 and Store 3 will have zero profits. Given that the price on the Internet is C, Store 1’s bricks-and-mortar store has a demand of D1 =

(C ⫺ P1) (1 ⫺ a + b) + , 2 · t · (1 ⫺ a ⫺ b) 2

(18)

and its equilibrium price and equilibrium profit are: 1 P1具Opt ⫺ Nash典 = · t · (1 ⫺ a ⫺ b) · (1 ⫺ a + b) + C 2

(19)

1 ␲1具Opt ⫺ Nash典 = · (1 ⫺ a + b)2 · (1 ⫺ a ⫺ b) · t > 0 8

(20)

Comparing Eq. (19) with Eq. (13), we can see that the price is lower in this scenario. Comparing Eq. (20) with Eq. (16), the profit is also lower. It suggests

Price Competition Between Pure Play Versus Bricks-and-Clicks e-Tailers

39

that when consumers perceive the bricks-and-mortar retailer’s Internet branch as no superior to the pure play Internet store, launching Store 3 will severely increase the competition on the Internet, which not only leads to lower price online, but also leads to lower price offline. As a result, Store 1, the bricksand-clicks retailer, will have lower profit compared to the case of not launching the Internet branch. Thus, we conclude that, when k = 1, the bricks-and-mortar retailer will not launch an Internet branch. We next analyze the case when 0 < k < 1. First, we make the assumption that to preserve channel integrity, the bricks-and-clicks retailer sets the same price at its two stores, the bricks-and-mortar store and the Internet branch, i.e. P1 = P3. This is supported by anecdotal evidence of the retail markets that shows that most bricks-and-mortar retailers have the same prices at the two channels (e.g. Best Buy, Circuit City, CompUSA, and Office Depot). However, this assumption is for simplicity and relaxing this assumption should not substantially change our results. We relax this assumption subsequently and discuss the case when the prices can be different at the two stores. When 0 < k < 1, the case is more interesting and also more complicated. The demand of each store is given by the following equations, i.e. when the full price at that store is lower than the full price at other stores. P1 + t · [x ⫺ (1 ⫺ b)]2 < Min[P2 + t · (x ⫺ a)2, P3 + k · t · (x ⫺ a)2] – Store 1’s demand (21) 2 2 P2 + t · (x ⫺ a) < Min[P1 + t · (x ⫺ (1 ⫺ b)] , P3 + k · t · (x ⫺ a)2] – Store 2’s demand (22) 2 2 P3 + k · t · (x ⫺ a) < Min[P2 + t · (x ⫺ (1 ⫺ b)] , P2 + t · (x ⫺ a)2] – Store 3’s demand (22) Figures 3a–3d illustrate the four cases of the demand structure. Notice that the bold solid curve is the full price at Store 3, and is flatter than the full price curves at the other stores, given its lower unit transaction cost. When P1 ⫺ P2 , (24) k>1⫺ t · a2 for consumers at the left end of the market, the full price at Store 3 is higher than that at Store 2, thus Store 3 cannot grab sales from Store 2.7 When k>

冋

册

(P1 ⫺ P2) ⫺ t · (1 ⫺ a ⫺ b)2 (P1 ⫺ P2) + t · (1 ⫺ a ⫺ b)2

2

,

(25)

40

XING PAN, VENKATESH SHANKAR AND BRIAN T. RATCHFORD

Fig. 3a.

Graphic Illustration of the Competition With Three Stores.

for consumers at the middle of the market, the full price at Store 3 is higher than that at Store 1 and Store 2, thus Store 3 cannot grab sales from the other stores. However, when k is smaller than any of the values in Eqs (24) and (25), Store 3 will have positive sales. Especially, when k is sufficiently small, i.e. when the superiority of the Internet branch of the bricks-and-mortar store relative to the pure play Internet store is sufficiently large, Store 3 will grab demand from both Store 1 and Store 2 (Fig. 3d). We use the case depicted in

Price Competition Between Pure Play Versus Bricks-and-Clicks e-Tailers

41

Fig. 3b.

Fig. 3d as an example to derive the equilibrium prices and then discuss the general case. We have the case in Fig. 3d when

冋

k < Min 1 ⫺

P1 ⫺ P2 , t · a2

冋

册册

(P1 ⫺ P2) ⫺ t · (1 ⫺ a ⫺ b)2 (P1 ⫺ P2) + t · (1 ⫺ a ⫺ b)2

2

.

(26)

42

XING PAN, VENKATESH SHANKAR AND BRIAN T. RATCHFORD

Fig. 3c.

In this case, the demand and profit functions for the pure play e-tailer are:

D2 = 2

冑 冋冑

P1 ⫺ P2 t · (1 ⫺ k)

␲2 = (P2 ⫺ C) · 2

(27)

册

P1 ⫺ P2 t · (1 ⫺ k)

(28)

Price Competition Between Pure Play Versus Bricks-and-Clicks e-Tailers

43

Fig. 3d.

The total demand for Store 1 and Store 3 (bricks-and-clicks e-tailer) is

D1 + D3 = 1 ⫺ 2 ·

冑

P1 ⫺ P2 , t · (1 ⫺ k)

(29)

44

XING PAN, VENKATESH SHANKAR AND BRIAN T. RATCHFORD

and the total profit for the bricks-and-clicks retailer is

冋 冑

␲1 = (P1 ⫺ C) · 1 ⫺ 2

册

P1 ⫺ P2 . t · (1 ⫺ k)

(30)

In a simultaneous Nash game, we have the two stores’ first order condition as 2 1 P2 = · P1 + · C 3 3

0=1⫺2·

冋

册 冋

(P1 ⫺ P2) t · (1 ⫺ k)

1/2

⫺

– Store 2 (P1 ⫺ C)

册

(P1 ⫺ P2) t · (1 ⫺ k)

1/2

· t · (1 ⫺ k)

– Store 1

Solving the two equations, we obtain the two stores’ optimal prices: 3 P1具Opt ⫺ Nash典 = C + · t · (1 ⫺ k) 25 P2具Opt ⫺ Nash典 = C +

(31)

2 · t · (1 ⫺ k) 25

(32)

(33) (34)

Thus, in Nash equilibrium, the price difference between the bricks-and-clicks retailer and the pure play e-tailer is (P1 ⫺ P2)具Opt ⫺ Nash典 =

1 · t · (1 ⫺ k) > 0. 25

(35)

We can also derive the following equilibrium profits.

冋 冑 册 冋冑 册

␲1具Opt ⫺ Nash典 = (P1 ⫺ C) · 1 ⫺ 2

␲2具Opt ⫺ Nash典 = (P2 ⫺ C) · 2

P1 ⫺ P2 9 = · t · (1 ⫺ k) t · (1 ⫺ k) 125

P1 ⫺ P2 4 = · t · (1 ⫺ k) t · (1 ⫺ k) 125

(36)

(37)

Similarly, we can solve a Stackelberg lead-follower game in which the bricksand-clicks retailer is the leader.8 In the Stackelberg game, the pure play e-tailer responds to the bricks-and-clicks retailer’s price according to its own first order condition. The bricks-and-clicks retailer knows the pure play e-tailer’s price

Price Competition Between Pure Play Versus Bricks-and-Clicks e-Tailers

45

response function and incorporates that in its own pricing decision. The equilibrium prices are: 1 P1具Opt ⫺ Leader典 = C + · t · (1 ⫺ k) > P1具Opt ⫺ Nash典 3

(38)

2 P2具Opt ⫺ Leader典 = C + · t · (1 ⫺ k) > P2具Opt ⫺ Leader典 9

(39)

Again, we find that in equilibrium, the bricks-and-clicks e-tailer charges a higher price than does the pure play e-tailer. 1 (P1 ⫺ P2)具Opt ⫺ Stackelberg典 = · t · (1 ⫺ k) > 0 9

(40)

Comparing the Stackelberg game with the Nash game, we find that both the stores charge higher prices in the Stackelberg game than in the Nash game and that the price difference between the two stores is larger in the Stackelberg game than it is in the Nash game. It suggests that sequential pricing allows the two stores to better observe each other’s strategy and better coordinate to reduce the degree of competition. By doing so, they get higher profits than in the Nash case. (P1 ⫺ P2)具Opt ⫺ Stackelberg典 > (P1 ⫺ P2)具Opt ⫺ Nash典

冋 冑 册 冋冑 册

␲1具Opt ⫺ Stackelberg典 = (P1 ⫺ C) · 1 ⫺ 2

␲2具Opt ⫺ Stackelberg典 = (P2 ⫺ C) · 2

(41)

P1 ⫺ P2 1 = · t · (1 ⫺ k) > ␲1具Opt ⫺ Nash典 (42) t · (1 ⫺ k) 9

P1 ⫺ P2 4 = · t · (1 ⫺ k) > ␲2具Opt ⫺ Nash典 t · (1 ⫺ k) 27

(43)

We have derived the equilibrium price difference between the two types of Internet retailers under certain conditions. We showed that in these cases, the bricks-and-clicks retailer charges a higher price than the pure play e-tailer. We do not solve the equilibria in all the different cases in Figs 3a–3d due to mathematical complexity. However, we offer the intuition behind the game and conclude that the bricks-and-clicks retailer charges a higher price than the pure play e-tailer in equilibrium. In the next discussion, we relax the assumption that the bricks-and-clicks e-tailer has same prices in its bricks-and-mortar and Internet stores. Let us assume that the opposite outcome is true, that is, the bricks-and-clicks retailer actually does not have a higher price than that at the pure play e-tailer,

46

XING PAN, VENKATESH SHANKAR AND BRIAN T. RATCHFORD

i.e. P3 ≤ P2. In such case, the pure play e-tailer’s full price will be higher than that at the Internet branch of the bricks-and-clicks retailer, so it will gain no sales and will be driven out of the market. To stay in the market, the pure play e-tailer will be willing to charge as low as its marginal cost C. Since we assume P3 ≤ P2, we know that the bricks-and-clicks retailer will also charge a price equal to the marginal cost C. However, by doing so, the bricks-and-clicks retailer will have lower profits compared to the case of not launching an Internet branch, similar to what we previously discussed in the case k = 1. In this case, the bricks-and-click retailer’s profit will be even lower than that described by Eq. (20) since its Internet branch can grab more sales from its bricks-and-mortar store (k < 1) and the Internet branch does not make profits by pricing at the marginal cost. As a result, the bricks-and-mortar retailer would not allow this to happen and we will always have P3 ≥ P2 ≥ C. 2.5. Summary In summary, we derived an analytic model of price competition based on the Hotelling modeling framework. We first analyzed the competition between a bricks-and-mortar retailer and a pure play Internet e-tailer. We then extended it to analyze the condition under which the bricks-and-mortar retailer should launch an Internet branch. We conclude that when consumers perceive that the Internet branch of the bricks-and-mortar retailer is not superior to the pure play e-tailer, the bricks-and-mortar retailer should not launch the Internet branch so as to avoid severe competition and maintain high profits. However, when consumers perceive the Internet branch of the bricks-and-mortar retailer to be superior to the pure play e-tailer, the bricks-and-mortar retailer will launch an Internet branch and charge a price higher than that at the pure play e-tailer.

3. DATA To empirically validate the result from the analytic model, we collected data from 905 e-tailers who sell items in eight product categories, including apparel, gifts and flowers, health and beauty, home and garden, sports equipment, computer hardware, consumer electronics, and office supplies. The data for this study are primarily drawn from a well-known price comparison web site, BizRate.com. Product, price, and deal information for a large number of e-tailers are searched and updated daily by BizRate.com. BizRate.com also surveys e-tailers’ customers and asks them to evaluate the e-tailers’ services. The survey results are published on BizRate.com’s web site, so we can use them to measure e-tailer service heterogeneity. Ten aspects of

Price Competition Between Pure Play Versus Bricks-and-Clicks e-Tailers

47

e-tailers’ services are evaluated using a ten-point scale and an overall measure of the average of the ten measures is also provided. The ratings of the retailers by Bizrate.com are recommended by Consumer Report as “unbiased” measures and are widely used in online markets. For example, shopper.com, shopping.com, and price.com, all cite BizRate.com’s ratings. In addition, many e-tailers who are BizRate.com’s certified Stores, also indicate this on their own websites (e.g. CircuitCity.com, Mercata.com, Motorola, CD Universe, Euclid Computers), which reflects the acceptance of BizRate.com’s authority. Thus, data from BizRate.com have high validity. To reduce the ten attributes (some of them highly correlated) of e-tailer services to a set of orthogonal e-tailer factors, we did a factor analysis of the ten variables. Four factors, namely, reliability in fulfillment, convenience, depth of information, and shipping and handling explained a dominant part of the variance (91%), so we retain these factor scores as measures of key e-tailer service quality factors for our empirical analysis.

4. EMPIRICAL MODEL AND ESTIMATION 4.1. Empirical Model To examine if the analytic results hold, we perform a cross-sectional empirical analysis of prices at pure play and bricks-and-clicks e-tailers. We first develop a two-equation model of price and web site traffic. The price equation is given by: PRICEij = a1j + a2jTRAFFICij + a3jRELij + a4jINFOij + a5jSHIPij + a6jCONVij + a7jETYPEij + a8jENTIMEij + a9jTRUSTij + ␧ij

(I)

where PRICE is perceived price of the e-tailer measured on a scale ranging from 1 to 10, TRAFFIC is the number of visits at the e-tailer’s web site, REL is the factor score of the e-tailer reliability, INFO is factor score of the depth of information available at the e-tailer, SHIP is the e-tailer’s factor score on shipping and handling, CONV is the factor score of shopping convenience at the e-tailer, ETYPE is a dummy variable representing e-tailer type – whether the e-tailer is a pure play e-tailer or a bricks-and-clicks e-tailer (1 if it is a pure play e-tailer, 0 otherwise), ENTIME is the time elapsed between the time of entry of the e-tailer and the entry time for the first e-tailer in the online market (consistent with Shankar, Carpenter & Krishnamurthi, 1999), TRUST is a composite variable representing the number of trust seals from BBB, Verisign, Truste, BizRate, and Gomez, ␧ is an error term, i is e-tailer, and j is product

48

XING PAN, VENKATESH SHANKAR AND BRIAN T. RATCHFORD

category (consistent with Pan, Ratchford & Shankar, 2001).9 In the absence of data on sales or transactions, we use traffic as the proxy for demand. The demand (traffic) equation is given by: TRAFFICij = ␤1j + ␤2jPRICEij + ␤3jRELij + ␤4jINFOij + ␤5jSHIPij + ␤6jCONVij + ␤7jNCATij + ␤8jENTIMEij + ␤9jTRUSTij + ␩ij

(II)

where NCAT is the number of product categories carried by the e-tailer, ␩ is an error term and the other terms are as defined before. The proxies for price and demand that we use are reasonable, given our research purpose. First, because our analysis is cross-sectional, the perceived price level measure allows us to compare prices across categories. Second, the price measure is a consumer level measure that better reflects the paid full prices rather than posted prices that may not be the same for all consumers. Third, the correlations between the perceived price levels and actual product prices were positive and significant, reinforcing the validity of consumer price perception data. Demand in terms of sales transactions or revenues by category are difficult to obtain given the confidential nature of such data for companies. E-tailer traffic is typically reflective of e-tailer sales or market share (Comscore Report 2002; Brynjolfsson and Smith 2000). Because we control for the number of categories sold by the e-tailer in our subsequent empirical analysis, traffic is a reasonable proxy for demand. Table 1 shows the operationalization of the model variables in the data. 4.2. Estimation We estimated the models by ordinary least squares (OLS), two-stage least squares (2SLS) and three-stage least squares (3SLS) methods for each product category as appropriate. We also performed a Hausman (1978) specification test of m-statistic to test for the endogeneity of price and traffic in the system of equations. In other words, we tested for the appropriateness of 2SLS and 3SLS estimations with respect to the OLS method. Because the price measure is standardized across categories (same scale) and most other variables are comparable across categories, we can also run a pooled model of price equation. For the traffic equation, we can estimate fixed effects for the product categories in the pooled model.

5. RESULTS The summary statistics of the model variables for each category appear in Table 2. The mean prices are quite comparable across the categories. The mean Web

Operationalization of Variables in the Data.

Variable

Definition and Scale

PRICE

Consumer perceived price level of the e-tailer on a scale of 1–10. The greater the number, the higher the price.

TRAFFIC

Number of visitors to the Web site.

REL

Reliability factor score.

CONV

Convenience factor score.

INFO

Information factor score.

SHIP

Shipping and handling factor score.

TRUST

Number of trust certificates that an e-tailer receives (Bizrate, BBB, Gomez, Truste, and VeriSign) ranging from 0 to 5.

ENTIME

Number of days elapsed between the e-tailer’s online entry into a category after the category online pioneer’s entry.

NCAT

Number of categories sold by the e-tailer.

ETYPE

Dummy variable for e-tailer type (pure play = 1, bricks-and-clicks = 0).

Price Competition Between Pure Play Versus Bricks-and-Clicks e-Tailers

Table 1.

49

50

Apparel N = 200 Variable NCAT PRICE TRAFFIC REL

Summary Statistics of the Data.

Gifts and Flowers N = 187

Health and Beauty N = 182

Home and Garden N = 200

Mean

Std

Min

Max

Mean

Std

Min

Max

Mean

Std

Min

Max

Mean

Std

Min

Max

2.41

1.49

1

8

2.30

1.46

1

8

2.46

1.63

1

8

2.34

1.42

1

8

1.71

0.53

0.5

3.1

1.67

0.50

0.5

2.8

1.57

0.54

0.6

3

1.49

0.51

0.6

5

36801 82230

3 649609

31277 108490

35 966977

61682 370882

5 4828880

36135 106720

3 966977

–1.19E-15

1 –3.87

2.37

2.35E-02

1 –3.41

1.84 –3.30E-15

1 –3.72

1.61 –5.12E-03

1 –3.02

1.99

CONV

2.49E-15

1 –3.31

2.78

3.67E-03

1 –3.74

1.93 –5.89E-15

1 –4.11

1.88 –5.58E-03

1 –7.19

1.96

INFO

–6.60E-15

1 –2.91

2.05 –1.75E-02

1 –4.05

2.83 –3.76E-15

1 –5.04

2.04

4.96E-03

1 –6.10

2.52

SHIP

–3.84E-15

1 –4.00

1 –2.34

2.77 –2.36E-15

1 –2.54

1 –2.58

2.13

2.59

8.92E-03

0.96

0

5

1.98

1693

674

0

3012

1789

0.55

0.50

0

1

0.63

TRUST

1.88

ENTIME ETYPE

1.01

0

5

1.73

638

0

2965

1791

0.49

0

1

0.65

2.50

1.35E-03

0

4

1.94

680

0

3117

1538

0.48

0

1

0.51

1.02

0.93

0

5

656

0

2718

0.50

0

1

XING PAN, VENKATESH SHANKAR AND BRIAN T. RATCHFORD

Table 2.

Sports and Outdoors N = 208 Variable

Mean

Std

NCAT

2.58

1.58

PRICE

1.62

0.55

TRAFFIC REL

69122 350204 –4.65E-03

Min

Continued

Computer Hardware N = 199 Max

Mean

Std

1

8

2.33

1.27

1

8

2.50

1.45

0.4

3.4

1.21

0.53

0.2

5

1.34

0.61

24044

65251

38287

89860

10 4828880

1 –4.36

1.51

3.82E-16

Min

Consumer Electronics N = 214 Max

5 581436

Mean

Std

Min

Office Supply N = 192 Max

Mean

Std

1

8

2.65

1.42

1

8

0.3

5.5

1.34

0.52

0.1

2.8

58303 359503

Min

Max

18

608574

1 –6.32

1.78

2.79E-02

1 –5.40

1.65

8.25E-16

1 –3.71

2 4828880 1.93

CONV

2.00E-03

1 –3.99

1.91

7.38E-15

1 –6.65

1.86 –1.26E-03

1 –7.32

1.74 –7.33E-15

1 –3.33

2.06

INFO

–3.48E-03

1 –2.99

2.06 –1.10E-14

1 –8.08

2.74

1 –8.05

2.85 –1.51E-15

1 –3.17

2.10

SHIP

–1.07E-03

1 –1.76

1 –2.91

2.78 –2.00E-02

1 –2.82

1 –2.56

2.38

1.97

1.86E-15

1.08

0

5

2.06

1595

677

0

2926

0.50

0.50

0

1

TRUST

1.86

ENTIME ETYPE

1.80E-04

2.60

9.09E-15

1.05

0

5

2.03

1868

678

0

3111

1812

0.65

0.48

0

1

0.71

1.01

0

5

2.10

1665

655

0

2842

0.77

0.42

0

1

0.96

0

5

707

0

3111

0.46

0

1

Price Competition Between Pure Play Versus Bricks-and-Clicks e-Tailers

Table 2.

51

52

XING PAN, VENKATESH SHANKAR AND BRIAN T. RATCHFORD

site traffic ranges from 24,044 for computer hardware to 69,122 for sports and outdoor products. The mean factor scores of e-tailer characteristics such as reliability, information, convenience and shipping are close to zero, as expected. The mean trust scores are also comparable across categories, between 1.73 and 2.10. The mean entry times of e-tailers in different categories range between 1538 to 1868 days. The average number of categories sold by e-tailers who sell a particular category is also consistent across categories. As one might expect, the proportion of pure play e-tailers is generally higher than that of bricks-and-clicks e-tailers ranging from 50% for sports and outdoors to 77% for computer hardware. The correlation matrixes of the model variables for each category showed that the correlations are generally not very high in any product category, so multicollinearity is not a problem in this data set. The results of t-tests of differences between mean prices at pure play and bricks-and-clicks e-tailers appear in Table 3. In all the categories, the average price at pure play e-tailers is lower than that at bricks-and-clicks e-tailers ( p < 0.10 and better). This result provides face validity to our analytical model Table 3.

T-tests of Equality of Prices between Bricks-and-Clicks and Pure Play E-tailers. Bricks-and-Clicks

Category Apparel Gifts & Flowers Health & Beauty Home & Garden Sports & Outdoors Computer Hardware Electronics Office Supply

Mean (Std. Dev.) 1.84 (0.48) 1.79 (0.48) 1.80 (0.51) 1.56 (0.46) 1.66 (0.56) 1.37 (0.52) 1.55 (0.70) 1.65 (0.52)

N 91 70 63 99 105 45 75 56

Pure Play Mean (Std. Dev.) 1.59 (0.55) 1.59 (0.50) 1.45 (0.52) 1.43 (0.56) 1.37 (0.51) 1.17 (0.53) 1.22 (0.52) 1.22 (0.46)

N

Difference in Mean t-value Significance

109

0.25

3.39

0.0080

117

0.19

2.58

0.0126

119

0.35

4.40

0.0001

101

0.13

1.81

0.0717

103

0.30

3.99

0.0001

154

0.20

2.25

0.0257

139

0.33

3.60

0.0005

136

0.43

5.61

0.0001

Price Competition Between Pure Play Versus Bricks-and-Clicks e-Tailers

53

conclusion. To examine differences between these two types of e-tailers after controlling for the effects of the other variables, we turn to the results of estimations of Eqs I and II. The results for the traffic and price equations are shown in Tables 4 and 5. Although we estimated the equations by OLS, 2SLS, and 3SLS methods, the Hausman (1978) specification test rejected the hypothesis of endogeneity of price and traffic, so we use the OLS estimates. First, we examine the results of the traffic equation. Because the models are cross-sectional, we do not expect high R2, so R2 is lowest (4%) for the Sports and Outdoor category and highest (21%) for the Home and Garden category. Price is not significant in most of the categories, so in general, e-tailer traffic does not depend on its perceived price levels. E-tailer factors such reliability, convenience, information and shipping are, in general, not significant drivers of traffic in most categories. Number of categories is positive and significant ( p < 0.05) in six out of eight categories, consistent with our expectation. That is, e-tailers who sell more product categories draw greater traffic to their Web sites. Time of entry is also negative and significant in seven out of eight categories ( p < 0.05). The later an e-tailer enters an online market, the smaller is its site traffic. Trust is significant for gifts and flowers and computer hardware categories ( p < 0.01), implying that Web sites in these categories with more trust certificates attract more traffic to their sites. We now present the results of the price equation. The R2 values in the price equation are much higher than those in the traffic equation, ranging from 31% for apparel to 48% for computer hardware. Traffic is insignficant in all categories except in apparel in which, greater traffic is associated with lower perceived price levels, consistent with conventional price-demand relationship. Time of entry is significant in only two categories, electronics and office supply ( p < 0.10). The earlier the e-tailer enters online in these categories, the higher is its perceived price level. Trust is not a significant driver of perceived price levels in any category. Reliability is associated with premium prices for computer harware and electronics, but negatively related to prices for health and beauty aids. Surprisingly, the other e-tail characteristic factors, namely, information, convenience and shipping are negative and significant in each category ( p < 0.10). These results suggest that e-tailers who offer deeper information, greater convenience and superior shipping and handling service, have lower perceived price levels. Finally, the perceived price levels at a bricksand-clicks stores are higher than those at a pure play stores in all categories except home and garden and computer hardware, in which the prices are not significantly different. This is our central finding in the empirical analysis.

54

Table 4.

Adjusted R2 N Intercept Price Reliability Information Shipping Convenience Number of categories Time of entry Trust

a

Apparel

Gifts and Flowers

Health and Beauty

Home and Garden

Sports and Outdoor

Computer Hardware

Electronics

Office Supply

19% 200 82106a (30160) –6607.08 (12270) 2604.00 (5413.56) –12765b (5261.54) –10278c (5990.84) –2373.38 (5864.63) 15366a (3998.85) –38.15a (8.08) –3409.48 (5909.03)

8% 187 13263 (48996) 29424 (20728) –4478.47 (7992.52) –156.29 (8061.53) 14228 (9517.20) –2684.42 (8458.99) 4722.68 (6024.72) –39.44a (12.61) 14463c (8448.24)

4% 182 146024 (152588) 10505 (65380) 16740 (27725) 15991 (29118) 14630 (30311) 29997 (32433) 54295a (20233) –100.87b (41.71) –31092 (30538)

21% 200 62548c (35890) 1940.49 (16834) –17442b (6820.61) –16778b (7417.03) 6932.75 (7574.45) –3299.07 (7938.03) 21531a (5515.43) –43.40a (10.56) –6658.5 (7874.6)

4% 208 136653 (130510) 11082 (53437) 13984 (25094) 1442.84 (26779) 30123 (38578) 7395.62 (26269) 49406a (17497) –105.26a (36.95) –24205 (25695)

12% 199 1751.81 (21954) 15989 (11420) –3380.2 (4465.42) 2891.4 (4458.40) 1387.77 (4401.77) –14709b (5925.23) 1362.87 (3701.96) –10.17 (6.95) 8120.27c (4605.04)

15% 214 46008 (30617) 20477c (12213) –3127.13 (6338.87) 5854.78 (5851.65) –2927.16 (5869.98) –9870.21 (7025.20) 10349b (4265.16) –28.91a (9.22) –3365.74 (5944.50)

5% 192 119846 (137164) 31008 (62373) 10990 (25730) 30942 (26171) 22367 (26793) 35207 (30679) 51101b (20484) –104.84a (38.28) –23983 (29965)

p < 0.01; b p < 0.05; c p < 0.10. Standard errors in parentheses. Numbers in bold represent significant estimates.

XING PAN, VENKATESH SHANKAR AND BRIAN T. RATCHFORD

Item

Results of the Traffic Equation.

Item Adjusted R2 N Intercept Traffic Reliability Information Shipping Convenience e-Tailer Type Time of entry Trust

Results of the Price Equation.

Apparel

Gifts and Flowers

Health and Beauty

Home and Garden

Sports and Outdoor

Computer Hardware

Electronics

Office Supply

31% 200 1.88a (0.12) –7.37E-07c (4.20E-07) –0.04 (0.03) –0.05c (0.03) –0.22a (0.03) –0.18a (0.03) –0.14b (0.07) –2E-05 (5.1E-05) –0.02 (0.03)

45% 187 1.83a (0.11) 4.00E-07 (2.67E-07) –0.01 (0.03) –0.14a (0.03) –0.25a (0.03) –0.15a (0.03) –0.12b (0.06) –3.9E-05 (4.67E-05) –0.02 (0.03)

42% 182 1.79a (0.11) 1.83E-08 (8.46E-08) –0.10a (0.03) –0.07b (0.03) –0.16a (0.03) –0.24a (0.03) –0.21a (0.07) –4.5E-05 (4.87E-05) –0.005 (0.03)

37% 200 1.56a (0.10) –1.54E-07 (3.02E-07) –0.02 (0.03) –0.11a (0.03) –0.18a (0.03) –0.23a (0.03) –0.08 (0.06) –9.6E-06 (4.93E-05) –0.01 (0.03)

35% 208 1.90a (0.11) 1.04E-08 (9.05E-08) –0.05 (0.03) –0.19a (0.03) –0.24a (0.05) –0.11a (0.03) –0.20a (0.07) –7E-05 (4.92E-05) –0.04 (0.03)

48% 199 1.32a (0.11) 5.64E-07 (4.57E-07) 0.08a (0.03) –0.07a (0.03) –0.05b (0.03) –0.33a (0.03) –0.11 (0.07) –4.4E-05 (4.51E-05) 0.02 (0.03)

42% 214 1.65a (0.13) 5.23E-07 (3.87E-07) 0.12a (0.03) –0.06c (0.03) –0.08b (0.03) –0.29a (0.03) –0.19a (0.07) –1.27E-04b (5.3E-05) 0.01 (0.03)

38% 192 1.62a (0.12) 7.16E-08 (8.43E-08) 0.03 (0.03) 0.08a (0.03) –0.12a (0.03) –0.22a (0.03) –0.21a (0.07) –7.6E05c (4.53E05) 0.002 (0.03)

Price Competition Between Pure Play Versus Bricks-and-Clicks e-Tailers

Table 5.

a

p < 0.01; b p < 0.05; c p < 0.10. Standard errors in parentheses. Numbers in bold represent significant estimates.

55

56

XING PAN, VENKATESH SHANKAR AND BRIAN T. RATCHFORD

Because the perceived price level measure reflects the perceived value in a sense, a reanalysis of the data using an adjusted price may be appropriate. Adjusted price is price adjusted for the possible correlation of other e-tailer factors such as reliability and information with perceived price level. To check the robustness of the results, we performed the following analysis with adjusted price. We regressed price on other e-tailer factors and used the residuals as adjusted prices in Eq. I (traffic equation). Since Eq. II (price equation) already has e-tailer factors, we omitted these variables when we used adjusted price as the dependent variable in this equation. The results were consistent with those from the original analysis – the signs of the coefficients in the traffic and price equations do not change. We do not report these findings to save space, but they reinforce our results from the original model. Because the category level regression coefficients are not that different, we can possibly parsimoniously estimate the coefficients through a pooled model. We tested for pooling through the Chow (1960) test. The test did not reject homogeneity of coefficients ( p < 0.05). The results of the pooled model appear in Table 6. These results are consistent with those results by category obtained in Tables 4 and 5. The central finding that price levels at pure play e-tailers are, on average, lower than those at bricks-and-clicks e-tailers also shows up in this table. In summary, the empirical results show that prices at pure play e-tailers are generally lower than prices at bricks and clicks e-tailers. E-tailers with high traffic do not necessarily command higher prices. E-tailers with high level of reliability, shopping convenience, and deep information, do not seem to draw high web traffic or enjoy high prices. But e-tailers with more trust certifications attract more traffic and in some categories, higher price levels. Early movers in online markets also enjoy greater traffic and higher prices in certain categories.

6. DISCUSSION, IMPLICATIONS, LIMITATIONS AND CONCLUSION Our results are consistent with prior research and extend it. Our analytic results extend the analyses of Balasubramanian (1998) (Druehl & Porteus, 2001) on differences between direct mail (pure play Internet) and bricks-and-mortar e-tailers and show that prices are lower at pure play e-tailers than they are at bricks-and-clicks e-tailers. The empirical results support Ancarani and Shankar (2002) and Tang and Xing (2001). While these studies studied only one category each, namely, books or DVDs, our results span eight categories. Pan, Ratchford and Shankar (2002) who analyzed eight categories, seven of which

Price Competition Between Pure Play Versus Bricks-and-Clicks e-Tailers

Table 6. Parameters

Pooled Estimation of Price and Traffic Equations. Traffic

Price

5%

32%

Adjusted R-square Intercept PRICE

81083.01** (33859.2) 15261.8 (12755.2)

TRAFFIC Reliability Convenience Information Shipping Time of Entry Trust Number of Categories

2481.35 (5755.8) 6022.27 (6522.4) 2439.11 (5745.4) 6966.90 (6234) –58.05*** (8.77) –7479.65 (6178.9) 27047.53*** (4278.5)

e-tailer Type Apparel Computer Electronics Gifts and Flowers Health and Beauty Home and Garden Office Supply

57

1.75*** (0.04)

6.46E-08 (5.16E-08) –1.26E-03 (0.01) –0.22*** (0.01) –0.08*** (0.01) –0.15*** (0.01) –7.00E-05 (1.80E-05) –0.01 (0.01)

–0.19*** (0.03) –23070.4 (22226.7) –26430.3 (22877) –6785.54 (22244.1) –18955.4 (22692.3) 6986.91 (22790.9) –27116.2 (22276.7) 5387.89 (22741.8)

Sports and Outdoors is the base category. Number of observations = 1582. Standard errors in parentheses. Numbers in bold represent significant estimates. * p < 0.10, ** p < 0.05, *** p < 0.01.

58

XING PAN, VENKATESH SHANKAR AND BRIAN T. RATCHFORD

were different from the categories we analyzed, found that in general, prices at pure play e-tailers are equal to or lower than those at brick-and-clicks e-tailers. In the electronics category that was common to our study, they found that prices are not significantly different across the two types of e-tailers, although the coefficient of pure play e-tailer dummy was negative in sign. Therefore, even in case of electronics, our results are not inconsistent with those of Pan, Ratchford and Shankar (2002). Thus, our conclusion that prices at pure play e-tailers are lower than those at brick-and-clicks e-tailers seems to be fairly robust. Our results have important implications. First, the existence of an equilibrium in which a pure play e-tailer charges a higher price than that by a bricks-and-clicks e-tailer suggests that bricks-and-clicks e-tailers can compete on other differentiating attributes such as brand equity and physical location. By the same token, a pure play e-tailer can also compete effectively through a low price strategy if its costs are lower than those of the bricks-and-clicks e-tailer. Second, marketing efforts aimed at generating traffic may not pay off in terms of increasing the prices that consumers are willing to pay. Third, improving e-tailer attributes such as reliability of service, navigation, information availability and interactivity may not yield high traffic and may not be associated with higher prices. However, improving trust and entering online markets early might result in greater traffic and possibly higher prices. Our research has some limitations that offer opportunities for future research. First, our analytic model is based on Hotelling’s linear city framework and is limited by its assumptions. Relaxing some of the assumptions can make the model richer, albeit more complex. Second, our empirical analysis is cross-sectional and does not capture the dynamics of price competition. A longitudinal or panel data analysis will offer richer modeling context and deeper insights. Third, due to data limitations, we used traffic as a proxy for demand and perceived price level as a proxy for price. The results from the empirical analysis were illustrative, but analysis of actual sales transactions and revenues and prices will form a stronger test of our hypotheses. The results, in particular, those on the relationships of e-tailer factors with traffic and price should be viewed with caution. In conclusion, we show both theoretically and empirically that prices at a pure play e-tailer are generally lower than those at a bricks-and-clicks e-tailer. We also find that e-tailers with high traffic may not command higher prices and that superior e-tailer attributes may not allow e-tailers to draw more traffic or enjoy higher prices. However, trust and early online entry can provide lifts in traffic and opportunities for premium pricing. These findings can help e-tail managers make better pricing decisions and compete smarter on the Internet.

Price Competition Between Pure Play Versus Bricks-and-Clicks e-Tailers

59

NOTES 1. The Hotelling model has also been analyzed in a circular market (e.g. Balasubramanian, 1998). Economides (1993) points out that “the simplicity and symmetry of the circular model cannot sufficiently compensate for its lack of the appropriate structure” and “for most goods, however, it is appropriate to use a line interval as the space of potential products.” 2. An Internet retailer may also accept orders by phone, fax, mail and other ways. In this study, we define a pure play Internet store as an Internet retailer who has no physical retail presence. 3. We discuss this issue subsequently in the paper. 4. In reality, the two stores may actually have different marginal costs due to different logistics structures. However, by assuming the same marginal cost, we can better focus on the competition due to demand side factors. 5. Similar results can be derived for a Stackelberg leader-follower game. When the bricks-and-mortar store is the leader, it will have a higher price if b > (a ⫺ 1/3) and when it is the follower, it will have a higher price if b > (a + 1/3). 6. This is because opening an Internet store is less expensive than opening a bricksand-mortar store. Brynjolfsson and Smith (2000) point out that opening an electronic storefront consists of creating web pages on the Internet (which is easy to build through services such as Yahoo! shopping and MSN shopping), so online market entry is easier than offline market entry. 7. Note that in the equations that follow, P1 = P3. 8. The bricks-and-clicks retailer can always be the leader since it can first price at the marginal cost to drive the pure play e-tailer out of market and then maintain a higher price until the pure play e-tailer can come back. 9. We also tried to include a variable to capture the effect of the number of links to external web sites, but this variable was extremely highly correlated to the traffic variable (over 0.95 for some categories), precluding its inclusion as an independent variable in a meaningful way.

ACKNOWLEDGMENT Xing Pan thanks the Economics Club of Washington for doctoral research fellowship award.

REFERENCES Ancarani, F., & Shankar, V. (2002). Price Levels and Price Dispersion: A Comparison of Pure Play vs. Bricks-and-Mortar vs. Bricks-and-Clicks e-Tailers. Working Paper, SDA Bocconi, Milan, Italy. Bailey, J. P. (1998). Electronic Commerce: Prices and Consumer Issues for Three Products: Books, Compact Discs, and Software. Organization for Economic Co-Operation and Development, OCDE/GD 98, 4. Balasubramanian, S. (1998). Mail vs. Mall: A Strategic Analysis of Competition between Direct Marketers and Conventional Retailers. Marketing Science, 17(3), 181–195.

60

XING PAN, VENKATESH SHANKAR AND BRIAN T. RATCHFORD

Bakos, Y. (1997). Reducing Buyer Search Costs: Implications for Electronic Marketplaces. Management Science, 43(12), 1676–1692. Baye, M. R., & Morgan, J. (2001a). Information Gatekeepers on the Internet and the Competitiveness of Homogeneous Product Markets. American Economic Review, 91(3), 454–474. Baye, M. R., & Morgan, J. (2001b). Price Dispersion in the Lab and on the Internet: Theory and Evidence. Working Paper, Indiana University. Baye, M. R., & Morgan, J. (2002). Information Gatekeepers and Price Discrimination on the Internet. Economics Letters, 76(1), 47–51. Brown, J. R., & Goolsbee, A. (2002). Does the Internet Make Markets More Competitive? Evidence from the Life Insurance Industry. The Journal of Political Economy, 110(3), 481–508. Brynjolfsson, E., & Smith, M. (2000). Frictionless Commerce? A Comparison of Internet and Conventional Retailers. Management Science, 46(4), 563–585. Carlson, J., & McAfee, P. (1983). Discrete Equilibrium Price Dispersion. Journal of Political Economy, 91(3), 480–493. Chow, G. C. (1960). Tests of Equality Between Sets of Coefficients in Two Linear Regressions. Econometrica, 28, 591–605. Clay, K., Krishnan, R., Wolff, E., & Fernandes, D. (1999). Retail Strategies on the Web: Price and Non-price Competition in the Online Book Industry. Working Paper, Carnegie-Mellon University. Clemons, E., Hann, I., & Hitt, L. (2002). Price Dispersion and Differentiation in Online Travel: An Empirical Investigation. Management Science, 48(4), 534–549. Cohen, M. (1998). Linking Price Dispersion to Product Differentiation – Incorporate Aspects of Customer Involvement. Applied Economics, 30, 829–835. Comscore Report (2002). Online Retail Activity. Reston, VA. Dahlby, B., & West, D. (1986). Price Dispersion in an Automobile Insurance Market. Journal of Political Economy, 94(2), 418–438. Dana, J. (1999). Equilibrium Price Dispersion Under Demand Uncertainty: The Role of Costly Capacity and Market Structure. Rand Journal of Economics, 30(4), 632–660. Druehl, C., & Porteus, E. (2001). Price Competition between an Internet Firm and a Bricks and Mortar Firm. Working Paper, Stanford University. Economides, N. (1993). Hotelling’s ‘Main Street’ with More Than Two Competitors. Journal of Regional Science, 33(3), 303–319. Ellison, G., & Ellison, S. (2001). Search, Obfuscation, and Price Elasticities on the Internet. Working Paper, Sloan School of Management, MIT. Erevelles, S., Rolland, E., & Srinivasan, S. (2001). Are Prices Really Lower on the Internet?: An Analysis of the Vitamin Industry. Working Paper, University of California, Riverside. Goolsbee, A. (2001). Competition in the Computer Industry: Online vs. Retail. Journal of Industrial Economics, 49(4), 487–500. Hausman, J. A. (1978). Specification Tests in Econometrics. Econometrica, 46 (November), 1251–1272. Hotelling, H. (1929). Stability in Competition. Economic Journal, 39 (March), 41–57. Kujala, J., & Johnson, M. (1993). Price Knowledge and Search Behavior for Habitual, Low Involvement Food Purchase. Journal of Economic Psychology, 14(2), 249–265. Lal, R., & Sarvary, M. (1999). When and How is the Internet Likely to Decrease Price Competition? Marketing Science, 18(4), 485–503.

Price Competition Between Pure Play Versus Bricks-and-Clicks e-Tailers

61

Messmer, E. (2001). Online Retailers Plan New Projects for Coming Year. Network World, 18(4), 10–11. Mitchell, W., & Sorensen, R. (1986). Pricing, Price Dispersion, and Information: The Discount Brokerage Industry. Journal of Economics and Business, 38(4), 273–282. Morgan, J., Orzen, H., & Sefton, M. (2001). An Experimental Study of Price Dispersion. Working Paper, Princeton University. Morton, F. S., Zettelmeyer, F., & Risso, J. S. (2001). Internet Car Retailing. Journal of Industrial Economics, 49(4), 501–519. Pan, X., Ratchford, B. T., & Shankar, V. (2001). Why Aren’t the Prices of the Same Item the Same at Me.com and You.com?: Drivers of Price Dispersion Among E-Tailers. Working Paper, University of Maryland, College Park, MD 20742. Pan, X., Ratchford, B. T., & Shankar, V. (2002). Can Price Dispersion in Online Markets be Explained by Differences in e-Tailer Service Quality? Journal of Academy of Marketing Science, 30(4), 429–441. Pratt, J., Wise, D., & Zeckhauser, R. (1979). Price Differences in Almost Competitive Markets. Quarterly Journal of Economics, 93 (May), 189–211. Rao, V. (1984). Pricing Research in Marketing: The State of Art. Journal of Business, 57(1), S39–S60. Ratchford, B. T., Agrawal, J., Grimm, P., & Srinivasan, N. (1996). Toward Understanding the Measurement of Market Efficiency. Journal of Public Policy and Marketing 15, 167–184. Ratchford, B. T., Pan, X., & Shankar, V. (2002). On the Efficiency of Electronic Markets for Consumer Goods. Working Paper, University of Maryland, College Park, MD 20742. Salop, S., & Stiglitz, J. (1982). The Theory of Sales: A Simple Model of Equilibrium Price Dispersion with Identical Agents. The American Economic Review, 72(December), 1121–1130. Shankar, V., Carpenter, G. S., & Krishnamurthi, L. (1999). Advantages of Entering in the Early Growth Stage: An Empirical Analysis. Journal of Marketing Research, 36 (May), 269–276. Shankar, V., Rangaswamy, A., & Pusateri, M. (2001). The Online Medium and Customer Price Sensitivity. Working Paper, University of Maryland, College Park, MD 20742. Smith, M., Bailey, J., & Brynjolfsson, E. (2000). Understanding Digital Markets: Review and Assessment. In: E. Brynjolfsson & B. Kahin (Eds), Understanding the Digital Economy. Cambridge, MA: MIT Press. Smith, M., & Brynjolfsson, E. (2001). Consumer Decision-Making at an Internet Shopbot. Quarterly Journal of Economics, forthcoming. Sorensen, A. (2000). Equilibrium Price Dispersion in Retail Markets for Prescription Drugs. Journal of Political Economy, 108(4), 833–850. Stigler, G. (1961). The Economics of Information. Journal of Political Economy, 69(3), 213–225. Tang, F., & Xing, X. (2001). Will the Growth of Multi-Channel Retailing Diminish the Pricing Efficiency of the Web? Journal of Retailing, 77, 319–333. Zettelmeyer, F. (2000). Expanding to the Internet: Pricing and Communication Strategies When Firms Compete on Multiple Channels. Journal of Marketing Research, 37(3), 292–308.

This Page Intentionally Left Blank

PRICE DISPERSION THEN AND NOW: EVIDENCE FROM RETAIL AND E-TAIL MARKETS Patrick Scholten and S. Adam Smith ABSTRACT This paper uses two datasets to examine price dispersion spanning a 24-year period. The first dataset permits us to compare levels of retail price dispersion in 1976 and 2000, while the second allows for a comparison of retail dispersion in 1976 with dispersion in e-tail markets in 2000. Our results indicate that price dispersion in 2000 for both retail and e-tail markets is comparable to that observed in 1976 retail markets. This suggests that, for the products in our sample, the Information Age has done little to reduce price dispersion in retail or e-tail markets.

1. INTRODUCTION Since Nobel Laureate George Stigler’s 1961 seminal article noted the ubiquity of price dispersion, it has become one of the most widely replicated findings in the economics literature. Indeed, many empirical studies have focused on documenting the presence of or identifying the causes of price dispersion in homogeneous product markets. Yet, no study has explicitly compared levels of price dispersion over a 24-year time horizon. This is the primary purpose of this paper. We compare the results from a previous study documenting price dispersion in retail markets in 1976 with a dataset of similar products collected from retailers in 2000. The Economics of the Internet and E-Commerce, Volume 11, pages 63–88. Copyright © 2002 by Elsevier Science Ltd. All rights of reproduction in any form reserved. ISBN: 0-7623-0971-7

63

64

PATRICK SCHOLTEN AND S. ADAM SMITH

While non-existent in 1976, electronic retail markets – henceforth, e-tail markets – are serving an increasingly important role in shaping how consumers make purchases. Given e-tail markets’ growing stature, we also compare price dispersion in 1976 with the dispersion observed in the e-tail markets of 2000. To describe relative price dispersion in retail and e-tail markets over this 24-year period, we assemble two datasets that are described in greater detail in Section 2: the first dataset is used to compare the level of retail price dispersion in 1976 with that in 2000, while the second is used to compare price dispersion at the retail level in 1976 with that in e-tail markets in 2000. Our general finding is that price dispersion in both retail markets and e-tail markets is at least as large in 2000 as it was in 1976. Stigler’s 1961 seminal article stimulated an important theoretical literature on equilibrium price dispersion. The rationales offered in this literature vary. One strand shows that equilibrium price dispersion arises when it is costly for consumers to observe individual prices (e.g. Reinganum, 1979; Burdett & Judd, 1983; Gatti, 2000). In these models, consumers weigh the cost of obtaining additional price information with the expected benefits. An alternative approach assumes that some consumers can view the entire distribution of prices from a clearinghouse. The leading theoretical justifications driving the dispersion in these models stems from asymmetries among consumers, or that it is costly to list or view prices posted at the clearinghouse.1 Identical firms, in these models, sell to two types of consumers: those who consult the clearinghouse to obtain the entire distribution of prices, and those who do not. Both approaches illustrate the importance that information plays in consumers’ optimal purchasing decisions. Price dispersion has been empirically documented in an array of seemingly homogeneous product markets. Early studies using descriptive statistics to document price dispersion for a variety of products are found in Pratt, Wise and Zeckhauser (1979) and Carlson and Pescatrice (1980) – henceforth PWZ and CP. These and other empirical studies use a variety of statistics to measure price dispersion. For instance, PWZ use the price range and find considerable dispersion in 39 retail product markets in the Boston area. They find price ranges (in 1979 dollars) between $0.04 to more than $400.2 In contrast, CP use the coefficient of variation (␴/␮) to measure price dispersion for 34 retail products in the downtown central business district in New Orleans and the surrounding neighborhoods. They report coefficients of variation ranging from 3.27% to 41.4%. While the estimates of price dispersion reported in PWZ and CP are not directly comparable and vary considerably across products, these two studies illustrate that price dispersion in the 1970s and 1980s was a pervasive phenomenon.

Price Dispersion Then and Now

65

In the present paper, we use the coefficient of variation to summarize the levels of price dispersion observed in both e-tail and retail markets in 2000.3 Our rationale for using this measure is two-fold. First, we wish to compare price dispersion in our datasets with that reported in an earlier study. Since the coefficient of variation is invariant to multiplicative changes in prices, like inflation, this methodology permits us to compare the dispersion in our data for 2000 with that reported for 1976 by CP. Second, the coefficient of variation can be meaningfully compared across products to test, among other things, the Stigler Hypothesis which is discussed in detail below. The Information Age has renewed researchers’ interest in price dispersion, as many in the popular press speculate that the Internet will lead to more competitive retail markets and e-tail markets. As a result, a host of academic papers have emerged documenting the levels of price dispersion observed in a wide array of e-tail markets, and a few papers that explicitly compare dispersion between retailers and e-tailers. Empirical estimates of price dispersion among e-tailers vary widely. Brynjolfsson and Smith (2000) were among the first to document price dispersion in e-tail markets. The average price range for books in their study is about 33%, while that for CDs is slightly lower than 25%. Moreover, Brynjolfsson and Smith (2000) find that, compared to retailers, price dispersion among e-tailers selling books is larger. However, they also find that dispersion among e-tailers selling CDs is comparable to the levels observed in retail markets. Thus, their results suggest that the Internet, at least to date, has not led to lower price dispersion. For another study comparing prices in retail and e-tail markets, see the chapter by Pan, Shankar and Ratchford (2002) in this volume. In contrast to Brynjolfsson and Smith (2000), Ellison and Ellison (2001) provide evidence of significant price competition among e-tailers selling computer memory modules. Despite e-tailers’ best obfuscation strategies, these researchers find an average price range of $4.33 among the 12 lowest-prices; a narrow price distribution for products that sell for over $100. The findings in Brynjolfsson and Smith (2000) and Ellison and Ellison (2001) are representative of the results that examine price dispersion in e-tail markets.4 The primary focus of the present paper is on whether price dispersion in 2000 is greater or less than it was in 1976. The results from our first dataset suggest that price dispersion in retail markets was 18% in 2000, compared with only 13% in 1976. Results from our second dataset reveal that prices are actually more dispersed in the e-tail markets of today than the retail markets of the 1970s. More specifically, we find price dispersion of 14.5% for e-tail markets in 2000, which is slightly higher than the 12% dispersion observed in

66

PATRICK SCHOLTEN AND S. ADAM SMITH

retail markets during 1976. Thus, our primary finding is that price dispersion is at least as large now as it was then. The remainder of this chapter is organized as follows. Section 2 discusses the methodology used to collect the two datasets and provides a description of the datasets. The main results of the paper are presented in Section 3, which presents a descriptive analysis of intertemporal price dispersion in retail markets and e-tail markets. Section 4 discusses the results and provides additional insights about how price dispersion differs between e-tail markets and retail markets in 2000. Section 5 concludes.

2. METHODOLOGY AND DATA Our datasets are motivated by CP, who analyze price dispersion for each of 34 products collected in the downtown central business district in New Orleans during the fall of 1976. To ensure that products were differentiated only by seller and location, they selected a particular brand and size of each product. We adopt the same methodology. In the following subsections, we describe the datasets used to compare 2000 levels of price dispersion with that observed by CP in 1976. 2.1. Retail Markets Our retail dataset consists of 136 price observations on 20 retail products. Price data were collected by visiting several retail outlets in Bloomington, Indiana between February 1, 2000 and February 4, 2000. Limiting our data collection to a four-day span reduces the likelihood of introducing systematic variations that may occur over time. Table 1 provides summary statistics for the retail products common to this study and the CP study. At first blush, it appears that the average nominal price of these 20 items declined over the past 24 years, from $30.95 to $23.81. However, even a casual look at Table 1 reveals that this is driven almost entirely by changes in camera technology over the past 24 years. In particular, comparing these 20 products in 1976 and 2000, the two products most likely to be differentiated over time – expensive and inexpensive cameras – account for the dramatic intertemporal price decline. Excluding these two products from our data and the CP data, the average price of the items actually increased from $2.44 in 1976 to $6.25 in 2000 – roughly a 4% annual inflation rate for these items over the 24 years. The non-inflation adjusted prices make meaningful intertemporal price comparisons difficult. Therefore, Table 1 also shows the relative position of

Summary Statistics for Matched 2000 Retail and 1976 Retail Data. 2000 Retail

Products lbs. Potatoes** Lettuce Stalk Celery Tea Deodorant Dozen Lemons Batteries Hair Spray Antacids Auto Polish Film Aspirin Hand Cream Razor Blades Male Contraceptives Contact Lens Solution Thermometer Inexpensive Camera Dozen Roses Expensive Camera

1976 Retail

Number of Firms

Average Price*

Standard Deviation

Position Rank

Number of Firms

Average Price*

Standard Deviation

Position Rank

4 4 4 9 7 4 7 6 7 6 8 7 7 7 13 8 6 7 12 3

$0.36 $1.02 $1.22 $2.38 $2.53 $2.92 $3.08 $3.24 $4.06 $4.18 $4.25 $5.43 $5.68 $6.58 $6.74 $7.22 $8.87 $10.48 $42.65 $353.30

0.03 0.21 0.09 0.26 0.40 0.37 0.34 0.44 0.59 1.67 0.83 0.72 0.71 0.75 2.51 0.71 1.79 3.04 15.48 5.77

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20

9 9 8 10 9 7 8 9 7 11 9 9 8 8 7 8 9 10 11 7

$0.29 $0.37 $0.37 $0.44 $1.36 $1.06 $0.16 $2.47 $2.44 $1.33 $1.54 $1.30 $1.60 $2.46 $1.13 $2.34 $1.44 $34.70 $21.91 $540.20

0.04 0.07 0.07 0.11 0.10 0.22 0.41 0.11 0.14 0.18 0.09 0.08 0.09 0.18 0.47 0.17 0.30 2.59 1.41 17.68

2 3 3 5 10 6 1 17 15 9 12 8 13 16 7 14 11 19 18 20

Price Dispersion Then and Now

Table 1.

67

68

Table 1.

Continued.

2000 Retail

Average (all products) Average (excluding inexpensive and expensive cameras)

Number of Firms

Average Price*

Standard Deviation

6.80

$23.81

6.88

$6.25

Position Rank

Number of Firms

Average Price*

Standard Deviation

1.84

8.65

$30.95

1.22

0.91

8.59

$2.44

0.32

* Prices are not adjusted for inflation. ** 2000 price data available in 10 lbs bags only. To obtain the per-pound price, we simply divided by 10.

Position Rank

PATRICK SCHOLTEN AND S. ADAM SMITH

Products

1976 Retail

Price Dispersion Then and Now

69

price in the list of products. In 1976, batteries were the least expensive product (position rank = 1) and by 2000 became the seventh most expensive product in the retail sample (position rank = 7). In contrast, expensive cameras remained the most expensive product in both samples. The position rank of four products in Table 1 remained unchanged between 1976 and 2000. The position rank of seven products increased and the remaining nine products experienced position rank declines. Of the 16 products experiencing a position rank change, nine changed positions by only one or two positions. Interestingly, the product experiencing the most dramatic position rank decline was hair spray,5 while the position rank of male contraceptives increased the most.6 Despite a few large changes, Table 1 illustrates that the position rank of most products in our sample changed very little over 24 years. While the original CP sample consists of 34 products, we chose to limit our retail sample to the 20 products listed in Table 1. There are two primary reasons for this decision. First, our empirical study is broader in scope than the CP study, as we estimate relative dispersion measures for both retail markets and e-tail markets. Limiting the number of products to 20 significantly reduced the cost of acquiring data. Second, and most importantly, some products that were popular at the time of the CP study were not generally available at retail outlets in 2000. For instance, black and white televisions were popular in the 1970s and 1980s, but by 2000 were virtually obsolete. Similarly, by the year 2000 disposable cameras and relatively inexpensive cameras with built-in flashes eliminated the need for disposable flash bulbs (and the cameras that required them). While the composition of products between the CP study and our study differ, we made several attempts to maintain the salient features of the products in their study. In particular, their study consists of products that vary with respect to consumer search costs. They conjectured that products like batteries, thermometers, and male contraceptives would be purchased by two types of consumers: those with an urgent need and consumers who anticipate their needs. The cost of an additional search to urgent-need consumers will likely be substantial. Consumers anticipating their needs, on the other hand, might be more willing to engage in additional searches to find lower prices. CP speculated that products with these characteristics would exhibit substantial dispersion among prices. In contrast, they hypothesized that products repeatedly purchased typically without great necessity – like deodorant, hand cream, and aspirin – would lead to a relatively narrow distribution of search costs and hence, prices. Their sample also includes both relatively expensive and inexpensive products. Products from the CP study exhibiting these

70

PATRICK SCHOLTEN AND S. ADAM SMITH

characteristics were also included in the retail (and e-tail) dataset assembled for this study. The position rank in the price list for most of the retail products in our dataset change very little over time. The products included in our sample retain the essential features of the CP sample and permit us to offer some insight about how price dispersion has changed over 24 years for both products that, a priori, might be expected to exhibit large price dispersion and for those expected not to exhibit much dispersion. A discussion of the methodology and data in e-tail markets follows in the next subsection. 2.2. E-tail Markets We assembled a second dataset to compare price dispersion in e-tail markets in 2000 with the dispersion observed in 1976 retail markets. The e-tail dataset consists of 70 price observations for 11 products. In addition to reasons noted above, the relative immaturity of some e-tail markets forced us to analyze a smaller set of products than CP examined in their study of 1976 retail markets. Data collection occurred concurrently with retail markets – February 1, 2000 to February 4, 2000. One advantage of acquiring price data from e-tail markets is that price quotes from several e-tailers can be obtained in a single search by using a price comparison service or shotbot. We adopted this strategy. Specifically, we utilized the services of mySimon.com to obtain the initial price quotes, and crosschecked a random sample of these prices with the e-tailers’ site. In every case, the price quoted by mySimon.com corresponded with the price quoted on the e-tailers’ site. Summary statistics of the 11 products common to our data on 2000 e-tail markets and the CP data on 1976 retail markets are provided in Table 2. Excluding cameras, the average nominal price of these products increased from $3.63 in 1976 retail markets to $9.59 in 2000 e-tail markets. The position rank of the price of hair spray experienced the largest decline. In 1976, hair spray was the ninth most expensive product in the sample and became the second least expensive product by 2000. Similarly, the position rank of male contraceptives increased the most over 24 years. In 1976, male contraceptives were the second most expensive product and increased to the seventh most expensive by 2000. For four of the 11 products in our sample the position rank in the price list did not change. The price rank for three of the 11 products declined, while the remaining four products experienced position rank increases. In contrast to our retail dataset, the magnitudes of the position rank

Summary Statistics for Matched 2000 E-tail and 1976 Retail Data. 2000 E-tail

Products Deodorant Hair Spray Batteries Antacids Aspirin Hand Cream Male Contraceptives Razor Blades Thermometer Dozen Roses Expensive Camera Average (all products) Average (excluding cameras)

Number of Firms 5 3 2 4 4 4 13 5 4 14 12 6.36 5.80

1976 Retail

Average Price*

Standard Deviation

Position Rank

$2.74 $2.86 $3.34 $4.44 $5.07 $5.85 $6.58 $6.61 $8.74 $9.68 $385.67

0.29 0.23 0.92 0.42 0.94 0.73 1.44 0.65 0.50 14.11 25.16

1 2 3 4 5 6 7 8 9 10 11

$43.78 $9.59

4.13 2.02

Number of Firms 9 9 8 7 9 8 7 8 9 11 7 8.36 8.50

Average Price*

Standard Deviation

Position Rank

$1.39 $2.47 $0.16 $2.44 $1.30 $1.60 $1.13 $2.46 $1.44 $21.91 $540.20

0.10 0.11 0.41 0.14 0.08 0.09 0.47 0.18 0.30 1.41 17.68

4 9 1 7 3 6 2 8 5 10 11

$52.41 $3.63

1.91 0.33

Price Dispersion Then and Now

Table 2.

* Prices are not adjusted for inflation.

71

72

PATRICK SCHOLTEN AND S. ADAM SMITH

changes in e-tail markets were large. This, however, is in part due to the smaller sample of products. The next section discusses our results.

3. RESULTS 3.1. Retail Markets Using the coefficient of variation (␴/␮), Table 3 compares relative price dispersion in retail markets in 1976 and 2000. The average coefficient of variation across all products was 18.19% in 2000, compared to 13.25% in 1976 – a 37% increase. In 1976, the coefficient of variation ranged from 3.27% to 41.38%. The smallest coefficient of variation in the 2000 retail sample of products was 1.63%, while the largest was 42.0%. Thus, the distribution of coefficients of variation is narrower for the 1976 retail data than the 2000 retail data. Strikingly, the coefficient of variation for over half of the products in the 1976 sample is less than 10%. In contrast, 80% of the products in the 2000 retail sample exhibit a coefficient of variation that is greater than 10%. This suggests that retail price dispersion is actually more prevalent today than in was in 1976. Indeed, of the 20 products common to our data and the CP data, 12 had higher coefficients in 2000 than in 1976. A coefficient of variation of zero indicates that prices are consistent with the “law of one price.” The results in Table 3 suggest that prices in these markets were generally inconsistent with the law of one price in both 1976 and 2000. Moreover, in retail markets, price dispersion was at least as large in 2000 as it was in 1976. 3.2. E-tail Markets Table 4 compares the levels of price dispersion in 2000 e-tail markets with that observed by CP in 1976 retail markets. On average, price dispersion in e-tail markets was 14.48%, compared to only 12.22% for retail markets in 1976. While the average coefficient of variation is larger in 2000 than 1976, unlike our retail dataset, the distribution of coefficients of variation across products is narrower in 2000. In 1976, the coefficient of variation ranges from just over 3% to over 41% and in 2000 ranges from slightly less than 6% up to about 28%. The distribution of coefficients of variation across products has shifted upward, despite its narrowing. Indeed, the coefficients of variation in the e-tail markets of 2000 are larger for all but two products compared to the products in 1976 retail markets.

Retail Dispersion in 2000 Compared to Retail Dispersion in 1976. 2000 Retail

Products Expensive Camera Stalk Celery lbs. Potatoes Contact Lens Solution Tea Batteries Razor Blades Dozen Lemons Hand Cream Aspirin Hair Spray Antacids Deodorant Film Lettuce Thermometer

1976 Retail

Number of Firms

Coefficient of Variation (Percent)

Position Rank

Number of Firms

Coefficient of Variation (Percent)

Position Rank

3 4 4 8 9 7 7 4 7 7 6 7 7 8 4 6

1.63 7.51 8.34 9.50 11.10 11.50 12.60 12.80 12.90 13.50 13.70 14.60 16.80 19.70 20.30 20.70

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16

7 8 9 8 10 8 8 7 8 9 9 7 9 9 9 9

3.27 18.35 14.94 7.27 23.97 26.21 7.42 21.08 5.63 5.84 4.47 5.55 7.01 5.94 18.24 21.15

1 15 13 9 18 19 10 16 4 5 2 3 8 6 14 17

Price Dispersion Then and Now

Table 3.

73

74

Table 3.

Continued.

2000 Retail

Dozen Roses Male Contraceptives Inexpensive Camera Auto Polish Average (all products) Average (excluding inexpensive and expensive cameras)

12 13 7 6

Coefficient of Variation (Percent)

Position Rank

36.10 37.30 41.30 42.00

17 18 19 20

Number of Firms 11 7 10 11

Coefficient of Variation (Percent)

Position Rank

6.44 41.38 7.48 13.34

7 20 11 12

6.00

18.19

8.65

13.25

6.83

17.83

8.44

14.12

PATRICK SCHOLTEN AND S. ADAM SMITH

Products

Number of Firms

1976 Retail

E-tail Dispersion in 2000 Compared to Retail Dispersion in 1976. 2000 E-tail

Products Thermometer Expensive Camera Hair Spray Antacids Razor Blades Deodorant Hand Cream Aspirin Male Contraceptives Batteries Dozen Roses Average

Number of Firms 4 12 3 4 5 5 4 4 13 2 14 6.36

1976 Retail

Coefficient of Variation (Percent)

Position Rank

5.72 6.53 8.08 9.47 9.89 10.49 12.45 18.53 22.17 27.52 28.40

1 2 3 4 5 6 7 8 9 10 11

14.48

Number of Firms 9 7 9 7 8 9 8 9 7 8 11 8.36

Coefficient of Variation (Percent)

Position Rank

21.15 3.27 4.47 5.55 7.42 7.01 5.63 5.84 41.38 26.21 6.44

9 1 2 3 8 7 4 5 11 10 6

Price Dispersion Then and Now

Table 4.

12.22

75

76

PATRICK SCHOLTEN AND S. ADAM SMITH

Across all products, the average change in the coefficients of variation is over 80% from 1976 to 2000. While, on average, the magnitude of the intertemporal change is large, the position rankings of coefficients of variation changes little. For five of the 11 products in this study, the coefficient of variation rank changed by one position or less. Similarly, the coefficient of variation rank for four of the 11 products changed by only two or three positions, while the rank of the remaining two products changed by five and eight positions. The results we presented in Table 3 revealed that retail price dispersion was at least as large in 2000 as it was in 1976. Similarly, Table 4 indicates that despite many of the predictions that e-tail markets will lead to more competitive product markets (and less price dispersion), the evidence suggests that price dispersion is just as prevalent in the e-tail markets of 2000 as it was in the retail markets of 1976. Based on the results in Tables 3 and 4, one might be tempted to conclude that price dispersion in 2000 e-tail markets was slightly lower compared to 2000 retail markets. This would be misguided, however, as the sets of products in these two samples are different. In the following section, we do offer insight on contemporaneous price dispersion in e-tail and retail markets, but the reader interested in a more formal analysis is referred to Pan, Shankar and Ratchford (2002) in this volume, or Brynjolfsson and Smith (2000).

4. ANALYSIS This section highlights a few potential explanations for the change in intertemporal price dispersion over the 24-year time span that this study covers and some potential determinants of contemporaneous price dispersion. In addition, we offer insight about the nature of contemporaneous price dispersion in retail markets and e-tail markets. While the potential explanations for price dispersion are voluminous, systematic differences between the Carlson and Pescatrice’s (1980) study and this study may explain some of the intertemporal change in price dispersion. The first and most obvious is the difference in the physical locations where data were collected. These geographical differences are likely to lead to structural differences. The downtown central business district and the immediately surrounding neighborhoods of New Orleans are more concentrated than the more suburban area of Bloomington. The demand curve retailers face for the products in these markets is likely to be very different. Moreover, given the more concentrated area where price data were collected in New Orleans,

Price Dispersion Then and Now

77

consumer search costs may be lower in New Orleans compared to Bloomington. Another important difference between the two studies is the composition of products. For each study, we recognize that homogeneity within each product is tenuous. Indeed, it is arguable whether products like lettuce, celery, and roses are of identical quality across retailers. Within each study, however, every attempt was made to limit product differentiation by collecting an identical brand and size of each product. But, between studies the composition of products is very different. CP use generic labels to describe the products in their study rendering it impossible to identically replicate the products in their study. Furthermore, even if the information about the brand and size of products in 1976 were available, it is unlikely that many of those products would exist in 2000. Using the coefficient of variation to study price dispersion makes this point mute, since it is a relative measure of price dispersion and is invariant to inflation. The important point is that a consumer who purchases, say, contact lens solution in 2000 faces at least as much price dispersion as the consumer who purchases contact lens solution in 1976. While systematic differences between this study and CP are likely to account for some of the intertemporal differences in price dispersion, it is naïve to dismiss systematic differences as the sole source. If the intertemporal changes stem solely from a systematic component, then we would observe similar changes in intertemporal dispersion across all products. This clearly contradicts the results in Tables 3 and 4. In particular, products like lettuce and contact lens solution in Table 3 exhibit increases in price dispersion on the order of 2%, while other products like roses, inexpensive cameras and auto polish experience increases of about 30%. Changes in price dispersion of this large magnitude suggests a more general structural change in the market. These changes may stem from technological advancements, consumer preferences or other ways in which the retailers in these markets compete. In his seminal 1961 article, Stigler conjectured that the expected savings to consumers who purchase an expensive product will be large, leading to a greater number of searches. As a result, contemporaneous price dispersion among expensive products will be less. Using Spearman’s rank correlation, CP find support for Stigler’s conjecture; namely, that there exists a negative and statistically significant relationship between the rank of the average price and the rank of the coefficient of variation. Table 5 replicates this test for the retail products common to this and the CP study and the e-tail products common to both studies. A casual comparison of columns 2 and 3 does not reveal the systematic relationship conjectured by

78

Table 5.

Comparison of Average Price Ranks and Coefficient of Variation Ranks. 2000 Retail

Products

1976 Retail

Average Price Rank

Coefficient of Variation Rank

Average Price Rank

Coefficient of Variation Rank

Average Price Rank

Coefficient of Variation Rank

9 12 10 7 16 5 6 19 20 11 8 13 18 1 2 15 14 3 4 17

12 10 20 6 4 18 13 8 17 1 14 11 9 19 3 15 7 2 5 16

4 5

4 8

3

10

1

6

10 11

11 2

2 6

3 7

8 7

9 5

9

1

15 8 9 1 14 10 6 18 20 12 17 13 19 2 3 7 16 3 5 11

3 5 12 19 9 20 8 16 7 1 6 2 4 11 13 14 10 15 18 17

PATRICK SCHOLTEN AND S. ADAM SMITH

Antacids Aspirin Auto Polish Batteries Contact Lens Solution Deodorant Dozen Lemons Dozen Roses Expensive Camera Film Hair Spray Hand Cream Inexpensive Camera lbs. Potatoes Lettuce Male Contraceptives Razor Blades Stalk Celery Tea Thermometer

2000 E-tail

Continued.

2000 Retail

Products Hypotheses Tests H0: ␮ and CV are independent. H1: ␮ and CV are not independent. Spearman’s rho Pr > |t|

Average Price Rank

Coefficient of Variation Rank

2000 E-tail Average Price Rank

Coefficient of Variation Rank

1976 Retail Average Price Rank

Price Dispersion Then and Now

Table 5.

Coefficient of Variation Rank

2000 Retail Markets Fail to Reject

2000 E-tail Markets Fail to Reject

1976 Retail Markets Reject

0.0947 0.6912

⫺ 0.0727

⫺ 0.4814

0.8317

0.0316

79

80

PATRICK SCHOLTEN AND S. ADAM SMITH

Stigler. This observation is confirmed by Spearman’s rank correlation at the bottom of Table 5 in columns 2 and 3. Specifically, we fail to reject the null hypothesis that the rank of the average price is independent of the rank of the coefficient of variation for the 2000 retail price data. Similarly, columns 4 and 5 of Table 5 indicate that we reject the null hypothesis of independence between the average price rank and the coefficient of variation rank in 2000 e-tail markets; however, the sign of Spearman’s rho is negative. In contrast, as shown in CP, 1976 retail markets tend to show support for Stigler’s hypothesis in Table 5. In this case, we reject the null hypothesis and conclude that a negative and statistically significant relationship between the average price and the coefficient of variation exists. The evidence presented in Tables 3 and 4 suggest that prices in 2000 are less dispersed in e-tail markets compared to retail markets in Bloomington. As previously mentioned, this comparison may be misleading since the composition of products in these two samples is different. While the focus of this paper is to describe intertemporal price dispersion, next we offer some insight on contemporaneous price dispersion in 2000 between e-tail markets and retail markets in Bloomington. To do this we assembled a third dataset consisting of an array of products that are popular in both e-tail markets and retail markets. These products consist of popular (at the time) book titles, music titles (CDs), movie titles, cameras and other computer-related items. These data were collected according to the same methodology and at the same time as data described in sections 2.1 and 2.2. Table 6 presents the average coefficient of variation in list prices for each product category in our third dataset that contemporaneously compares price dispersion in e-tail markets and retail markets. The average coefficient of variation among list prices in e-tail markets is remarkably similar to those in retail markets; 12.87% in e-tail markets and 12.83% in the retail markets of Bloomington. About half of the product categories in this sample – five of 11 product categories – have an average coefficient of variation that is greater in e-tail markets compared to retail markets. The Spearman rank correlation tests presented in Table 5 provide conflicting evidence on whether price dispersion is negatively related to the average price of a product. The results presented in Table 6, however, seem to provide casual support for this hypothesis. The average coefficient of variation for product categories with an average price that is greater than $100 is just over 5% in e-tail markets and about 3.5% in retail markets. In contrast, when the average price is less than $100, the average coefficient of variation is about 17% in e-tail markets and about 18% in retail markets. Thus, while price dispersion is,

Price Dispersion Then and Now

Table 6.

81

Contemporaneous Dispersion Among List Prices. 2000 E-tail Average Coefficient of Variation

2000 Retail Average Coefficient of Variation

Books Cameras Compact Discs Computers Flowers Fragrance Movies Personal Items Printer Scanner Software

20.27% 5.56% 9.56% 2.82% 28.40% 15.98% 11.66% 13.68% 8.51% 3.76% 21.41%

10.83% 1.33% 10.69% 2.71% 36.30% 21.84% 14.84% 16.63% 10.11% 0.14% 15.72%

Average (all products) Average (Average Price < $100) Average (Average Price > $100)

12.87% 17.28% 5.16%

12.83% 18.12% 3.57%

Products

on average, about the same magnitude in e-tail markets and retail markets, there appears to be some casual support for Stigler’s hypothesis that price dispersion will be less in markets for relatively expensive items. Our Table 6 results are consistent with previous studies comparing price dispersion in e-tail markets and retail markets. Like Brynjolfsson and Smith (2000), we find that book prices are generally more dispersed among e-tailers than retailers, but that for CDs dispersion in e-tail markets is approximately the same, and perhaps slightly lower. However, while differences in the level of price dispersion are observed in e-tail markets and retail markets across product categories, on average, these results suggest that price dispersion is about the same among e-tailers and retailers. Two fundamental and competing differences are readily observable between e-tail markets and retail markets.7 First are taxes. To encourage the development and use of e-tail markets, and Internet markets in general, the U.S. government enacted the Internet Tax Freedom Act in 2001 that extends a 1998 moratorium on “new, special, and discriminatory Internet taxes.” Thus, e-tail markets allow consumers to avoid paying taxes on purchases. There is a caveat. To take advantage of tax-free purchases a consumer must purchase from an e-tailer located outside of the state in which the consumer resides. In the

82

PATRICK SCHOLTEN AND S. ADAM SMITH

analysis that follows, we assume that consumers purchasing from e-tailers deliberately avoid paying taxes by adhering to the caveat, but that retail purchases incur a 5% sales tax; the rate prevailing at the time when and in the state – Indiana – where data were collected. The second observable difference is the cost of acquiring products. E-tailers typically charge shipping costs that consumers pay. Shipping costs often vary from e-tailer to e-tailer. However, to obtain actual shipping costs consumers often have to enter their shipping address and credit card information. Since doing this is an extremely time intensive process, we instead utilized the shipping cost schedule posted by the e-tailer that was most representative of the observable shipping costs in the sample – Amazon.com. Depending on the product category, shipping costs may exhibit economies of scale when multiple items are purchased during a single shopping episode, as illustrated in Table 7. For most of the products in the sample, shipping costs consist of a base fee and either a per-unit or per-pound fee. Thus, for certain products purchasing multiple items during a single shopping episode may lower shipping cost compared to purchasing items individually. For example, Table 7 illustrates that the base shipping cost associated with purchasing a book is $3.00 and the per-unit shipping cost is $0.99. Therefore, purchasing a single title results in a total shipping cost of $3.99, and drops to $1.99 and $1.59 when purchasing three or five books, respectively. In contrast, consumers purchasing a product from a local retailer incur a transportation cost. To calculate the transportation cost of purchasing from a retailer, we use the government standard reimbursement rate of $0.32 per mile and the average number of miles to each retailer (from one of the author’s homes, which is centrally located in Bloomington). These results are presented in Table 7. A typical consumer in Bloomington travels, on average, between five and seven miles to visit a retailer in our sample. Consumers can reduce their transportation costs by purchasing multiple units from the same retailer in a single shopping episode. Neither shipping costs nor transportation costs fully capture the exact transaction cost; however, they at least serve as a proxy. These (proxies for) transaction costs are likely to change the full cost (listed price + (proxies for) transaction costs) of acquiring products from e-tailers and retailers and may have a profound impact whether price dispersion is observed in e-tail markets and retail markets. In fact, to the extent that price dispersion is observed among e-tailers and retailers, some have argued that accounting for transaction costs will explain away dispersion. The following analysis offers some insight on this conjecture.

Summary Statistics for Shipping Costs and Transportation Costs, 2000. Shipping Cost in E-tail Markets

Products Book Titles Cameras Compact Discs Computers Flowers Fragrances Movie Titles (VHS) Movie Titles (DVD) Personal Care Items Printers Scanners Software

Transportation Cost in Retail Markets

Base Shipping Cost

Per-unit Shipping Cost

Per-unit Shipping Cost 1 Item

Per-unit Shipping Cost 3 Items

Per-unit Shipping Cost 5 Items

Average Miles (round trip)

Per-unit Transportation Cost 1 Item

Per-unit Transportation Cost 3 Items

Per-unit Transportation Cost 5 Items

$3.00 4.95 2.00 4.95 8.00 3.95 3.00 2.00 3.75 4.95 4.95 4.00

$0.99/item $0.50/lbs $0.99/item $0.50/lbs – – $0.99/item $0.99/item $0.47/item $0.50/lbs $0.50/lbs $0.99/item

$3.99 6.20 2.99 22.45 8.00 3.95 3.99 2.99 4.22 12.45 10.95 4.99

$1.99 – 1.66 – 2.67 1.32 1.99 1.66 1.72 – – 2.33

$1.59 – 1.39 – 1.60 0.79 1.59 1.39 1.22 – – 1.79

5 6 5 6 7 7 5 5 7 6 6 6

$1.60 1.92 1.60 1.92 2.24 2.24 1.60 1.60 2.24 1.92 1.92 1.92

$0.53 0.64 0.53 – 0.75 0.75 0.53 0.53 0.75 – – 0.64

$0.32 0.38 0.32 – 0.45 0.45 0.32 0.32 0.45 – – 0.38

Price Dispersion Then and Now

Table 7.

83

84

Table 8.

Contemporaneous Price Dispersion Including Transaction Costs. 2000 Retail

Products

Average Coefficient of Variation Including Shipping Costs 1 Item 3 Item 5 Item

Average Coefficient of Variation Including Transportation Costs 1 Item 3 Item 5 Item

Books Cameras Compact Discs Computers Flowers Fragrance Movies Personal Items Printer Scanner Software

16.8% 5.5% 7.7% 2.7% 24.5% 14.1% 9.3% 7.6% 7.8% 3.5% 19.2%

Average (all products) Average (Average Price < $100) Average (Average Price > $100)

10.8% 14.2% 4.9%

18.3%

18.7%

8.5%

8.6%

27.0% 15.3% 10.3% 10.2%

27.5% 15.6% 10.5% 11.0%

20.3%

20.6%

10.1% 1.3% 9.7% 2.7% 34.6% 20.3% 13.8% 11.4% 9.9% 0.1% 15.1%

15.7% 15.7%

16.1% 16.1%

11.7% 16.4% 3.5%

10.6%

10.7%

10.3%

10.5%

35.7% 21.3% 14.5% 14.4%

35.9% 21.5% 14.6% 15.2%

15.5%

15.6%

17.5% 17.5%

17.7% 17.7%

PATRICK SCHOLTEN AND S. ADAM SMITH

2000 E-tail

Price Dispersion Then and Now

85

Table 8 illustrates that contemporaneous price dispersion persists among etailers and retailers even if (proxies for) transaction costs are taken into consideration. On average, the coefficient of variation that accounts for transaction costs in e-tail markets is about 11% and almost 12% in retail markets. Compared to the average coefficient of variation among only list prices, including transactions costs slightly reduces the average coefficient of variation by about 1%. For each product category, transaction costs lower the coefficient of variation. The magnitude of transaction cost effects on the coefficient of variation ranges from about 0% to about 5%. As transaction costs are spread over multiple-item purchases within the same shopping episode, the coefficient of variation monotonically increases. The efficiency gains from multiple-item purchases tend to increase the coefficient of variation. Thus, on average, transaction costs do not appear to materially impact price dispersion. These results are perhaps not surprising given the methodology we adopt to account for transaction costs. Indeed, it is the differences in transaction costs that may lead to a decline in price dispersion. In particular, evidence provided by Ellison and Ellison (2000) suggest that e-tailers may engage in obfuscation strategies in attempt to conceal its price. Shipping costs (transaction costs in general) are but another variable that e-tailers can use to conceal the true purchase price of a good. Since we are not able to fully address this issue with these data, this area is left for future research.

5. CONCLUSIONS The primary purpose of this paper is to explore the persistence of price dispersion over time. Along the way, ancillary information about price dispersion was presented. In particular, our dataset comparing contemporaneous price dispersion in e-tail markets and retail markets suggests dispersion is about the same in each of these markets: prices are dispersed by about 13%, although the level of price dispersion depends on the products in the sample. Our results for books and CDs are consistent with previous research; namely, that book prices in e-tail markets tend to be more dispersed than in retail markets while CD prices are about equally dispersed among e-tailers and retailers. Moreover, transaction costs were found to reduce the amount of dispersion observed among prices, but the magnitude of these effects were on average relatively small – on the order of 1% in both e-tail markets and retail markets.

86

PATRICK SCHOLTEN AND S. ADAM SMITH

More importantly, we find evidence that price dispersion in 2000 is at least as great as it was in 1976, regardless of whether 2000 e-tail or retail markets are used to compute measures of relative dispersion. This finding is fairly remarkable given the evolution of markets over the past 24 years. Indeed, by 2000 the Internet has changed the way in which many consumers purchase products. Many pundits predict that the Internet will lead to more informed consumers resulting in greater competition among and between e-tail markets and retail markets. This suggests that price dispersion should diminish. The data presented in this paper indicates that the Information Age has yet to reduce the levels of price dispersion. Using 1976 as a benchmark, we conclude that levels of retail price dispersion “then and now” are roughly the same.

NOTES 1. See Salop and Stiglitz (1977), Shilony (1979), Varian (1980), Rosenthal (1980), Narasimhan (1988), Stahl (1989), Stahl (2000), Baye and Morgan (2001), and Janssen and Moraga (2001) for examples of this literature. 2. As a percentage of the average price, these ranges were between 10% and 198%. 3. More recently, Baye, Morgan and Scholten (2001a) offer a third measure of price dispersion – the price gap – that measures the difference between the two lowest prices in the market. 4. See Brown and Goolsbee (2000); Clemons, Hann, and Hitt (2000); Morton, Zettlemeyer, and Risso (2000); Smith (2000); Baye, Morgan, Scholten (2001a, 2001b, 2003, 2004); Baylis and Perloff (2001); Clay, Krishnan, and Tay (2001); and Pan, Ratchford, and Shankar (2002). 5. A potential explanation for the large relative price decline for hair spray during this period was an international agreement calling for a worldwide ban on the primary chemical used in hair spray, and many other aerosol products, chlorofluorocarbons (CFCs). In response, many aerosol manufacturers abandoned aerosols altogether opting instead for cheaper alternatives such as pump-based spray bottles or aerosols containing carbon dioxide or other hydrocarbons. 6. The spread of AIDS and other sexually transmitted diseases is consistent with the relative large price increase experienced in male contraceptives. 7. Unobservable differences also exist. For instance, e-tail markets offer consumers the convenience of searching and purchasing products 24-hours a day from any location with Internet access. Consumers valuing this convenience are typically willing to pay higher prices. Thus, compared to retailers, e-tailers may be able to charge relatively higher prices. In contrast, marginal selling costs to e-tailers are likely to be lower since fewer sales personnel are required. Thus, whether prices in e-tail markets will be lower compared to retail markets will depend on which of these two effects dominates. Other unobservables are the true search costs and transaction costs. The marginal cost of search is presumably less in e-tail markets compared to retail markets, but by how much is unknown. Similarly, the true transaction cost is unknown since it varies with each consumer’s opportunity cost of time.

Price Dispersion Then and Now

87

ACKNOWLEDGMENTS We are especially grateful to our mentor, Michael Baye, for the many helpful discussions and his guidance throughout the stages of this project. We gratefully acknowledge Indiana University’s financial assistance during this project. Furthermore, we would like to thank Josh Altschuler and Scott Bartfield for their valuable assistance in collecting data.

REFERENCES Bakos, Y. (1997). Reducing Buyer Search Costs: Implications for Electronic Marketplaces. Management Science, 43, 1676–1692. Bakos, Y. (2001). The Emerging Landscape of Retail E-commerce. Journal of Economic Perspectives, 15, 69–80. Baye, M. R., & Morgan, J. (2001). Information Gatekeepers on the Internet and the Competitiveness of Homogeneous Product Markets. American Economic Review, 91, 454–457. Baye, M. R., Morgan, J., & Scholten, P. (2001a). Price Dispersion in the Small and in the Large: Evidence from an Internet Price Comparison Site. Mimeo. Baye, M. R., Morgan, J., & Scholten, P. (2001b). Pricing and Reputation in an Online Consumer Electronics Market. Mimeo. Baye, M. R., Morgan, J., & Scholten, P. (2003). The Value of Information in Online Markets: Theory and Evidence. Journal of Public Policy and Marketing, forthcoming. Baye, M. R., Morgan, J., & Scholten, P. (2004). Persistent Price Dispersion in Online Markets. In: D. Jansen (Eds), The New Economy. University of Chicago Press, forthcomimg. Baylis, K., & Perloff, J. M. (2001). Price Dispersion on the Internet: Good Firms and Bad Firms. Mimeo. Brown, J. R., & Goolsbee, A. (2000). Does the Internet Make Markets More Competitive? Evidence from the Life Insurance Industry. NBER Working Paper 7996. Brynjolfsson, E., & Smith, M. (2000). Frictionless Commerce? A Comparison of Internet and Conventional Retailers. Management Science, 46, 563–585. Burdett, K., & Judd, K. (1983). Equilibrium Price Dispersion. Econometrica, 51, 955–969. Carlson, J., & Pescatrice, D. (1980). Persistent Price Distributions. Journal of Economics and Business, 33, 21–27. Clay, K., Krishnan, R., & Wolf, E. (2001). Prices and Price Dispersion on the Web: Evidence from the Online Book Industry. NBER Working Paper 8271. Ellison, G., & Ellison, S. (2001). Search, Obfuscation, and Price Elasticities on the Internet. Mimeo. Clemons, E. K., Hann, I., & Hitt, L. M. (1998). The Nature of Competition among Online Travel Agents: An Empirical Investigation. The Wharton School of the University of Pennsylvania Working Paper. Gatti, R. (2000). Equilibrium Price Dispersion with Sequential Search. Mimeo. Janssen, M., & Moraga, J. (2001). Pricing, Consumer Search and the Size of Internet Markets. Mimeo. Morton, F., Zettlemeyer, F., & Risso, J. S. (2001). Internet Car Retailing. Mimeo. Narasimhan, C. (1988). Competitive Promotional Strategies. Journal of Business, 61, 427–449.

88

PATRICK SCHOLTEN AND S. ADAM SMITH

Pan, X., Shankar, V., & Ratchford, B. (2002). Price Competition between Pure Plays vs. Bricksand-Clicks E-tailers: An Analytical and Empirical Analysis. Advances in Applied Microeconomics, 11. Pratt, J., Wise, D., & Zeckhauser, R. (1979). Price Differences in Almost Competitive Markets. Quarterly Journal of Economics, 93, 189–211. Reinganum, J. (1979). A Simple Model of Equilibrium Price Dispersion. Journal of Political Economy, 87, 851–858. Rosenthal, R. (1980). A Model in Which an Increase in the Number of Sellers Leads to a Higher Price. Econometrica, 46, 1575–1580. Salop, S., & Stiglitz, J. (1977). Bargains and Ripoffs: A Model of Monopolistically Competitive Price Dispersion. Review of Economic Studies, 44, 493–510. Shilony, Y. (1979). Mixed Pricing in Oligopoly. Journal of Economic Theory, 14, 373–388. Smith, M. D. (2001). The Law of One Price? Price Dispersion and Parallel Pricing in Internet Markets. Carnegie Mellon University Working Paper. Stahl, D. (1989). Oligopolistic Pricing with Sequential Consumer Search. American Economic Review, 79, 700–712. Stahl, D. (2000). Strategic Advertising and Pricing in E-commerce. Advances in Applied Microeconomics, 9, 69–100. Stigler, G. (1961). The Economics of Information. Journal of Political Economy, 69, 213–225. Varian, H. (1980). A Model of Sales. American Economic Review, 70, 651–659.

BUSINESS-TO-BUSINESS E-COMMERCE: VALUE CREATION, VALUE CAPTURE AND VALUATION Luis Garicano and Steven N. Kaplan ABSTRACT This paper presents a framework to analyze the potential changes in transaction costs due to the introduction of e-commerce on transactions between businesses. It then illustrates and applies this framework using internal data from an Internet-based firm to measure process improvements, marketplace benefits, and motivation costs. We find that process improvements and marketplace benefits are potentially large, while little evidence exists of increases in motivation costs. Finally, we use the framework to help discuss why valuations of Internet companies were so high at the end of 1999 and why they have declined so precipitously since then.

1. INTRODUCTION In this paper, we study the economic impact resulting from the introduction of the Internet in transactions between firms (i.e. business-to-business (B2B)

The Economics of the Internet and E-Commerce, Volume 11, pages 89–125. © 2002 Published by Elsevier Science Ltd. ISBN: 0-7623-0971-7

89

90

LUIS GARICANO AND STEVEN N. KAPLAN

e-commerce). We present a framework that describes the potential changes in transaction costs caused by transferring a transaction from a physical marketplace to an Internet-based one. Following Milgrom and Roberts (1992), our framework differentiates between coordination costs and motivation costs. We argue that it is likely that B2B e-commerce reduces coordination costs and increases efficiency. We illustrate and apply this framework using detailed internal data from one Internet-based firm to measure process improvements, marketplace benefits, and motivation costs. Our results suggest that process improvements and marketplace benefits are potentially large. We find little evidence that informational asymmetries are more important in the electronic marketplace we study than the existing physical ones. Finally, we use the framework to help discuss why valuations were so high at the end of 1999 and why they have declined so precipitously since then. We also speculate that the long-term real effects of B2B and the Internet are likely to be quite favorable.

2. MEASURING VALUE CREATION IN B2B E-COMMERCE As mentioned above, B2B e-commerce has the potential to substantially reduce transaction costs in inter-firm trade. Following Milgrom and Roberts (1992), we classify transaction costs in two categories: costs associated with the problem of coordination and costs associated with the problem of motivation. Shifting a transaction from a physical environment to the Internet has the potential to affects both types of transaction costs. Coordination Costs Coordination costs are “related to the need to determine prices and other details of the transaction, to make the existence and location of potential buyers and sellers known to one another, and to bring the buyers and sellers together to transact.” We find it useful to classify the effects of the Internet on coordination costs into two general categories: process improvements and marketplace benefits. Below, we describe the potential Internet-based improvements in these coordination costs. It is important to recognize (and we then discuss) that reductions in transactions costs are likely to lead to additional direct and

Business-to-Business E-Commerce

91

indirect benefits. We use this framework in later sections to study the gains attained in some examples. Process Improvements B2B e-commerce can improve efficiencies by reducing the costs involved in an existing business process. Such an improvement may take place in two basic forms. First, it may simply reduce the cost of an activity already being conducted, as when a transaction that is currently conducted by phone or fax is automated. In other instances, the Internet provides an opportunity to redesign the existing process. The methodology we use to measure or estimate the value of process improvements is straightforward. First, we describe and measure the costs of the activities involved in the existing process in detail. Second, we describe and measure the costs of the process using B2B e-commerce. The difference, if any, is the value of the process improvement. Marketplace Benefits We classify the second way in which B2B e-commerce can reduce coordination costs as marketplace benefits (or direct information improvements). These benefits come in some of the following forms. The Internet potentially reduces a buyer’s cost of finding suppliers because it is less expensive to search for products and compare prices over the Internet than it is to read catalogs and make phone calls. Conversely, sellers can reach more potential customers at lower cost. As a result, buyers will find sellers they might not have otherwise found. EBay is an example of this on the consumer side. (eBay is C2C – consumer-to-consumer.) Second, the Internet potentially provides buyers with better information about product characteristics (including prices and availability) because it is less expensive to obtain. Finally, the Internet also potentially provides better information about buyers and sellers. On the other hand, conducting the transaction over the Internet may increase these transaction costs, due to the buyers’ inability to physically inspect the merchandise object of the exchange. This may be the case when buyers need to match their needs for objects based precisely on a characteristic that requires physical inspection. For example, consider the second hand car example that we explore in depth later. Suppose that dealers in a particular location sell cars to a lower income, older consumer who takes good care of the cars, while dealers in another location cater to lower income handy-men.

92

LUIS GARICANO AND STEVEN N. KAPLAN

Holding all the observable characteristics constant, dealers in the first location will be looking for cars in perfect condition; while dealers in the second location will be looking for cars in bad, but repairable condition. If the condition of the car is hard to communicate without hearing the motor and looking at the car, it will be difficult to distinguish between these cars in an Internet auction. As a consequence, the matching of cars with buyers may be worsened. It is important to note that this effect takes place regardless of the fact that the composition of supply of cars is unchanged (no adverse selection). Estimating these costs and benefits is appreciably more difficult than estimating the process improvement benefits. One place to look – and one for which we have data – is at the buyer’s willingness to pay for each object. Higher willingness to pay by buyers for a particular item is evidence of better matching. Other places to look include the amount of trade and prices sellers receive. If B2B e-commerce delivers marketplace benefits, trade should increase. Ebay is a clear example of this in that trade occurs that would not otherwise occur. Higher prices for sellers would represent better matching. It is likely, on the other hand, that lower customer acquisition costs would reduce prices. Direct and Indirect Effects of Coordination Costs Reductions Clearly, any reduction in coordination costs results in direct economic gains through a reduction in the cost of undertaking these transactions. It is possible, however, that other indirect benefit also will arise. As the costs of undertaking spot market transactions decreases, participants in these transactions may adjust their behavior and realize further efficiency gains. Although estimating these effects is beyond the scope of this paper, we discuss briefly here the effects of the two main sources of these changes: better information processing, and changes in organizational form. Better information about future demand through B2B e-commerce may allow a seller to improve its demand forecasts, and use that information to change its production decisions to better match demand. Conversely, a buyer might obtain better information about existing (and future supply), and use that information to change its inventory decisions. Second, make or buy decisions are likely to be affected. If the Internet is able to produce important decreases in the costs of carrying out transactions in the market, the transaction costs economizing paradigm (Coase, 1939; Williamson, 1985) leads us to predict that fewer transactions will be undertaken inside firms and more will be undertaken in the market.

Business-to-Business E-Commerce

93

Motivation Costs Milgrom and Roberts (1992) distinguish two types of motivation-related transaction costs: those associated with informational incompleteness and asymmetries, and those associated with imperfect commitment. Informational Incompleteness and Asymmetries These types of transaction costs are present “when the parties to the transaction do not have all the relevant information needed to determine whether the terms of an agreement are acceptable and whether they are actually being met.” To the extent that physically observing the merchandise to evaluate its condition is valuable to the buyer, some of that information is lost through the conduct of the transaction through an electronic format. This loss of information about the object of the exchange may translate into an efficiency loss if adverse selection worsens in virtual transactions. Consider, for example, the original lemons issue in second hand automobile markets (Akerlof, 1970), which will later be our example. Holding observable characteristics constant, sellers might try to sell cars with strange sounding motors exclusively thorough the Internet. If sellers offer this type of object more frequently over the Internet, buyers willingness to pay for the average object decreases, leading sellers of higher (unobserved) quality to withdraw from the market. Transaction Costs That Arise from Imperfect Commitment Milgrom and Roberts (1992) define these costs as deriving from “the inability of parties to bind themselves to follow through on threats and promises that they would like to make but which, having made, they would like to renounce.” B2B e-commerce has the potential to increase or decrease these costs. First, by standardizing processes and by leaving an electronic trail, the Internet has the potential to reduce the costs of imperfect commitment. Alternatively, a buyer may avoid intermediary fees by viewing the product over the Internet, but contacting the seller directly. Value Capture After applying the framework, it should be possible to understand the effect of a new technology or process on transaction costs. If the technology does reduce transaction costs, it is potentially viable/valuable. The question then becomes who will capture the reduction in transaction costs. If the technology is unique or difficult to imitate, the innovator should be able to capture some of the

94

LUIS GARICANO AND STEVEN N. KAPLAN

improvements and become valuable. On the other hand, if the technology can be easily imitated by competitors, the customers will capture most of the benefits.

3. THE FRAMEWORK IN ACTION: THE CASE OF AUTODAQ Impact on Coordination costs (1): Process improvements In this section, we compare the time and economic costs involved in the Autodaq/Internet process with those in the physical auction process. In the physical world, when a large volume seller needs to dispose of a car, the seller stores the car and then has it transported to a physical auction site. At the physical auction site, the car is described and inspected. The car may also be reconditioned by the auction site operator. Reconditioning involves repairing minor flaws in the car’s exterior – dents, scratches, etc. When a sufficient number of cars are physically at the auction site, an auction is held. Dealers travel to the physical auction site and bid on the car. After the auction, the car is transported again to the winning dealer. The winning dealer performs any necessary maintenance or repairs and any additional reconditioning needed to retail the car. In the Autodaq/Internet system, Autodaq contracts with an inspector who inspects, describes, and photographs the car. For cars coming off lease, this occurs at the turn-in dealer. For cars coming from rental fleets, this occurs at the fleet marshalling yard. The car is then put up for sale in an online auction. Dealers bid on the car over the Internet from their computers. The car is transported to the winning dealer. The winning dealer performs any necessary maintenance, repairs and reconditioning. If the car does not sell over the Internet, the car continues through the physical auction process.1 Unlike physical auctions, which are run as ascending oral auctions, Autodaq auctions employ a second price auction in the form of a “proxy bidding” mechanism. With a proxy bid, dealers submit the highest price they would be willing to pay and Autodaq automatically increases their bid in the presence of other bids by just enough to become the leading bid. The auction format also allows dealers to directly purchase the car by accepting the ask price given by the seller. Table 1 compares the physical auction process to the Autodaq Internet process, both in terms of time and money. The comparison is made for a typical car coming off lease or from a rental fleet. The table measures time from the day the car comes off lease or is retired by the rental car company to the day

Process cost of physical auction vs. Internet auction based on used car auctions from 1999 to 2000. Physical Auction Industry Estimate Sample Results Time Dollars Time Dollars (days) (days) (1 (2) (3) (4) Wait for pick-up Ship to auction Ready for sale Ready for sale until sale Ship to dealer Total Time

9 to 15 2 10 5 to 15 2 28 to 44

Capital Cost of Time (36 days) Depreciation Cost of Time (36 days) Inspection Cost Shipping Cost Reconditioning Cost Dealer Travel Cost Dealer Time Not Bidding Total Economic Cost Seller Fee Buyer Fee

0.5 Hours 0 to 4 Hours

N.A. N.A. N.A. N.A. 2 37

Internet Auction Autodaq Estimate Sample Results Time Dollars Time Dollars (days) (days) (5) (6) (7) (8) 0 0 2 2 3 7

N.A. N.A. N.A. N.A. 3 17

$107

$110

$21

$51

$188

$193

$37

$89

$5 $220 X

$5 $220 X

$60 $137 X

$60 $137 X

$20

0.5 Hours 0 to 4 Hours

Business-to-Business E-Commerce

Table 1. Process Cost of Physical Auction vs. Internet Auction Process Per Car.

$20

$540 + X

$548 + X

$255 + X

$337 + X

$100 $200

$100 $200

$100 $175

$100 $175

95

96

Table 1.

Continued.

Assumptions: Times: Industry Estimates: Industry estimates for physical auction were obtained from Tom Kontos at ADT Automotive and confirmed by other sources. Autodaq estimates for Internet auction were provided by Autodaq. Sample Results: Sample median for physical auction is time from inspection to time of sale for 9205 cars sold by lessors through physical auction process augmented by two days for dealer shipment. Sample median for Internet auction is time from inspection to time of sale for 694 cars sold through Autodaq process augmented by three days for dealer shipment.

Used cars in our sample have an average sale value of $13,600. Interest rate/cost of capital assumed to equal 8%. Each day, therefore, costs seller 8% ⫻ $13,600/365 = $2.98 per day in capital costs. The table assumes that used car values decline or depreciate in value by 14% per year. Each day, therefore, costs seller 14% ⫻ $13,600/365 = $5.22 per day in depreciation costs/forgone sales price. In the data provided by Autodaq, the sales price declines by 14.8% (with a standard error of 1.7%) per year. Inspection cost for physical auction assumes 15 minutes at a cost of $20 per hour; for Autodaq, is the cost to Autodaq. Dealer travel cost assumes that dealer travels a total of two hours and buys four cars for an average of 0.5 hours per car. Dealer time is valued at $40 per hour. Shipping cost is two shipments at $110 each for the physical auction; one shipment at $137 for the Autodaq process. Based on Autodaq and industry interviews. Reconditioning costs assumed to be the same for both processes

LUIS GARICANO AND STEVEN N. KAPLAN

Wait for pick-up is time from lessee delivery of car to dealer until car is picked-up by physical auction. Ready for sale is time from delivery at physical auction site to the time car is ready for sale. Includes time to recondition.

Business-to-Business E-Commerce

97

the car arrives at the buying dealer. The table measures costs as the economic costs of the process. It does not measure the benefits to a seller from moving from a physical process to the Internet process. We report both estimated times involved in the physical auction process and in the Internet process and actual times for both processes from a sample provided by one of the sellers that used the Autodaq process. The estimates for the physical auction process in columns (1) and (2) were provided by Autodaq and by Tom Kontos of ADT Automotive. As mentioned earlier, ADT Automotive was the second largest competitor in the physical auction business.2 We obtained similar estimates in interviews with other industry participants. Column (1) reports that the physical auction process takes from 28 to 44 days. We also estimate these costs directly from a sample of cars sold through the physical auction process provided to us by one of the lessors that provided Autodaq cars to sell. Our analysis is in columns (3) and (4) of Table 1. The information provided by the seller allows us to calculate time to sale from: (1) lessor inspection date; and (2) lease end for cars sold through the physical auction and for cars sold through Autodaq. Neither date is ideal. According to Autodaq, a car was typically inspected before the lessee turned it in. Time to sale from lessor inspection date, therefore, overstates the time from turn-in to sale. A car sold through Autodaq was inspected an average of 9 days before the car was turned in. According to Autodaq, the overstatement is slightly worse for the cars sold through Autodaq because all such cars were inspected before they were turned in. While most of the cars sold through the physical auction were inspected before they were turned in, a small number were inspected at the physical auction. The comparisons between Internet and physical auction processes, therefore, will slightly understate the advantage of the Internet. Time to sale from lease end also is problematic because cars are sold both well before the lease end date and well after the lease end date. On this dimension, we do not know if there is a bias between the cars sold through Autodaq and cars sold through the physical auction. In our analysis, we use the time from inspection to sale because: (1) it appears to be a more reliable measure of the disposition process; and (2) we have inspection dates for all cars, but do not have lease-end dates for all cars. The results are qualitatively similar using both dates. For all 9,205 cars, we calculate the time that elapsed from lessor inspection date to the date the car was sold. The median time is 35 days. We add two days to this to estimate the delivery time from the auction to the purchasing dealer. As we report in Table 1, the median elapsed time is 37 days. This is close to

98

LUIS GARICANO AND STEVEN N. KAPLAN

the midpoint of the range provided by ADT Automotive. 37 days also is consistent with the estimates we obtained from other interviews. On the other hand, if these cars were inspected 9 days before they were turned in, the median time would be more like 28 days, which is at the low end of the range in column 1.3 Column 5 in the table reports Autodaq’s estimates of the time that is involved in the Internet auction process. Autodaq believes that the Internet process should take 7 days compared to the 28 to 44 days in the physical auction process. The potential time reductions come in several areas. First, it typically takes 9 to 15 days before lessors and fleet owners ship a car to the physical auction site. Part of the reason for the delay is that the physical auction company does not pick the car up immediately. The other reason is that the seller may attempt to sell the car to the original dealer, but must take some time attempting to determine the appropriate price. It is not entirely clear that all of the savings here are Internet specific. It would seem possible for the lessors to contract with a physical auction site to reduce this time as well. It remains to be seen whether Autodaq can reduce this time. Second, it typically takes 15 to 25 days from the time a car arrives at a physical auction site until it is sold. On the Internet, Autodaq estimates this time can be reduced to 4 days. One reason for the delay in the physical auction process is that the car generally waits some time before it is reconditioned and reconditioning takes some time.4 The more important reason, however, is that the physical auction sites try to make each individual auction somewhat homogeneous in terms of the cars available. In other words, they attempt to sell largely Fords in one auction; largely Toyotas in the next. This is done because dealers typically look for particular types of cars. As a result, the physical auctions will wait until they have a critical mass of a particular car type or brand before holding an auction. This is not a consideration for Autodaq because the dealer does not have to physically go to the Autodaq auction site. Autodaq’s estimates make two optimistic assumptions. First, the estimates assume that the cars sell quickly on the Internet, which implies a liquid market. Second, the estimates assume that the cars are listed for sale almost immediately after they come off lease which assumes sophisticated and timely tracking and inspection processes. We interpret Autodaq’s estimates, therefore, as the likely process costs of a liquid Internet market. To obtain a more neutral estimate of the gains generated by the Internet, we calculated the actual time that elapsed from the day a car was inspected by the lessor to the day the car was sold for a sample of 694 cars sold over the Internet by Autodaq. The median time is 14 days. We add three days to this to estimate

Business-to-Business E-Commerce

99

the time until the car is delivered to the purchasing dealer. Column 7 reports that the median actual elapsed time is 17 days.5 The “Dollars” columns in Table 1 attempt to value the economic costs of the two processes. The most important costs are the costs of capital, depreciation, and transportation. The cost of capital is relatively straightforward. The typical car (in our sample) sells for $13,600. Each day the car is not sold, the seller is not able to deploy that capital elsewhere. We assume a cost of capital of 8%. This is essentially a debt cost of capital (and, as such, may understate the true cost of capital for a seller). The cost of depreciation is based on the fact that sale prices for used cars depreciate with the age of the car. We assume a depreciation rate of 14%. This reflects the fact that in the data provided by Autodaq, the sales price declines by 14.8% per car-year (with a standard error of 1.7%). Autodaq and the industry experts with whom we spoke estimated that it costs $110 to ship a car from the lessor to a physical auction and then an additional $110 to ship a car from the physical auction to the buyer. The transportation cost to ship a car from the lessor directly to a local buyer was estimated at $137. The difference reflects the absence of economies of scale in shipping directly. Autodaq estimates that a dealer travels one hour each way to an auction and buys four cars. This translates to one-half hour of travel time per car purchased. Conservatively valuing a dealer’s time at $40 per hour6 this translates into $20 per car. Autodaq assumes that a dealer spends five hours at the physical auction. We assume, conservatively, that the dealer does not waste any of these hours at the auction. Finally, we assume that reconditioning costs are the same for the physical auction as for the Internet auction. This also is likely to be conservative in that cars bought in a physical auction are usually reconditioned again by the buyer after they are bought. To account for this, we have not added any extra time to the Internet process for reconditioning. Based on these assumptions, we estimate in column (2) that the physical auction process has an economic cost of $540 per car (not including reconditioning) given the industry estimates of the time costs in column (1). Using our sample results rather than the industry estimates, we obtain an almost identical cost of $548 per car (not including reconditioning) in column (4). Under the assumption of a liquid market and using Autodaq’s estimates, the Autodaq/Internet process has an economic cost in column (6) of only $255 per car (without reconditioning) – a $285 reduction from the industry estimates. Using the costs implied by the actual 17 days elapsed from inspection to

100

LUIS GARICANO AND STEVEN N. KAPLAN

delivery in our sample, the Internet process has a total economic cost in column (8) of $337 (without reconditioning) – a $211 reduction from the physical auction sample results. The difference would be at least as large if we measured the costs from turn in to delivery because the average time from inspection to turn in is at least as large for the Internet sample as for the physical auction sample. Both of the estimated reductions ($211 from the sample and $285 from industry estimates) are conditional on both markets being liquid. In the Autodaq sample, this was not the case – the probability of a sale was 24% not 100%. As a result, the (conditional) process savings overstate actual savings. We estimate the actual savings using the following assumptions. The seller attempts to sell a car over the Internet. If a sale occurs, it occurs in a median 5 days.7 In the 76% of cases in which a sale does not occur, the seller decides after 5 days to sell the car through a physical auction process. The car then takes 28 days before it is delivered to a purchasing dealer.8 For these cars, the lessor incurs 5 additional days of interest and depreciation costs that we estimate to be $41. In our sample, therefore, relative to the physical auction process, the Autodaq/Internet process provides a 24% likelihood of a $211 reduction in process costs and a 76% likelihood of a $41 increase in process costs. The net effect is an average decrease in process costs of $19 per car. This analysis highlights that liquidity is important in an Internet market not only to deliver attractive pricing, but also to deliver savings in process costs. Overall, the results in Table 1 indicate moderate reduction in process costs for cars sold using the Internet in our sample. The results suggest potentially substantial reductions in process costs as the Internet market becomes more liquid. Not including reconditioning, the reductions in a liquid market of more than $200 are on the order of 40% of the total economic cost. Multiplied over an annual market of 5 million cars, the analysis implies potential process cost reductions on the order of $1 billion per year. Impact on Coordination costs (2): Marketplace Benefits A second potential benefit of B2B is the extension of the market it provides. Both buyers and sellers can search a larger number of counterparts, and, as a consequence, may find goods and services that they would not otherwise have found. In the case of used automobiles, this seems likely to be an advantage. Used automobile dealers require an appropriate mix of inventory in their dealerships. Obtaining that mix is the main reason they purchase at auctions. In this section, we attempt to estimate the marketplace benefits in the Autodaq/Internet process. Our goal is to assess how much more a dealer would

Business-to-Business E-Commerce

101

be willing to pay in the Internet market (vs. the physical market) for a car that better matches the dealer’s desired inventory. This is not possible to estimate directly because it is not observable. On the other hand, because marketplace benefits are potentially large on the Internet, getting some grasp on the magnitude of this gain is important. In what follows, we propose a simple method that exploits: (1) the geographic rollout used by Autodaq; and (2) a no-arbitrage argument on the seller side. Under reasonable conditions, this method places a lower bound on the dealer’s willingness to pay for access to the larger marketplace. Autodaq’s rollout followed a predetermined pattern. Between the end of October of 1999 and the end of February 2000, the buyers were almost exclusively in California. The sellers, on the other hand, were three large leasing companies that sold cars coming off lease throughout the U.S. Cars sold in California in that period, therefore, included cars from the Southern, Midwestern and Western U.S. The type of sellers implies that the cars were, from their perspective, commodities up to their physical characteristics.9 The willingness to pay by buyers for each car differs widely, as it depends on the quality of the match of the particular car with the needs of the dealership. Suppose that a dealer has a choice between two cars that are from the seller’s perspective identical, but that are valued differently by the dealer because of the dealer’s particular requirements. Suppose, first, that both of these cars can be purchased over the Internet, but one is geographically further away. If we observe a dealer buying a car that is not from California, the dealer must have viewed that car as a particularly attractive match in order to incur the additional transportation costs. From a seller perspective, cars are indistinguishable. The difference in transportation costs, therefore, provides a lower bound estimate of the difference in willingness-to-pay for cars that are purchased from out-of-state. This provides an estimate of the marketplace benefit for those cars. In Table 2, we report the transportation costs for 586 cars sold in the Autodaq Internet auction. The transportation costs are the actual costs paid by the buyers. Table 2 shows that transportation costs average $465 for out-of-state cars and only $223 for California cars, implying a transportation cost differential of $242 per car. It is important to note that we cannot say with certainty how much value was created from this improved matching in our sample. In the extreme, it is possible that the buyer values an out-of-state car at exactly $242 more than an in-state car and pays the entire differential in transportation costs, leaving the buyer with no surplus. It seems reasonable to argue, however, that with a liquid Internet market, dealers in California will be able to buy cars in California over

102

Table 2. Transport Costs Incurred by Buyers.

Transport Cost

Number of Cars

Cars transported from outside of California in Autodaq Auction Cars transported within California in Autodaq Auction Difference

$465 $223 $242

411 175

Estimated Transport Cost in Physical Auction

$110

LUIS GARICANO AND STEVEN N. KAPLAN

Transport costs incurred by buyers in Autodaq Internet auctions from 1999 to 2000. Transport costs are actual transportation costs paid by purchasers. Estimated transportation cost in physical auction provided by Autodaq.com and corroborated by interviews with industry participants.

Business-to-Business E-Commerce

103

the Internet and capture more, if not all, of the gains from improved matching. We also can attempt to estimate the marketplace benefits relative to a physical auction. For a sale to take place on the Internet, the Internet price must be at least equal to the physical auction price less the process cost savings. Thus the sum of the marketplace benefits and the process improvements is at least equal to the difference in transportation costs caused by the additional shipping distance of the Internet auction vs. the physical auction. As before, it may be the case that, at the current stage of development of the Internet, the total increase in surplus is small, if the transport costs are equal to the efficiency gains. In our sample, we can estimate differential transportation costs caused by the additional shipping distance. Autodaq and industry analysts we spoke to estimated that the buyer pays roughly $110 to transport a car it buys from a physical auction site to its dealership. As we reported above, Table 2 shows that average transportation costs are $465 for cars transported from out of state to dealers in California. This suggests that the average car from out-of-state purchased on the Internet is transported a much greater distance than the average car purchased at a physical auction. The extra transportation cost of $355 suggests that the sum of the marketplace benefit and process cost reductions exceeded $355 on average for out-of-state cars sold on the Internet.10 Again, in a more liquid market, the distance required to obtain improved matching should decline, and more of the benefit should accrue to buyers and sellers. Overall, our results suggest substantial marketplace benefits to the Internet auction in the wholesale used car market. These benefits are potentially of the same order of magnitude as the process improvements. Asymmetric Information in Physical and Internet Automobile Auctions Quality Information in Car Auctions While in a physical auction a buyer can obtain an independent indication of the condition of the car (by self-inspection), the Internet auction relies exclusively on information that can be observed in the database. As a consequence, informational asymmetries between sellers and buyers may be more pronounced in Internet auctions. In the wholesale used car market, however, the potential informational loss may be small as information in physical auctions is usually restricted. In describing the physical auctions, Genesove (1993) points out that bidders have limited access to the cars:

104

LUIS GARICANO AND STEVEN N. KAPLAN

Prior to the bidding, the car is parked outside, where potential bidders can examine its exterior. They are prohibited from opening the doors or raising the hood. Mileage and options are chalked on the car’s windows. When the car’s turn approaches, it is driven into the appropriate lane and then, before bidding is concluded on the previous car, driven up to the auction block. Now the hood is raised and dealers are permitted to enter the car. There is time to check the odometer, to ensure that the air conditioner works (but, in the summer months at least, not the heater) and to take a look at the running motor. But there is no opportunity to test the brakes or any number of other things that a consumer might check out in a drive around the block (. . .) On top of the auction block stands the auctioneer and, beside him, the seller, who under the rules of the auction must be present. The auctioneer announces any major defects in the car, of which the seller has informed him. Bidding is oral and ascending. When bidding will go no higher, the seller is asked to accept or reject the winning bid. About 60% of the time he accepts. The car will have been driven away before the bidding is concluded. From the time it arrived at the auction block until the time it is driven away, a minute and a half will have passed.

Internet-based auctions such as those run by Autodaq, on the other hand, do not allow any physical inspection of the cars by the buyer.11 Instead, the seller and the third-party inspection made available by Autodaq provide extensive information on the car’s options and all other measurable aspects of the car condition, such as its mileage, the damages suffered, age, etc. Importantly, Autodaq does not preclude buyers and sellers from participating in physical auctions. This raises the possibility of sellers offering only those cars that are in a relatively worse unobservable condition through this channel. Possibly attenuating adverse selection in our data is the fact that Autodaq is primarily directed at lessors and fleet owners. Individual used car dealers have only recently started selling cars. Only 571 out of 3552 cars auctioned in our sample, and 111 out of 864 cars sold where auctioned by a dealership. To understand the implications of the coexistence of these two markets, we take as our starting point a variant of the simple model of adverse selection of Akerlof (1970). Suppose that, conditional on all the observable characteristics, there are two types of cars, G (good) valued by consumers at PG and lemons L, valued at PL with the proportion of good cars sold in a particular market given by q.12 First, consider the physical market. Assume there is no asymmetric information in the physical market, so that good cars can be sold at price PG and lemons at price PL there. Suppose the higher cost of the physical market mechanism is C, so that the value of the sale to the seller is PG ⫺ C if the car is not a lemon, and PL ⫺ C if it is. The average price of cars sold in the physical market is then Pp = q PG + (1 ⫺ q) PL. Now introduce a competitive electronic market. Here, both classes of cars cannot be distinguished, as consumers cannot physically inspect the cars. There

Business-to-Business E-Commerce

105

are two types of outcomes in this market, depending on the cost of the informational asymmetries relative to the benefit of using an electronic market medium: (1) When the cost of the physical market mechanism is high enough relative to the asymmetric information costs, so that the average price is higher than the net profit from selling a known good car in the physical market, i.e. if q PG + (1 ⫺ q) PL > (PG ⫺ C) or, equivalently, C > (1 ⫺ q)(PG ⫺ PL), both types of cars are sold in the electronic market, at a price Pe = q PG + (1 ⫺ q) PL. In this case, the ratio of average physical market price to electronic market price is 1. (2) If the cost imposed by the presence of lemons on the sellers of good cars is higher than the gain from using an electronic market C, or formally if C < (1 ⫺ q)(PG ⫺ PL), no transactions of good cars take place, as good cars are withdrawn and sold in the physical market. In this case, adverse selection exists in the electronic market. The observed average market price, reflecting the lower average quality of cars transacted, is Pe = L. Adverse selection translates in this case to the withdrawal from the electronic market of cars with relatively good unobservable characteristics within each class in favor of the physical world auction. If adverse selection is present, we would expect to see a lower average price, conditional on observable characteristics, for cars sold over the Internet. High quality cars for each level of observable characteristics would have a low probability of being sold,13 given that the seller would demand high average prices for the average condition that buyers expect to find in the market. Apart from this implication for relative price levels, adverse selection also has implications on the price structure. If adverse selection is important, Internet prices will be lower relative to physical world prices when adverse selection risk is larger. When a low risk of adverse selection exists, i.e. when the variance in the condition of cars is small, the difference between the physical world and the Internet prices will be small. On the other hand, when the adverse selection risk is high, this spread will be large. This different risk is to a large extent predictable. The variance in the unobservable condition of the car is largely a consequence of the unobservable care by the owner. Thus the more that the quality of care affects the value of the car, the larger the risk of adverse selection.14 Genesove’s (1993) study of adverse selection in used car markets is the most notable precedent for our research. He tests for adverse selection by analyzing the effects of the identity of the seller on prices. He expects systematic differences between the incentives of used and new car dealers to sell used cars

106

LUIS GARICANO AND STEVEN N. KAPLAN

to show up in differences in prices if they are selling different quality cars. Our study differs from his in that, rather than examining adverse selection in one market, our focus is on comparing adverse selection in two different markets where we expect, a priori, to find different degrees of informational asymmetry in them. However, we rely on Genesove’s insights to examine the extent of adverse selection in the Internet market in itself. Data Our sample consists of 3552 sold and unsold cars on auction at Autodaq for a period in 1999 and 2000. These are all of the cars that were auctioned by Autodaq at least once in the period we study, except for those that were withdrawn by their owners without completing a three-auction cycle.15 For most of our sample period, all cars were put through a maximum of three oneday auction cycles.16 The construction and content of most of the variables in our sample is selfexplanatory. One exception is the ratio of Internet to physical price, which is intended as a proxy for “how much a buyer would have been willing to pay for this car in the physical world.” We construct this as the ratio of two variables: the price at which a car was actually sold; and the price at which an average car with similar observable characteristics was sold in that month in physical auctions. Autodaq provided these estimates using as a complete data set of physical auction sales. While matching these prices generates rich information, the inferences we can draw from the matches are limited, because the matches are not as precise as we would wish. In particular, the algorithm takes into account motorization, drive (2-wheel, 4-wheel), style, model, and model year, but does not differentiate the matches by mileage and option data. For this reason, part of our analysis also relies on the wholesale Kelley Blue Book (KBB) prices of the cars in the sample. The KBB is an industry guide of wholesale and retail prices for vehicles. The KBB price uses the full physical observable characteristics of the car (year, make, model, series, engine, drivetrain, options, mileage, etc.). We do not use the KBB for any analysis of price levels, but only for our analysis of the changes in relative prices in response to changes in the cars’ physical characteristics. Descriptive statistics for our sample are given in Table 3. The table shows that 24.3% of the 3552 cars offered for sale on the Internet were sold over the period at an average sell price of roughly $13,600. Estimated physical auction values for the cars were available for a subsample of 3001 observations. The average physical auction value for these cars was $14,200, while the average KBB value for these cars was substantially higher at $15,500. According to

Descriptive statistics for Autodaq Internet auctions from 1999 to 2000. Sold equals one if a car put up for sale in an Autodaq auction was sold. Internet price is the price the car sold for in the Internet auction. Estimated physical auction value is the value the Autodaq estimates the car would have sold for in a physical auction based on data from physical auctions. Book price is the price of a similar car according to the Kelley Blue Book. Variable

Obs

Mean

Std· Dev·

Vehicle Sold Internet Price [$1000] Estimated Physical Auction Value [$1000] Blue Book Price [$1000] Mileage (1000) Age (2000-year) Dollar of damages (1000$) Auction Date Seller = Rental Company Seller = Individual Dealer Seller = Leasing Company

3552 865 3001

0.24 13.6 14.2

0.43 5.3 5.6

0.00 3.1 0.0

1.00 32.7 74.9

1.43 1.50 0.20 0.13 0.46 0.20 0.48

196.00 0.00 0.00 0.67 0 0 0

10.83 7.00 2.22 2.10 1 1 1

3552 3552 3552 616 3552 3552 3552

3.57 2.15 0.128 1.04 0.30 0.05 0.65

Min

Business-to-Business E-Commerce

Table 3. Descriptive Statistics.

Max

107

108

LUIS GARICANO AND STEVEN N. KAPLAN

Autodaq and the industry sources we spoke to (including one competitor), the roughly 10% differential is an industry standard.17 Share of damages is the ratio of the estimated dollar value of damages suffered to the book value of the car. Our main objective is to use the data to measure the importance of informational asymmetries in Internet auctions. We try to do this using three elements in our data: the difference in price levels between the Internet and the physical world, the structure of relative prices, and the actual probability of sale of individual cars on the Internet. We expand on our use of these three pieces of information below. Empirical Specifications and Results (i) Relative Price Levels. Assuming that the physical and Internet markets are competitive, we expect to see lower prices relative to the physical market when the quality of the cars sold on the Internet is worse than the quality of the cars sold in the physical real world. We can directly test this implication by comparing the average prices attained by the auctioned cars in the Internet market with the average price they would have attained had they been auctioned in the physical market. Table 4 presents this test. The data in the table reject the hypothesis (at the 95% level) that cars attain lower prices over the Internet than they would have attained in the physical world. In fact, the data suggest that the Internet prices are significantly higher than the prices in the physical world. There are two caveats to this interpretation. First, the price in the physical market is for an average car within a model-year-motor-drivetrain cell. A finding that the average price of a car in the Internet is higher than in the physical market could be compatible with adverse selection, if cars in the Internet have less mileage for a given number of years driven, or if they have more options. Regrettably, the micro-data on the physical market do not allow us to draw such distinctions. Second, the differences in average prices may respond to factors other than the average quality of the car sold. In particular, given the lower transaction costs using the Internet, dealers may be willing to pay higher prices for Internet-based transactions. For this reason, we think the evidence on the structure of relative prices in both markets is a better gauge of the extent to which adverse selection matters. We turn to this issue next. (ii) Relative Price Structure. As noted above, adverse selection between the physical market and the Internet is more likely to be a problem the higher the proportion of “lemons” and the lower their value. Accordingly, if adverse selection is a problem, the effect should increase as the variance of a car’s condition increases. As the variance of a car’s condition increases, the adverse

Business-to-Business E-Commerce

109

Table 4. Price Levels: Internet Auctions vs. Physical Auctions. Price level of internet auction relative to physical auctions Autodaq Internet auctions from 1999 to 2000. Internet price is the price the car sold for. In the Autodaq auction. Estimated physical auction value is the value the Autodaq estimates the car would have sold for in a physical auction based on data from physical auctions. The data correspond to a matched sample of internet sales with the price of an average car of identical model year, motorization and drive in physical auctions. ** is significantly larger than 1. Model Year of Manufacture

Internet price/ physical auction value

N (number of Internet Cars)

95% Confidence

Interval

1995

1.176** (0.035)

31

1.104

1.247

1996

1.114** (0.022)

80

1.07

1.158

1997

1.018** (0.005)

418

1.008

1.028

1998

1.049** (0.014)

42

1.02

1.078

1999

1.062** (0.021)

23

1.019

1.105

selection should manifest itself in a larger decrease in the Internet price than in the physical world price. We test for adverse selection by assuming that conditional on model characteristics, the variance of the value of a newer car, of a car with low mileage or of one with a good observable condition is likely to be small. In contrast, the variance of the value of an older car, or one with more miles will vary more depending on the care taken by its user. In other words, if the quality of care can only be (partially) gauged from direct observation of the car, the cars in the second category, for which care is likely to matter more, will be subject to relatively more important adverse selection problems over the Internet. To test this hypothesis, we could turn again to the most direct data available, i.e. the price on the Internet relative to the price in the physical world. Because of the limitations in matching Internet prices to physical auction prices, we turn

110

LUIS GARICANO AND STEVEN N. KAPLAN

to the KBB prices as our proxy for the price that the car would have attained in a physical auction; these prices do differentiate cars by mileage, condition and options. As long as the relation between the KBB price and the price that an average car with the same characteristics would have obtained in the physical auction is constant, this is appropriate for our purposes.18 We test the hypothesis that as the variance in the condition of the car increases – as given by mileage, age and% of damages suffered by the car – the Internet price relative to the physical price decreases. The first set (a) of specifications in Table 5 presents the OLS evidence conditional on the car being sold. The evidence on the existence of more adverse selection in the electronic auction is mixed. Holding constant the KBB price, a car does not lose significantly more value in the Internet than in the physical market. Similarly, higher mileage does not appear to decrease the value of the car more in the Internet than in the physical market. In fact, the Internet price declines significantly more slowly with miles than the KBB price would predict, rejecting the hypothesis of adverse selection. On the other hand, each car-year reduces the Internet price by around 2% points more than the KBB price, suggesting that this could be a mechanism through which adverse selection is observed. A problem with those specifications is that the Internet sample is censored by the reserve price, as we do not observe transaction prices for unsold cars. For this reason, the second set of specifications (b) in Table 5 repeats the analysis using all of the observations in the sample, with a censored normal regression where the reserve price is the censoring point. The evidence in favor of adverse selection in these specifications is also weak. Including seller fixed effects, the age and the mileage do not appear to reduce the price that a car could attain on the Internet relative to the price that would attain in a physical world auction as measured by the KBB price. (iii) Identity of the Seller. Most cars sold in the Autodaq auctions are sold by leasing companies or rental car companies. For the last part of our sample period, however, individual dealers sold cars on the Internet. Following Genesove (1993), we exploit the difference in incentives between the three types of sellers (dealers, leasing companies and rental car companies) to uncover evidence of adverse selection. We expect adverse selection to be most important for cars sold by individual dealers. Dealers have greater incentives and opportunities: (1) to check the quality of care and condition of each individual car; and (2) to select those cars to sell on their lot and those to sell on the Internet. As a consequence, after controlling for physical characteristics of these cars, we expect the identity of the seller of the car to matter. Individual

Regressions for Autodaq Internet auctions from 1999 to 2000. Internet price is the price the car sold for. Book price is the price of a similar car according to the Kelley Blue Book. Independent Variables

Dependent Variables

Log Blue Book Price

Mileage (000)

Damages ($1000)

Age

Auction Date

(a) Log Internet Price (OLS)

1.0505** (0.0076)

0.0077** (0.0026)

0.0000 (0.0000)

–0.0154** (0.0039)

0.0003** (0.0000)

1.0502** (0.0078)

0.0091** (0.0026)

0.0000 (0)

–0.0209** (0.0043)

0.0001 (0.0001)

(b) Log Internet Price (Censored Normal regressions)

1.0137** (0.0072)

–0.011** (0.0025)

0.0001** (0.000)

0.0383** (0.0034)

0.9945** (0.0071)

–0.0038 (0.0024)

0.0001* (0.000

–0.0039 (0.0039)

(c) Log Internet Price (OLS)

1.0615** (0.0075) 1.0514** (0.0077)

(d) Probability Car Sold (Probit)

Seller = Individual Dealer

Seller = Rental Company

yes

0.0002** (0) 0.0001* (0.0001)

0.0084** (0.0025)

0.0000 (0.0000)

–0.0205** (0.0042)

0.0002** (0)

–0.03 (0.0238)

–0.0005** (0.0001)

0.2963** (0.0308)

–0.0022** (0.0003)

0.0066 (0.0246)

–0.0002* (0.0001)

0.1734** (0.0408)

–0.0069** (0.0007)

yes

yes

Constant

R Sq (N )

–0.2924** (0.028)

0.9628 (843)

–0.2475** (0.0342)

0.9637 (843)

–0.3141** (0.0251)

(3147)

–0.1175** (0.0324)

(3147)

0.0978** (0.0103)

0.0489** (0.016)

–0.3083** (0.0203)

0.9604 (843)

0.0384** (0.0141)

–0.0389* (0.0193)

–0.2759** (0.0288)

0.9636 (843)

–0.8456** (0.1073)

(3552)

0.539* (0.2174)

(3552)

111

Standard errors in parenthesis. * Significant at 5% level; ** Significant at 1% level.

Seller Fixed Effects

Business-to-Business E-Commerce

Table 5. Performance of Internet Auctions.

112

LUIS GARICANO AND STEVEN N. KAPLAN

dealers should obtain lower prices for their cars, holding everything else equal. The effect of rental car companies is more ambiguous. Holding all else constant, rental car company cars have been through many more users. Their unobservable quality should be lower and, as a result, their average price should be lower. On the other hand, selection should be less important for rental cars, as rental car companies have a policy of selling all their cars after some fixed period of time. The set of specifications (c) in Table 5 tests these hypotheses. The regressions reject the hypothesis that individual dealers are perceived to sell lower quality cars over the Internet than institutional sellers. The first column controls only for the book price of the car, and shows a significant effect of individual dealer on price, but exactly of the opposite sign as the one predicted by the theory. Controlling for a car’s physical characteristics, the effect decreases, but the dealer effect is still positive and significant. Holding all physical characteristics constant, a car sold by a individual dealer earns a premium over the KBB price that is 4% higher than that of the leasing companies (the excluded category). The lack of evidence of adverse selection in the dealer market is important for another reason. One might argue that our results for lessor or fleet sales are biased because: (1) the buyers know the identity of the sellers; and (2) lessor sellers get good prices for their cars in general. The similar results for the dealer market suggest that our results are not biased for this reason. (iv) Probability of Sale. A final hypothesis concerns the probability that a car is sold. This contains no information on the comparison of adverse selection in the Internet market relative to physical markets, but may contain some information on whether adverse selection exists at all. Clearly, adverse selection implies that cars with good (online) unobservable condition should be relatively less likely to sell on the Internet market. For this difference to be translated into an actually lower probability of sale, however, it would be necessary that cars in relatively good unobservable condition do ‘show-up’ in the Internet market, likely with a higher reservation price, only to be later withdrawn from auction. In our data, it is possible to assume that sellers initially attempt to sell all cars that they are planning to sell in auction on the Internet. First, that is the arrangement between the firm in our study and the sellers. Second, the sellers can choose a high reservation price for even their best-conditioned cars. As a result, there is at worst only a small opportunity cost of trying the Internet.

Business-to-Business E-Commerce

113

Accordingly, if adverse selection is important on the Internet, we expect that the greater the variance in a car’s condition (which would increase the incidence of adverse selection), the lower the probability that the auction is successful. In our sample, cars with more miles, higher age, and a history of more accidents in the past would have a lower probability of sale if adverse selection exists. The alternative hypothesis is that adverse selection is not a particular problem in Internet markets. A reason for this in Autodaq’s case is that individual dealer sales over the Internet were limited and under stringent conditions. To the extent that such dealers are more likely to try to dump their lemons, it may actually be that the Autodaq market is more, rather than less, efficient than the physical auction market. The final set of specifications (d) in Table 5 present the analysis of the probability that a sale actually takes place using a Probit model. The evidence is inconsistent with the existence of important adverse selection in these markets. Damaged cars do seem to be somewhat less likely to be sold, but neither older cars nor cars with more miles are less likely to be transacted. In fact, there is a significantly positive effect of the age of the car on the probability of a sale. Caution must also be exercised in interpreting these results. Only if dealers do not withdraw their ‘good condition’ cars prior to sale, but rather post them with a higher reservation price, do we expect to see adverse selection manifested in a lower probability of sale for cars more affected by adverse selection. (v) Is Adverse Selection a problem over the Internet? Overall, we find little evidence consistent with the hypothesis that adverse selection is more pervasive in the Internet market than in the physical world, or even that adverse selection is a problem at all in the Internet marketplace we study.19 As we observed before, this conclusion necessarily must be qualified by the measures that Autodaq has taken to reduce the incidence of adverse selection in this particular instance.

4. FROM VALUE CREATION TO VALUE CAPTURE: THE RISE AND FALL OF B2B VALUATIONS It is well-known that publicly-traded Internet firms achieved levels that were extraordinary by most standards. For example, Ofek and Richardson (2001) show that in the aggregate, Internet firms traded at roughly 35 times revenue at the end of 1999. If those firms had achieved industry-average net income margins at the time, they would have had price-earnings (P/E) ratios of 605.

114

LUIS GARICANO AND STEVEN N. KAPLAN

Ofek and Richardson (2001) also estimate the growth rates that would have been required to justify such high P/E ratios and find that such rates are extremely high by historical standards. Cooper et al. (2001) find that firms that announce name changes to include “dotcom” experience abnormal returns of 74% over this period. In the two years since the end of 1999, Internet valuations have declined precipitously. From February 2000 to December 2000, Ofek and Richardson report that the value of these firms declined by an average of 80%. That decline has continued in the subsequent months. In this section, we discuss what the market appears to have believed when Internet valuations peaked. we then use the framework of the previous section to discuss why those beliefs turned out to be so wrong. Why Were Valuations So High? Valuations of B2B e-commerce were based on very aggressive growth assumptions. One B2B e-commerce firm, Chemdex, attained a market capitalization of $11 billion with $2 million of true revenues. Rajgopal et al. (2000) also find that B2B valuations related to alliances, acquisitions, customer acquisition, but not to earnings. The rational story for these companies is that investors assumed that: (1) the businesses delivered large reductions in transaction costs; (2) business customers would adopt quickly, i.e. a large volume of activity would move to the internet; (3) competition would be slow and network effects would emerge; and (4) the B2Bs would be able to capture a meaningful portion of transaction cost savings. Why Are They So Low Now? Why have the valuations of Internet companies decline so precipitously since March 2000? Clearly, the market’s expectations of growth have declined a great deal. Ofek and Richardson (2001) argue that part of the reason for the decline was an increase in the number of selling shareholders driven by expiring lockup agreements. In this section, we present some additional thoughts concerning the downward revisions in growth expectations. For B2B as well as business to consumer (B2C) e-commerce businesses, the market greatly reduced its expectations of (some combination of) future growth, of the extent of transaction cost reductions, the ability to capture those reductions, the speed of adoption, the ability to take advantage of network effects, the extent of competition, and (for B2C) the extent to which traffic

Business-to-Business E-Commerce

115

could be transformed into revenues. Is the change in the market’s expectations for B2C and B2B companies surprising? It is worth considering the framework from Section 2. Many B2C companies are simply improved catalogs. Such businesses reduce transaction costs for individual consumers –the Internet can make it easier to find items (like books) and easier to order them (books and stocks) – and for the cataloger – order taking and order fulfillment are less costly. However, this is not an earth shattering change. The introduction of catalogs brought with them transaction cost reductions, but not extraordinary valuations. Catalogs (and brokerage firms) also regularly face competition. It is hard to imagine a rational story for such high B2C valuations for e-commerce companies. One exception is a company like eBay. EBay does provide a service that is not available offline. It also benefits from network effects because it connects many buyers to many sellers. Sellers know they are more likely to find buyers at eBay. That attracts more sellers. Buyers know they are more likely to find sellers at eBay. This attracts more buyers. Buyers and sellers are less likely to make good matches through other companies. As more buyers and sellers use eBay, the advantage of eBay over other companies increases. Consistent with this, eBay’s value has only declined by slightly more than 50% of its peak value. Is the change in the market’s expectations for B2B companies surprising? The extent of the decline in B2B was more of a surprise to us. It was not surprising to see some decline. It was surprising to see a large fraction of these companies fail. Based on the framework, it was more plausible that B2B companies reduced transaction costs substantially. B2B business models also were more likely (than B2C models) to rely on business models that utilized network effects, matching many buyers to many sellers in the way that eBay did. What went wrong? In some markets, companies have obtained transaction cost reductions, but B2B companies have not been able to capture much of this reduction because of competition. This is arguably true in the procurement area where the a number of companies have been able to provide software and procurement processes that are not largely differentiated from each other. Network effects have not materialized in those markets. There was also a belief in a number of markets that B2B companies would be able to charge a percentage of the transaction value, rather than a fixed transaction fee. This reflected a misunderstanding of the nature of the transaction cost savings. In many cases, the transaction cost savings is a fixed amount – time spent punching in data – rather than a percentage of the transaction value.

116

LUIS GARICANO AND STEVEN N. KAPLAN

Finally, in some markets, companies just have not adopted the new technologies. This occurred for two reasons. First, some companies, particularly suppliers, were not interested in using internet marketplaces because they did not want to put an intermediary between their customers and themselves. Second, companies have been able to use the Internet without having to commit. That is, it is possible to use the internet to get price information, but then go to traditional suppliers for execution. Did Sophisticated Investors “Know” Prices Were Too High? Answering whether people knew prices were too high is, of course, very difficult. Ofek and Richardson present evidence and argue that the decline in Internet stocks is related to short sales constraints and the expiration of IPO lock-ups. They argue that the rise and fall of Internet stocks can be explained by an initial relative oversupply of optimistic investors who drove prices up followed by the arrival of more pessimistic investors – insiders – who drove prices down. The Ofek and Richardson story suggests that sophisticated investors – like venture capitalists – believed a bubble existed. While this story is plausible, there are some pieces of evidence that are not consistent with this explanation. At the same time that venture capitalists were some of the insiders who sold shares after lock-ups expired, the venture capitalists also sharply increased the amount of money they raised and the pace of their investments in new Internet and technology related start-ups. Figure 1 shows the large increase in funds committed to VC funds while Fig. 2 shows the huge increase in investments by VCs in 1999 and 2000. Much of this investment went into New Economy investments. Hendershott (2001) documents a similar pattern for pure Internet investments. Presumably the VCs who made these investments believed that the investments would be profitable on average. To believe the investments would be profitable, the VCs must have believed, on average, that the companies they invested in would be viable and valuable. In other words, such a large increase in investment seems inconsistent with a pessimistic view of the New Economy companies. Furthermore, the VCs received most of their capital commitments from large institutional investors – pension funds, endowments, etc – who also must have been optimistic about these investments. One might argue that the VCs and institutional investors made these investments with the expectation of flipping their private investments to

Business-to-Business E-Commerce

Fig. 1. Fundraising by Venture Capital Partnerships 1980–2001. (in $ billions) Source: Private Equity Analyst.

117

118

Source: Venture Economics.

LUIS GARICANO AND STEVEN N. KAPLAN

Fig. 2. Venture Capital Financing 1990–2001.

Business-to-Business E-Commerce

119

irrational public investors. This argument, however, would require the VCs to have believed that stock prices would remain irrationally high for at least two years. That is, even under optimistic conditions, it still would take that time for the VC to invest in an early stage company, take it public, wait for the lock-up period to end, and then sell the shares. This argument also runs into difficulty in that it assumes that the investors in public securities would be irrational. Yet, a substantial number of investors in public securities were the same institutions who invested in the VC funds. Figure 3 sheds some light on this. Figure 3 presents a time series of VCbacked IPOs and first VC round investments (based on data from Venture Economics). First VC round investments provide a measure of the number of new companies backed by VCs. VC-backed IPOs provide a measure of the number of VC companies that succeed. Figure 3 shows that it was reasonable for VCs to assume there would be 200 to 250 VC-backed IPOs per year. At the same time, Fig. 3 shows an incredible increase in VC funded first rounds in 1999 and, particularly, 2000. The large increase in VC investments without a concomitant increase in the number of IPOs is certainly consistent with VCs and institutional investors believing that stock prices would remain high. Figure 3 does leave us with a puzzle. The huge increase in number of companies funded suggests that competition would be a huge problem. Yet it is difficult to justify the high valuations in 1999 and early 2000 without assuming that competition would be modest. Two other observations are relevant. First, buyout investors made large and high profile investments in B2B and other technology companies. For example, Forstmann Little, Hicks Muse, and KKR, among others, invested and have subsequently lost hundreds of millions of dollars in such companies. These sophisticated buyout investors must have believed that the investments had a positive expected value at the time. Second, infrastructure companies like Cisco, Lucent, and others also have lost a large fraction of their values. This is important to mention because their securities were liquid throughout the rise and fall.20 We draw the following conclusion from these observations. Insiders and sophisticated investors – including VCs and some buyout investors – may have believed that many of their individual stocks were overvalued when Internet valuations were high. As a result, they sold shares. At the same, however, those same investors believed that the New Economy companies were viable entities and that there were opportunities to create more New Economy companies. Furthermore, some of these sophisticated investors believed that some of these companies were undervalued – particularly the buyout investors who invested in telecommunications.

120

Source: Venture Economics.

LUIS GARICANO AND STEVEN N. KAPLAN

Fig. 3. VC 1st Rounds Versus VC-Backed IPOs.

Business-to-Business E-Commerce

121

5. CONCLUSION AND IMPLICATIONS Summary In this paper, we present a framework to evaluate the impact of B2B (and other Internet/New Economy) businesses on transaction costs. We apply this framework to one particular business and find that process improvements and marketplace benefits are potentially large – on the order of 5% of the automobile value and a much large fraction of the total transaction cost. Moreover, we do not find evidence that the Internet increases adverse selection costs. We then use the framework to consider the rise and fall of B2B (and other Internet) valuations. High valuations were fueled by beliefs that B2Bs would grow significantly and would deliver larger reductions in transaction costs. There also was an implicit assumption that competition would be weak, possibly because of network effects. Valuations fell as the market began to realize that those beliefs and assumptions would not be validated. We then discuss the implications of the rise and fall of valuations. It is simplistic to argue that smart, informed individuals took advantage of naïve public investors. Sophisticated and previously successful venture capital and buyout investors behaved as if they believed that Internet and New Economy companies would be much more successful than they have been. What Are the Real Effects of the Internet/New Economy Likely to Be? We have seen a boom and then a bust in B2B, Internet, and technology valuations. Stock market investors obtained terrific returns and then horrific ones. In April 2002, the S&P 500 stands at roughly 1100 while the NASDAQ Composite rests at roughly 1750. These are the same levels these indices registered in early 1998. In other words, the stock market has roughly stood still (ignoring modest dividends) overall in the last four years. The results in Hendershott (2001) suggest that the overall return on investment in Internet companies also was roughly breakeven. The question, then, is whether the investments in B2B (as well as the New Economy and technology in general) had a similar negligible effect on the overall economy. It is here that the real effects on the economy need not be the same as the effects on the stock market. It is our sense that the B2B and other related technology investments have generated and will continue to generate substantial improvements in productivity. The favorable productivity numbers

122

LUIS GARICANO AND STEVEN N. KAPLAN

since the mid-1990s and continuing in the recent downturn certainly are consistent with this. The Internet allows companies to substantially alter many of the processes by which they do business. For example, B2B and other technologies allow large reductions in transaction costs in areas like procurement, accounts payable, and human resources. Many of these are labor intensive functions that can be outsourced or automated. Consistent with this, an increasing number of companies move tasks and processes like data entry, simple programming, and call center services from the United States to India and other lower wage countries. Much of this would not be possible without the New Economy investments and technologies. General Electric (GE) provides an interesting example.21 In the late 1990s, Jack Welch challenged his employees to move everything they could to the Internet. They found that while they could not move transactions so quickly to the Internet, they could move a large number of internal and support processes. And they could do so with “simple Web application [software] supported by email.” GE expects that transactions will gradually move to the Internet as software evolves and other companies move more toward the Internet. GE also expects to develop Web-based customer systems that monitor how GE equipment is performing and, therefore, improve the performance of that equipment. We have not attempted to estimate the overall or macro implications of all this. Casual empiricism suggests that there are still a large number of existing processes for which New Economy technology can reduce transaction costs substantially. The implementation of these transaction cost reductions will be gradual as they require some up-front investment and adjustment costs. It is possible, therefore, that the New Economy technology can generate strong productivity increases at the same time that the companies and technologies that enable them do not earn much profit and the corporations that implement them do not earn much additional profit. Competition and the ability to copy drive profits down for the enablers. Competition among the companies that implement the improvements drives prices down for end users. In the end, the end users/consumer benefit as measured by the productivity increases despite the fact that the stock market does not.

NOTES 1. This process is not unique to Autodaq. Several competitors exist. In particular, the largest operator of physical auctions, Manheim, has an Internet based subsidiary – Manheim Online. Manheim Online differs from Autodaq in that it uses the Internet to list the cars that it has for sale at its physical auction site. In its current incarnation,

Business-to-Business E-Commerce

123

therefore, Manheim Online, potentially reduces buyer transaction costs, but does not change seller transaction costs. 2. ADT has subsequently merged with Mannheim, the largest competitor in the physical auction business. 3. For 7,221 cars, we can measure the time from lease end to sale. The median time from lease end to sale in our sample is 36 days (compared to 35 days from inspection). 4. It is important to note that in both the Autodaq system and the physical auction, buying dealers typically perform reconditioning (despite the fact that some reconditioning is performed by the physical auction). It is possible that the reconditioning time is greater in the Autodaq system, although Autodaq claims that this is not the case. 5. For the 270 cars with a lease-end date, the median time from lease end to sale is 20 days. 6. This is conservative as mechanics probably cost more than this. 7. This assumes that the car is inspected 9 days before it is turned in. 8. Again, this assumes that the car is inspected 9 days before it is turned in. 9. In this analysis, we assume that adverse selection problems are absent. Our analysis below confirms that adverse selection is likely not an issue here. 10. This calculation is also conditional on an interstate sale. 11. Both Autodaq and the physical auctions do inspect the cars and describe them for buyers. Autodaq argues (and we agree), that the information in electronic form is richer and more useful as it allows buyers to search more efficiently for their desired cars and options. 12. Consistent with our previous discussion, the entire surplus is captured by the seller. We now ignore the transport costs considerations to simplify the discussion. 13. Note that the seller could just not bring high quality cars to the auction block. Given that the cost of merely posting the car on the electronic market is very low, and that the reserve price can be used to avoid selling it cheap, we can expect even cars with very good unobservable characteristics to be posted. In fact, the opposite is likely to have occurred. Autodaq screened out (i.e. did not list) cars with excessive mileage and excessive known damage. 14. There is another type of adverse selection in this market, unrelated to the quality of care, but related to the quality of manufacturing. This is unlikely to be an issue here for two reasons: first, initial defects are not frequent; and, second, all the models are less than 4 years old so manufacturer warranties typically cover such defects. 15. None of our results are sensitive to including those cars as not sold. 16. This constraint was lifted later in the sample, but to little effect: only six cars were sold after three cycles. 17. The difference between book value and selling price and its magnitude are also present in Genesove (1993). He finds that the book value is an imperfect predictor of the selling price, but does not document any systematic relation between the bias and the age, mileage or other characteristics of the car. 18. Note that the use of the KBB rather than the average physical auction price biases the results in favor of finding that adverse selection is more pronounced in the Internet than in physical markets. This is true because more information goes into nonauction sales in the physical world that are the bases for the KBB than at auctions. Using the estimated physical auction car of an equivalent car rather than the KBB as the counterfactual does not affect our results.

124

LUIS GARICANO AND STEVEN N. KAPLAN

19. In turn, previous studies of physical motor vehicle market by Bond (1982) and Genesove (1993), have found little evidence of adverse selection in these markets. 20. We thank John Cochrane for this observation. 21. The quote and the information in this paragraph are taken from the Wall Street Journal, May 8, 2001, p. A1.

ACKNOWLEDGMENTS This research has been supported by the Center for Research in Security Prices, the Kauffman Foundation, the Graduate School of Business at the University of Chicago and the Lynde and Harry Bradley Foundation and the Olin Foundation through grants to the Center for the Study of the Economy and the State. We thank Severin Borenstein, Judy Chevalier, David Genesove Charles Morris, Rod Parsley, Erik Peterson, Jagadish Turimella, Frank Wolak, and seminar participants at IESE (Barcelona), University Pompeu Fabra, and the NBER E-Commerce project for helpful comments and discussions. Address correspondence to Luis Garicano, Graduate School of Business, The University of Chicago, 1101 East 58th Street, Chicago, IL 60637 or e-mail at [email protected]. ©2002 Luis Garicano and Steven Kaplan.

REFERENCES Akerlof, G. (1970). The Market for ‘Lemons’: Quality Uncertainty and the Market Mechanism. Quarterly Journal of Economics, 222, 488–500. Bond, E. W. (1982). A Direct Test of the “Lemons” Model: The Market for Used Pick-Up Trucks. American Economic Review, 72(4), 832–880. Demers, E., & Lev. B. (2001). A Rude Awakening: Internet Value-Drivers in 2000. Working paper, NYU. Garicano, L., & Kaplan, S. N. (2001). The Effects of Business-to-Business E-Commerce on Transaction Costs. Journal of Industrial Economics, 49(4), 463–485. Genesove, D. (1993). Adverse Selection in the Wholesale Used Car Market. Journal of Political Economy, 101(4), 644–665. Hand, J. R. M. (2000). The role of economic fundamentals, web traffic, and supply and demand in the pricing of U.S. Internet stocks. Working paper, University of North Carolina. Hendershott, R. (2001). Net Value: Wealth Creation (and Destruction) during the Internet Boom. Working Paper, Leavey School, Santa Clara University. Jorion, P., & Talmor, E. (2001). Value relevance of financial and non-financial information in emerging industries. Working paper, UC-Irvine. Milgrom, P., & Roberts, J. (1992). Economics, Organization and Management. Englewood Cliffs, NJ: Prentice Hall. Ofek, E., & Richardson, M. (2001). Dotcom mania: The rise and fall of internet stock prices. NBER Working Paper #8630. Rajgopal, S., Kotha, S., & Venkatachalam, M. (2000). The relevance of web traffic for Internet stock prices. Working paper, University of Washington. Sawhney, M., & Kaplan, S. N. (1999). Let’s Get Vertical. Business, 2.0 (September).

Business-to-Business E-Commerce

125

Trueman, B., Wong, M. H., & Zhang, X. (2001). The eyeballs have it: Searching for the value in Internet stocks. Working paper, UC-Berkeley. Williamson, O. (1985). The Economic Institutions of Capitalism. New York: Free Press.

This Page Intentionally Left Blank

TRUST AMONG STRANGERS IN INTERNET TRANSACTIONS: EMPIRICAL ANALYSIS OF eBAY’S REPUTATION SYSTEM Paul Resnick and Richard Zeckhauser ABSTRACT One of the earliest and best known Internet reputation systems is run by eBay, which gathers comments from buyers and sellers about each other after each transaction. Examination of a large data set from 1999 reveals several interesting features. First, despite incentives to free ride, feedback was provided more than half the time. Second, well beyond reasonable expectation, it was almost always positive. Third, reputation profiles were predictive of future performance, though eBay’s net feedback statistic is far from the best predictor available. Fourth, there was a high correlation between buyer and seller feedback, suggesting that the players reciprocate and retaliate.

1. INTRODUCTION Reputations in a classic marketplace spread overwhelmingly by word of mouth, on a “retail” basis. The news of a dissatisfied customer at a store, or even a dozen dissatisfied customers, is not likely to reach the ears of most potential customers. There are some coordinated rating agencies in regular markets; one example is Zagats. Even though Cousin Larry may be more reliable than the The Economics of the Internet and E-Commerce, Volume 11, pages 127–157. © 2002 Paul Resnick. Published by Elsevier Science Ltd. All rights of reproduction in any form reserved. ISBN: 0-7623-0971-7

127

128

PAUL RESNICK AND RICHARD ZECKHAUSER

unknown masses who provide ratings to Zagats, the guide has an edge (at least for many diners) because it reports on dozens of opinions on each of hundreds of restaurants. The difference between Zagats in the bookstore and Zagats on the web is also instructive. The book costs money, but most all of the same information is available on the web for free. With the web Zagats, one can instantly get information on restaurants in far-away cities, whereas printed Zagat books on distant locales are harder to come by. While one wordof-mouth review reaches perhaps dozens, Zagats books reach thousands, and Zagats on-line reaches hundreds of thousands. The potential of the last rating system – which capitalizes on the extraordinary reach of the web – has not begun to be met for most goods and services. By virtually any metric, orders of magnitude more people are connected to each other, and communicate cheaply with each other, than at any time in history. However, many of these communications are among strangers, people who do not know each other before they receive a communication, learn little about each other from the communication, and do not encounter each other again. Where such encounters involve merely mail or chat room messages, such exchanges are not surprising. Risks are small so not much trust is required. What is surprising is the vast shuttling of both new and second hand goods among distant strangers on the Internet, through such mechanisms as eBay and the Yahoo auction site. Buyers, who must pay before inspecting or receiving their items, must put considerable dollars at risk. This paper seeks to explain why buyers trust unknown sellers in this vast electronic garage sale. For data, we shall be drawing on all the transactions on the eBay auction site from February through June 1999. The scale of these operations is impressive. eBay’s site now boasts an average of more than five million active auction listings and though there has been growth over time our complete data set consists of millions of items.1 How is trust traditionally created when goods are exchanged? We identify eight factors; readers would add more: (1) Most retail transactions are conducted locally, which gives individuals the opportunity to inspect them, as say with fruit in a rural market. If quality is discernible, no trust is needed; (2) Retail operations tend to be large relative to their local market, be they vegetable sellers or the local department store. Buyers have frequent interaction with the same seller, and learn whom they can trust; (3) Even when one’s personal interactions are limited, given that a retailer’s sales are concentrated in a locale makes it easy to develop reputations so customers learn about retailers from their peers;

Trust Among Strangers in Internet Transactions

129

(4) Retailer reputations are borrowed from other contexts. For example, retailers are likely to be pillars of the church and community, and would be highly reluctant to sacrifice the status that comes from such reputations.2 (5) Reputations are built over many years; at an extreme, witness the reputations of Rothschild and Sons and the Bank of Scotland, leading financial institutions, which have built reputations over hundreds of years; (6) Reputations are borrowed from others. Thus celebrities will attest to the quality of products; (7) New goods benefit from established brand names, and policing of quality by those who own them. The product, not the retailer, wins the reputation; (8) Significant expenditures – e.g. building a fancy store on Manhattan’s Fifth Avenue3 – indicates that one will be reliable, lest this expenditure be wasted, a form of signaling. Internet auctions have none of these mechanisms available. Sellers are not met, and little or nothing is known about their characteristics, or even their location beyond their city. Customers rarely repeat, and they do not run into each other. Putting items on the web is a cheap activity. Some goods that are traded are not brand name, and when they are, there is a risk of being counterfeit. Measured in relation to the age of significant retail operations, all of the sellers are new. No one attests about the sellers. Firms like eBay do not stand behind their auctioneers. Yet millions of transactions have taken place. Cheap collection, distribution, and computation, however, can accomplish a lot. If one has a bad experience in the village, one may tell one’s friends, at a cost of a few minutes for each telling. A bad experience with an Internet seller can be recorded in less than a minute, and spread to millions of potential customers. Each of those potential customers is much less likely to ever encounter the blameworthy seller, but they will have the relevant information on a seller when they need it. Through computational power, vast amounts of even snippets of information can be aggregated to inform buyers about patterns of seller trustworthiness. A reputation system collects, distributes, and aggregates feedback about participants’ past behaviors. The task of this paper is to determine, as best as is possible with data provided by eBay, how the system is working. Resnick et al. (2000) identify three challenges for a reputation system. It must: (1) provide information that allows buyers to distinguish between trustworthy and non-trustworthy sellers; (2) encourage sellers to be trustworthy; and (3) discourage participation from those who aren’t. In the terminology of asymmetric information, the second and third criteria are that a reputation system must deter moral hazard and

130

PAUL RESNICK AND RICHARD ZECKHAUSER

adverse selection on the part of sellers (e.g. Milgrom & Roberts, 1992, Chaps 5 and 6). The eBay reputation system is applied to buyers as well. However, buyers’ reputations matter substantially less, since sellers can hold goods until they are paid. The greatest risk is that they will not get paid, in which case they can turn to the second high bidder. Moreover, even if sellers wished to rely on buyers’ reputations it would do little good, since it is not possible to exclude buyers with bad reputations from one’s auction.4 It is worth noting at the outset that the system need not be theoretically sound in order to work. It may only be necessary that both buyers and sellers believe that the system or some part of the system works. There is little published literature on the effective workings of reputation systems on the Internet, so it seems extremely unlikely that many participants are aware of frequency of feedback, disproportions in feedback among those having positive and negative experiences, etc. What matters therefore, is not how the system works, but how its participants believe it works, or even whether they believe it works even if they have no concern about why. To invoke an analogy drawn from grander considerations, the behavior of man in a world without a God might be fully moral and God-fearing if its denizens believed there was a God who would judge them and possibly punish them in the hereafter. Let us illustrate how a system might deter moral hazard and adverse selection, even if it did not allow buyers to actually distinguish trustworthy from untrustworthy sellers. Suppose that negative feedback is rarely given even when buyers are dissatisfied, suggesting that the system would not work if this fact were widely known. Say that new sellers have to pay an entry or initiation fee in terms of reduced prices and reduced frequency of sale/item listed. If unreliable sellers know that they will have to pay their dues at the outset, and if they believe that the feedback system is likely to give them poor ratings, they will be deterred from participating. They will not make the investment of entering in the first place. We begin our analysis with general characteristics of the eBay marketplace and reputation system. Section 2 describes the mechanical workings of the feedback system. Section 3 describes the data sets used in the remainder of the paper. Section 4 presents the distribution of feedback profiles among buyers and sellers and examines the extent of ongoing trading relationships versus single transactions. In Section 5, we address the provision of feedback. Information about the behavior of others is a public good, and whether the information is bad or good, there is little incentive to provide it. Even keystrokes are costly. To counteract underprovision, small payments could be provided to raters. Avery et al. (1999)

Trust Among Strangers in Internet Transactions

131

explored alternative mechanisms for organizing such a market for evaluation. However, even without such payments, we found empirically that buyers left feedback for more than half of the transactions. In Section 6, we examine whether feedback profiles are informative to buyers. Sellers might manipulate their profiles by handling small transactions honestly but cash in their reputations on high value items. Or they might ask friends to help artificially inflate their profiles. eBay’s summary statistic, the number of positive feedbacks minus number of negatives, is highly inappropriate in an information theory sense, because it underweights negative feedback, which is quite rare. We find, however, that a seller’s prior feedback profile as a whole is somewhat predictive of future performance. Section 7 turns to incentive effects in deterring moral hazard. We do not consider here whether better reputations are rewarded with price premiums. A number of recent studies have investigated this question through observational studies (e.g. Ba & Pavlou, forthcoming; Bajari & Hortacsu, 2000; Dewan, 2001; Eaton, 2002; Houser & Wooders, 2000; Kalyanam & McIntyre, 2001; Kauffman & Wood, 2000; Lee et al., 2000; Lucking-Reiley et al., 2000; McDonald & Slawson, 2001; Melnik & Alm, forthcoming). Resnick et al. (2002) recently conducted a controlled field experiment, but no clear consensus has yet emerged from these studies about the effect of reputation on seller profits. We do examine one bit of evidence about the incentive effects for sellers, their propensity to monitor their own feedback and respond to negative comments. Section 8 then returns to questions of feedback provision. Assuming feedback is provided, will it be without bias? The disincentive to provide negative information may be far stronger, with the potential for lawsuits, and for retaliatory negative feedback.5 Once again, the reality may not matter. Even if there has never been a lawsuit, participants may feel the threat.6 Beyond these concerns, most people do not like to provide negative feedback. And there may be a bias to the positive side, particularly if the seller goes first, if the feedback process is basically an exchange of courtesies: “You behaved well. You also behaved well.” But why should anyone consider providing feedback at all? Why shouldn’t they merely free ride? We suspect that many people do it as part of some quasicivic duty. It is an encouraged activity, and does not cost much. Others do it as a courtesy. They have had a successful transaction and want to say thanks. Some expect reciprocity. Indeed, numerous sellers communicate with buyers that they always provide feedback for a successful transaction, and they hope the buyer will do so as well. These reasons do not apply, or apply much less forcefully, when an experience was bad. Bad evaluations, theory would

132

PAUL RESNICK AND RICHARD ZECKHAUSER

suggest, are much less likely to be given, unless revenge is a strong motivating force.7 On the other hand, a phenomenon of stoning may emerge, where buyers are much more willing to give negative feedback to sellers who have recently received other negative feedback. In ongoing work with Miller (2002), we are considering how payments tied to the quality of evaluations (as measured by agreement with future evaluations) might overcome these natural distortions in the elicitation of feedback.

2. THE eBAY FEEDBACK SYSTEM Before turning to our actual data, it is worth looking at how the eBay reputation system actually works. Before participating in any auctions, as either buyer or seller, people (henceforth called users, as in users of the auction service) register, providing information to eBay, including name and contact information. The only information that eBay verifies, at least for buyers, is that the email address is valid.8 Not all of the information provided to eBay is made visible to other users. In particular, as part of the registration process, the user chooses an online pseudonym, or ID. This, rather than the full name provided to eBay, is shown to other users when buying and selling. Some users choose to use their email address as their pseudonym but many others choose other short monikers. Other users can request the email address associated with any pseudonym, but must reveal their own email address in order to do so. eBay does not reveal the real names and physical addresses provided during registration to other users. Since there are many ways to sign up for free email accounts at services like HotMail and Yahoo, this system means that anyone who wants to remain anonymous has the option to do so. Buyers and sellers can leave comments about each other after transactions, but are not required to do so. Each comment consists of one line of text, plus a numeric rating of + 1 (positive), 0 (neutral), or ⫺ 1 (negative). Initially, any user could leave feedback about any other user. Beginning in February, 1999, all negative feedback had to be tied to a particular transaction. Beginning in February, 2000, all feedback had to be tied to a transaction: i.e. only the seller and winning bidder can leave feedback about each other. Users have the option of whether to make their feedback visible to other users or not, but this decision must be made about the entire feedback profile, not on a case by case basis. By default, the feedback profile is publicly visible, and we have never encountered a user who had chosen to hide their feedback profile. When a buyer searches for items, she sees a summary listing of items, with item titles and current prices, but not the sellers’ pseudonyms or any indication

Trust Among Strangers in Internet Transactions

133

of their prior feedback (Fig. 1). After clicking through to the item’s full description page, the seller’s ID and a summary feedback score are displayed (Fig. 2). The feedback summary statistic is the number of unique (distinct) users who left positive feedback, minus the number of unique users who left negative feedback. If the buyer chooses to click on the summary feedback score, she sees details of the seller’s feedback profile (Fig. 3). This includes a breakdown of positives, neutrals, and negatives, and also breakdowns for the most recent week, month, and six-month period. The buyer can then scroll down to view the text of the actual comments, most recent comments first (Fig. 4). It is possible to scroll through any or all of the comments, but eBay does not provide an easy way to search for the negative and neutral comments. After placing a bid on an item, the seller can view the list of bidders and click through to see their feedback profiles. Sellers are permitted to cancel particular bids, but they do not have a way to proactively prohibit future bids from the same bidder, or all bids from bidders having bad feedback profiles. Since many bids arrive in the last few minutes of auctions (as analyzed by Roth and Ockenfels, forthcoming), it is often impossible for a seller to avoid unwanted buyers.9 A user can change the pseudonym presented to others at any time, but the previous feedback follows the user. It is possible, however, to get a new email address, and register completely anew, thus leaving behind one’s previous feedback. Over time, eBay has taken efforts to make it increasingly difficult to return with a new identity (e.g. when registering with a hotmail.com email address, a user has to provide a credit card number and eBay may notice when the same number is used again). It is not, however, completely impossible to do so.

3. DATA SETS Our primary data source consists of transactional data from eBay from February 1 – June 30, 1999, and all feedback data up to June 30, 1999. This data set contained more than 20 gigabytes, or the equivalent of more than 13 million pages of double-spaced text. In order to protect the privacy of users, the data set does not include email addresses, pseudonyms, or any other personally identifying information.10 In order to protect eBay’s commercial interests, we are not able to report data that could be used to infer revenue or profits, such as average number of items sold or average selling price. In order to make the data set manageable, we received only the item titles, and not the more extensive text descriptions of items that appeared on the item detail page (e.g. Fig. 2). A few other data fields are also not available to us, in order to protect privacy interests of eBay users, or commercial interests of the company. For the

134 PAUL RESNICK AND RICHARD ZECKHAUSER

Fig. 1. The Auction Summaries Listing.

Trust Among Strangers in Internet Transactions

Fig. 2. Detail About One Item, Including Feedback of Seller and Highest Bidder. Note: User Names are Disguised in this and Other Figures to Protect the Privacy of eBay Users.

135

136 PAUL RESNICK AND RICHARD ZECKHAUSER

Fig. 3. Feedback Profile of the Seller.

Trust Among Strangers in Internet Transactions

Fig. 4. Some of the Individual Comments about the Seller. Note that the seller‘s response to the one negative feedback alleges that it was retaliatory in nature rather than a real indicator of the seller’s performance.

137

138

PAUL RESNICK AND RICHARD ZECKHAUSER

most part, however, we have fairly complete data about items listed, bids placed, and feedback given and received. Several smaller data extractions form the basis for analyses in this paper. The first we refer to as the listed items data set, LI. It is a sample of single items (eBay also permits Dutch auctions of multiple units of the same item) that were open for public bidding (eBay also permits private sales). 36,233 of these items attracted a bid at least as high as both the starting bid and the reserve price and hence were officially sold (some buyers back out after winning auctions, which is the primary source of seller complaints against buyers). All the selected items were listed on February 20, 1999, but this was not a complete sample of items sold on that day.11 There were 13,695 distinct sellers who listed items for sale in the LI data set (we refer to them as sellers even if the item did not sell), and 25,103 distinct buyers. This data set was small enough to permit, for each transaction, computation of buyer and seller feedback profiles prior to the transaction and extraction of feedback about the transaction. The second data extraction, which we refer to as the longitudinal listings data set, LL, contains all the items listed by a sample of 1000 sellers, during the full time period February 1 – June 30, 1999. The LL data set contains 168,680 item sales. It is too large to permit detailed analyses, but is useful for assessing the extent of repeated interactions between particular buyers and sellers. A third data extraction focuses on negative feedback in particular, and we refer to it as NF. It consists of a random sample of 1,580 negative feedbacks entered on May 1, 1999. For each recipient of this negative feedback, all transactions from February 1- June 30, 1999 were also extracted. This data set is considered in Section 7.

4. GENERAL CHARACTERISTICS Before looking more specifically at the provision and effects of feedback, we first examine three characteristics of the eBay marketplace. First, are most transactions really between strangers, or do users develop ongoing trading relationships. Second, what is the distribution of prior feedback for buyers and sellers? Third, do most buyers also tend to sell and vice versa, or are the buying and selling roles distinct? To assess the frequency of ongoing trading relationships, we examined the trading histories of 1,000 sellers in the LL data set. Overall, these sellers sold 168,680 items during the five-month period from February 1 to June 30, an average of 169 items each. There were 121,564 distinct buyers. There were 138,458 distinct seller-buyer combinations. This indicates that 17.9% of all sales involved a buyer and seller who had done business with each

Trust Among Strangers in Internet Transactions

139

other before. However, 89.0% of all seller-buyer pairs conducted just one transaction during the time period, and 98.9% conducted no more than four. In the vast majority of cases, multiple transactions between a seller and buyer occurred within a few days of each other (sellers often offer reduced shipping costs to buyers who buy several items that can be shipped together). In most cases, even these multiple transactions are best thought of as a single interaction. Thus, performance in the current transaction will rarely be directly remembered by future buyers who are, after all, different from the current buyer: an explicit reputation system is needed to make current performance visible to other buyers. Sellers tend to be far more experienced, on average, than buyers. Figure 5 shows histograms of the overall feedback scores (unique positives minus unique negatives) of the two groups for the LI data extraction. The sellers are much more likely to have very high feedback scores. The sellers’ median was 33 (the score of the median item was 82, reflecting that high feedback sellers listed more items in the data set). The buyers’ median was 8. It’s unlikely that many people accumulate scores of 55 (ln[score] of 4) or 148 (ln[score] of 5) selling items from their attics, so eBay must be attracting professional sellers who are deliberately acquiring items from other sources in order to sell them on eBay. On the other hand, about a quarter of items were listed by people with net scores of less than 20, and 18% were listed by sellers with scores of 10 or less. Thus, there is also a large contingent of amateur sellers. Among the buyers, the concentration of low feedback scores is much higher. Overall, then, eBay is not only a c2c (consumer to consumer) marketplace but also a b2c (business to consumer) marketplace. While some eBay users are traders, most are primarily either sellers or buyers. We traced the selling and buying of all users who listed or bought items in the LI extraction, for the entire period from February 1 through June 30, 1999. For the sellers in the LI extraction, the median number of items sold during the five-month period was 76 but the median number of items bought was just 9. For the buyers in the LI extraction, the median number of items sold was 0 (three quarters sold fewer than 10 items) and the median number of items bought was 28.

5. PROVISION OF FEEDBACK AT eBAY The first and most basic questions about feedback concern the extent to which it is provided, and its content when provided. Clearly, there are some users with hundreds or even thousands of feedback points, so feedback is given at least some of the time. How often is feedback positive? A brief perusal of the ratings

140 Each bar shows the number of buyers and sellers, respectively, with ln(positives-negatives + 1) defining the range of feedback scores grouped in that bar. A ln[score] of 2 corresponds to a score of 7, 4 to 55, and 6 to 403. The sellers group includes all users who listed items, whether or not the items sold.

PAUL RESNICK AND RICHARD ZECKHAUSER

Fig. 5. Histograms of Feedback Profiles for Distinct Buyers and Sellers in the LI Data Set.

Trust Among Strangers in Internet Transactions

141

on ebay or Yahoo suggests that negative feedback is rare, but hardly unknown.12 The LI data extraction of 36,233 sold items permits more systematic examination of these questions. During this time period, feedback was sometimes but not always explicitly marked as relating to a particular transaction. Using computerized matching, we were, however, able to find, for each transaction, the first feedback from the buyer about the seller that occurred in the six weeks following the transaction.13 Table 1 summarizes the descriptive statistics for this data set. Buyers commented on sellers for 52.1% of the items, sellers on buyers 60.6% of the time. Thus sellers – many of whom have quite extensive experience – will have at least some expectation that there will be feedback on their transactions. Of feedback provided by buyers, 0.6% of comments were negative, 0.3% were neutral, and 99.1% were positive. Sellers were slightly less happy (the most common complaint by sellers is that the winning bidder simply doesn’t follow through on the transaction). Taken at face value, the overwhelming majority of transactions lead to happy outcomes. A cursory examination of the follow-up comments after negative feedbacks suggests that even these transactions sometimes end with happy buyers. Analysis of the text comments accompanying the 111 negative and 62 neutral comments revealed many reasons for dissatisfaction, as shown in Table 2. While both neutrals and negatives generally indicate dissatisfaction (except for 8 of the neutrals), they tended to be used for different kinds of complaints. Surprisingly, items that did not match their description were somewhat more likely to receive neutral than negative feedback, perhaps reflecting that buyers may have thought discrepancies were honest mistakes on the part of sellers. Similarly, slow shipment was more likely to lead to a neutral

Table 1.

The Frequency of Feedback After Transactions in the LI Data Extraction. Buyer of Seller

Seller of Buyer

Frequency

Percent

Frequency

Percent

negative neutral positive none

111 62 18,569 17,491

0.3 0.2 51.2 48.3

353 60 21,560 14,260

1.0 0.2 59.5 39.4

Total

36,233

36,233

142

PAUL RESNICK AND RICHARD ZECKHAUSER

Table 2.

Buyers’ Most Common Complaints. Neutral feedback

Negative feedback

24

17

Arrived in poor condition, item not as advertised, replica rather than original Backed out of transaction (did not contact or respond to high bidder) No item received after sending payment Other communication problems Slow shipping Positive about transaction

1

30

2 17 13 8

38 18 6 0

All

62

111

Note that a single comment may be classified in more than one way, if it mentions more than one problem with the transaction.

than negative feedback. Not following through on a sale, or worse, not sending the item after receiving payment, tended to yield negative rather than neutral feedback.

6. DOES PRIOR FEEDBACK PREDICT FUTURE PERFORMANCE? What is the information content of a seller’s reputation? In particular, does a seller’s feedback profile predict future performance? As suggested in the introduction, the feedback system may be useful even if it is not predictive of future performance, so long as people believe it is. There are also game theoretic models in which buyers should punish sellers for prior negative feedback (or punish newcomers for lack of prior positive feedback) even if prior feedback does not indicate future performance. Still, it is interesting to investigate whether prior feedback is predictive. Table 2 indicated that neutral comments are typically used for slightly problematic transactions (delays, poor communication) while negative comments are used for very problematic transactions (never shipped, broken, counterfeit, etc.). Since both are rare, we have grouped them together as problematic transactions. Negative feedback appears to be less frequently directed to experienced sellers than to those who are less experienced. Again considering the buyers and sellers involved in the transactions from the LI data extraction, Table 3 shows the percentage of negative feedbacks among users with varying amounts

Trust Among Strangers in Internet Transactions

Table 3.

The Percentages of Prior Problematic Feedback for Buyers and Sellers with Different Experience Levels.

Group 0–9 positive 10–49 positive 50–199 positive 200–999 positive 1000 + All

143

N (Sellers)

Percent neutral and negative (Sellers)

N (Buyers)

Percent neutral and negative (Buyers)

4,018 3,932 3,728 1,895 122

2.83% 1.25% 0.95% 0.79% 1.18%

13,306 7,366 3,678 738 15

1.99% 1.09% 0.76% 0.60% 0.92%

13,695

0.93%

25,103

0.83%

of positive feedback. Among sellers with fewer than 10 positive feedbacks, 2.83% of all feedbacks were neutral or negative. This figure steadily declines as sellers get more experienced, until the group with more than 1,000 positives.14 Removing one outlier – with 180 negatives, 125 neutrals and 3,681 positives – reduces the ratio of problematic transactions to 1.04%, still a lot higher than the 200–999 group and somewhat higher than the 50–199 group.15 The feedback profiles of buyers follow a similar pattern, although they have somewhat less problematic feedback overall than sellers. One possible explanation is that those who were buyers in the transactions in our sample had much more of their prior feedback from purchases than did our sellers. Since buyers offer a standard good (money) and offer it first, one might expect buyers to receive less negative feedback overall than sellers. We next turn to the question of whether prior feedback could be used to help buyers avoid problematic transactions, again using the LI data extraction. The overall probability of a neutral or negative feedback was 0.46% (0.89% of evaluations given). This may be an underestimate of the true percentage of problematic transactions, because problems may be resolved before users contribute feedback and even if the problem is not resolved, some buyers may not leave negative feedback. It could be a substantial underestimate if participants are hesitant to give negative feedback for fear of retaliation. We model the feedback the buyer provides about the transaction as a function of the seller’s previous feedback profile. It is not obvious exactly what features of the feedback profile should be most diagnostic of performance. The analysis above suggests that more experienced sellers are better, up to a point. We compute the log of the number of positive feedbacks, adding 1 to avoid the possibility of taking the log of 0.16 To measure previous seller problems, we

144

PAUL RESNICK AND RICHARD ZECKHAUSER

combined the neutral and negative feedbacks as an indicator of problematic transactions, added one and computed the log. A logistic regression the following coefficient estimates17: The coefficients are all highly significant.

ln

冉

Pr (problematic) 1 ⫺ Pr (problematic)

冊

= ⫺ 3.9404 + 0.7712 8 ln (PROBLEMS + 1) ⫺ 0.5137 * ln (POS + 1) Thus, for a newcomer with no previous feedback, this estimates the probability of a problematic transaction at 1.91%. A veteran with 100 positive and no negative feedbacks would have an estimated probability of a problematic transaction of 0.18%. 100 positives and three negatives would yield an estimate of 0.53%. It is worth noting here that since negative feedback is rare, for experienced buyers the positive feedback score is almost the same as the net score of positives minus negatives. Thus, predicting performance based only on the net score that eBay computes would treat the sellers with 100 positives and 0 or 3 negatives as almost the same, while this model suggests that the risk of problematic transactions is quite different for the two profiles. The model accounts for only a small part of the variability in which transactions are problematic, but it does provide some value, as summarized in Fig. 6 and Table 4.18 For example, someone using this model who was willing to forego bidding on about half of the items could avoid more than four-fifths of transactions that were reported as problematic. This would cut the probability of a problematic transaction by almost two-thirds (from 0.48% to 0.18%).

7. INCENTIVE EFFECTS: DO SELLERS PAY ATTENTION TO THEIR FEEDBACK PROFILES? A necessary (though not sufficient) condition for feedback to have an incentive effect, i.e. to encourage better behavior from sellers, is for them to care about their feedback profiles. Anecdotally, we hear that sellers do indeed care, subtly or not so subtly encouraging buyers to provide positive feedback at the end of successful transactions, and going to great lengths to make buyers happy in order to avoid negative feedback. But perhaps these stories reflect only a small minority of sellers. To assess this question more systematically, we examine the propensity of sellers to respond to negative feedback. Starting in February 1999, eBay offered users the opportunity to respond to feedback. When feedback comments are displayed, the text of any response is displayed right below the

Trust Among Strangers in Internet Transactions

145

Fig. 6. The Tradeoff Between Sensitivity and Specificity for Predicting Problematic Transactions Based on Previous Feedback of Seller Listing the Item.

Table 4. Some Particular Values for the Tradeoff Between Reducing the Danger of Problematic Transactions and the Percentage of Unproblematic Transactions That Would Be Rejected. 1-specificity (% of unproblematic transactions rejected) 75 % 50 % 25% 10 % 0%

Sensitivity (% of problematic transactions rejected)

Cutoff predicted probability

% of accepted transactions that are problematic

94.2% 81.5% 57.2% 32.4% 0%

0.20% 0.31% 0.54% 1.09% Accept all

0.11% 0.18% 0.27% 0.36% 0.48%

The third column gives the cutoff: if the model predicts a probability less than this cutoff, then the transaction would be accepted. Since some of the accepted transactions have predicted probabilities far less than the cutoff, the figures in the right column tend to be somewhat lower.

146

PAUL RESNICK AND RICHARD ZECKHAUSER

text of the feedback. (The original issuer of feedback has the further opportunity to respond to the response, but the dialog ends there.) If sellers were concerned about how negative feedback would affect their reputations, we might expect them to provide explanations in response to negative feedback. Since the opportunity to respond was introduced during the month of February, 1999, we selected a sample of 1,580 negative feedback events from May 1 to generate the NF data extraction, well after the feature was introduced, so that participants could have learned how to play. Of these, the recipient entered explanatory text in 457 cases, 29% of the time. In future work, we plan to analyze whether sellers tend to alter their behavior in any detectable way after receiving a negative. For example, they might tend to buy a few items or sell lower-priced items in order to cause the negative feedback to no longer appear at the top of their feedback profile, or simply wait a week or a month, so that it no longer appears as recent feedback. We also plan to analyze whether negative feedback is correlated with decisions to stop participating altogether.

8. EQUILIBRIUM CONSIDERATIONS Economists are used to studying interactive situations, setting up the game, and describing the equilibrium or equilibria that result. Strangers buying and selling on the Internet do not have the luxury of having a keen game theorist on hand. Many will not understand the game they are playing. Others would understand, but do not have the information to draw appropriate conclusions. No doubt, different people – based in part on presuppositions and part on personal experience – believe the games are quite different. Bearing these qualifications in mind, our empirical evidence tells a great deal about the nature of the game being played. We turn now to a discussion of considerations that likely inform the perceptions of the players and the performance of the feedback market. High Courtesy Equilibrium There are two surprising facts about the feedback game on eBay: (1) the high rate of providing evaluations; and (2) the extreme rarity of neutral or negative evaluations. The first suggests that free riding is overcome, the second that buyers are grading generously, or saying nothing after bad experiences. We think of this as a High Courtesy Equilibrium. Manners frequently lead people to make small cost efforts, even when dealing with strangers that one will never again encounter, that promote general welfare and a sense of comity (Martin, 1985). Such behaviors, say waving a car ahead of you at an

Trust Among Strangers in Internet Transactions

147

intersection, can be self-reinforcing if people take satisfaction in doing the right thing, or experience discomfort from violating a possible norm. If buyers think it is the courteous or right thing to do to provide feedback, that is what they will do, if it is not too expensive. The frequency of positive feedback can also be a result of High Courtesy. As children we have been told that: “If you can’t say something nice about someone say nothing at all.” If it appears that everyone, or near everyone, on eBay is behaving this way, then it is natural that we should as well. Etiquette is the art of getting people to coordinate on social conventions, and the convention of being positive in casual encounters is well known. eBay itself has done what it could to create an environment where people will be strongly positive. For example, in computing an overall score, eBay merely subtracts the number of negatives from the number of positives, despite the former being much rarer and hence presumably more informative. eBay does not provide a facility for searching for a seller’s negative feedback comments, leaving potential buyers with the task of scrolling through all the other feedback to find the rare but informative complaints. Finally, eBay encourages buyers to contact sellers to try to resolve problems and leave negative feedback only as a last resort. Indeed, eBay may produce additional inducement of positive feedback through its policy of allowing sellers to provide feedback on buyers as well as the reverse. If a seller provides positive feedback, it may create a reciprocal obligation to provide positive feedback in return.19 There may also be a fear of retaliation for negative feedback.20 Table 5 shows the distribution when both parties provided feedback. As before, we lump negative and neutral into the “problematic” category. There are two striking features about the table. First, there is a noticeable correlation in the propensity to provide feedback at all. Second, there is a very strong correlation of buyer and seller feedback when both provide feedback.

Table 5.

Patterns of Reciprocation.

Buyer feedback about seller Problematic Problematic Positive None

54 17 342

Seller feedback about buyer Positive 35 15,122 6,403

None 84 3,430 10,746

148

PAUL RESNICK AND RICHARD ZECKHAUSER

The seller is positive 99.8% of the time when the buyer is positive but only 39.3% of the time when the buyer is neutral or negative. Similarly, the buyer is positive 99.7% of the time when the seller is, but only 23.9% of the time when the seller is neutral or negative. This overwhelming correlation is presumably due to two factors: (1) Some transactions just work out poorly – say a shipment that gets delayed and elicits complaints, and both parties get upset. (2) Even when one side of the transaction has been without blame, there is retaliation. Why do these features emerge? We might expect sellers to provide more feedback. First, since they engage in many more transactions, and they have more to do for each transaction, we would expect them to be more automated. Second, reputations count much more for them. If there is some degree of reciprocity, providing positive feedback early could be a good seller strategy (although retaining the threat of giving negative feedback would be a reason for sellers to delay giving feedback). Cutting in the opposite direction, buyers have much more to provide feedback about, promptness of shipment, adequacy of wrapping, and whether the product was as described, including sins of omission. In fact, sellers do provide feedback more often, 60.6% as opposed to 51.7% for buyers. To address the other questions, we provide more disaggregated data, including information about whether the buyer or seller first provided feedback, in Table 6. We might expect that in most cases sellers would provide feedback first, since the buyer’s responsibilities – namely payment – are completed first, and reputations are much more important to sellers, who therefore might provide early positive feedback to elicit reciprocity. Overall, buyers are the first provider of feedback 37.6% of the time, sellers are first 32.8% of the time, and neither provides feedback 29.7% of the time. Surprisingly, when both provide feedback, buyers actually go first about twice as often as sellers. Can an early positive feedback by a seller strongly discourage a negative rating of the seller? When a satisfied seller provided feedback first, buyers gave them only 8 problematic (negative or neutral) feedbacks, a rate of 0.07%. By contrast, when the seller was satisfied but waited (either providing no feedback or positive feedback after the buyer evaluated) there were 111 problematic feedbacks, a rate of 0.81%, or 11 times as high. Of course, providing early feedback may simply be a covariate with some other cause of buyer satisfaction (e.g. good communication skill on the part of the seller). Still, the numbers suggest that early positive feedback may discourage negatives from buyers, even if they are dissatisfied.

Timing of Reciprocation.

Seller feedback about buyer Seller first (or only)

Buyer first (or only)

No feedback

Buyer feedback about seller

neg

neut

plus

neg

neut

plus

none

none neg neut plus

300 20 2 0

42 1 4 6

6,403 3 5 5,091

18 5 8

0 4 3

11 16 10,031

58 26 3,430

none

Trust Among Strangers in Internet Transactions

Table 6.

10,746

149

150

PAUL RESNICK AND RICHARD ZECKHAUSER

Similarly, does the threat of retaliatory negative feedback from sellers discourage a negative rating of sellers? Sellers are far more likely to provide negative feedback after a buyer negative or neutral (19.4%) than they are overall (1.2%). Again, however, the evidence is not conclusive, since a botched transaction (e.g. lost in the mail) is likely to leave both parties unsatisfied. In such cases, reciprocal feedback may reflect misplaced blame by both buyer and seller, rather than retaliation.21 Skunk at the Garden Party, A Few Bad Apples Assume that our High Courtesy equilibrium is in effect. Might not a few disreputable sellers seek to capitalize on the situation – e.g. negative feedback unlikely – and begin a process akin to Akerlof’s (1970) lemons model that ruins the market for all? Suppose, for example that there were two types of people, good and bad apples, with two types of sales, satisfactory and unsatisfactory. Suppose good apples produced an unsatisfactory sale 10% of the time due to chance variation. Bad apples are totally strategic. They behave like goods when it is desirable to do so. But they strategically exploit the system and reap profits from unsatisfactory sales, say from excessively generous descriptions of items. If there were an infinite supply of bad apples, it would seem, and if the vast majority of buyers were willing to forego bad evaluations despite an unsatisfactory sale, then a bad apple could cruise along for a while before being detected. Consider a world filled with goods. With only 0.3% bad evaluations, this suggests that a negative evaluation conditional on an unsatisfactory sale is only about 3% as likely as a positive evaluation after a satisfactory sale (based on our assumption of 10% of sales being unsatisfactory and the observed levels of negative feedback at eBay). Thus, a bad apple could imitate goods half the time, presumably on cheaper items, and build up a reputation of 30 positives on average before receiving his first negative. Understanding this, more and more bad apples will come into the market, bad evaluations will increase in frequency, buyers will become less satisfied with the distribution, and presumably the whole process could spiral downward. However, we do not observe such a downward spiral. We identify two phenomena that may help maintain the High Courtesy equilibrium while deterring bad apples: paying initiation dues and stoning bad behavior.22 Paying Initiation Dues Many organizations have initiation dues to be paid before one can become a member, and this may well be the case on eBay, even though eBay does not

Trust Among Strangers in Internet Transactions

151

charge an entry fee. For example, a new seller may find it difficult to make a sale, or may get a significantly lower price until his reputation is built. That is, buyers may consider the extent of a seller’s record before deciding whether to do business with him, and what price to offer. If this is so, it may be worthwhile for goods to enter the market and pay the initiation dues. Their long-term returns from such dues may be greater than the amount the bad apples can get from exploiting the system for a short period of time. It may also be cheaper for them to engage in good behavior. An honest dealer will find it cheaper to sell lithographs, for example, than will a bad apple whose principal profits come from selling forgeries. Initiation dues in effect are a way of creating trust. Such dues are to sales on the Internet the equivalent of what a massive advertising campaign is for a new retail brand, or a fancy new store is to a retailer. They show the seller is willing to invest a lot of money, which suggests that he has a high quality product that he knows will sell well in the future. In fact, Friedman and Resnick (Friedman & Resnick, 2001) found that in environments where players can assume new identities and thus shed previous reputations, some form of initiation dues is inevitable if trust among veteran players is to be maintained. It has to be more valuable for veterans to maintain good behavior than to defect and then start afresh. The dues may be monetary. Or they may come in the form of worse treatment for newcomers, as we found with the lower probability of sale for newcomers. Stoning Bad Behavior Initiation dues, by themselves, may not be sufficient to deter bad apples, assuming that they had the ability to get away with their unsatisfactory behavior on a probabilistic basis. It is possible, however, that there is a contagion effect in reaping negative feedbacks. Once one has a black mark by one’s name, others may be much more willing, indeed eager, to cast stones. Given the same level of service, they might become far more critical. Thus, the 3% number we mentioned above (P(neg|unsat)/P(pos|sat) might rise to 20%, going still higher after a second or third negative. If one seeks to exploit on a sustained basis, one will almost certainly get eliminated from the market. If there is stoning, a bad apple might expect that he would have a very limited run of positive-profit exploitation after his introductory era of paying his dues and before getting stoned out of the market.23 Future empirical work will examine the frequency of negative feedback after a range of feedback profiles, and the likelihood of being driven quickly from the market. It will also examine differences in the contagion effects for different types of feedback. Some unsatisfactory experiences might be viewed

152

PAUL RESNICK AND RICHARD ZECKHAUSER

as likely stemming from bad luck, say a delayed shipment. But others might indicate a difference in types, e.g. bad apple behavior of inflating the description of an item. If the goal is to deter bad apples from playing in this market, it would be desirable if stoning were reserved for behaviors that are deliberate rather than merely careless, and where the perpetrator benefits significantly from the action that led to an unsatisfactory transaction.

9. CONCLUSION The Internet, despite its vast scale and intense use, is still in its infancy. Moreover, its form emerged from hundreds of thousands of private decisions, much as a city develops, rather than being designed by some central coordinator (see Johnson, 2001). It seems likely that many aspects of reputation systems for the web have yet to develop. Still, it is worth examining empirically the functioning of the most prominent system deployed to date. Despite the lack of incentives to provide feedback, especially negative feedback, the system appears to work reasonably well. We consider two explanations: (1) The system may still work, even if it is unreliable or unsound, if its participants think it is working. Thus, if sellers believe poor behavior will elicit negative feedback, and that buyers depend strongly on reputations, then sellers will behave well and bad sellers will be deterred. It is the perception of how the system operates, not the facts, that matters. We suspect that few participants have conducted even cursory versions of the types of analyses conducted here. (2) Even though the system may not work well in the statistical tabulation sense, it may function successfully if it swiftly turns against undesirable sellers, a process we call stoning; and if it imposes costs for a seller to get established, what we label initiation dues. We also suspect that norms drawn from elsewhere in society help enable Internet reputation systems to work. To illustrate, we mention merely the norm of reciprocity: The frequency of feedback from little-to-gain buyers on auction sites may derive from the positive but low value feedback to them by sellers. They reciprocate a low value favor with the only favor they have available. On the other hand, sellers gave feedback first only about half the time at eBay, despite the fact that they received the buyer’s money before the buyer received the goods, suggesting that such seller-initiated reciprocity can not be the only reason that buyers provide feedback. It is interesting to speculate about whether eBay would be better off with a system where mildly dissatisfied buyers recorded their dissatisfaction more than they do now. It would help buyers to differentiate among sellers,24 perhaps

Trust Among Strangers in Internet Transactions

153

creating greater faith in the effectiveness of the feedback system. On the other hand, making dissatisfaction more visible might destroy people’s overall faith in eBay as a generally safe marketplace. Strangers encounter each other regularly on the hyperactive crossroads of the Internet. In some encounters – e.g. discussing books in a chat room – trust is natural, since neither party has much to gain by dissembling. With auctions of objects, however, sellers have strong incentives to exaggerate the quality of or misrepresent the authenticity of their items. Yet judging by the volume of transactions, sellers successfully build reputations of trust. This analysis provided some data on how seller reputations are created, and sketched some mechanisms that are turning such reputations into trust.

NOTES 1. We are not permitted to reveal the exact number of transactions per day in our data set. 2. In recent years, a literature on social capital has documented the many positive effects of civic activities and informal social ties, including their effects on trust building. Robert Putnam (2000) reviews this literature and also documents declining social capital in the United States from the 1960s to the mid-nineties. 3. Recently, H&M, a Swedish retailer of high quality but little known in the United States, had round the block lines at the opening of such a store. 4. In fact buyers’ reputations are slightly better than sellers’ reputations overall, as discussed in Section 6. 5. A concern about retaliation for negative feedback is prominent in eBay’s discussion forums and eBay has publicly acknowledged the concern, although their only remedy thus far is to exhort users to provide honest feedback despite the risk of retaliation. For example, founder Pierre Omidyar posted a public letter to that effect on June 9, 1998. 6. Note the willingness of credit card companies to cover losses when one’s card is stolen or used inappropriately. Presumably the expected cost is low. But assuring customers on this score would be far less effective than merely providing coverage and charging a bit more for the card. The advantage is not due to customer risk aversion, but rather that credit card companies could never assure customers that the level of risk is really low. 7. Personal experience suggests that the threat of providing negative feedback may be employed when seeking to rectify a transaction (a counterfeit watch), but such threats do not always work, and the person issuing the threat (RJZ) does not always follow through. 8. As part of the registration process, eBay sends an email to the user, which authenticates that the email contact information is a working email account that the person registering has access to. eBay does not require buyers to provide a credit card number. Beginning October 22, 1999, eBay required sellers to provide a credit card number or go through an alternative identity verification process.

154

PAUL RESNICK AND RICHARD ZECKHAUSER

9. On the online forum that eBay runs, sellers complain about their inability to avoid problem bidders, especially those who do not pay after winning auctions. For example, see the discussion in Appendix B. 10. In our data set, users are identified with numeric identifiers, enabling tracking of user activity over time without revealing who the users are. eBay has lookup tables to match the numeric identifiers with the registration information for users, but this information was not provided to us. 11. Given the possible concern that the selection procedure may have been incompletely random or that Feb. 20 may have been unusual, we repeated the analysis of Sections 4, 5, and 6 for all listings on Feb. 19. The results were quite consistent, although particular numbers varied somewhat. 12. Informal surveys of economists, all of whom knew about eBay but only a small fraction had purchased from it, suggests that theory alone provides little indication to the answers to these questions. There was considerable variation in answer to the questions of: How frequently is feedback given, and how often is it negative? None of our respondents got close to the actual frequency of negative feedback. 13. It is possible that such a feedback could be about a different transaction, if the buyer and seller engaged in multiple transactions during the time period. Thus, the data reported here probably slightly overestimates the percentage of transactions that get positive, neutral, and negative feedback. 14. There are many possible explanations for the experience/frequency of feedback relationships. For example, experience may make someone a better seller. Only sellers with stellar reputations will stay long-term on the system; with a bad record one comes back with a new identity. Buyers may be hesitant to give negative feedback to a long/ strong record, perhaps doubting their own judgment. 15. In future work, we hope to explore the decline in performance once an extremely large number of positives has been received. Initial hypotheses include that it becomes too expensive to leave the system once one has acquired so many evaluations, sellers are counting on many buyers not looking beneath the aggregate score, or that sellers become “worse” because they capitalize or ride on their established reputations. 16. We also included the square of ln(pos) as a variable, to possibly account for the fact that performance declines with very high feedback, but its coefficient was not significant and was of the same sign as ln(pos), so we have omitted it from the results reported here. 17. We considered other models, including variants on the fraction of negative feedback, but the coefficient on this variable was never significant. We speculated that the value of the transaction may be correlated with the probability it will be problematic, since sellers can profit more from cheating on a high-priced item and the cost of a negative is independent of the price. But the effect of price was tiny and not statistically significant in models either with or without the prior feedback as covariates. 18. More sophisticated models might do even better at predicting future performance. For example, Zaccaria et al. (2000) have analyzed adaptive scoring functions that weight more recent activity more heavily. Dellarocas (2000) has examined models that might discount the effects of a few ratings that were entered in a deliberate effort to manipulate reputations. The results from the simple model in this paper merely demonstrate that there is enough information to do predictions and should motivate further development of prediction functions.

Trust Among Strangers in Internet Transactions

155

19. Robert B. Cialdini explores the norm of reciprocation in his book Influence: Science and Practice (Cialdini, 1993, pp. 19–49). At eBay, some sellers stimulate the reciprocity norm by emailing their buyers and assuring that they themselves always provide feedback. In other venues, shrewd marketers take advantage of reciprocation norms. For example, charities mail out address labels or holiday seals, hoping that respondents will find a need to reciprocate with a contribution. 20. The fear of retaliation for negative feedback is mentioned frequently on eBay message boards. The situation is apparently serious enough that eBay founder Pierre Omidyar posted a message on June 9, 1998 exhorting users to give negative feedback when it was warranted, in spite of such fears. 21. One of our students, Ko Kuwabara, is conducting further theoretical and empirical analysis of the patterns of reciprocity and retaliation. 22. Thomas Schelling (1978) addresses these possibilities. If good sellers can still make a profit despite the suspicion raised by the presence of bads, and if bads hurt bads more than goods, there may be a stable equilibrium with goods and bads together in a stable mix. An increase in the number of bads might also make buyers attend more to details of reputations, e.g. the nature of negative feedback, which would help goods. 23. In situations where initiation does vary by player type (some types find it more difficult to carry out honest sales in order to establish good reputations) and players can trade names and hence reputations, stoning might be especially important. It could deter bad apples from acquiring and spending down good reputations: reputations would be spent down too quickly. Tadelis (1999, forthcoming) analyzes models of name trading, finding that reputations can still convey some information to buyers, but does not explicitly consider the effects of stoning. 24. Dellarocas (2001) analyzes a model in which, so long as harshness in interpretation of feedback is appropriately tied with leniency in giving feedback, leniency has some advantages in deterring seller opportunism even though information about particular mildly unsatisfactory events is not shared.

ACKNOWLEDGMENTS We gratefully acknowledge financial support from the National Science Foundation under grant number IIS-9977999. We also thank eBay for providing data, and especially Patrick Firouzian and his team for doing the data extraction in the midst of many other responsibilities. Mihir Mahajan, Ko Kuwabara, and Kate Lockwood provided valuable research assistance. The participants in the NBER workshops on empirical studies of ecommerce provided useful comments at several stages in the development of the paper.

REFERENCES Akerlof, G. A. (1970). The Market for “Lemons”: Quality Uncertainty and the Market Mechanism. Quarterly Journal of Economics, 84(3), 488–500. Avery, C., Resnick, P., & Zeckhauser, R. (1999). The Market for Evaluations. American Economic Review, 89(3), 564–584.

156

PAUL RESNICK AND RICHARD ZECKHAUSER

Ba, S., & Pavlou, P. A. (forthcoming). Evidence of the Effect of Trust Building Technology in Electronic Markets: Price Premiums and Buyer Behavior. MIS Quarterly. Bajari, P., & Hortacsu, A. (2000). Winner’s Curse, Reserve Prices and Endogenous Entry: Empirical Insights from eBay Auctions. Working paper available on-line at http:/ /siepr.stanford.edu/papers/pdf/99–23.html. Cialdini, R. B. (1993). Influence: Science and Practice. New York: Harper Collins. Dellarocas, C. (2000). Immunizing Online Reputation Reporting Systems Against Unfair Ratings and Discriminatory Behavior. Proceedings of the 2nd ACM Conference on Electronic Commerce EC-00. New York: ACM Press. Dellarocas, C. (2001). Analyzing the Efficiency of eBay-like Online Reputation Reporting Mechanisms. Proceedings of the 3rd ACM Conference on Electronic Commerce EC-01. New York: ACM Press. Dewan, H. (2001). Trust in Electronic Markets: Price Discovery in Generalist vs. Specialty Online Auctions. Working paper available on-line at http://databases.si.umich.edu/reputations/bib/ papers/Dewan&Hsu.doc. Eaton, D. H. (2002). Valuing Information: Evidence from Guitar Auctions on eBay. Working paper available on-line at http://databases.si.umich.edu/reputations/bib/papers/eatonpaper.pdf. Friedman, E., & Resnick, P. (2001). The Social Cost of Cheap Pseudonyms. Journal of Economics and Management Strategy, 10(2), 173–199. Houser, D., & Wooders, J. (2000). Reputation in Auctions: Theory, and Evidence from eBay. Working paper available on-line at http://bpa.arizona.edu/ ~ jwooders/ebay.pdf. Johnson, S. (2001). Emergence: The Connected Lives of Ants, Brains, Cities, and Software. New York: Scribner. Kalyanam, K., & McIntyre, S. (2001). Returns to Reputation in Online Auction Markets. Working paper W-RW01-02, Santa Clara, CA, Santa Clara University. Kauffman, R. J., & Wood, C. A. (2000). Running up the Bid: Modeling Seller Opportunism in Internet Auctions. Working paper available on-line at http://misrc.umn.edu/wpaper/ WorkingPapers/MISRC_RunningUpBid.pdf. Lee, Z., Im, I., & Lee, S. (2000). The Effect of Negative Buyer Feedback on Prices in Internet Auction Markets. 21st International Conference on Information Systems. Lucking-Reiley, D., Bryan, D., Prasad, N., & Reeves, D. (2000). Pennies from eBay: The Determinants of Price in Online Auction. Working paper available on-line at http:// www.vanderbilt.edu/econ/reiley/papers/PenniesFromEBay.pdf McDonald, C. G., & Slawson Jr., V. C., (2001). Reputation in an Internet Auction Market. Working paper available on-line at http://papers.ssrn.com/sol3/papers.cfm?abstract_id = 207448. Martin, J. (1985). Common Courtesy: In which Miss Manners solves the problem that baffled Mr. Jefferson. New York: Atheneum. Melnik, M. I., & Alm, J. (forthcoming). Does a Seller’s Reputation Matter? Evidence from eBay Auctions. Journal of Industrial Economics. Milgrom, P., & Roberts, J. (1992). Economics, Organization, and Management. Englewood Cliffs, NJ: Prentice-Hall. Miller, N., Resnick, P., & Zeckhauser, R. (2002). Eliciting Honest Feedback in Electronic Markets. Working paper available on-line at http://www.si.umich.edu/ ~ presnick/papers/elicit/. Putnam, R. (2000). Bowling Alone: The Crumbling and Revival of American Community. New York: Simon & Schuster. Resnick, P., Zeckhauser, R., Friedman, E., & Kuwabara, K. (2000). Reputation Systems: Facilitating Trust in Internet Interactions. Communications of the ACM, 43(12), 45–48.

Trust Among Strangers in Internet Transactions

157

Resnick, P., Zeckhauser, R., Swanson, J., & Lockwood, K. (2002). The Value of Reputation on eBay: A Controlled Experiment. Working paper available on-line at http://www. si.umich.edu/ ~ presnick/papers/postcards/. Roth, A., & Ockenfels, A. (forthcoming). Last Minute Bidding and the Rules for Ending SecondPrice Auctions: Evidence from eBay and Amazon on the Internet. American Economic Review. Schelling, T. C. (1978). Micromotives and Macrobehavior. New York: W W Norton & Company. Tadelis, S. (1999). What’s in a Name? Reputation as a Tradeable Asset. American Economic Review, 89(3), 548–563. Tadelis, S. (forthcoming). The Market for Reputations as an Incentive Mechanism. Journal of Political Economy. Zacharia, G., Moukas, A., & Maes, P. (2000). Collaborative Reputation Mechanisms in Electronic Marketplaces. Decision Support Systems, 29(4), 371–388.

This Page Intentionally Left Blank

TRANSACTION INNOVATION AND THE ROLE OF THE FIRM Daniel F. Spulber ABSTRACT Advances in computers and telecommunications and the development of the Internet have been applied to develop new transaction technologies that lower transaction costs. By lowering transaction costs, these changes in transaction technology make possible transaction innovation, that is, the development of new types of market transactions. Transaction innovation changes not only business methods and organizational design, but the content of transactions and the way that markets are organized. I consider the impact of transaction innovation on the role of the firm in various economic activities including market clearing, auction design, price adjustment, quality certification, and agency.

1. INTRODUCTION Advances in computers and telecommunications and the development of the Internet have been applied to develop new transaction technologies. Changes in transaction technology reduce the costs of conducting business with the firm’s customers and suppliers, as well as the costs of interaction within the organization. By lowering transaction costs, innovations in transaction technology make possible transaction innovation, that is, the development of new types of market transactions. Transaction innovation refers to the content of transactions rather than the technology of transacting. Thus, transaction The Economics of the Internet and E-Commerce, Volume 11, pages 159–189. Copyright © 2002 by Elsevier Science Ltd. All rights of reproduction in any form reserved. ISBN: 0-7623-0971-7

159

160

DANIEL F. SPULBER

innovation is fundamentally different from process innovation that changes production methods or from product innovation that changes product features. Transaction innovation results in new types of contracts, new combinations of buyers and sellers, and new forms of market microstructure. This paper examines some of the ways that transaction innovation affects the economic role of the firm. I consider the impact of transaction innovation on market clearing, competition between auction markets and search markets, competition between dealers and search markets, quality certification by online intermediaries, and the costs of using economic agents. The economic role of the firm is to create and coordinate market transactions. The economic role of the firm is much more complex than production of goods and services, which has been the main concern of neoclassical economics. Firms establish and operate markets creating the microstructure of markets, see Spulber (1998, 1999). Firms add value by acting as intermediaries between buyers and sellers, see Spulber (1996a, b). Firms organize the production of goods and services both through internal management processes and through market transactions. A firm plays an economic role if it can create transactions with both buyers and sellers that have lower transaction costs than direct exchange between the buyers and sellers. Firms incur many of the costs of buying and selling, including searching for trading partners, establishing prices, communicating price and product information, negotiating and writing contracts, arranging payments, recording exchange data, and monitoring contractual performance. Firms coordinate a complex set of external market transactions such as capital financing, hiring labor services, procuring resources, land, parts and capital equipment, purchasing technology, and marketing and sales of output. In addition, firms manage transactions within their organizations, including budgeting, compensation of employees and managers, allocation of intermediate outputs, inventory control, and transfer pricing. The commercial use of the Internet and related technologies for information processing and communications can substantial lower the costs of these activities and therefore enable new forms of transactions. My concept of the firm builds on and extends Ronald Coase’s (1937) pathbreaking concept of transaction costs. In his classic article The Nature of the Firm, Ronald Coase explained that firms choose their activities by comparing the costs of engaging in market transactions with the costs of internal resource allocation. For Coase, firms arise as a means of economizing on market transaction costs by internalizing those transactions within the organization. I build on Coase’s fundamental transaction cost idea by showing that firms economize not only by internalizing transactions but by establishing and

Transaction Innovation and the Role of the Firm

161

operating markets. I draw a distinction between the firm as an intermediary, and Coase’s analysis of the make-or-buy decision. I show that buyers and sellers can choose between vertical integration (the make decision), bilateral exchange (the buy decision), and in addition, transacting through intermediaries. Thus, in my framework, the firm arises as a means of reducing transaction costs between buyers and sellers. I identify some of the ways that the Internet and electronic commerce lower the transaction costs for firms acting as intermediaries relative to the transaction costs of direct exchange between individual buyers and sellers. Then, I consider how lower transaction costs lead to transaction innovation. The paper is organized as follows. Section 2 sets out the economic role of the firm by considering ways that firms acting as intermediaries lower transaction costs in comparison with direct exchange between buyers and sellers. The next four sections consider transaction innovation, that is, new developments in the content of transactions. Section 3 considers transaction innovation that alters the role of the firm in matching demand and supply. Section 4 looks at transaction innovation that affects competition between auction markets and search markets. Section 5 examines the implications of transaction innovation for competition between dealers that post prices and search markets. Section 6 considers transaction innovation such quality certification by online intermediaries as a means of alleviating adverse selection. Section 7 examines the impact of transaction innovation on the costs of employing economic agents. Section 8 concludes.

2. TRANSACTION COSTS AND THE ECONOMIC ROLE OF THE FIRM This section presents the basic framework for examining the economic role of the firm. There is a role for the firm as an intermediary if the transaction costs of intermediated exchange are less than those of bilateral exchange between a buyer and seller. The role of the firm is defined in a pure exchange setting. The role of the firm that I identify departs fundamentally from classical and neoclassical economics. In traditional economics, the firm operates its technology taking prices as external givens. Prices are adjusted by a hypothetical auctioneer whose activities are exogenous to firms. In the neoclassical framework, markets are established effortlessly for any and all goods and markets clear automatically – so that transactions are costless. Neoclassical economics emphasizes the firm’s manufacturing function because it ignores the need for economic actors to arrange transactions. It would be misguided to assert that the neoclassical model applies because creating

162

DANIEL F. SPULBER

transactions is just another good produced by the firm. Rather, the firm in my framework creates transactions that operate the market itself. Consider two consumers with endowments of various goods who wish to exchange one good for another. To simplify matters, suppose that one good serves as a medium of exchange so that one consumer is the buyer and the other consumer is the seller. The buyer obtains value V from the transaction and the seller has cost C from providing the good, so that gains from trade equal V – C. The analysis can be extended to an environment with more general endowments and consumer preferences. Suppose that both the buyer and seller incur transaction costs that sum to T. Transaction costs are the costs associated with completing a trade. The costs are incurred by the buyer and seller in searching for each other, evaluating the goods to be exchanged, negotiating over the division of the surplus and keeping track of the details of the exchange, such as handling the payment and verifying the delivery of the good. The buyer and seller expend time and effort as well as other goods in completing the exchange. These costs are measured in terms of the medium of exchange. Transaction costs are represented as distinct from the buyer’s enjoyment of the good and from the seller’s cost of providing the good. These costs could be represented equivalently as a reduction in the buyer’s enjoyment or an increase in the seller’s costs. However, the usefulness of treating such costs separately will become apparent when firms are introduced. The net gains from trade between the buyer and seller equal the gains from trade less the costs of trade. The buyer and seller choose to trade if and only if they obtain net gains from trade, V ⫺ C ⫺ T ≥ 0.

(1)

If the net gains from trade are negative, the buyer and seller choose not to trade, that is, autarky results. The transaction costs T are the total costs of direct exchange for a buyer-seller pair since the particular allocation of these costs is not important for the purpose of the present discussion. A firm creates a transaction by purchasing from the seller and reselling to the buyer. The firm obtains the good from the seller and in return makes payment to the seller. The firm supplies the good to the buyer and collects a payment from the buyer. The firm’s intermediation activities consume resources. The product is not transformed but only transferred from the seller to the buyer so that the firm acts as a merchant rather than a manufacturer. Let K represent the cost to the firm of providing a transaction between the buyer and the seller. For ease of presentation, suppose that the firm bears all of the transaction costs in its dealings with the buyer and the seller.

Transaction Innovation and the Role of the Firm

163

Suppose that buyer value and seller cost are the same in direct exchange and intermediated exchange, with any difference in transactions affecting transaction costs. Then, the buyer and seller will trade through the firm only if the firm’s transaction cost is less than or equal to costs that the buyer and seller incur in trading with each other, K ≤ T. There is an incentive for the firm to enter the market only if the firm lowers transaction costs relative to direct exchange. The firm is able to enter the market profitably if the firm is able to lower transaction costs in comparison with direct exchange and if the gains from trade between the buyer and seller are greater than the firm’s transaction costs, V ⫺ C ⫺ K ≥ 0. If the firm is more efficient at creating transactions, it is possible that the firm could be profitable even if the buyer and seller would not choose to trade directly, see Fig. 1. If the gains from trade are greater than or equal to K but less than T, the firm makes transactions feasible that would not otherwise occur. Thus, the firm is able to operate profitably if and only if the firm’s transaction costs are less than or equal to the minimum of the transaction costs of direct exchange and the buyer and seller gains from trade, K ≤ min{T, V ⫺ C}. This fundamental inequality defines the economic role of the firm. Ronald Coase (1937) attributes the nature of the firm to the make-or-buy decision: “a firm will tend to expand until the costs of organizing an extra transaction within the firm become equal to the costs of carrying out the same

Fig. 1. The Transaction Triangle: The Firm Intermediates Transactions Between a Seller and a Buyer if and only if K ≤ min{T, V ⫺ C}.

164

DANIEL F. SPULBER

transaction by means of an exchange on the open market or the costs of organizing in another firm.” The activities carried out by firms are those that are less costly to handle within an organization as compared to market contracts. Coase (1937, p. 390) identified “a cost of using the price mechanism” and noted that firms carry out various production tasks to reduce the costs of searching for prices, negotiating individual transactions, and specifying contingencies in long term contracts. The Internet and computer technology have certainly affected costs within organizations. A comparison of the relative effects of the Internet on organizational costs vs. market transaction costs is beyond the scope of the present paper. In contrast to the make-or-buy decision, I wish to highlight the effect of electronic commerce on the role of the firm as an intermediary in its external transactions. The role of the firm as a creator of transactions, rather than simply as a maker of products, accords well with empirical observation. Firms in the retail, wholesale and financial sectors account for approximately a quarter of gross domestic product in the United States (Spulber, 1996a). Retailers are intermediaries between final consumers and wholesalers and manufacturers. Two thirds of wholesale transactions are sales made by wholesale merchants (including also distributors, jobbers, drop shippers, import/export merchants, grain elevators and farm product assemblers), and agents (including also brokers, commission merchants, import/export agents and brokers, auction companies, and manufacturers’ agents). The remaining third of wholesale transactions are sales conducted through manufacturers’ sales branches and offices to wholesalers, retailers and other manufacturers (U.S. Census Bureau, 2000). Financial firms, including banks, securities brokerages, mutual funds and insurance companies, also primarily create transactions. The analysis applies to the intermediation activities of attorneys, sales agents, real estate brokers and other specialized agents. Moreover, it explains significant costs of sales, purchasing, financing and other market functions of manufacturing firms, as well as their internal costs of management, accounting and information technology. As Alfred D. Chandler (1990) points out, beginning in the last decades of the nineteenth century, companies seeking to benefit from technological change invested not only in production facilities, but also in distribution and in management.

3. EFFECTS OF CHANGES IN TRANSACTION TECHNOLOGY ON TRANSACTION COSTS Changes in transaction technology lower transaction costs and foster the development of innovative transactions. In this section, I examine how changes

Transaction Innovation and the Role of the Firm

165

in transaction technology reduce transaction costs. I consider economies of scale and scope in transactions. I then examine problems that arise in measuring transaction costs in electronic commerce. 2.1. Changes in Transaction Technology Changes in transaction technology reflect underlying technological change in computer hardware and software and communications. Technological change in computers, communications and the Internet affect transaction costs in several ways. First, automation of transactions substitutes capital for labor services in the production of transactions, such as travel websites in which travelers can reserve a flight without a travel agent. Second, computation of transactions uses information technology to process data from transactions, to calculate equilibrium prices on auction websites or to combine transaction data with enterprise software to update inventories and production plans. Third, communication of transactions combines data transmission with automation and computation so that buyers and sellers can transact with companies from remote locations and at different times. Automation of transactions allows the substitution of capital for labor services in the production of transactions. Thus, a buyer can place an order from an online catalog without dealing directly with sales personnel. The capital used to automate transactions consists of computer hardware and software and data transmission networks. Such capital substitutes at the margin for costly labor services that are applied to routine commercial tasks including the time that employees spend communicating with customers and suppliers regarding prices, product availability, ordering, billing, and shipping, and the costs of managing those employees. Thus, advances in transactions technology allow firms to reduce their transaction costs by choosing the optimal capitallabor mix in their commercial activities. Costs are lowered in comparison with alternative transaction technologies that rely primarily on sales personnel to process transactions. Application of computation to transaction data is another innovation in transaction technology that potentially lowers transaction costs. By applying information technology to process data from transactions, companies can lower the transaction costs of complex transactions, such as real-time price adjustments on auction websites such as eBay. Also, companies can match large numbers of buyers and sellers on electronic stock exchanges such as electronic communications networks (ECNs). Computation of transactions data further allows companies to link external transaction systems with their internal computer systems, thus increasing the frequency, rapidity and accuracy of

166

DANIEL F. SPULBER

communication and allowing links to production and inventory management systems within each organization. Companies have made increasing use of socalled enterprise software to manage their operations and back-office systems. Companies can reduce their transaction costs by using the information from market transactions to update inventories and production plans and reduce the costs of engaging in purchasing and sales. Communication of transaction information allows buyers and sellers to transact with the firm at remote locations and at different times. Thus, the buyers and sellers in an auction on eBay need not be present at the same location and can participate in the auction at different times. This reduces transaction costs by avoiding the costs of travel and the costs of holding meetings, whether those costs would be borne by the firm or its customers and suppliers. Thus, technological change in information processing and communications result in innovations in the technology of producing commercial transactions. 2.2. Economies of Scale and Scope in Transactions As is the case with production economies, economies of scale and scope in transactions are achieved when an increase in volume of transactions or in the number of types of transactions allows the application of automation. For one or two transactions, it may be best to use a sales representative or a purchasing agent. When the firm engages in many transactions, automation using computers and Internet technology becomes worthwhile, lowering the cost per transaction. Thus, a small bookstore may find it worthwhile to use traditional sales methods. With a very high volume of transactions, it becomes possible to use automated sales methods as does the online bookseller Amazon.com, which uses websites, computer ordering and automated warehouse systems. Even without automation, there are economies of scale in providing transactions from such effects as specialization and division of labor in sales and purchasing activities. The intermediary reduces the time costs of trade by speeding up communication through standardized transaction and information processing. Retail stores have fixed costs of transactions, that is, the costs do not depend on the volume of transactions, including information-processing equipment such as computers, cash registers, bar coding and point-of-sale terminals. Banks have economies of scale in handling accounts and processing transactions, see empirical studies by Sealey and Lindley (1977), Gilligan et al. (1984) and Ferrier and Lovell (1990). Advertising agencies have economies of

Transaction Innovation and the Role of the Firm

167

scale and scope deriving from transactions in placing media advertisements for their clients, see Silk and Berndt (1994). Suppose that there are many buyers and many sellers and that the intermediary’s cost of transacting depends on the number of transactions K(N). The transaction cost of an individual direct exchange equals T. Then, because firms trade a larger volume of goods than individual buyers and sellers, the firm derives an advantage over individual transactions. Given that the average costs of transactions, K(N)/N is decreasing in the number of transactions over some range, then if the number of transactions N is sufficiently large, there is a role for the firm based on transaction technology, K(N)/N ≤ min{T, V ⫺ C}.

(3)

Let N* be the critical number of transactions such that the intermediary’s average cost equals the unit cost of direct exchange, K(N*)/N* = min{T, V ⫺ C}. Then, an intermediary will be viable if the number of buyers and the number of sellers exceed N*. Transaction costs have important implications for pricing. Let W represent the firm’s payment to the seller and let P represent the payment that the firm receives from the buyer. Consider two polar cases, one in which the firm is a monopolist and the other in which the firm competes with other firms. If the firm is a monopolist, the firm can extract rents from the buyer and seller up to their net gains from trading directly with each other. Assume that the buyer and seller would choose to trade in the absence of the firm, that is, they have net gains from trade that are greater than or equal to zero. If T ≤ V ⫺ C, the monopoly firm will increase its bid-ask spread only to the point where it equals the transaction costs to the buyer and sellers from direct exchange, P ⫺ W = T.

(4)

Given a price spread less than or equal to the transaction costs of direct exchange, the buyer and seller will choose to deal with the firm rather than transacting directly with each other. Then, the economic role of the firm is to reduce the costs of exchange. Now suppose that there is competition between firms that provide similar intermediary services with the same transaction costs. Competitive entry will drive the bid-ask spread toward average cost, P ⫺ W = K(N)/N.

(5)

168

DANIEL F. SPULBER

The returns to competitive intermediation result from innovative transactions, that is, if the firm can lower transaction costs relative to competing alternatives. There are a number of reasons that firms acting as intermediaries would be expected to incur lower transaction costs than individual buyers and sellers. Such advantages explain the formation of firms as a means of handling transactions. Retailers and wholesalers have realized these types of economies of scale by standardizing transactions and using generic business forms for bills and receipts. By organizing their work force to handle transactions, companies achieve scale economies from division of labor and specialization of labor, just as in Adam Smith’s example of pin manufacturing. Retailers and wholesalers achieve economies of scale in communication with customers and suppliers through standardization of communication, advertising and specialized sales forces. Retailers and wholesalers also realize economies of scale in handling transactions by investing in specialized capital equipment. For example, retailers employ cash registers, bar code scanners and point-of-sale computer equipment that yield economies of scale in handling repeated transactions. Wholesalers use bar code scanners, automated warehouses, and computerization of back-office systems to achieve economies of scale. Intermediaries that offer a variety of goods and services realize similar economies of scope. Thus, an intermediary that offers multiple products will realize cost efficiencies in comparison with direct exchange of fewer units or a narrow range of products. Such economies help to explain the existence of retailers and wholesalers who offer tens of thousands of products. For example, Ingram Micro is a computer industry wholesaler that offers 145,000 products. Internet-related technologies are likely to affect both the cost of an individual transaction and economies of scale in transactions. There are a number of potential sources of economies of scale that would favor e-commerce through intermediaries rather than direct exchange. Use of the Internet for communication is itself subject to economies of scale because of the fixed costs of servers and other computer equipment and the low marginal cost of sending and receiving multiple messages. There are substantial fixed costs associated with formulating a standardized message while the marginal cost of sending the message to an additional recipient over the Internet is constant and relatively low. Chuang and Sirbu (2001) demonstrate that there are economies of scale associated with alternative multicasting approaches and protocols that might be used to transmit information to multiple recipients. Online systems for billing and invoicing exhibit economies of scale as a result of substantial fixed costs of hardware and software and relatively constant costs of processing an additional transaction. Integration of online transactions with backoffice

Transaction Innovation and the Role of the Firm

169

systems could entail fixed costs of hardware, software and management systems but yield a lower and relatively constant marginal cost of updating inventories and accounting information. Specialized intermediaries can capture some of these types of economies of scale. For example, online payment intermediaries take advantage of economies of scale in digital payment systems, just as Visa or MasterCard offer consumers and merchants scale economies associated with credit card payment systems. Chuang and Sirbu (2000) consider economies of scale in networkbased document delivery systems that result from transmitting bundles of documents, such as might be used to sell academic journals online. They suggest that if individuals value elements of the bundle much more than others it may be worthwhile unbundling despite production scale economies. However, if scale economies that result from bundling are substantial, it may be worthwhile offering bundles even if consumers are only interested in using parts of the bundle. By transacting with many buyers and sellers, firms benefit from network effects. The firm that serves many buyers and sellers creates a network that gives its customers access to many suppliers and its suppliers access to many customers. The firm as intermediary is thus at the center of a hub-and-spoke network. Such hub-and-spoke economies are well-known in marketing, Alderson (1954), see also Townsend (1978). Transaction costs create some advantages to concentration in comparison to interconnected networks. Banks benefit from processing payments between their own accounts, known as “on us” transactions, rather than payments between banks which double the number of transactions, referred to as “transit transactions,” Shaffer (1997). By centralizing communication and contracting, intermediaries achieve economies from centralization. These well-understood network effects create advantages for many types of intermediaries, such as retailers and wholesalers. Suppose that there are N buyers and M sellers and that the transaction costs of contact between an individual seller and an individual buyer are T per contact. With decentralized exchange, there would be a total of N ⫻ M potential transactions, so that the total transaction costs from bilateral exchange would be T(N ⫻ M). The firm is able to carry out the same volume of transactions by contacting each of the buyers and sellers, so that the firm need only make N + M contacts. If the firm’s cost per transaction is K, the firm incurs total cost K(N + M), which is less than T(N ⫻ M) for any T if N and M are sufficiently large. The use of Internet communications lowers the costs of multipoint communications and thus can increase the cost advantages from trading networks.

170

DANIEL F. SPULBER

2.3. Measuring the Impact of Technological Change on Transaction Costs The comparison of the transaction costs of direct and intermediated exchange helps to define the impact of transaction innovation on the role of the firm. The use of Internet technology in electronic commerce potentially affects the transaction costs of both direct and intermediated exchange. Therefore, electronic commerce can change the relative levels of these costs, creating new roles for the firm in some industries and eliminating existing roles of the firm in other industries. Estimation of the size of electronic commerce faces substantial measurement problems. This is due in part to the fact that electronic commerce represents a change in the production technology for carrying out transactions rather than a more easily measured product or service. Thus, it is relatively easy to measure the output of computers and other information technology but it is more difficult to measure usage of their services as capital equipment inputs. Moreover, since electronic commerce refers to a production method, it may not be desirable to distinguish electronic commerce from commerce generally. Transactions may vary in the degree to which they take advantage of electronic commerce technology. For example, a transaction at a retail store employs electronic commerce in verifying the customer’s credit card and in ordering additional inventory, and perhaps in providing additional service and information to the customer through a website, even if the initial transaction was not conducted online. The Census Bureau defines electronic commerce as the value of goods and services sold over computer mediated networks, (Mesenbourg, 2001). Based on this definition, which focuses on the method of making the final sale, electronic commerce is a small subset of retail, wholesale and financial transactions, and selected other sectors of the economic such as transportation. The definition does not count traditional retail, wholesale or financial transactions that might have substantial components that rely on computer networks. In addition, the Census Bureau includes as electronic business any process that a business organization conducts over computer mediated networks, (Mesenbourg, 2001). The Census Bureau provides a broad definition of electronic business infrastructure, including “hardware, software, telecommunications networks, support services, and human capital used in electronic business and commerce” (Mesenbourg, 2001). In what follows, I use the term electronic commerce to apply broadly to both final sales and business processes. Measured in terms of sales methods, electronic commerce is a small share of economic activity, measuring less than 1% of retail sales for the year 2000. Based on 1999 data, e-commerce accounted for about 5% of sales of merchant

Transaction Innovation and the Role of the Firm

171

wholesalers. Finally, also based on 1999 data, e-commerce shipments by manufacturers were estimated at 12% of total manufacturing shipments. This data suggests that most electronic commerce is interbusiness transactions, (Mesenbourg, 2001). Measurement of the impact of electronic commerce on transaction costs will require detailed analyses of the structure of transactions within different industries. L. Jean Camp (2000, p. 32) identifies seven stages of a transaction: account acquisition, browsing or discovery, price negotiation, payment, merchandise delivery, dispute resolution, collections and final settlement. She points out that the Internet reduces costs at each stage. Thus, the means of payment are facilitated by Internet-based transaction providers, credit cards or digital cash. Product discovery is enabled by Internet-based advertising, web pages, online catalogs, search engines and Usenet groups. The costs of price negotiation are reduced by electronic communication such as e-mail or instant messaging. Payment can be simplified by digital payment systems and automated provision of bills and receipts. The cost of merchandise delivery is substantially reduced for information-based goods and services that can be delivered directly online, such as documents and software. The costs of arranging merchandise delivery of products requiring shipment can be reduced by automated shipping and tracking software. Camp observes that in electronic commerce, dispute resolution, collections and final settlement can be more or less costly than traditional alternatives depending on the nature of payment mechanisms. What are the transaction cost implications of electronic commerce? Patricia Danzon and Michael Furekawa (2001) suggest that the cost of processing a health insurance claim potentially can be reduced from $10–15 per claim to 2–4 cents per claim. Eric Clemons and Loren Hitt (2001) observe significant cost savings in securities trades. Charles Fine and Daniel Raff (2001) examine possible efficiency gains resulting from Internet-based auto sales in comparison to traditional dealer sales methods. These observations suggest that industrylevel and even firm-level data are needed to identify the effects of electronic commerce on businesses processes and transaction costs. The difficulties in measuring the effects of electronic commerce on transaction costs are compounded when making comparisons between direct exchange and intermediated exchange. The same types of business processes can potentially be applied to different types of transactions. Accordingly, comparing the changes in the transaction costs of firms as intermediaries, K, with changes in the costs of transaction costs of direct exchange, T, is likely to prove challenging. As a result, theoretical analysis of the impacts of transaction cost changes on the institutions of exchange, that is on market microstructure,

172

DANIEL F. SPULBER

is necessary to shed light on the effects of electronic commerce. Changes in the market microstructure of key sectors of the economy should reflect the impact of electronic commerce. By understanding how transaction costs affect the economic role of the firm, inferences about the effects of electronic commerce can be derived from observed changes in market microstructure. Such information can supplement more standard measurements of electronic commerce based on market shares in the retail, wholesale, financial, and manufacturing sectors.

3. AGGREGATION AND MARKET CLEARING Reductions in transaction costs allow various types of transaction innovation. An important type of transaction innovation is the development of new ways of aggregating demands and supplies and clearing markets. An important aspect of the firm’s economic role is to aggregate its customers’ demand and its sellers’ supply. The firm adjusts prices to maximize profits. The firm adjusts prices such that its purchases and sales clear the market, so that the firm performs the market clearing function that neoclassical economics ascribes to the mythical Walrasian auctioneer. The price adjustment function has a number of advantages for individual buyers and sellers. Transacting at posted prices avoids the transaction costs of bilateral exchange, including risk due to imperfect information about trading partners and the costs of negotiation. In securities markets, financial intermediary firms such as stock specialists smooth the pattern of exchange, creating market liquidity by holding inventories, see Stoll (1985). Demsetz (1968) notes that in securities markets “the ask-bid spread is the markup that is paid for predictable immediacy of exchange in organized markets; in other markets it is the inventory markup of retailer and wholesaler.” In product markets, dealers, wholesalers and retailers also provide immediacy services by holding inventories of goods on hand and standing ready to sell to customers. They further have cash on hand and stand ready to buy from suppliers. This avoids the problem of the coincidence of wants, in which a buyer and a seller need to want to transact with each other at the same time, see also Clower and Leijonhufvud (1975). By aggregating demands and suppliers, firms pool risk thus reducing transaction costs relative to individual buyers, see Lim (1981) and Spulber (1985). The immediacy services provided by firms lower the transaction costs of buyers and sellers relative to decentralized bilateral exchange. Consider an example with two consumers each with a per-unit value of V. One consumer only wants one unit of the good and the other consumer wants two units of the good. There are two sellers each with a per-unit cost of C. One

Transaction Innovation and the Role of the Firm

173

seller has a capacity of one unit of the good and the other seller has a capacity of two units of the good. If the sellers and buyers are matched properly, three units will be exchanged and total gains from trade will equal 3(V ⫺ C). Otherwise, the matches will be either demand or supply constrained, only two units will be exchanged and total gains from trade will equal 2(V ⫺ C). With random matching, the two outcomes are equally likely, so that the expected gains from trade will be (5/2)(V ⫺ C). The low-demand consumer and the lowcapacity seller are never capacity constrained, so that their expected gains from trade each equal (V ⫺ C)/2. The high-demand consumer and the high-capacity seller are constrained half of the time, so that their expected gains from trade each equal (3/4)(V ⫺ C). An intermediary can clear the market by aggregating both demand and supply. Suppose that the intermediary rations the long side of the market. In the present example there are finite numbers of buyers and sellers. Accordingly, a buyer or seller deciding whether or not to transact with the intermediary takes into account the effect of their decision. Thus, a buyer or seller compares the returns from transacting with the intermediary with the returns they would obtain if they entered the direct exchange market. The intermediary offers no advantage for the low-demand consumer and the low-capacity seller since they are never capacity constrained, so that trying to attract them will not be profitable for the intermediary. The profit-maximizing intermediary will choose a bid-ask spread to attract the high-demand consumer and the high-capacity seller. In equilibrium, the intermediary sets prices so that both the high-demand consumer and the high-capacity seller are indifferent between dealing with the intermediary and entering the direct exchange market. Since the intermediary buys two units from the high-demand consumer and sells two units to the high-capacity seller, the intermediary’s equilibrium prices satisfy the following two equations, 2(V ⫺ P) = (3/4)(V ⫺ C),

(6)

2(W ⫺ C) = (3/4)(V ⫺ C).

(7)

Therefore, the intermediary’s equilibrium prices are as follows, P = (1/8)(5V + 3C).

(8)

W = (1/8)(3V + 5C).

(9)

The intermediary offers a price spread of P ⫺ W = (1/4)(V ⫺ C). The intermediary’s prices will not attract the low-demand consumer and the low-capacity seller since they would obtain less surplus from dealing with the intermediary than they obtain in the matching market. Accordingly, in

174

DANIEL F. SPULBER

equilibrium, the matching market continues to function with one unit traded between the low-demand consumer and the low-capacity seller. The intermediary buys and sells two units, so that in equilibrium the total volume of trade equals three units and efficiency is achieved. The profit of the intermediary equals the efficiency loss that would have occurred with random matching, which is equal to T = 3(V⫺C)⫺(3/2)(V⫺C) = 1 (2)(V ⫺ C). If the intermediary has a transaction cost K, then the intermediary is viable only if 1

K ≤ (2)(V ⫺ C).

(10)

Electronic commerce provides opportunities for lowering the costs of aggregating demand and supply and establishing markets. Because of the economies of scale in distributed information, the intermediary can communicate the same information about prices and availability with many buyers and sellers simultaneously. Moreover, individual buyers and sellers can communicate information to the intermediary who then relays the information at low average cost to all other buyers and sellers.

4. COMPETITION BETWEEN AN AUCTION AND A SEARCH MARKET Reductions in transaction costs have important implications for transaction innovation in the organization of markets. The institutions and rules of exchange constitute the market microstructure. By affecting the relative costs of alternative institutions of exchange, transaction innovation can change the microstructure of markets. For example, if auctions become relatively less expensive than a process of search and bilateral bargaining, there may be more transactions conducted through auctions than through search and direct exchange. A company holding an online auction can receive bids from individual buyers and sellers, determine the highest bid as bids are received, and communicate the information to all interested buyers and sellers. There are hundreds of thousands of online auctions taking place each day with billions of dollars of trading per year, see Lucking-Reiley (2000). Online auctions have advantages over auctions that require buyers and sellers to be in attendance at a specific location by avoiding the costs of travel. Moreover, online auctions greatly reduce the time costs of bidding because they allow buyers and sellers to submit their bids remotely and monitor the progress of the auction.

Transaction Innovation and the Role of the Firm

175

Auctions have an advantage over decentralized exchange by avoiding the time costs of search and bargaining. Moreover, buyers benefit from receiving bids from multiple sellers, and sellers benefit from receiving bids from multiple buyers. To illustrate the benefits of an auction, consider a simple case of a single buyer and two sellers. Each seller’s cost is either c or C, where c < C. Either type is equally likely, and the costs of the two sellers are independent. The buyer’s valuation V is common knowledge, and is greater than the cost of either seller, c < C < V. Suppose that the buyer and the sellers engage in search and the buyer only meets one of the sellers. After the buyer and one of the sellers meet, the buyer observes the seller’s costs and trade takes place. Because of the time costs of search and bargaining, the buyer is not able to interact with more than one seller. The buyer and the seller evenly divide the gains from trade. The buyer’s expected benefit from direct exchange is V/2 ⫺ (c + C)/4. Suppose that an intermediary holds a second-price auction. The sellers choose their bids to maximize their net payoff from the auction. The winning seller is the one that makes the lowest bid. The winning seller receives the next highest bid. By standard arguments due to Vickrey, it follows that it is a dominant strategy for each of the sellers to exactly bid their cost, no matter whether they have a high or a low cost. The auction yields a high cost C with probability 3/4 and a low-cost c with probability 1/4. Therefore, the expected price from the auction is P = c/4 + 3C/4 The buyer’s expected benefit from the auction is thus V ⫺ P = V ⫺ c/4 ⫺ 3C/4. It follows that the buyer’s benefit from participating in the auction is greater than the buyer’s benefit from entering the matching market. This reflects the fact that the buyer benefits from competition between sellers. When both have high cost, the buyer pays the high cost but does not have to share the gains from trade with the seller. Also, when both sellers have low costs, the buyer pays only the low cost. When only one seller has high cost, the buyer pays the higher cost and thus shares the gains from trade with the low-cost seller. The buyer’s net expected benefit from participating in the auction vs. entering the search market equals (V ⫺ C)/2. The intermediary who runs the auction can impose a fixed fee on the buyer equal to that amount. The

176

DANIEL F. SPULBER

auctioneer is viable if and only if the transaction costs of holding the auction do not exceed the benefits of participation for the buyer, K ≤ (V ⫺ C)/2. Because online auctions lower the transaction costs of buyer and seller participation in auctions as well as the costs of holding auctions, auctions become a more attractive alternative relative to search and matching. This is reflected in the growth of online auctions. The largest online market, eBay, provides an example of aggregation and market clearing through auctions. At eBay.com, sellers offer items for sale, and eBay.com holds auctions with bidding by buyers. In the year 2000, eBay listed about 265 million items for sale in over 8,000 categories, including eBay.com, the company’s international sites, and Half.com, with an aggregate value of goods sold in that year of $5.4 billion. According to the company, its traditional competitors were “classified advertisements, collectibles shows, garage sales, flea markets, and other venues such as auction houses.” As the company has expanded, it identified a new set of competitors including “distributors, liquidators, retailers, import and export companies, catalog and mail order companies, and virtually all online and office commerce participants.” The company maintains that it has advantages over the other trading forums, because in the other forums “it is difficult for buyers and sellers to meet, exchange information, and complete transactions.” Moreover, individual buyers and sellers have a limited variety of goods, high transaction costs, and are inefficient in setting prices (Annual Report, 2000, pp. 1–3). Half.com, which eBay acquired in July 2000, brings together many buyers and many sellers by letting individual sellers post prices. Buyers purchase items at the posted prices and pay Half.com. In turn, Half.com pays the seller minus a 15% commission. Sellers post item descriptions that can include such information as model numbers for electronics or ISBN numbers for books and the condition of the item. In February 2002, the company had over 100 million items listed for sale (www.half.com/index.cfm). FreeMarkets provides auction services in which suppliers bid to provide a service. According to the company, it has created online markets for over $30 billion of goods and services in over 195 industrial categories. The company reports that over 19,000 suppliers from more than 70 countries have actively bid through FreeMarkets since the company was founded in 1995. Suppliers include BP Amoco, Dana Corporation, Eaton Corporation, GlaxoSmithKline, H. J. Heinz Company, John Deere, Marathon Oil Company, The Quaker Oats

Transaction Innovation and the Role of the Firm

177

Company, Raytheon Company, The Royal Bank of Scotland, Schering-Plough Corporation, Singapore Technologies Engineering, Visteon Corporation, and Welch Foods. Acting as a buyer, electrical tool and hardware manufacturer Cooper Industries used FreeMarkets to obtain domestic airfreight services in an auction that received 57 bids from 11 suppliers, see http://www.freemarkets.com/corpinfo/.

5. COMPETITION BETWEEN A DEALER AND A SEARCH MARKET Reductions in transaction costs also can affect market microstructure by changing the relative transaction costs of using dealers in comparison with the costs of search and bilateral bargaining. If changes in transaction technology allow dealers to achieve lower transaction costs relative to decentralized search by buyers and sellers, there will be a greater role of specialized intermediaries. When buyers and sellers search for each other, they face the time costs of delay, travel costs and the cost of learning about trading partners. Accordingly, buyers and sellers may settle for a match that does not yield the highest possible gains from trade to avoid the costs of further search. The cost of random search are manifested in terms of inferior trades, the time costs of shopping and making price comparisons, and equilibrium price dispersion. Firms acting as intermediaries reduce the inefficiencies of consumer search by centralizing exchange and adjusting prices to clear markets. By posting prices, intermediaries divert buyers and sellers from search markets. These buyers and sellers are able to buy and sell at posted prices and avoid the uncertainty of the search market. Moreover, posted prices improve the efficiency of the search market itself by diverting the high-value buyers and sellers and discouraging buyers and sellers whose values are outside the bid-ask spread, see Gehrig (1993) and Spulber (1996a, 1999, 2002) for further discussion. The advantage of intermediated exchange can be illustrated by means of a simple example with two buyers and two sellers. The example is based on Gehrig (1993) and Spulber (1999). The buyers have a willingness to pay v and 1 respectively, where v < 1. The sellers have costs 0 and c respectively, where 0 < c < 1. Before entering the matching market, the buyers and sellers do not know the type of their trading partner. After a buyer and seller decide to trade, they learn each other’s type. At that point, trade occurs if and only if they have gains from trade. The buyer and seller split the gains from trade evenly.

178

DANIEL F. SPULBER

To illustrate the effects of search costs, suppose that the buyers and sellers can only make one match. A high-value buyer can trade with both types of sellers and a low-cost seller can trade with both types of buyers. The market outcome will depend on whether or not a low-value buyer can trade with a high-cost seller, that is, whether v is greater than or less than c. If v is greater than c, all matches will result in trade. This situation corresponds to a case where the demand curve is everywhere above the supply curve. There are no efficiency losses from random search so there is no economic role for the intermediary. Assume that the low-value buyer cannot trade with the high-cost seller, that is v < c. In an efficient market, the high-value buyer would be matched with the low-cost seller, with the other types excluded from trade. Therefore, in an efficient market, the total gains from trade equals 1. By standard supply-anddemand reasoning, the market-clearing price should be greater than v and less than c, so that the low-value buyer and high-cost firm are excluded from the market. Random search with bilateral exchange is inefficient because it can result in too much trade. With random search the high-value buyer is able to meet and trade with the high-cost seller and the low-value buyer is able to meet and trade with the low-cost seller. Thus, with random search, there are two possibilities. First, there is an efficient outcome in which the high-value buyer meets the lowcost seller, and the other two types do not trade, so that total gains from trade equals 1. Second, there is an inefficient outcome in which the high-value buyer meets and trades with the high-cost seller and the low-value buyer meets and trades with the low-cost seller, so that total gains from trade equals (1 ⫺ c) + v < 1. Since these two possibilities are equally likely, the total 1 1 expected gains from trade under random search equals 2 + 2[(1 ⫺ c) + v]} = 1 ⫺ (c ⫺ v)/2. Since this is less than one, it follows that random search is inefficient due to the possibility of excessive trade. Consider the market outcome with an intermediary. The intermediary posts a bid and an ask price. The intermediary rations the long side of the market, so that the intermediary’s purchases and sales balance. In equilibrium, the intermediary offers a bid-ask spread that attracts the high-value buyer and the low-cost seller. If the high-value buyer and the low-cost seller trade with the intermediary, there is no search market since the low-value buyer and the high-cost seller cannot trade. In the present example, with a finite number of buyers and sellers, the buyers and sellers take into account the consequences of their choices on the market equilibrium. If the high-value buyer and the lowcost seller do not trade with the intermediary, a search market forms and the low-value buyer and the high-cost seller also enter the market.

Transaction Innovation and the Role of the Firm

179

The intermediary’s prices must give the high-value buyer and the low-cost seller at least what they could obtain in the search market, 1 ⫺ p = 1/4 + (1 ⫺ c)/4

(11)

w = 1/4 + v/4

(12)

Therefore, the bid-ask spread offered by the intermediary is p ⫺ w = (1 + c ⫺ v)/4.

(13)

Only the high-value buyer and the low-cost seller are active in equilibrium. The total volume of trade equals the efficient amount of one unit. The intermediary earns a return from posting prices since the high-value buyer and the low-cost seller prefer the certainty of posted prices to the uncertainty of the search market where they may be paired with a lower value trading partner. The intermediary is active if the returns to posting prices exceed the intermediary’s transaction cost, (1 + c ⫺ v)/4 > K. Electronic commerce impacts the pricing role of intermediaries in two ways. First, the cost of search are potentially reduced, since buyers and sellers can make price comparisons with greater ease. This can enhance opportunities for direct exchange, allowing buyers and sellers to sample multiple alternatives quickly thus reducing the inefficiencies of the search market. This implies that an intermediary faces greater competition from the search market. Second, the transaction cost of intermediaries are reduced thus increasing competition between intermediaries, but also making intermediaries more competitive with the search market. Because electronic commerce potentially lowers both the costs of direct and intermediated exchange, one cannot conclude that the result will be fewer intermediaries despite common predictions of disintermediation. Moreover, even without intermediaries posting prices, buyers and sellers engaging in search obtain price information through specialized intermediaries that provide price information. Thus, the competition is between dealers who post prices and intermediaries who supply price information. The presence of lower search costs does not mean that search costs are eliminated or that electronic commerce markets are frictionless. Buyers and sellers still face time costs of evaluating and comparing alternatives. Moreover, with uncertainty about the characteristics of trading partners some uncertainties of the search process remain even if price comparisons are made easier. This helps to explain observed price dispersion in electronic commerce markets. Price dispersion can result from decentralized search among buyers and sellers

180

DANIEL F. SPULBER

seeking to transact directly. Alternatively, buyers and sellers search among multiple intermediaries, see Spulber (1996b). Evidence from electronic commerce suggests that price dispersion persists, see surveys by Smith, Bailey and Brynjolfsson (1999) and Bakos (2001). For example, Baye, Morgan and Scholten (2001) examine price dispersion on Shopper.com, a price-comparison site on the Internet, using 4 million price observations for consumer electronics. They find that the greater the number of firms that are competing, the smaller is the difference between the two lowest prices. Also, with more firms, the range of prices increases. They cannot support the notion that prices on the Internet are converging according to the so-called law of one price. Instead, they find that price dispersion is common with the average range in prices equaling about 40%. Moreover, the gap between the two lowest prices remained stable at about 5%. There are many theoretical explanations for price dispersion. For example, if there is asymmetric information about costs, firms will engage in price competition resembling an auction, with higher cost firms making higher bids, see Spulber (1995). If some proportion of consumers are informed about prices while others are uninformed, and it is costly to list prices with an Internet gatekeeper, then not all firms list their prices on the Internet and equilibrium price dispersion occurs, see Baye and Morgan (2001). In contrast to the disintermediation predicted for electronic commerce, many markets have multiple intermediaries. For example, there are many retailers on the Internet selling the same products such as books, music recordings, or consumer electronic products. There are multiple wholesale intermediaries engaged in business-to-business electronic commerce, see Lucking-Reiley and Spulber (2001). To examine search with competing intermediaries, consider the model presented in Spulber (1996b). In that model, buyers search across intermediaries for the lowest price and suppliers search across intermediaries for the highest price. Intermediaries have different per unit transaction costs k. The equilibrium consists of a price distribution of ask prices p and a price distribution of bid prices w. Search is costly because it is time consuming for buyers and sellers who discount future returns. As the discount rate tends to zero, that is, as the time costs of search fall, price dispersion is eliminated and prices converge to the Walrasian price. In contrast, as the discount rate tends to infinity, that is, as the time costs of search rise, price dispersion increases, and in the limit each firm with cost k offers prices equal to bid-ask spread of a monopolist with cost k. With higher time costs of search, more intermediary firms operate in equilibrium. This is consistent with the observation of Baye, Morgan and Scholten (2001) that with more firms competing, the range of prices is greater.

Transaction Innovation and the Role of the Firm

181

6. ADVERSE SELECTION AND QUALITY CERTIFICATION Advances in transaction technology, allow the development of transactions that lower the cost of adverse selection. Firms traditionally play a role in reducing problems associated with asymmetric information. Retailers and wholesalers often certify the quality of products that they resell. Manufacturers certify the quality of the inputs included in their products and provide guarantees for the performance of their products. Electronic commerce provides opportunities for enhancing quality certification. Internet intermediaries that can attract greater numbers of buyers and sellers as a result of lower costs of communication and operation can then lower the average cost of certification. If an intermediary reaches a larger set of buyers and sellers online, there are greater returns to building a reputation for accurate certification. The low costs of communication also allow the intermediary to obtain feedback from buyers and sellers and to provide that information to others. For example, eBay allows buyers and sellers to rate the performance of trading partners and makes the information available to prospective buyers and sellers. In Akerlof’s (1970) market-for-lemons model, low-quality cars drive out high-quality cars because buyers are only willing to pay the expected value of a car while sellers are not able to demonstrate whether their cars are of low or high quality. Sellers of high quality cars are driven from the market because the expected value of a car is less than their opportunity cost of selling a car. In Akerlof’s model, buyers have asymmetric information in comparison to sellers presumably because the cost of observing quality exceeds the value of the exchange. In contrast, Biglaiser’s (1993) examines the possibility that an intermediary such as a used car dealer is able to invest in expertise to verify the quality of cars at a fixed cost of K. Since the intermediary sells multiple cars, it obtains economies of scale in quality verification that are not available to individual buyers and sellers. Consider an example based on Biglaiser’s (1993) model. The number of sellers with high-quality cars equals N, the number of sellers with low-quality cars equals M and the number of buyers equals N + M. The proportion of high quality cars in the population is N/(N + M). Buyers value a high-quality car at V and a low-quality car at v, and sellers value a high-quality car at C and a lowquality car at zero. Assuming the V > C and v > 0, there are gains from trade for any buyer and seller match. Suppose that the gains from trade from a highquality car is greater than for a low-quality car, V ⫺ C > v. Akerlof’s condition

182

DANIEL F. SPULBER

for low-quality cars to drive out high-quality cars is (NV + Mv)/(N + M) < C. Suppose that in equilibrium only low-quality cars are sold and that a buyer and a seller evenly divide the gains from trade with each obtaining v/2. Consider now the intermediary that is able to verify the quality of cars. Sellers with a low-quality car need not sell through the intermediary since they credibly claim their car is of low quality. The intermediary resells only highquality cars. Suppose that buyers consider the possibility of being rationed on the used-car search market, but in equilibrium that possibility does not occur. Also, since there are more buyers than there are sellers of good cars, the intermediary will ration the excess buyers. Thus, in equilibrium, buyers purchasing from the intermediary are indifferent between purchasing a highquality car from the intermediary and directly buying a low-quality car, V ⫺ P = v/2. So, the intermediary’s ask price is P = V ⫺ v/2. The intermediary pays highquality car sellers their opportunity cost, W = C. The intermediary’s return is equal to N(V ⫺ v/2 ⫺ C). The intermediary is viable if the average cost of quality certification is less than or equal to the buyer’s benefit from buying from the dealer minus the cost of the high-quality car, K/N ≤ (V ⫺ v/2 ⫺ C). Thus, if there is a sufficiently large proportion of high-quality cars to lower the average cost of certification, the intermediary earns positive profit and solves the adverse selection problem. The lemons problem suggests that there is an important role for the firm in markets with adverse selection. In the case of used cars, both retail and wholesale dealers provide both market clearing and quality certification. Garicano and Kaplan (2001) examine the effects of electronic commerce on a particular used-car wholesaler, Autodaq. They find little evidence that the Internet worsens adverse selection problems due to asymmetric information. They suggest that the use of Internet auctions offers substantial transaction efficiencies and cost savings in comparison with physical used-car auctions, as much as 5% of the automobile’s value. Because products and services are not readily observable online in comparison with physical displays by retailers or wholesalers, there can be a need for companies that specialize in quality certification without being a party to the transaction. Although buyers and sellers obtain information through brand names and catalog descriptions, additional certification can add value particularly in international trade. Some firms known as quality systems registrars help companies to obtain certification for their manufacturing

Transaction Innovation and the Role of the Firm

183

processes to determine whether they conform to accepted product quality standards. The International Organization for Standardization (ISO) is a worldwide federation of national standards bodies from some 140 countries that helps to establish international standards for practically all manufactured products (http://www.iso.ch/iso/en/ISOOnline.frontpage). An exception is electrical products, which are governed by another standard setting body, the International Electrotechnical Commission (IEC). According to the ISO, tens of thousands of businesses are implementing a set of standards known as ISO 9000 that applies to quality management. A quality systems registrar can certify that a company’s products and quality management systems conform to ISO 9000 standards. Technical standardization is essential to the development of electronic commerce, particularly in transactions between businesses. The International Telecommunication Union (ITU) on March 24 2000 signed a Memorandum of Understanding on electronic business along with the ISO, the IEC, and the United Nations Economic Commission for Europe (UN/ECE) (see http://www.itu.int/home/index.html). The agreement specifies standards for exchanging data electronically between companies as well as product design standards.

7. AGENCY Improvements in transaction technology lead to transaction innovation in the form of increased use of specialized intermediaries. Firms play important roles both as agents and as principals. As principals, firms employ sales representatives, purchasing agents, and other intermediaries. Firms also act as agents, representing customers and suppliers by serving as their representative in transactions. Principals employ agents to reduce the transaction costs of dealing with third parties, relying on the expertise and efficiency of specialized representatives, see Spulber (1999, Chaps 11 and 12). Transaction innovation can take the form of new tasks for specialized agents. The purpose of agents is closely tied to the question of the role of the firm. Because agents act as intermediaries between the principal and the third party, the purpose of using an agent is to reduce the transaction costs relative to direct exchange, see Fig. 2. Let K represent the transaction cost of the agent and let T represent the transaction costs of direct exchange. The role of the firm as an agent depends on a comparison of the transaction costs of the agent and the transaction costs to the principal and the third party of engaging in direct exchange. The principal will employ an agent if K ≤ T. The advent of electronic

184

DANIEL F. SPULBER

Fig. 2. The Firm as Agent: The Principal Chooses to Use the Agent if K ≤ T.

commerce affects the role of firms both as agent and as principal by changing the transaction costs of direct and intermediated exchange. Trust is an important element of the principal-agent relationship because an agent represents the principal in transactions with third party, see CasadesusMasanell and Spulber (2002). If the agent acts under the direction of the principal and represents the interests of the principal accurately, the principal is bound by the terms of transactions with the third party. Accordingly, agents are said to be fiduciaries, that is, they are in a position of trust. Agency law specifies duties that the agent has to the principal. The agent must represent information accurately and act in the interest of the principal. An important development in electronic commerce is the software agent, an automated agent that represents the principal in transactions over the Internet. Vulkan (1999) characterizes the use of software agents as the second generation of electronic commerce. Software agents act as intermediaries for buyers or sellers, carrying out some type of economic activity for their principal. Because electronic commerce automates parts of transactions, many types of software can be considered agents. For example, Internet auctions offer buyers the possibility of specifying a maximum bid. Then, the software agent automatically raises the buyer’s bid by the standard increment in response to competing bids until the maximum bid is reached. Thus, the software provided by the auction site acts for the buyer. Effectively, the buyer is committed to the bid of the software agent up to the specified maximum.

Transaction Innovation and the Role of the Firm

185

Search engines are also examples of automated agents. As with human agents, buyers need to trust that these types of software agents report information accurately and act in their best interest. Various services search the Internet and offer buyers comparisons of products and prices, without committing the buyer to a transaction. There are dozens of such services including mySimon, DealTime, PriceScan, Cangetit.com, Best Web Buys, Bottom Dollar, Shopbot.com, and Robo Shopper. Biz Rate compares prices and customer ratings of over 1500 Internet businesses. The service Anything on Earth offers customized searches: “We have searched for and located airplanes, islands, sports franchises, antiques of all descriptions, electronics, cars, jewelry, designer clothes and thousands of more items.” Specialized search services exist for specific products such as automobiles, travel, insurance and loans. The strength of electronic commerce is also its potential weakness. The issue of trust is particularly important in electronic commerce since transactions are often automated. Buyers and sellers interact with commercial websites cannot verify the identity of trading partners in the same way as with face-to-face interaction. Moreover, dealing with trading partners in remote locations differs from interaction with traditional retailers and wholesalers at a specific geographic location. According to Camp (2000, p. 184): “Reliability, security, and privacy are critical” in electronic commerce, particularly with regards to payment systems. Camp identifies two categories of payment systems: notational currency, in which instructions for payment are recorded with a trusted institution, and token currency, in which payment is made in digital cash generally through a bank. She observes that the many different forms of money in electronic commerce have distinct implications for trust between buyers and sellers. Many of the electronic commerce payment systems, especially token currency, rely on the presence of a trusted intermediary such as a bank. The issue of trust in payment systems is affected by the types of information shared directly by buyers and sellers and the information that is provided to the payment intermediary. This creates a role for firms including banks to provide guaranties of reliability, privacy and security as payment intermediaries. The volume of transactions in electronic commerce depends significantly on the extent to which buyers and sellers trust payment intermediaries.

9. CONCLUSION The role of the firm is to reduce the myriad costs of exchange including search, purchasing, marketing, negotiation, sales, accounting, handling payments, and writing contracts. Reducing these costs entails continual innovation in business

186

DANIEL F. SPULBER

methods and enhanced efficiency in communication between businesses and between businesses and consumers. Innovations in information processing and communications reduce the costs of economic transactions between the firm and its customers and suppliers and within the organization. Lowering transaction costs in turn stimulates transaction innovation. New forms of transactions will create additional economic roles for the firm and lead to both changes in the organization of firms and changes in the microstructure of markets. Although transaction costs of retailers, wholesalers, financial institutions, and manufacturers are certainly quantifiable, understanding the effects of technological change requires theoretical analysis of market institutions. Empirical studies of electronic commerce should examine how changes in transaction costs affect the design of market transactions. The intermediation model of the firm presented here suggests a broad set of empirical questions regarding the determinants of transaction costs under different institutions of exchange. Among its many activities, the firm obtains economies of scale and scope by aggregating transactions. The firm aggregates demand and supply and clears markets. The firm alleviates search costs by matching buyers and sellers. Finally, the firm acts as agent for buyers or sellers. By better understanding the role of the firm in creating transactions, economists stand to gain a deeper appreciation of the impact of technological innovation on economic activity.

ACKNOWLEDGMENTS Daniel F. Spulber is the Elinor Hobbs Distinguished Professor of International Business. Research support under a grant from the Searle Fund is gratefully acknowledged.

REFERENCES Akerlof, G. A. (1970). The Market for ‘Lemons’: Quality Uncertainty and the Market Mechanism. Quarterly Journal of Economics, 84, 488–500. Alderson, W. (1954). Factors Governing the Development of Marketing Channels. In: R. M. Clewett (Ed.), Marketing Channels for Manufactured Products. Homewood, IL: Irwin. American Law Institute (1958). Restatement Second of Agency (Vols 1 and 2). St. Paul, MN: American Law Institute Publishers. Bakos, Y. (2001). The Emerging Landscape for Retail E-Commerce. Journal of Economic Perspectives, 15 (Winter), 69–80. Baye, M. R., Morgan, J., & Scholten, P. (2001). Price Dispersion in the Small and in the Large: Evidence from an Internet Price Comparison Site. Indiana University Working Paper, July.

Transaction Innovation and the Role of the Firm

187

Baye, M. R., & Morgan, J. (2001). Information Gatekeepers on the Internet and the Competitiveness of Homogeneous Product Markets. The American Economic Review, 91 (June), 454–474. Biglaiser, G. (1993). Middlemen as Experts. Rand Journal of Economics, 24 (Summer), 212–223. Brown, S. A. (1997). Revolution at the Checkout Counter: The Explosion of the Bar Code. Cambridge, MA: Harvard University Press. Boemer, C. S., & Macher, J. T. (2001). Transaction Cost Economics; An Assessment of Empirical Research in the Social Sciences. Georgetown University Working Paper. Camp, L. J. (2000). Trust and Risk in Internet Commerce. Cambridge, MA: MIT Press. Casadesus-Masanell, R., & Spulber, D. F. (2002). Agency Revisited. Northwestern University Working Paper. Chandler, A. D. (1990). Fin de Siècle: Industrial Transformation. In: M. Teich & R. Porter (Eds), Fin de Siècle and its Legacy (pp. 28–41). New York: Cambridge University Press. Chuang, J. C.-I., & Sirbu, M. A. (2001). Pricing Multicast Communications: A Cost-Based Approach. Telecommunication Systems, 17 (July), 281–297. Chuang, J. C.-I., & Sirbu, M. A. (2000). Network Delivery of Information Goods: Optimal Pricing of Articles and Subscriptions. In: B. Kahin & H. Varian (Eds), Internet Publishing and Beyond: The Economics of Digital Information and Intellectual Property (pp. 138–166). Cambridge, MA: MIT Press. Clemons, E., & Hitt, L. (2001). The Internet and the Future of Financial Services: Transparency, Differential Pricing, and Disintermediation. In: Brookings Task Force on the Internet. The Economic Payoff from the Internet Revolution (pp. 87–128). Washington, D.C.: Brookings Institute Press. Clower, R., & Leijonhufvud, A. (1975). The Coordination of Economic Activities: A Keynesian Perspective. American Economic Review, 65, 182–188. Coase, R. H. (1937). The Nature of the Firm. Economica, 4, 386–405. Coase, R. H. (1960). The Problem of Social Cost. Journal of Law and Economics, 3, 1–44. Coase, R. H. (1988). The Nature of the Firm: Origin, Meaning, Influence. Journal of Law, Economics & Organization, 4. Reprinted in: O. E. Williamson & S. G. Winter (Eds) (1991). The Nature of the Firm: Origin, Meaning, Influence (pp. 34–74). Oxford: Oxford University Press. Danzon, P., & Furekawa, M. (2001). E-Health: Effects of the Internet on Competition and Productivity in Health Care. In: Brookings Task Force on the Internet. The Economic Payoff from the Internet Revolution (pp. 189–234). Washington, D.C.: Brookings Institute Press. DeMott, D. A. (1998). A Revised Prospectus for a Third Restatement of Agency. U. C. Davis Law Review, 31 (Summer), 1035–1062. Demsetz, H. (1968). The Cost of Contracting. Quarterly Journal of Economics, 87 (February), 33–53. eBay (2001). Annual Report 2000. http://www.shareholder.com/ebay/annual/2000_annual_ 10K.pdf. Ferrier, G. D., & Lovell, C. A. K. (1990). Measuring Cost Efficiency in Banking: Econometric and Linear Programming Evidence. Journal of Econometrics, 46, 229–245. Fine, C., & Raff, D. (2001). Internet-Driven Innovation and Economic Performance in the American Automobile Industry. In: Brookings Task Force on the Internet. The Economic Payoff from the Internet Revolution (pp. 62–87). Washington, D.C.: Brookings Institute Press.

188

DANIEL F. SPULBER

Garicano, L., & Kaplan, S. N. (2001). The Effects of Business-to-Business E-Commerce on Transaction Costs. Working Paper, July. Graduate School of Business, University of Chicago. Gehrig, T. (1993). Intermediation in Search Markets. Journal of Economics and Management Strategy, 2, 97–120. Gilligan, T., Smirlock, M., & Marshall, W. (1984). Scale and Scope Economies in the MultiProduct Banking Firm. Journal of Monetary Economics, 13 (May), 393–405. Lim, C. (1981). Risk Pooling and Intermediate Trading Agents. Canadian Journal of Economics, 14 (May), 261–267. Lucking-Reiley, D. (2000). Auctions on the Internet: What’s Being Auctioned, and How? Journal of Industrial Economics, 48 (September), 227–252. Lucking-Reiley, D., & Spulber, D. F. (2001). Business-to-Business Electronic Commerce. Journal of Economic Perspectives, 15 (Winter), 55–68. Lucking-Reiley, D. (2000). Auctions on the Internet: What’s Being Auctioned, and How? Journal of Industrial Economics, 48 (September), 227–252. Mesenbourg, T. L. (2001). Measuring Electronic Business. Working Paper, U.S. Bureau of the Census, August. Washington, D.C. Osborne, M. J., & Rubinstein, A. (1990). Bargaining and Markets. San Diego: Academic Press. Sealey, C., & Lindley, J. T. (1977). Inputs, Outputs, and a Theory of Production and Cost at Depository Financial Institutions. Journal of Finance, 32, 1251–1266. Shaffer, S. (1997). Network Diseconomies and Optimal Structure. July, Working Paper No. 97–19. Federal Reserve Bank of Philadelphia. Silk, A. J., & Berndt, E. R. (1994). Costs, Institutional Mobility Barriers, and Market Structure: Advertising Agencies as Multiproduct Firms. Journal of Economics & Management Strategy, 3 (Fall), 437–480. Smith, M., Bailey, J., & Brynjolfsson, E. (1999). Understanding Digital Markets: Review and Assessment. In: E. Brynjolfsson & B. Kahin (Eds), Understanding the Digital Economy (pp. 99–136). Cambridge, MA: MIT Press. Spulber, D. F. (2002). Market Microstructure and Incentives to Invest. Journal of Political Economy, 110 (April), 352–381. Spulber, D. F. (1999). Market Microstructure: Intermediaries and the Theory of the Firm. New York: Cambridge University Press. Spulber, D. F. (1998). The Market Makers: How Leading Companies Create and Win Markets. New York: McGraw Hill/ Business Week Books. Spulber, D. F. (1996a). Market Microstructure and Intermediation. Journal of Economic Perspectives, 10 (Summer), 135–152. Spulber, D. F. (1996b). Market Making by Price Setting Firms. Review of Economic Studies, 63, 559–580. Spulber, D. F. (1995). Bertrand Competition when Rivals’ Costs are Unknown. Journal of Industrial Economics, 43, 1–12. Spulber, D. F. (1985). Risk Sharing and Inventories. Journal of Economic Behavior and Organization, 6, 55–68. Stoll, H. R. (1985). Alternative Views of Market Making. In: Y. Amihud, T. S. Y. Ho & R. Schwartz (Eds), Market Making and the Changing Structure of the Securities Industry (pp. 67–92). Lexington, Mass.: D. C. Heath. Townsend, R. M. (1978). Intermediation with Costly Bilateral Exchange. Review of Economic Studies, 55, 417–425.

Transaction Innovation and the Role of the Firm

189

United States Census Bureau (2000). 1997 Economic Census, Wholesale Trade, Geographic Area Series. Washington, D.C.: U.S. Department of Commerce, March. Vulkan, Nir (1999). Economic Implications of Agent Technology and E-Commerce. Economic Journal, 109 (February), F67–F90.

This Page Intentionally Left Blank

COMBINATORIAL AUCTIONS IN THE INFORMATION AGE: AN EXPERIMENTAL STUDY John Morgan ABSTRACT In private values settings, the Vickrey-Clarke-Groves (VCG) mechanism leads to efficient auction outcomes, while the theoretical properties of the Simultaneous Ascending (SA) auction are not well understood. This leads us to compare the properties of an SA and a VCG auction in an experimental setting with private values for multiple objects having complementarities. Statistically, we find little to distinguish the two auctions with both auction forms achieving more than 98% efficiency and extracting roughly 95% of the available surplus. Finally, in contrast to experimental results in single object VCG settings, the theoretical prediction of demand revelation in the multiple object VCG auction is largely supported in our experiments.

1. INTRODUCTION The rapid decline in the cost of and gains in the efficiency of information technology over the last two decades continue to change the landscape in which items are bought and sold. Auctions, in particular, have gained much greater prominence as a means of determining prices at which goods are bought and sold, largely as a result of improved communication and information The Economics of the Internet and E-Commerce, Volume 11, pages 191–207. Copyright © 2002 by Elsevier Science Ltd. All rights of reproduction in any form reserved. ISBN: 0-7623-0971-7

191

192

JOHN MORGAN

processing capabilities of the personal computer and the Internet. At a consumer level, the auction site eBay has transformed the market for buying and selling collectibles as well as a host of other products. EBay, with a market capitalization of over $17 billion and sales of $0.75 billion per year continues to grow rapidly (132% increase in net income year on year in 2001) despite the general slowdown in online markets. As Lucking-Reiley (2000) points out, the ubiquity of eBay has fundamentally changed the scope for using auctions as a means of allocating items by dramatically expanding the potential market for an item. On the consumer side, for instance, the collectibles market has expanded enormously in scale and scope as a result of the online auctions. As Lucking-Reiley puts it, “an item which might have been relegated to the trash heap in Shreveport, for lack of local interest, can now find its way to an enthusiastic collector in Boise.” Nor is eBay even close to being the largest source of auction transactions – it is dwarfed in sales volume by business to business and business to government transactions conducted via auction. It is not just in the expansion of the set of potential buyers and sellers that information technology has changed the nature of auction. The complexity of feasible auction mechanisms has also changed considerably owing to huge gains in processing power. Nowhere is this better illustrated than in the construction of the auction mechanism for the sale of bandwidth by the U.S. and other governments in recent years. Governments would hire auction theorists for advice about how to auction licenses and, remarkably, the somewhat complicated schemes that the theorists came up with were incorporated in practice. In view of these changes in the landscape in which auctions are conducted, it seems appropriate to re-evaluate the practicality of certain auction mechanisms which were earlier ruled out or discarded owing to their perceived complexity. First among these is the theoretically desirable auction mechanism first proposed by Vickrey. The focus of this paper is to study the efficiency properties of the auction mechanism used by the U.S. and other countries in auctioning telecoms licenses – the simultaneous ascending auction – as compared to Vickrey’s mechanism. In particular, we focus on efficiency and revenues of these competing auction forms in a complicated bidding environment characterized by demand for multiple objects and synergies for bidders depending on the combination of objects acquired. The idea behind this structure is to mimic some of the key strategic elements present in the auctions of spectrum rights held in the U.S. and elsewhere. Before proceeding, it is useful to present a bit of background on some of the earliest of these auctions. In August of 1993, Congress granted the FCC the authority to auction off thousands of personal communications services (PCS) licenses, rights to use

Combinatorial Auctions in the Information Age

193

the electromagnetic spectrum defined by both wavelength and geographic coverage. In granting the FCC this authority, Congress charged them with the task of allocating the licenses in a manner encouraging the “efficient and intensive use of the electromagnetic spectrum” [italics added]. The FCC settled on using a simultaneous ascending (SA) auction to allocate spectrum rights.1 McAfee and McMillan (1996) have argued that the principal reasons for choosing the SA auction over sealed bid or sequential auction forms arise largely because of interdependencies over the objects being offered. That is, some licenses may represent close substitutes for one another while others represent complements.2 While the theoretical properties of the simultaneous ascending auction have not been characterized, a competing alternative with desirable theoretical properties is a multiple object version of the second-price sealed-bid mechanism proposed by Vickrey (1961) (which is a special case of the Groves (1973)-Clarke (1971) class of mechanisms). This mechanism is known to implement efficient allocations in dominant strategies; hence its use would seem to be consistent in achieving the efficiency objective set forth by Congress in allocating spectrum rights. In the parlance of auction theory, the Vickrey-Clarke-Groves mechanism (hereafter ‘VCG auction’) is a combinatorial auction. That is, bidders reveal values, not just for each of the licenses being allocated separately, but also for each combination of licenses being allocated. However, the precise allocation procedure in the VCG auction differs from previously considered (and rejected) combinatorial bid mechanisms surveyed in McMillan (1994). Specifically, the combinatorial bid mechanism examined in Palfrey (1983) has the auctioneer deciding strategically which objects to bundle together and which to sell separately. In the VCG auction, the decision to bundle is determined by the revealed demands of the bidders rather than through strategic considerations of the auctioneer. The VCG auction is closer in spirit to the iterative Vickrey Groves (IVG) mechanism considered in Banks, Ledyard, and Porter (1989). However the VCG auction differs along two lines: first, its allocation procedure is determined after a single round rather than through iterations; and second, its rules for awarding the objects are somewhat more complicated than the adding up rules used in Banks et al. (1989). Thus, to our knowledge, this paper represents the first experimental examination of the properties of the VCG auction relative to the SA auction. While revenue maximization was not an explicit goal of the FCC in allocating spectrum rights, it seems sensible to think that, given a choice of efficient auction forms, the FCC would likely choose a form yielding more revenues rather than fewer. The VCG auction has the desirable theoretical

194

JOHN MORGAN

property that, in the class of efficient Bayesian mechanisms, it maximizes expected revenue. This result, due to Krishna and Perry (1997), extends the more familiar Green and Laffont (1979) result to the case of vector valued private information implementation problems. The theoretical revenue properties of the SA auction are not known. Although demonstrably efficient in theory, the VCG auction has been criticized in practice along several grounds. First, even in the simple single object case, the theoretical prediction of demand revelation often does not occur. For instance, it has been frequently observed that in single object second price sealed bid auctions with independent private values, subjects persistently bid above the dominant strategy prediction (see Kagel (1995) for a survey).3 Such violations of the dominant strategy prediction often result in inefficient allocations. Second, with interdependencies among the objects, the vector of valuations which bidders are required to submit in the VCG auction increases geometrically in the number of objects. The complexity of administering a VCG auction grows rapidly with the number of objects being auctioned. This complexity has been an important objection to its practical use. However, advances in processing speed and power in recent years suggest that even at the scale contemplated by the FCC, computing power is sufficient to handle the complexity of the mechanism for most auctions currently being contemplated. A further objection to the sealed-bid version of the VCG auction is that it does not allow for bidders to incorporate new information from others bids into their own bids. This objection, however, may be overcome by running a dynamic version of the VCG auction along the lines suggested by Ausubel (2000). Finally, with multiple object demand, the VCG auction can have undesirable equity properties – the bidder bidding the highest for the object of objects may end up paying a lower price than some other bidder. We return to this issue in Section 2. In this paper, we compare the efficiency, revenue and bidding properties of the VCG and the SA auctions in the presence of complementarities among the objects being allocated. Specifically, we conduct laboratory experiments allocating three objects with private values and complementarities in a VCG auction and in a stylized version of the simultaneous ascending auction. Statistically, we find little to distinguish the two auctions, with both auction forms achieving more than 98% efficiency and extracting roughly 95% of the available surplus. In addition, unlike most of the findings in the survey by Kagel (1995), the theoretical prediction of demand revelation in the VCG auction is largely borne out in our experiments. The remainder of the paper proceeds as follows: Section 2 presents the information and valuation structure used in the experiment and establishes

Combinatorial Auctions in the Information Age

195

the theoretical properties of the VCG auction as well as the main features of the SA auction. Section 3 gives an overview of the experimental design. Section 4 reports the main results from the experiments. Section 5 concludes.

2. THEORY We consider a model in which fifteen bidders {1, 2, . . . , 15} compete for three objects {A, B, C} under some auction form ␣. Bidder i receives a vector of private signals, v¯ i, about her value for all possible combinations of the objects. Let V¯ i denote the set of all possible values of v¯ i . Let v = {¯v1, v¯ 2, . . . , v¯ 15} denote the matrix of values for all bidders, and let v ⫺ i = {¯v1, v¯ 2, . . . v¯ i ⫺ 1, v¯ i + 1, . . . , v¯ 15} denote the matrix of values of all bidders ¯ ¯ except i. Likewise, let V = x15 i = 1Vi and V ⫺ i = xj ≠ iVj . BC AC ABC } In this set of experiments, v¯ i consists of {v Ai, v Bi, v Ci, v AB i , vi , vi , vi where for j = {A, B, C} v ji ~ Uniform(10, 20) jk j k v i = v i + v i + Uniform(0, 5) for j, k = {A, B, C}, j ≠ k jk l v ABC = max (v + v ) + Uniform(0, 2) i i i j≠k≠l for all i = {1, 2, . . . , 15}. Information about the distribution of valuations is common knowledge among the bidders; however, the particular realizations of the random variables are private information to each bidder. It is useful to notice that this structure of valuations leads to complementarities between and among the objects. The model also includes an auctioneer who is commonly known to value the objects at zero. An auction form ␣ is efficient if it allocates the three objects in a way which maximizes the aggregate value of the objects to the bidders. Let K denote the set of possible allocations of the objects among the bidders. A particular allocation k = {k¯ 1, k¯ 2, . . . , k¯ 15} is a 15 ⫻ 7 matrix where element k ji = 1 if object j苸{A, B, C, AB, BC, AC, ABC } is assigned to bidder i and zero otherwise. An allocation rule is a mapping ␬ : V → K. Definition 1: An allocation rule K* is ex post efficient if for all v苸V

冘 15

␬*(v)苸arg max k苸K

v¯ i k¯ i

i=1

2.1. VCG Auction In the VCG auction, bidder i submits a bid vector b¯ i for all combinations of the objects. Define b to be the concatenation of b¯ i for all i = 1, 2, . . . , 15, and b ⫺ i

196

JOHN MORGAN

to be the matrix b excluding the ith vector. The objects are awarded to bidders such that the sum of the accepted bids is maximized. That is

冘 15

␬*(b)苸arg max k苸K

b¯ i k¯ i .

i=1

It is useful to define the allocation rule when bidder i is not present at the auction. This is given by ␬*(b ⫺ i )苸arg max k苸K

冘

b¯ j k¯ j ;

j≠i

that is, the allocation maximizing aggregate surplus under the constraint that no objects be allocated to bidder i. Winning bidders pay the highest value that would have been bid for the set of objects awarded to the winner, had the winning bidder not been present at the auction. That is, the payment rule ti(b) for bidder i is given by ti(b) =

冘

b¯ j · ␬*j (b ⫺ i ) ⫺

j≠i

冘

b¯ j · ␬*j (b)

j≠i

Proposition 2: The dominant strategy truth-telling equilibrium of the VCG auction is ex post efficient. Proof: Bidder i ’s optimization is to choose a bid vector b¯ i to maximize v¯ i · ␬*i (b¯ i, b ⫺ i ) ⫺

冘 j≠i

b¯ j · ␬*j (b ⫺ i ) +

冘

b¯ j · ␬*j (b¯ i , b ⫺ i ).

j≠i

Since, for a given allocation, the payment by bidder i is independent of his bid, bidder i is only affected by his bid through changes in the allocation. It then follows by standard arguments that truth-telling is a dominant strategy, that is, b¯ i = v¯ i . By construction, under truth-telling leads to the ex post efficient allocation. 䊏 Finally, it is a simple matter to verify that participation constraints are likewise satisfied since bidding zero for all objects guarantees a non-negative payoff. Despite its efficiency properties, the VCG auction, by not tying the price paid directly to a bidder’s bid amount, can have the undesirable property that a high bidder ends up paying less for the object than some lower bidder. To see this, consider the following simple example involving only two objects: Bidder 1 values object A at 100 and object B at 0, but values the pair of objects at 106.

Combinatorial Auctions in the Information Age

197

Bidder 2 values object A at 0, object B at 8, and the pair at 10. Under the VCG auction, each bidder truthfully reveals his value and the objects are assigned efficiently. In this case, that means object A is assigned to bidder 1 and object B to bidder 2. Bidder 1 pays the loss to bidder 2 of his participation in the auction, which in this example is 2. Likewise, bidder 2 pays the loss to bidder 1 of his participation, which in this example is 6. Thus, bidder 1 pays only 2 for the object that he values at 100, whereas bidder 2 pays 6 for an object whose value is only 8. From a practical perspective, explaining why bidder 1 got the obviously very valuable object A for a pittance might be difficult. 2.2. Simultaneous Ascending Auction In the simultaneous ascending auction, we attempted to capture a stylized version of the main elements of the final rounds of the true FCC implementation. Since the rules for the true FCC auction procedure comprise more than 150 pages, simplifications were required for experimental implementation. Below, we briefly explain the auction procedures used in the experiment. The auction consists of an unknown number of rounds, where, in each round, eligible bidders make separate bids on any or all of the objects (but not combinations of the objects). When only one active bidder remains for each of the objects being allocated, the auction ends and the high bidders for each of the objects receive the objects and pay their amounts bid. In the first round, consistent with FCC practices, there was a minimum opening bid of $10. This reservation price was chosen to be consistent with the stipulated goal of the FCC that the “opening bid . . . provide bidders with an incentive to start bidding at a substantial portion of license value, thus ensuring a rapid conclusion of the auction” (FCC, 1994). Our reserve price lies (weakly) below the lowest value any bidder might place on any object; thus, we avoid any strategic tradeoff between efficiency and revenues as in Riley and Samuelson (1981). In subsequent rounds, a bid increment rule was instituted for each object requiring that the minimum bid must be 50 cents higher than the previous round’s high bid. Finally, an activity rule stated that once a bidder chose not to bid for an object in a given round, subject to opening bid and bid increment requirements, she was no longer eligible to bid for the object in any subsequent rounds. Our activity rule is a considerable simplification over FCC procedures which use an up-front payment to determine initial eligibility and then use activity rules to adjust these eligibility levels. Our simplified design is intended to model the

198

JOHN MORGAN

final stage of an FCC auction, which has the effect of making a bidder ineligible to bid on an object if he fails to bid on that object in any round. To derive results on efficiency, we model the SA auction as a simultaneous English auction for all objects. To be concrete, as long as the bid increment is small enough, we may think of the SA auction as consisting of a clock starting with a low price and slowly ascending. The clock shows the current price for all objects. Bidders then decide at what point to drop out of the auction for each object. When only one bidder has not dropped out, the price for that object is set at the price of the last person to drop out, and so on. The main result here is to establish that allocations are not, in general, efficient in this auction. To establish this result most starkly, consider the simplest case of 2 bidders and 2 objects. Suppose that the valuations are as follows: Bidder 1 values object A at 7, B at 2, and the pair at 9 (i.e. there is no synergy for bidder 1). Bidder 2 values A at 5, B at 0, and the pair at 8 (i.e. object B is only useful to bidder 2 in conjunction with object A). Notice that the efficient allocation in this case is for bidder 1 to receive both objects. Suppose that bidder 2’s strategy calls for him to drop out of the market for B at price p. If p > 2, then 2 will inefficiently win object B. Thus, for an efficient equilibrium to exist, it must be the case that 2 drops out of the B market at a price below 2. But now suppose that bidder 1’s value for object A is 4 instead of 7 while all of the other valuations remain the same. In particular, since bidder 2’s valuations are unchanged, his bidding strategy prior to any bidder exiting the auction must likewise be unchanged. In this case, the efficient allocation calls for bidder 2 to receive both objects. However, since in the putative equilibrium bidder 2 is dropping out at a price below 2, this means that bidder 1 will inefficiently receive object B. Since all of the inequalities in the example are strict, it occurs with positive probability. Thus, we have shown: Proposition 3: The SA auction is not ex post efficient. The degree of inefficiency of course depends on the magnitude of the synergy terms. When there are no synergies, both auction forms are efficient whereas positive synergies lead to the possibility of inefficiency. The magnitude of the synergy terms is considerable in our experiment thus leading to the possibility of inefficiency.

3. EXPERIMENTAL PROCEDURES The experiment consisted of two sessions conducted at Princeton University in the Spring semester of 1997. Subjects were recruited from the undergraduate

Combinatorial Auctions in the Information Age

199

population using posters advertising payoffs between $10 and $25. In the first session, subjects participated in six auctions consisting of three VCG auctions and three simultaneous ascending auctions. The order in which the auction forms were performed was determined randomly prior to the start of the first session. In the second session, the order was reversed to alleviate any presentation effects in the design. To ensure that both sessions used the same procedures, we adopted a written protocol. In all sessions, the subjects were seated in a large room, read a set of instructions, and given an opportunity to ask questions. Throughout each session, the only permitted communication between subjects was via their formal decisions. These decisions involved indicating a bid for each object (or combination of objects in the VCG auction only) on a bid submission form. Monitors waited until all subjects had made their decisions before collecting these forms – this preserved the simultaneity of moves required by the theoretical model. The monitors then entered the bids in a spreadsheet and then executed a macro to calculate: (a) the winning bidders and the prices paid in the VCG auction; or (b) the bid increment rule and the set of active or winning bidders in the simultaneous ascending auction.4 In the event of a tie, a winner was determined randomly from among high bidders. Before being asked to bid, subjects received a handout listing their valuations for all possible combinations of the objects in all six auctions. To ensure comparability, identical realizations of the random valuations (permuted by bidder number) were used in the second session and across auction treatments. The feedback received by the subjects naturally differed between the auction forms. In the VCG auction, subjects were informed of the objects they received and the prices paid. They then recorded this information. In the simultaneous ascending auction, subjects received feedback at the end of each round regarding the high bidder for each object. This information was also recorded. Throughout, monitors verified that the calculations entered were correct. At the conclusion of the experiment, the subjects were paid in private and in cash according to their final balance sheet cash balance. This consisted of a $5 show-up fee, a $10 starting balance, plus (minus) the proceeds from the six auctions. To aid in retaining experimental control, subjects were informed in the instructions that in the event of losses in excess of their initial $10, they would be required to leave the session. This bankruptcy event did not occur in any of the sessions.

200

JOHN MORGAN

4. RESULTS Throughout we present the pooled results of both sessions in terms of efficiency, revenues, and bidding behavior. Since we observed few differences in behavior across sessions, pooling seems justified. 4.1. Efficiency and Revenue We begin by considering how each auction fared in achieving the efficient allocation of the objects. Efficiency is calculated by dividing the value of the objects actually realized by the bidders by the theoretical maximum obtainable. The associated standard deviation (␴) is given below each of the mean figures. A visual inspection of the efficiency column of Table 1 highlights the fact that the experimental efficiency results of the two mechanisms are quite close to one another. We can make this more precise by using a Wilcoxon matchedpairs signed-ranks test by matching the six pairs in our sample for each treatment. In testing efficiency in this manner, we obtain a z statistic of 0.342; thus we fail to reject the null hypothesis that mean efficiency is the same under the two treatments against both one and two-sided alternative hypotheses. Likewise, we fail to reject the null hypothesis that the efficiency attained by the VCG conforms to theoretical predictions. Revenue, calculated as a fraction of what would be obtained under full extraction, is given in the last column of Table 1. Again, visual inspection reveals little difference in the revenue properties of the two mechanisms. Analogous to our efficiency calculations, we can test the equality of median revenues across the two auction forms by using a Wilcoxon matched-pairs signed-ranks test. Testing the equality of revenues yields a z statistic of –0.943 which is also not significant at conventional levels against one and two-sided alternative hypotheses.

Table 1. Auction Form SA (␴SA) VCG (␴VCG)

Comparisons of Efficiency and Revenue. Efficiency

Revenue

98.5% (0.036) 98.4% (0.026)

94.7% (0.029) 95.8% (0.032)

Combinatorial Auctions in the Information Age

201

Table 2 presents the efficiency and revenue achievement measures for all 12 auctions. There seems to be no correlation between the sequence of the auction and efficiency or revenue achievement. Regressing revenue on efficiency with all 12 auctions we get a very small but negative (–0.076) coefficient. However, in SA auction the coefficient is positive (–0.276) and in VCG auctions, it is negative (–0.742). Nevertheless, they are insignificant at 95% confidence level in all cases. 4.2. Bidding The VCG has the theoretical prediction that it is a dominant strategy for bidders to submit bids equal to their values for the objects. For each of the objects, the dominant strategy prediction was the modal bidding response. Table 3 shows the frequency that subjects’ bids were exactly equal to their underlying valuations for the VCG auction. To see whether there is a pattern of systematic underbidding or overbidding, it is useful to consider bids as a percentage of underlying values in each of the auction forms. To compute bids for AB, AC, etc., we simply summed the final bids for each bidder for each of the individual objects. This is given in Table 4. In contrast to single object second price sealed bid auctions in which subjects persistently overbid; here, we see underbidding for the object. Likewise, the first-price nature of the SA auction also leads to downward shading of bids relative to underlying valuations. Figure 1 presents scatterplots of bids against values for each of the object combinations {A, B, . . . , ABC} for the SA and VCG auctions respectively. Visual inspection of the VCG plots suggests that underbidding in the VCG auction is mainly due to the submission of a few “non-serious” bids close to zero. Apart from these outliers, the theoretical prediction of full demand revelation appears to be borne out in the data. In the SA auction, there is far more dispersion of bids relative to underlying valuations. We can use a signed-ranks to test the theoretical prediction that median bids are equal to underlying valuations in the VCG auction. Specifically, for each bidder in each VCG auction, we compute the difference between bids and underlying values for each object or group of objects. We then compute the median of these differences for each bidder and compare this to the theoretical prediction. Table 5 presents the signed-ranks test results for both VCG and SA auctions.5 Here, p-values are from Pr(stat ≤ | z |). Only the bid for object A is less than its value at the 5% significance level. For all other combinations of objects we fail to reject the null hypothesis that

202

Table 2.

Summary of Efficiency and Revenue in Each Auction. Session 1

Auction

Session 2

1

2

3

4

5

6

1

2

3

4

5

6

SA

SA

VCG

VCG

VCG

SA

SA

VCG

VCG

VCG

SA

SA

Efficiency

100%

100%

97.9%

100%

93.3%

100%

91.1%

99.2%

100%

100%

100%

100%

Revenue

95.0%

90.2%

97.8%

92.7%

99.4%

95.4%

92.7%

93.8%

92.5%

99.1%

97.5%

97.6%

Type

JOHN MORGAN

Combinatorial Auctions in the Information Age

Table 3. Object

Frequency of Dominant Strategy Bidding in the VCG Auction. A

B

C

AB

AC

BC

ABC

All

42%

39%

43%

26%

38%

37%

40%

39%

Table 4. Object

203

Bids as a Percentage of Values.

A

B

C

AB

AC

BC

ABC

SA

87.2%

86.7%

89.7%

81.0%

81.9%

81.1%

79.6%

VCG

92.5%

91.7%

92.0%

94.1%

92.9%

93.6%

96.3%

bids are equal to the underlying values. This is in contrast to Banks et al. (1989) who obtained overbidding in a variant of the VCG auction. A similar test may be conducted by comparing the mean bid to the mean valuation for each bidder. However, as Fig. 1 shows, there are a number of “non-serious” outlier bids near zero which have the potential to significantly affect the results of the test. Thus, we eliminate all bids less than or equal to one dollar from the sample. Using this procedure, the sign test results are not qualitatively different from results in Table 5. One might speculate that full revelation, rather than reflecting a strategic decision on the part of the bidders, in fact, reflects a heuristic bidding strategy which happens to coincide with the equilibrium. Were this the case, one might expect to see something similar in the SA auction. The results for the SA auction are quite different. Here, for all combinations of objects, we reject the hypothesis of demand revelation. The results for the

Table 5.

VCG SA

Comparisons of Demand Revelation (P Values).

A

B

C

AB

AC

BC

ABC

All

0.03

0.46

0.53

0.31

0.93

0.93

0.39

0.29

0

0

0

0

0

0

0

0

204

JOHN MORGAN

Fig. 1. Bids in the SA auction are shown in the left column, and bids in the VCG auction are shown in the right column.

SA auction suggest the need for caution in attributing demand revelation observed in the VCG auction purely to heuristics such as bidding one’s value. Kagel, Harstad, and Levin (1987) speculate that the overbidding observed in single object VCG auctions results from defective reasoning by subject along the following lines: by bidding above value, subjects correctly believe that their chances of winning the auction are increased; however, subjects place no probability on the event that by overbidding and winning the auction, losses

Combinatorial Auctions in the Information Age

205

might occur. In contrast, the complexity of the multiple object VCG payment scheme perhaps forces subject to recognize the possibility that overbidding may lead to losses and hence offsetting the perceived benefit of the increased probability of winning the auction. Notice that despite the differences in the bidding functions of the two auctions, the monotonicity of bids relative to values and the differences in the payment rules are such that both auctions lead to essentially the same allocations and generate the same revenue. This is broadly consistent with implications of the revenue equivalence theorem (Myerson, 1981).

5. CONCLUSION Our experimental results show that the revenue and efficiency properties of the VCG auction are comparable to a stylized version of the SA auction currently being employed by the FCC. Moreover, in contrast to findings for single object Vickrey auctions, the VCG auction leads bidders to truthfully reveal their values for combinations of objects with complementarities. Despite its administrative complexity, the VCG auction is implementable in a laboratory setting (at least for three objects) and resolves the auction outcomes more quickly than the multiple round SA auction (taking roughly half as long as an SA auction to conduct). The sealed bid nature of the VCG auction might also serve to circumvent signaling and other attempts at what might be viewed as collusive behavior in the SA auction. In short, our results (combined with recent allegations of collusion in the existing auction framework) suggest a reconsideration of the merits of the VCG auction. Several important extensions are still needed in evaluating the merits of the VCG vs. the SA auction. First, robustness checks in the magnitude and direction of the externality terms as well as the underlying information structure would seem useful. Since our experimental auctions were conducted under an independent private values with complementarities framework, we have, in some respects, minimized problems associated with the winner’s curse. As McAfee and McMillan (1996) point out, the ascending nature of the SA auction is designed precisely to mitigate winner’s curse problems often associated with sealed bid auctions, such as the VCG. Thus, it would be useful to compare the SA and VCG auctions in a common (or affiliated) values setting where winner’s curse effects are present. Such an extension, however, will dramatically affect the equilibrium properties of the VCG auction in that demand revelation will no longer be a weakly dominant strategy.

206

JOHN MORGAN

NOTES 1. This mechanism was proposed independently by Milgrom and Wilson as well as McAfee. 2. Licenses covering a similar geographic region at certain bandwidths are substitutes for one another; whereas those covering contiguous geographic regions may be viewed as complements in enabling their owner to obtain a larger geographic “footprint”. 3. In the open outcry version of this auction, such overbidding is not present. 4. Detailed descriptions of the precise allocation and payment determination algorithms are available upon request from the author. 5. Sign test results are qualitatively similar.

ACKNOWLEDGMENTS This paper has benefited greatly from the comments provided by Rachel Croson, Vijay Krishna, David Lucking-Reiley, Martin Sefton, Tim Van Zandt, and Robert Willig. The author thanks Tanjim Hossain for his superlative research assistance. He thanks Princeton University for their financial support of this research. The author also gratefully acknowledges support from the National Science Foundation provided under grant SBR 9618648.

REFERENCES Ausubel, L. (1997). System and Method for an Efficient Dynamic Auction for Multiple Objects. U.S. Patent Number 6,026,383, issued 15 February, 2000. Banks, J., Ledyard, J., & Porter, D. (1989). Allocating uncertain and unresponsive resources: an experimental approach. RAND Journal of Economics, 20, 1–25. Clarke, E. (1971). Multipart pricing of public goods. Public Choice, 2, 19–33. Green, J., & Laffont, J-J. (1979). Incentives in Public Decision Making. North-Holland: Amsterdam. Groves, T. (1973). Incentives in teams. Econometrica, 41, 617–631. Kagel, J. (1995). Auctions: A survey of experimental research. In: Kagel & Roth (Eds), Handbook of Experimental Economics. Princeton: Princeton University Press. Kagel, J., Harstad, R., & Levin, D. (1987). Information impact and allocation rules in auctions with affiliated private values. Econometrica, 55, 1275–1304. Krishna, V., & Perry, M. (1997). Efficient mechanism design. Mimeo. Lucking-Reiley, D. (2000). Auctions on the Internet: What’s Being Auctioned, and How? Journal of Industrial Economics, 48(3), 227–252. McAfee, P., & McMillan, J. (1996). Analyzing the airwaves auction. Journal of Economic Perspectives, 10, 159–175. McMillan, J. (1994). Selling spectrum rights. Journal of Economic Perspectives, 8, 145–162. Myerson, R. (1981). Optimal auction design. Mathematics of Operations Research, 6, 58–73.

Combinatorial Auctions in the Information Age

207

Palfrey, T. (1983). Bundling decisions by a multiproduct monopolist with incomplete information. Econometrica, 51, 463–483. Riley, J., & Samuelson, W. (1981). Optimal auctions. American Economic Review, 71, 381–392. Vickrey, W. (1961). Counterspeculation, auctions, and competitive sealed tenders. Journal of Finance, 16, 8–37.

This Page Intentionally Left Blank

ANALYZING WEBSITE CHOICE USING CLICKSTREAM DATA Avi Goldfarb ABSTRACT This paper estimates demand for Internet portals using a clickstream data panel of 2654 users. It shows that familiar econometric methodologies used to study grocery store scanner data can be applied to analyze advertising-supported Internet markets using clickstream data. In particular, it applies the methodology of Guadagni and Little (1983) to better understand households’ Internet portal choices. The methodology has reasonable out of sample predictive power and can be used to simulate changes in company strategy.

1. INTRODUCTION The growth of the Internet has provided economists, marketers, and statisticians with a potentially rich and informative data source. Since everything on the Internet is necessarily digital, all activity can be easily recorded and stored in a database for future examination. This data has found disparate uses, from advertisement targeting to law enforcement. One prevalent but relatively under-used example of such data is clickstream data. This data consists of each website visited by a panel of users and the order in which they arrive at the websites. It is often accompanied by the time of arrival at and departure from the website as well as the degree of activity at the website and the demographic characteristics of the users. Examples of companies that collect clickstream data based on broad panels are Netratings Inc., The Economics of the Internet and E-Commerce, Volume 11, pages 209–230. Copyright © 2002 by Elsevier Science Ltd. All rights of reproduction in any form reserved. ISBN: 0-7623-0971-7

209

210

AVI GOLDFARB

MediaMetrix Inc., and Plurimus Corp. This paper uses data from Plurimus Corp. to analyze user choice of Internet portals. It will show that commonly used econometric models for examining grocery scanner data can be applied to clickstream data in advertising-based online markets. Following Hargittai (2000, p. 233), I define an Internet portal as “any site that classifies content and primarily presents itself as a one-stop point-of-entry to content on the Web.” Portals, such as Yahoo, Altavista, and MSN have search engine capabilities, but they also have other features. These may include email, news, and a link-based directory to the web separate from the search service. There are few, if any, pure search engines remaining. I narrow Hargittai’s definition further. I am interested in the portal as a starting point and not as a destination, and I therefore look at the use of portal main pages, directory pages, and search pages, but not at email and shopping pages. The methodology used here closely mimics that of Guadagni and Little’s (1983) paper that estimates a multinomial logit model with scanner data to examine consumer coffee purchases. It shows that the model has reasonably good out-of-sample predictive ability. Furthermore, informative simulations can be conducted on the effects on market share of changing a variable. For example, it can derive an estimate of the impact on number of visits of increasing advertising by one dollar. Developing a framework to study consumer choices of free (advertisingsupported) websites is an essential step to better understanding user behavior on the Internet. According to the data set used in this study, more than twothirds of all consumer Internet traffic is at advertising-supported sites. With the exceptions of Amazon and EBay, the top twenty sites in terms of unique visitors are all advertising-supported. The literature on this important aspect of the Internet is sparse. Three studies that focus on advertising-supported websites are Adar and Huberman (1999), Gandal (2001), and Goldfarb (2001). Adar and Huberman (1999) show that portals can discriminate between users as those looking for certain topics are willing to spend more time. This means that search engines can capture more consumer surplus (in the form of advertising revenue) by forcing consumers that are willing to spend more time to view more pages and advertisements. Gandal (2001) examines market share at an aggregate level to try to examine the portal market. He finds that early entrants have an advantage and that certain features matter more than others. Goldfarb (2001) examines concentration levels in advertising-supported Internet markets. Lynch and Ariely (2000) is one of few Internet studies that looks at choicespecific data. They construct a simulated environment for the purchase of wine and examine purchase choice. Like Lynch and Ariely’s study, this paper takes

Analyzing Website Choice Using Clickstream Data

211

advantage of the choice-specific data. Unlike their study, I look at the choice of free web sites using actual user clickstreams. The main data for this study was supplied by Plurimus Corporation. It is a clickstream data set consisting of every website visited by 2654 users from December 27, 1999 to March 31, 2000. It also contains data on the time of arrival at and departure from each site. In total, the data set contains 3,228,595 website visits, of which 859,587 (2622 households) are to Internet portals. Using this data, I construct measures of past search success, past time spent searching, whether a site is an individual’s starting page, whether an individual has an email account at the site, and the number of pages viewed at each site. A considerable section of this paper is dedicated to explaining the construction of these variables from the raw data. I link the Plurimus data to monthly advertising spending data from J. Walter Thompson Company and media mentions data found through the Lexis-Nexis Academic Universe. The next section of the paper provides a brief history of Internet portals. Section 3 describes the application of the methodology used by Guadagni and Little to the present problem, and Section 4 explains the data set. Section 5 presents the results, tests the model’s predictive ability, and examines market response to changes in the control variables. The paper concludes by summarizing the key results and proposing several potential areas for future research.

2. THE INTERNET PORTAL MARKET Portals operate in a peculiar environment. They compete in two distinct markets. They compete in quality for users, but users provide no direct revenue. The revenue comes from advertising. Hotwired magazine pioneered this business model in October 1994, when the first banner advertisement (for Zima alcoholic beverages) appeared on their website. The advertising market is largely competitive: portals compete with thousands of other websites, as well as television, radio, newspapers, magazines, and billboards. In their first generation, Internet portals were search engines. They maintained large databases of websites and allowed users to search them. In exchange, users viewed banner advertisements. By late 1996, it became obvious that there were too many undifferentiated search engines competing for the same users and the same banner advertisements. OpenText, once a large search engine, closed in the middle of 1997. Webcrawler and Magellen were taken over by Excite and neglected. The more successful search engines began to provide proprietary content in an attempt to differentiate themselves from their rivals. As search engines began offering email accounts, stock quotes, and

212

AVI GOLDFARB

news services, they became known as ‘portals’ because they were gateways to the Internet. Over the next few years, these portals bought other content companies and used that content to generate traffic and revenue. Yahoo expanded aggressively, buying broadcast.com, Geocities, and dozens of other smaller companies. Lycos bought Gamesville and quote.com. Excite bought Bluemountainarts (then the largest e-card company) and classifieds2000 (then the largest provider of online classified listings). Search engines were differentiated by the proprietary content they provided following a search. The richer content generated more banner advertising, but many looked to new revenue streams. ‘Partnerships’ were the first of these new revenue streams to succeed. In July 1997, Amazon became the ‘preferred book merchant’ to Excite and Yahoo. Preferred sponsor programs soon arrived at AOL and Lycos. Unlike banner advertisements, sponsors’ names appeared in response to certain keyword searches. Furthermore, their logos were placed prominently in the middle of the website. Goto.com (now Overture) took the idea of sponsorships one step further. Founded in 1998, Goto’s Internet database consists only of paid sponsors. Advertisers bid on keywords. The advertiser who pays most for a given keyword is listed first in the search results, the advertiser who paid second-most is listed second, and so on. Today, Goto continues to thrive and many other portals, including looksmart, about.com, and iwon.com, use its technology. The results in Goto’s database are supplemented by another search engine’s results if there are not enough paid listings. There are several other revenue streams that portals use. Pop-up advertisements have become widespread. Many portals charge businesses a fee to be included in their directory manually. For example, a business can pay Yahoo to get listed immediately, or wait and hope that Yahoo’s directory editors stumble across the website at some point in the future. Today, the Internet portal market is stabilizing. The last major entries occured in 1999 with Google and Iwon.com. Since then, exit has been much more common. NBCi, Go.com, and Excite have all folded. Yet, quality improvements in search technology and in content continue. Google recently added ‘pdf’ files to its search capabilities, and MSN recently entered a partnership with ESPN. One finding in this paper is that both search efficacy and usable content are important to driving users to portals.

3. USING THE MULTINOMIAL LOGIT WITH CLICKSTREAM DATA Internet users choose which website to visit just as they make several other economic choices: given the alternatives available and the information they

Analyzing Website Choice Using Clickstream Data

213

have about those alternatives, they choose the alternative that will give them the highest utility. In terms of grocery products such as coffee (studied by Guadagni and Little), this means that households buy the product that has the best attributes for the lowest price. In terms of portals, this means that households will use the portal that will allow them to maximize the probability of finding what they seek and minimizing the time spent. Conceptually, I assume households are exogenously given a “goal” when they go online. They go to the portal that they expect will help them achieve that goal in the least time with the most accuracy. In the multinomial logit model, the expected utility of the portal is based on past history, several website characteristics (that may vary over time), outside influences such as advertising and media mentions, and an idiosyncratic error term. Formally, household i visits website j on choice occasion t when (1) Euijt ≥ Euikt for all k ≠ j. Here Euijt is defined by (2) Euijt = Xijt ␤ijt + ␧ijt Xijt may include variables that change over any or all of i, j, and t. ␤ may vary over i, j, or t, implying household heterogeneity, brand heterogeneity, time (choice occasion) heterogeneity or any combination of the three. In this chapter, Xijt will never vary over just t, just i, just t and i, or just t and j. It will vary over just j in the form of portal-specific dummy variables. ␤ will be assumed constant. There are I households, J websites, and Ti choice occasions for each household. It is expected utility to the user, not to the observer, that is of interest. It is assumed that the user knows ␧ijt. The expectation is taken over relevant variables that the user may not know the value of before visiting the website. For example, the user does not know how long she will spend on the website. She does, however, have an expectation of how long it will take based on her past experience at that website. In order to get the multinomial logit form, ␧ijt is assumed to be independently distributed random variables with a type II extreme value distribution. Given the above assumptions, the probability of household i choosing brand j at choice occasion t can be expressed as: exp(Xijt ␤ijt ) (3) Pit( j | Xijt , ␤ijt ) = J

冘

exp(Xikt ␤ikt )

k=1

The model, as expressed above is a combination of Theil’s (1969) multinomial logit and McFadden’s (1974) conditional logit. It is commonly referred to as a

214

AVI GOLDFARB

mixed logit or as a multinomial logit. Since this paper assumes ␤ is fixed, the model here is a conditional logit. The log likelihood function is as follows:

冘冘冘 I

Ti

J

i=1

t=1

j=1

dijt ln Pit ( j | Xijt , ␤)

(4)

where dijt is equal to one if alternative j is chosen by individual i at time t, and is equal to zero otherwise. A significant potential problem with this framework is that it implies an assumption of independence of irrelevant alternatives (IIA). If a household is offered a new alternative that is almost identical to one of the current alternatives, say k, then this new alternative should be expected to only draw buyers from k; however, under IIA, the new alternative draws buyers from all the other alternatives. IIA is not a major factor in the questions addressed in this study. Furthermore, it complicates the econometric analysis considerably. In this model, the researcher observes the choice by each household on each choice occasion. Let yijt = 1 if household i chooses website j on choice occasion t and let yijt = 0 otherwise. The researcher also observes the characteristics of each website at that choice occasion for that household Xijt.

4. DATA 4.1. Raw Data Sources and Description The main data set consists of 3,228,595 website visits by 2654 households from December 27, 1999 to March 31, 2000. Also included in the initial data set was the time of arrival at and departure from a website, the beginning and end of each online session, and the number of pages visited at that site. This data, collected by Plurimus Corporation, was used to construct a data set of 859,587 portal choices by 2622 households. This study uses only 2008 of these households and keeps the others to test the model out of sample. Furthermore, it only looks at the eight most frequently used portals comprising 80% of all portal visits. Therefore the final data set consists of 519,705 portal choices by 2005 households. Plurimus has an anonymizing technology that allows them to collect information about users without needing the users’ permission. Plurimus avoids significant privacy concerns because the users are anonymous and the data cannot be traced to any actual person. They are regularly audited by PriceWaterhouseCoopers in order to ensure they exceed the privacy requirements of the FCC guidelines. Unlike volunteer panel data, behavioral records

Analyzing Website Choice Using Clickstream Data

215

from anonymized users are not biased by the wish to be seen in a socially desirable light. Moreover, there is no selection bias into the sample itself, yielding a sample from a broader spectrum of socioeconomic status than is typically available from panel studies. This data, however, has five limitations that need to be considered when extending the results of this study to the entire Internet. First, the geographic distribution of the sample is considerably biased. New York, Chicago, and Los Angeles are under-represented. Roughly half the sample comes from the Pittsburgh area. Another quarter is from North Carolina and another eighth is from Tampa. This problem is not as severe as it may first appear because portals are a national product. The second limitation is that it does not collect data on America Online (AOL) users. Since AOL subscribers make up roughly 50% of all American home Internet users, this could bias the results. AOL, however, provides a different product from the other Internet service providers. AOL users are encouraged to stay within the gated AOL community and they generally do not venture out onto the rest of the Internet. Moreover, preliminary surveys commissioned by Plurimus show that when AOL users do leave the gated AOL community, they have similar habits to other web users. This data limitation will, however, put a downward bias on visits to the AOL portal. Third, the data contains information on few users at work. Online habits at work are likely different from those at home; however according to a study by Nie and Erbring (2000), 64.3% of Internet users use the Internet primarily at home; just 16.8% use it primarily at work. Few data sets, however, contain reliable at-work panel data. With the exception of AOL, Plurimus’ market share numbers for Internet portals are well within the range of the other companies. The correlation coefficients for monthly market shares from January to March 2000 are 0.90 for Plurimus and MediaMetrix and 0.78 for Plurimus and PC Data Online. Since the numbers are generally quite close, the above issues with the data may not be important for understanding portal choice by users who are not AOL subscribers. The fourth limitation is that the data is collected at the household level rather than at the individual level. If two people in a given household have considerably different habits this will show up as one person with widely varying habits. While this makes it difficult to assess the extent of learning over time, it is a standard problem in consumer panels. Fifth, it does not contain information on households from the first time they go online. Therefore initial conditions are potentially a problem. Although the observations may not be independently and identically distributed, this problem

216

AVI GOLDFARB

may be partially alleviated by the law of large numbers due to the number of observations per household in the data set. More than 79% of the households in the final data set make 30 or more choices. The mean household makes 259 portal choices and the median household makes 120 portal choices. Together, these five data limitations mean that results should be extended to different geographic distributions, AOL users, and at-work users with caution. Furthermore, the fourth and fifth limitations mean that understanding learning behavior is not possible. I join this clickstream data set with two other data sets. The first is an advertising data set provided by J. Walter Thompson Company. This data set consists of all advertising spending by each of the portals used in this study on a monthly basis. The spending is determined by a thorough sampling of television, radio, newspaper, magazine, outdoor, and Internet advertising by each of the portals. The number of advertisements is then multiplied by the average cost of advertising in each medium (at the program level in television and the issue level in magazines). Since this data is not individual-specific, it will likely underestimate the impact of advertising. The methodology used in this paper, however, can easily be adapted to individual-specific advertising data. I also constructed a data set of ‘media mentions’ for each of the relevant companies. If a company is mentioned on network television news (ABC, CBS, or NBC), in the Wall Street Journal, in the New York Times, or in USA Today on a given day or the day before then the media mentions variable is equal to one. Otherwise it is equal to zero. Unfortunately, I do not know which individuals were actually watching or reading which media. It is likely, however, that mentions in these media are highly correlated with mentions in other media such as local newspapers. In the data set, several dozen portals are observed to be chosen. For computational feasibility, I limit the number of portals to the eight with the most visits (in order): Yahoo, Microsoft Network (MSN), Netscape, Excite, AOL, Altavista, Iwon, and Lycos. These eight make up 80% of all visits and all portals with more than 2.5% of total visits. There was a natural break after Lycos because the ninth most visited portal, MyWay, is a site that is the default of several Internet Service Providers and is rarely chosen as anything but a start-up page. Go.com is not included because, although it is commonly ranked in the top five portals, a large percentage of those visits are to destination websites such as ESPN.com, Disney.com, and MrShowbiz.com. The Go.com portal page itself ranks tenth in total visits. Qualitative results, however, do not change with the addition of more portals.

Analyzing Website Choice Using Clickstream Data

217

4.2. Data Set Construction I used the above information to construct several variables from the raw clickstream data. Table 1 shows a sample of ten lines of raw data. Using only this information, I constructed the following variables: email, goal of search, start page, view length at the portal, links, repeated search, whether a portal was the first visited in the search process, and Guadagni & Little’s weighted loyalty variable. I will describe the derivation of each in turn. A household was considered to have an email account at a site if the household used the email feature at that site more than that at any other portal. I know that a household used email at a given site because the ‘host’ in the data would reveal this. For example, ‘com.yahoo.mail’ is Yahoo’s email provider and ‘com.hotmail’ is MSN’s email provider. No household used more than one email account a large number of times, so I did not allow for households to have more than one portal as an email provider. Many households did not use a portal email provider. This same email variable is potentially endogenous when individual heterogeneity is not taken into account because users will set up an email account at their favorite portal. As such it can be used as a proxy for some individual heterogeneity. Furthermore, if the goal is to predict future choices or to simulate changes, then this endogeneity is not relevant. It was the initial decision to use the email that was endogenous; once that account is set up, each choice of portal is based on the existence of the email account. I cannot identify search failure exactly, but I proxy it with a repeated search variable. If a household visited two portal sites in a row, and there was less than five minutes between visits, then the first search is likely a failure. Furthermore, if the household conducts a search and then searches again for the same goal1 (at the same site or at a different one) within five minutes of the first search then the repeated search variable is equal to one. While five minutes is arbitrary, extending the time to ten minutes or shortening it to three did not change the number of repeats much. As with time spent, it is whether previous searches at a portal were repeated that matters. Also as with time spent, more complicated functions of past repeated searches do not yield qualitatively different results. I call this variable last search repeated. A portal is considered to be a household’s start page if at least 50% of all online sessions begin with that page. An online session is considered to end if a user does not do any activity for thirty minutes. While imperfect, this method determines a starting page for almost all of the households. Like, same email, start page is potentially endogenous. People often change their start page to

218

Table 1. USER

HOST

1 1 1 1 1 1 1 1 1 1

com.yahoo com.allrecipes com.ivillage com.allrecipes com.allrecipes com.excite com.adobe gov.nara gov.nara com.allrecipes

Clickstream Data Sample.

START TIME

END TIME

BYTES FROM

BYTES TO

# PAGES VIEWED AT HOST

14MAR00:08:42:55 14MAR00:08:45:28 14MAR00:08:55:00 18MAR00:12:27:10 21MAR00:12:31:01 28MAR00:13:13:59 28MAR00:13:15:06 28MAR00:13:19:38 28MAR00:13:34:09 30MAR00:16:44:18

14MAR00:08:45:28 14MAR00:08:50:59 14MAR00:09:09:48 18MAR00:12:34:46 21MAR00:12:36:51 28MAR00:13:15:22 28MAR00:13:19:39 28MAR00:13:21:57 28MAR00:13:38:00 30MAR00:16:52:05

196593 65825 541337 75403 75873 105884 70732 1259 60155 86186

34484 656 72005 4454 658 4006 11988 2340 9074 1857

3 12 53 5 2 4 9 1 13 4

AVI GOLDFARB

Analyzing Website Choice Using Clickstream Data

219

their favorite website. Again like same email, this can proxy individual heterogeneity and the endogeneity is not relevant if the goal is to predict future choices or to simulate changes. Twenty-eight percent of households have their start page at a portal. This is likely lower than the general population due to the lack of AOL users. The view length spent at a portal is the time of departure minus the time of arrival (in seconds). Recall that it is time spent during previous visits that is important for whether a household returns to that portal. The number of pages viewed at a portal may reflect the depth of search. While individuals likely want to minimize time spent generally, search depth may be an important control factor. As with view length, it is number of pages viewed during previous visits that is important for whether a household returns to that portal. This study only reports results from a one period lag on last view length and last number of pages. More complicated functions of past time spent and previous number of pages viewed do not yield qualitatively different results. Links were determined by visiting each portal and recording which websites were directly linked to the main page. I recorded links in early April for each of the portals. While it is possible that several of the links changed, there were no relevant changes in partnerships over that time. If the site that an individual visited following a portal visit was linked to a portal, the link variable takes on a value of one. Otherwise, it equals zero. Note that the link variable can equal one even if the household did not visit that portal. For example, a household could search for financial information on Yahoo and the search may turn up information on MSNmoneycentral. The link variable serves as a proxy for portal features. Instead of listing whether a portal has features, this variable proxies whether people actually use these features. In other words, if people use a link, it means they are using a feature at that site, rather than the search capabilities. If a portal was the first visited in the search process, then firsttryijt = 1. If an individual has already searched, then firsttryijt = 0. This paper mimics Guadagni and Little’s methodology for constructing their ‘loyalty’ variable almost exactly. In their paper, loyalty is considered to be a weighted average of past purchases of the brand, treated as dummy variables. Let portsameijt = 1 if household i bought brand j as its previous purchase and zero otherwise. loyaltyijt ⬅ ␣loyaltyijt ⫺ 1 + (1 ⫺ ␣)portsameijt

(5)

220

AVI GOLDFARB

Rather than estimate ␣ by maximum likelihood which would significantly complicate the computational problem they calibrate ␣ based on dummies for lags of length one to ten. In the present study, the value for alpha that minimizes the sum of the difference between the actual dummy coefficients and the loyalty function above was 0.7782. This loyalty variable can be a result of either individual preferences for a given portal or from some kind of lock-in. I do not separate these two effects, but the variable is an important predictor of portal choice. In a recent study, Abramson, Andrews, Currim, and Jones (2000) find this to be the best loyalty measure they tried. Defining loyalty as portsame rather than GL Loyalty does not change the qualitative results. In this study, I define the portsameijt variable to depend on the previous portal visited of any kind, not just the previous of the eight portals used in this study. Therefore, if a household visits Yahoo then About.com and then Yahoo again, portsameijt on the second visit to Yahoo is equal to zero, even though only two observations are included in the data set. This means that a household is not considered brand loyal if it went to a rival portal’s website, even if that rival portal is not in the sample. If I only include the sample, the coefficient on the loyalty variable increases slightly but its significance falls slightly. The initial conditions problem frequently encountered in this literature does not apply here due to the large number of observations per household. How much time a household’s previous visit to a portal took and whether that search was repeated are only observed when the household has visited that portal previously in the data set. Since not every household visits every portal, these variables are missing for a large number of observations. I therefore created a dummy variable for missing data. I also interact one minus the missing data variable with the view length of previous search and the repeated search variables. This overcomes the significant potential bias of assuming a value for the missing data or of ignoring it entirely. The missing data dummy has no economic interpretation. Table 2 contains descriptive statistics of the final data set. Yahoo has over twice the market share of its closest competitor, MSN. This table suggests that Yahoo’s success may be largely a function of the features of its website. Searches are repeated much less often on Yahoo than elsewhere, it is the start page for the largest number of users, and it is the email provider for the second largest number of users. Furthermore, Yahoo advertises heavily, is frequently mentioned in the media, has frequently used links, and does not take long to search. Lycos searches, on the other hand, are repeated frequently, Altavista’s links are rarely used, and only Yahoo and MSN have a large number of email users.

Summary Statistics.

Percent share of visits to top 8 portals

Average time spent at site (in seconds)

Percentage of times search repeated

Percentage of households with portal as start page

Percentage of households with same email

Percentage of visits using a link

Percentage of days with media mentions

Average monthly advertising spending (thousands of dollars)

Yahoo MSN Netscape Excite AOL Altavista Iwon Lycos

42.0 20.9 13.5 6.5 5.5 5.0 3.6 3.0

96.67 116.72 114.0 93.21 93.89 109.7 152.0 96.21

7.03 12.10 13.33 11.28 11.11 14.41 14.81 31.55

9.76 7.17 5.38 1.29 0.75 0.30 0.30 0.20

19.92 32.97 4.38 2.39 4.48 0.40 1.59 4.63

3.20 4.41 3.62 2.57 2.78 0.17 0.69 1.82

58.33 6.35 13.54 15.63 82.29 5.21 1.04 16.67

2361.5 277.4 198.7 397.7 7263.4 1161.0 0 1570.2

Mean over all observations Standard deviation over all observations

N/A

105.45

15.30

2.41

11.32

1.87

33.92

1772.5

N/A

171.59

36.01

15.34

31.72

13.53

47.34

2389.6

Portal

Analyzing Website Choice Using Clickstream Data

Table 2.

221

222

AVI GOLDFARB

5. RESULTS 5.1. Coefficients Table 3 presents the main results of the paper. Model (1) presents the base model. Here, the potentially endogenous variables of same mail, link, and start page are not included. The variables all have the expected signs, although last view length is barely significant: loyalty, advertising, and media mentions are all correlated with a higher probability of search. Last view length and last search failed are all correlated with a lower probability of search. The positive sign on last view length squared suggests that the effect of last view length is concave. There was no expectation on the sign of missing data. The coefficient on advertising likely underestimates the actual effect of advertising as the data is aggregated over the month rather than actual advertising viewed by the user. Model (2) adds same email and link with the expected results. Taking these into account makes last view length significant. Model (3) adds last number of pages and first try. Last number of pages is found to have an increasing and concave relationship with choice probability. This is consistent with the assumption that pages viewed proxy depth of search. In this regression, last view length is significant at the 99% confidence level. Thus, controlling for depth, households prefer to spend less time at a portal. First try reveals that Netscape and MSN are preferred as first pages in a search than as later pages. This makes sense as they are the pages that appear when using the search function in the Netscape Navigator and Microsoft Internet Explorer browsers. They are also often default start pages, but the results do not change in models (4) through (6) which control for the start page. Model (4) adds the start page variable to model (2). The coefficient on this variable is very large compared to the other dummy variables and the likelihood improves more for this variable than for any others; however, the coefficient is not significantly different from zero as it has an extremely high standard error. Model (5) is the same as model (4) except that is adds the interaction variable of media mentions and loyalty. Of particular interest here is the increase in the significance of media mentions. This suggests that being mentioned in the media has a larger effect for households that are less loyal to the brand. Model (6) is the ‘kitchen sink’ regression in that it includes all of the variables in the study. The coefficients and their significance are similar to models (1) through (5).

Variable GL Loyalty Missing Data Last view time at that site Last view time squared Last search failed Advertising ($000) Media Mentions Media Mentions*loyalty Same email Link Last number pages viewed at that site Last number of pages squared Start page

Model coefficients (with standard errors in parentheses). Model (2)

Model (3)

Model (4)

Model (5)

Model (6)

1.35*** (0.00235) –2.35*** (0.0126) –1.90E-05∧ (1.34E-05) 2.08E-09** (9.89E-10) –0.476*** (0.00608) 5.89E-06* (3.01E-06) 0.0137** (0.00667)

1.31*** (0.00245) –2.32*** (0.0127) –2.20E-05* (1.35E-05) 2.31E-09** (9.87E-10) –0.440*** (0.00618) 6.08E-06** (3.07E-06) 0.0136** (0.00680)

1.32*** (0.00247) –2.28*** (0.0129) –0.000120*** (1.59E-05) 6.69E-09*** (9.90E-10) –0.425*** (0.00620) 6.17E-06** (3.09E-06) 0.0124* (0.00683)

1.21*** (0.00261) –2.26*** (0.0129) –2.60E-05* (1.41E-05) 2.43E-09** (9.87E-10) –0.451*** (0.00645) 4.59E-06∧ (3.17E-06) 0.0109∧ (0.00712)

0.166*** (0.00511) 1.98*** (0.0109)

0.174*** (0.00513) 2.02*** (0.0110) 0.0103*** (0.000710) –6.70E-05*** (9.19E-06)

0.174*** (0.00544) 2.05*** (0.0113)

1.27*** (0.00368) –2.24*** (0.0130) –2.60E-05* (1.41E-05) 2.49E-09** (9.93E-10) –0.452*** (0.00646) 5.30E-06* (3.16E-06) 0.129*** (0.00857) –0.144*** (0.00590) 0.181*** (0.00544) 2.06*** (0.0113)

34.12 (146.12)

41.11 (247.87)

1.27*** (0.00368) –2.21*** (0.013165) –0.000110*** (1.67E-05) 5.95E-09*** (9.97E-10) –0.451*** (0.00646) 5.53E-06* (3.16E-06) 0.128*** (0.00857) –0.143*** (0.00590) 0.181*** (0.00544) 2.05*** (0.0113) 0.00875*** (0.000726) –5.10E-05*** (8.38E-06) 36.11 (203.90)

223

Model (1)

Analyzing Website Choice Using Clickstream Data

Table 3.

224

Table 3.

Continued.

Model (1)

Model (2)

Model (3)

Model (4)

Model (5)

Model (6)

Altavista

–0.530*** (0.0103) –0.571*** (0.0169) –0.479*** (0.00971) –0.415*** (0.0135) –0.686*** (0.0105) –0.0270*** (0.00953) –0.157*** (0.0101)

–0.494*** (0.0105) –0.700*** (0.0173) –0.612*** (0.0101) –0.430*** (0.0138) –0.808*** (0.0108) –0.174*** (0.00971) –0.261*** (0.0104)

–0.287*** (0.0141) –0.764*** (0.0202) –0.548*** (0.0145) –0.662*** (0.0204) –0.489*** (0.0147) –0.592*** (0.0128) –0.695*** (0.0144) –0.393*** (0.0169) 0.0924*** (0.0167) –0.126*** (0.0171) 0.321*** (0.0219) –0.580*** (0.0195)

–0.248*** (0.0142) –0.726*** (0.0205) –0.540*** (0.0147) –0.633*** (0.0205) –0.494*** (0.0149) –0.654*** (0.0133) –0.779*** (0.0150) –0.345*** (0.0170) 0.135*** (0.0169) –0.153*** (0.0176) 0.361*** (0.0221) –0.468*** (0.0197)

–0.246*** (0.0142) –0.769*** (0.0205) –0.543*** (0.0147) –0.639*** (0.0206) –0.499*** (0.0148) –0.674*** (0.0133) –0.791*** (0.0150) –0.353*** (0.0171) 0.137*** (0.0168) –0.165*** (0.0176) 0.357*** (0.0223) –0.474*** (0.0196)

–0.258*** (0.0142) –0.779*** (0.0205) –0.553*** (0.0147) –0.662*** (0.0207) –0.496*** (0.0148) –0.670*** (0.0133) –0.798*** (0.0151) –0.353*** (0.0171) 0.139*** (0.0168) –0.168*** (0.0177) 0.361*** (0.0223) –0.475*** (0.0196)

AOL Excite Iwon Lycos MSN Netscape First Try (Altavista) First Try (AOL) First Try (Excite) First Try (Iwon) First Try (Lycos)

AVI GOLDFARB

Variable

Variable

Model (1)

Model (2)

First Try (MSN) First Try (Netscape) Log likelihood

–442,856

–425,651

Continued. Model (3)

Model (4)

Model (5)

Model (6)

0.631*** (0.0123) 0.646*** (0.0144)

0.632*** (0.0129) 0.668*** (0.0154)

0.632*** (0.0129) 0.665*** (0.0154)

0.633*** (0.0129) 0.667*** (0.0154)

–421,531

–386,956

–386,659

–386,581

Analyzing Website Choice Using Clickstream Data

Table 3.

*** significant at a 1% level in a two-tailed test. ** significant at a 5% level in a two-tailed test. * significant at a 10% level in a two-tailed test. ∧ significant at a 10% level in a one-tailed test.

225

226

AVI GOLDFARB

Another interesting aspect of all of the models is that there is a clear brand preference for Yahoo over the others. Models (1) through (3) have negative coefficients for all brand dummies (Yahoo is the base). Models (4) through (6) also have negative dummies for Yahoo. While others may seem preferred on the first try, adding the coefficients together leaves a negative number meaning that Yahoo is generally preferred even on the first try. The Akaike information criterion revealed that last view length squared, last number of pages squared, and media mentions*loyalty should be included. Other variables such as advertising squared and advertising*loyalty did not satisfy the Akaike information criterion. Note that including start page increases the likelihood a great deal, even though the effect is statistically insignificant. Any variables included in this study that satisfy the Akaike information criterion also satisfy the Bayesian information criterion. The model also displays good out-of-sample predictive power. I calculated predicted and actual market shares using an outside sample of roughly 600 households. For the predictions, I used model (1) and model (6). I found that the predictions match actual market shares closely. The correlation coefficients between predicted and actual market shares of the eight portals over fourteen weeks (n = 112) are 0.78 for model (1) and 0.86 for model (6). Most fluctuations in market share are captured and, while not perfect, this model has significant predictive poower. It could be used to explore how policies in one market would work in another.

5.2. Market Response to Variable Changes

Table 4 explores the market responses to variable changes in model (2) assuming no competitive response. This table presents elasticities in the form of changes in number of total visits over the three month period, assuming that there are a total of 43.3 million online households, Plurimus’ estimate for the month of February 2000. Taking the results at face value, if MSN users’ searches were repeated just 1% less often, MSN would get almost 1.7 million more site visits. If each site visit is worth five cents (about the revenue received from the five advertisements seen over typical two page views at a typical search engine), then it would be worth it for MSN to implement this change as long as it cost less than eighty-five thousand dollars over three months. Links, search time, and search failure (proxied by repeated search) matter, suggesting that usable content and search efficacy are important factors in driving users to portals.

Altavista AOL Excite Iwon Lycos MSN Netscape Yahoo

Increase in Number of Site Visits Over Sample Period Due to Small Changes in Variable*. Increase advertising by one dollar

One more media mention

Searches take one second less on average

Searches repeated 1% less often

Links used 1% more often

3.70 3.52 12.60 0.0160 1.91 11.65 8.42 15.02

13,137 1,296,860 164,221 7819 18,951 630,456 3,344,629 1,780,132

6761 8088 8551 2385 4134 30,042 17,572 48,501

352,195 329,865 399,711 116,384 310,342 1,676,960 873,186 1,368,822

175,599 3,154,163 3,532,417 532,908 1,143,310 16,626,927 9,856,314 160,735

Analyzing Website Choice Using Clickstream Data

Table 4.

* Assumes 43.3 Million total online households. This is Plurimus’ estimate of the total number of online households in February 2000.

227

228

AVI GOLDFARB

The advertising results are perhaps the most interesting. An increase in advertising by one dollar would bring 3.7 more visits to Altavista but 15 more to Yahoo. Therefore, Altavista should increase its advertising if each new site visit brings in twenty-seven cents of revenue and Yahoo should increase its advertising if each new site visit brings in just 6.7 cents of revenue. More effective links and more media mentions are other ways companies can drive traffic to their websites. Caution should be used in interpreting these results because of the lack of IIA and because the functional form of the error term is important to deriving these results. While the numbers themselves may not be completely accurate, it is likely that an extra dollar of advertising by Yahoo has a larger effect than an extra dollar of advertising by Altavista. The current exercise should be viewed as an approximation that demonstrates potential marginal gains from the variables. Another way to simulate policy changes by the firms is to change the underlying data and reestimate the market shares given the known coefficients. This method underestimates changes because it does not count dynamic effects. It does, however, provide a lower bound for the impact. Again using model (2), I undertook this exercise for several variables. If MSN advertised as much as AOL, then MSN would gain 7,843,659 more visits assuming 43.3 million households. If, on the other hand, Iwon advertised as much as AOL then it would only gain 1,617,622 visits. If Lycos searches were successful as often as Yahoo searches, Lycos traffic would rise by 14,561,538 or 4%. If Altavista had the same links as MSN then it would get 56,005,914 more visitors or 10%. The exact quantities of these predictions should be interpreted with caution. The general trends, however, are informative.

6. CONCLUSION This study has provided a preliminary look at estimating demand for advertising-supported Internet websites based on clickstream data. The methodology provides a reasonable fit to the actual patterns in the data. It has good predictive power and is informative about the potential impact of various policy changes. With respect policy implications, the study provides a framework for understanding policy effects. The simulations in section 5.3 show the impact of potential policy changes on market shares. While they do not take into account supply side reactions or individual heterogeneity, they do give better estimates of policy effects than currently exist. More detailed policy analysis can also be

Analyzing Website Choice Using Clickstream Data

229

explored in this framework. For example, a portal could simulate a link to a commonly used website, say americangreetings.com. It could then determine the effect of this link on market share. The actual increase in share resulting from this change would be no more than the simulated level. It may be less because it may be that people who go to a given portal are also the kind of people who like the links it has. Thus the effects of the new link may be less than predicted. Because it does not account for individual heterogeneity, this model does not provide an effective framework for examining the effects of major industry changes such as bankruptcies, nor does it provide a way to look at the welfare impact of improved technology. The main purpose of this study was to show that demand for free online services can be estimated using methodologies that are common in both the economics and the marketing literature. The coefficients on the variables in the study have the expected signs and the predictive ability of the model, though not perfect, captured the major trends. Furthermore, I present informative simulations about the effects on share of changing variable values. Clickstream data will be an important tool in understanding online demand. This study has shown that the standard econometric methods that have previously been applied to grocery scanner data can successfully be applied to clickstream data. By bringing more econometric sophistication to this analysis, economists and marketers can gain a better understanding of online user behavior.

NOTE 1. The goals were divided into roughly one hundred overlapping categories including news, music, email, shopping for computers, automotive information and travel. I did not include goal of search in the final analysis because including it did not satisfy the Akaike information criterion or the Bayesian information criterion for goodness of fit.

ACKNOWLEDGMENTS This research was supported by the Social Science Research Council through the predissertation fellowship of the Program in Applied Economics and by a Plurimus Corporation Research Fellowship. I would like to thank Plurimus Corporation for providing me with the clickstream data and J. Walter Thompson Company for providing me with advertising data. I would also like to thank Shane Greenstein, Charles Manski, and Robert Porter for helpful comments. All remaining errors are my own.

230

AVI GOLDFARB

REFERENCES Abramson, C., Andrews, R. L., Currim, I. S., & Jones, M. (2000). Parameter Bias from Unobserved Effects in the Multinomial Logit Model of Consumer Choice. Journal of Marketing Research, 37, 410–426. Adar, E., & Huberman, B. (1999). The Economics of Surfing. Working Paper No. 42, Center for eBusiness at MIT. Gandal, N. (2001). The Dynamics of Competition in the Internet Search Engine Market. International Journal of Industrial Organization, 19, 1103–1117. Goldfarb, A. (2001). Concentration in Advertising-Supported Online Markets: An Empirical Approach. Unpublished Working Paper, University of Toronto. Guadagni, P. M., & Little, J. D. C. (1983). A Logit Model of Brand Choice Calibrated on Scanner Data. Marketing Science, 2, 203–238. Hargittai, E. (2000). Open Portals or Closed Gates? Channeling Content on the World Wide Web. Poetics, 27, 233–254. Lynch, J. G., & Ariely, D. (2000). Wine Online: Search Costs Affect Competition on Price, Quality, and Distribution. Marketing Science, 19, 83–103. McFadden, D. (1974). Conditional Logit Analysis of Qualitative Behavior. In: P. Zarembka (Ed.), Frontiers of Econometrics (pp. 105–142). New York: The Academic Press, Inc. Nie, N. H., & Erbring, L. (2000). Internet and Society. Unpublished Working Paper, Stanford Institute for the Quantitative Study of Society. Theil, H. (1969). A Multinomial Extension of the Linear Logit Model. International Economic Review, 10, 251–259.

CONSUMER ACQUISITION OF PRODUCT INFORMATION AND SUBSEQUENT PURCHASE CHANNEL DECISIONS Michael R. Ward and Michelle Morganosky ABSTRACT We examine consumer use of the Internet as a product information gathering tool (distinct from the use of it for transaction completion). We use data from surveys to estimate how consumer use of different marketing information channels (Internet, print, catalog, broadcast) affects the choice of channel for purchasing the good (Internet, retail, direct mail). Across many product categories, we find that Internet product information gathering increases the likeliness of purchase in other channels. Similar effects from other informational channels are not observed. Our findings have implications for measuring the retail impact of the Internet, the assumption of better informed consumers, the competitiveness of off-line, as well as online markets, and the management of “channel conflict.”

INTRODUCTION Retailers have understandable concerns about the profitability of marketing online. Developing and maintaining an Internet presence with current product information and the ability for customers to source goods online entails a The Economics of the Internet and E-Commerce, Volume 11, pages 231–255. Copyright © 2002 by Elsevier Science Ltd. All rights of reproduction in any form reserved. ISBN: 0-7623-0971-7

231

232

MICHAEL R. WARD AND MICHELLE MORGANOSKY

non-trivial expense for most retailers. The presence of new marketing channels could allow retailers to reach new sets of customers who would otherwise not make purchases with the retailer and, as such, could provide a means for increasing sales volume. On the other hand, customers purchasing through the online channel might be drawn from a company’s existing customer base. In this case, additional costs are borne with limited additional sales volume and these at possibly lower profit margins (Brynjolfsson & Smith, 1999). Therefore, if inter-channel intra-firm consumer substitution is likely, cannibalization of traditional marketing channels could render online retailing less than profitable. We investigate consumer search and purchase behaviors in the context of a multiple channel retail environment. A model of consumer information and purchase decisions across channels is presented and empirically estimated. The model addresses how search behavior and channel familiarity many affect purchase decisions across different channels. As modeled, consumers make purchase decisions aligned with their information search strategies and channel familiarity. However, we allow for consumer information search in one channel, to lead to purchases in another (e.g. online search leading to store purchase). We hypothesize that inter-channel complementarities are likely to play a role, especially for online searching. Our approach is essentially an economic one. We assume that consumers choose a purchase channel so as to maximize the expected benefits of purchases to be made. Benefits include finding a product that best matches their preferences and minimizes their transaction costs. These transaction costs include both price and other non-pecuniary costs incurred by the shopper (e.g. delivery costs, time spent on product search, riskiness of the seller, and mental effort in choosing). The focus is on comparisons between channels and the benefits that various channels may provide. While our approach is primarily based on economic theory and methods, there are important marketing applications. A fundamental question for most retailers is how consumers respond to various marketing offers made by the firm. According to stimulus-response models (Schiffman & Kanuk, 1997) marketing stimuli influence consumer buying decisions as reflected in consumer response to marketing offers. We analyze consumer responses in the context of a retail environment that consists of multiple channel choices for both search and purchase decisions. Models of consumer behavior typically include consideration of the buyer (consumer), the exchange process, the marketer’s strategy, individual influencers, and environmental influencers. Resources exchanged between buyers and

Consumer Acquisition of Product Information

233

sellers often include goods, services, information, time, and money. While exchange is considered the core concept of marketing (Bagozzi, 1975), it is important to remember that exchange involves more than simple transfers of money for products between buyers and sellers. In fact, consumers often bring a unique combination of “costs” into the exchange including monetary, time, psychic, and energy costs (Woodruff, 1997). Consumers are continually exposed to information that may influence both search and purchase decisions. Therefore, marketers and researchers are interested in knowing how consumers respond to marketing offers made in the context of different channels. According to Akhter (1989), marketers who send messages to consumers are source channels and media used to communicate with consumers are mode channels. Alternatives used by consumers to purchase or place orders are response channels. Consumers may possess schemas in relation to each of these which generate favorable or unfavorable responses. Although a schema may exist independently, its content is influenced by interactions with the other schemata. Furthermore, schematic representations may be modified in light of new information or new channels (e.g. Internet). For example, Szymanski and Hise (2000) found that shopping online reduces cognitive effort, thereby improving consumers’ satisfaction with the online experience. How do consumers adjust their thinking and decision making in the context of new and emerging channel choices? Consistency theories suggest that consumers behave so as to maximize the internal consistency of their cognitive systems, lowering the overall “costs” associated with search and purchase. Accordingly, frames of reference tend toward maximal simplicity and change in the direction of increased congruity with an existing frame of reference. Typically, the magnitude of shift is inversely proportional to the intensity of interacting evaluations (Fishbein & Ajzen, 1975). Thus, evaluations tend to change toward the direction of congruity.

MODEL We model how consumer heterogeneity can lead to different consumer choices and may produce distinct substitution patterns across channels. The source of the heterogeneity in consumers could be their preferences, as suggested by Pashigian and Bowen (1994). However, differences in information gathering costs, as suggested by Benabou (1990, 1993) and Carlson and McAfee (1983) are also possibilities. Information about a consumer’s channel usage for other

234

MICHAEL R. WARD AND MICHELLE MORGANOSKY

product categories is used here to model channel usage in a particular product category. In this regard, the model is similar to Ainslie and Rossi (1998), Rossi, McCulloch, and Allenby (1996), and Messinger and Narasimhan (1997). In our model, consumers attempt to maximize expected future benefits by choosing the amount and type of search that tends to best match product choices with preferences and reduces transaction costs. Gathering information from a particular source affects transaction costs for the associated marketing channel as consumers become more familiar with the exchange patterns particular to that channel. We call these congruent effects. Information gathering could also lead to a preferred product selection independent of purchase channel choice. Therefore, the information has value regardless of how or where the product is eventually purchased. Our model allows for this possibility of information gathering from one source affecting purchase behavior in non-congruent marketing channels. In particular, we wish to determine if searching for product information online leads to more purchases online at the expense of purchases via other channels or if it encourages purchases in these other channels. A product, X, can have varying amounts of attributes so that Xj = (Xj1Xj2, . . . , XjN ). Attributes include features and characteristics as well as the product’s price and the transactions costs of obtaining the good. Consumers wish to maximize their net utility by choice of product, max U(Xj ) ⫺ TC(Xj ) w.r.t. j. Unless they undertake efforts to investigate potential product purchases, consumers do not know the actual bundle of attributes that any particular product possesses. The set of possible goods over which consumers choose is determined by the amount of product search they undertake. Let S be this choice set. The elements of S are random but the number of items in S is increasing in the amount of product information gathering I. In expected value terms, utility increases as more choices are added to S but at a decreasing rate. Information gathering is costly. For tractability, assume that these costs are linear:

冋

册

max max U e(Xj ) ⫺ TC e(Xj ) ⫺ ␣I I

j苸S(I )

This is solved by backward induction. First, the consumer maximizes expected utility minus transactions costs by choice of product for different amounts of information gathering, I, and then maximizes the resulting function of I by choice of information gathering intensity. Under fairly general conditions,

Consumer Acquisition of Product Information

235

these will be smooth functions yielding an interior solution for I. That is, consumers will conduct some product search but will be less than fully informed about all product choices. Essentially, more information increases the possible choices in S, which make it more likely to find a product j that increases utility, or U(Xj ) ⫺ TC(Xj ) becomes U(I ) where U ⬘ > 0 and U ⬙ < 0. The first-order condition for a maximum is U ⬘(I *) = ␣ and the choice of j depends on the realization of S(I *). Now consider that information can be obtained from multiple different sources. In our application, there will be four information sources: Online (O), News and Magazines (N ), Direct Mail (M ) and Television and Radio (T ). The consumer makes four separate information gathering decisions, IO, IN , IM , and IT . The set S increases with each of these information gathering strategies, possibly at different rates, S = S(IO, IN , IM , IT ). Moreover, each of these strategies can have different costs, ␣O, ␣N , ␣M , and ␣T . The same product can be purchased through any of several marketing channels. In this application, there will be three marking channels for purchase: Online (O), Off-line Retail (R), and Catalog (C ). While the physical product attributes may not differ across marking channel, some other attributes might. In particular, the price and the transactions costs of purchasing are likely to differ. In this application, we are not able to distinguish between the transactions costs of purchasing different products within the same channel, but transactions costs varying across channels and with information gathering will prove to be important. Therefore, we denote TC(Xj ) as ␤c + ⌺c␤cl Icl where c denotes the purchase channel and l denotes the information source. Similar to above, consumers first choose their information gathering strategy, the values of IO, IN , IM , and IT . Based on this information, they simultaneously choose the which product j and which marketing channel O, R, or C. Formally, this becomes: max

IO, IN , IM , IT

冋

max

c = O, R, D j苸S(IO, IN , IM , IT )

U e(Xj ) ⫺ ␤c ⫺

冘

l = O, N, M, T

␤cl Il ⫺

冘 册 ␣l Il

l = O, N, M, T

Note that the transactions costs could depend on the channel used for purchase, but that the utility obtained from consumption is independent of the channel used for purchasing the product. Again, under general conditions, first-order conditions will yield Ul(Il ) = ␤cl + ␣l and a solution with Il > 0 for at least some of l = O, N, M or T. Again the consumer typically will be less than fully informed about all product choices. The choice of purchase channel c is given

236

MICHAEL R. WARD AND MICHELLE MORGANOSKY

by the minimum of (␤c + ⌺c␤cl Icl ). We are most concerned with two things: the formation of consumers’ information gathering strategies and how these strategies affect consumers’ choices of marketing channel used to purchase the product. The actual choice of product, j, while important for a particular marketing application, is beyond the scope of this paper. Two factors could affect a consumer’s choice of intensity of product search for a particular information source – its cost-effectiveness and its congruity with expected purchase channel choice. Cost-effectiveness, in this context, could mean that the extent to which the set S increases relative to search costs with more information gathering. Higher costs of gathering information, ␣l, will lead to less product search. Across sources, it may be less costly to find information online than through newspapers and magazines, ␣O < ␣N. If so, consumers will tend to gather more information online than through printed material. This difference may be observed by comparing two different types of products. Some products benefit greatly from database lookup functions provided online (air travel arrangements) while others may not (apparel). Across individuals, it may be less costly to find information online for some individuals than others. In particular, consumers who have developed more human capital specific to that information source, HIic, will have lower information gathering costs or ␣⬘c(HIic) < 0. Hence consumer i will tend to gather more information online than will consumer k when HIic > HIkc. For example, consumers who are more computer literate or who have developed more human capital specific to online searching will tend to have lower values of ␣O and, consequently, higher values of IO. By congruity, we mean the extent to which information from a particular source consistently affects a particular purchase channel. For example, while information about products obtained from direct mailings could be used to select a product regardless of purchase channel, it will quite often lower the transactions costs of catalog shopping by a larger amount than other purchase channels. As a consequence, consumers who know they are more inclined toward catalog shopping tend to invest more information gathering effort toward direct mail sources. In terms of the model, purchase channel c and information source l are congruent if transactions costs for channel c fall with more information from source l, or ␤cl < 0. We expect that consumers with more human capital specific to purchase channel c, HPic, will tend to gather more information from those sources that are congruent to c. In general, we expect direct mail to be congruent with catalog purchasing, online search to be congruent with online purchasing and news, magazine, television, and radio to be congruent with traditional retail.

Consumer Acquisition of Product Information

237

In general, theory predicts that product information gathering from a source is a function of human capital specific to the various sources and to purchase channels. These behavioral relationships are represented in vector notation by: ⭸U i⫺ 1 ([␤l(HPid ) + ␣l(HIic)]) (1) [Iil ] = ⭸Il where l in O, M, N, T. We expect a positive effect of HIl on Ic if c = l and a positive effect of HPd on Il if information source c is congruent with purchase channel d. The choice of marketing channel for purchase is affected by both human capital and congruity. First, consumers who have more general experience with a particular channel will be more likely to favor that channel for any particular purchase decision. Essentially, the human capital acquired by the consumer that is specific to that channel, HPic, tends to be transferable across product categories. This human capital will tend to lead to lower transactions costs for purchases made through this channel, ␤⬘c(HPic) < 0, and thus more purchase through this channel. Second, as mentioned above, information gathering from a particular source will lower transactions costs for a particular marketing channel if ␤cl < 0. For example, online information gathering is likely to lower the transactions costs of online purchasing than any other marketing channel. Evidence of behavior that links an information gathering source to its related marketing channel provides support for congruity theory. To sum up, the model predicts that choice of purchase channel is a function of human capital specific to the purchase channels and of the source of product information. Let Pic denote whether consumer i purchases through channel c. Then, for any given purchase decision, these behavioral relationships are represented in vector form by: Pic = 1

iff

((␤c([HPi]) ⫺ ␤d ([HPi])) +

冘

(␤cl ⫺ ␤dl )Iil < 0)

(2)

for all d in O, R, C. That is, the consumer choose the purchase channel with the least transactions costs given her human capital specific to purchase channels, HPiO, HPiR, HPiC , and her prior search, IiO, IiN , IiM , IiT . We expect a positive effect of HPid on Pic if c = d and a positive effect of Iil on Pic if information source l is congruent with purchase channel c. Of particular interest, however, is evidence of non-congruent linkages. In particular, does online information gathering increase the probability that a consumer will purchase in traditional retail outlets or through catalogs? Is this effect larger than the comparable effect from printed material to online purchasing? Are these patterns consistent across different product categories?

238

MICHAEL R. WARD AND MICHELLE MORGANOSKY

DATA DESCRIPTION AND EMPIRICAL SPECIFICATION To empirically examine inter-channel search and purchase patterns, we analyzed survey data from Georgia Institute of Technology’s Graphics, Visualization, and Usability Center (GVU) on e-commerce usage and similar survey data from the Illinois Project on Internet Economics (IPIE). The GVU data are from their 9th and 10th surveys of e-commerce, traditional retailing, and direct-mail usage conducted during the Spring and Fall of 1998 and, combined, include responses from 1,786 participants. Since the GVU surveys were discontinued, IPIE conducted a similar survey of 1,237 respondents during from November, 2001 through March, 2002. The specific questions asked of respondents and used in our analyses are presented in Table 1. The percentage of total “respondents’ ” information searching and purchasing by product category are summarized in Tables 2 and 3. The GVU surveys asked separate questions for computer hardware and software less than and greater than $50. For analysis purposes, the two dollar amounts were collapsed into one category. As expected, Table 2 indicates that consumers tend to purchase most often through traditional retail and least often through catalogs and that catalogs are especially unlikely choices for travel and investments. This sample tended to obtain product information online about as often as it did from newspapers and magazines. A similar table for the IPIE data is presented in Table 3. It includes two additional product categories and an additional

Table 1.

Search and Purchase Survey Questions.

Product Search Question: “Before making a purchase, you might seek information from a variety of external sources. Listed below are several types of products/services. For each product or service please indicate whether you do (check the checkbox) or do not (leave checkbox blank) consult the information sources.” Choices are: On-Line, Newspapers/Magazines, Direct Mail (and TV/Radio in the IPIE survey) Product Purchase Question: “When buying an item, we usually have several choices regarding who and where to buy it from. Listed below are several products and services that you might have bought in the last six months, and several types of shopping outlets. For each product or service please indicate whether you have bought it within the past six months (check the checkbox) or not (leave checkbox blank) from each of the types of outlets.” Choices are: On-Line, Catalogs, Retail Stores

Consumer Acquisition of Product Information

Table 2.

239

Product Purchase Channel and Information Gathering Source for GVU Respondents (1998) by Product Category.

Comp. Hardware Comp. Software Books Music Investments Travel

Product Purchase Traditional Online Retail Catalog

Information Gathering Newspapers Direct Online & Magazines Mail

43.1% 55.9% 52.5% 39.9% 16.7% 38.7%

83.6% 84.5% 72.8% 62.8% 42.2% 70.1%

72.0% 69.1% 78.3% 72.6% 14.5% 35.8%

32.3% 31.2% 22.9% 24.2% 2.7% 3.3%

76.9% 72.5% 54.8% 45.4% 41.2% 42.6%

34.5% 34.4% 26.5% 24.0% 10.9% 13.9%

information source, television and radio. The proportion of respondents using direct mail and purchasing through catalogs is much lower in this sample and is more likely to correspond current patterns. In this sample, consumers still purchase at traditional retail outlets more often than online, but now appear to gather information online more often than through newspapers and magazines. A purchase in one product category may represent a larger dollar transaction than another. Table 4 reports evidence of the size of an online transaction across product categories. While many consumers have purchased books and music online, few have spent over $500. On the other hand, less than half as many

Table 3.

Product Purchase Channel and Information Gathering Source for IPIE Respondents (2001–2002) by Product Category. Product Purchase Information Gathering Traditional Newspapers Direct Television Online Retail Catalog Online & Magazines Mail & Radio

Hardware Software Books Music Investments Travel Apparel Home Electronics

37.7% 45.1% 50.4% 49.2% 20.4% 44.0% 35.4% 35.0%

56.5% 56.3% 63.5% 60.1% 13.7% 19.8% 69.4% 61.8%

8.4% 7.4% 8.2% 9.1% 2.0% 1.4% 19.7% 8.6%

69.0% 71.7% 61.2% 67.0% 39.8% 62.8% 45.4% 64.7%

43.6% 43.2% 42.8% 33.8% 31.5% 23.0% 39.2% 44.1%

11.3% 10.7% 16.8% 14.1% 7.4% 8.2% 22.9% 17.8%

11.6% 11.0% 13.9% 30.8% 11.7% 7.8% 15.0% 19.9%

240

Table 4.

MICHAEL R. WARD AND MICHELLE MORGANOSKY

Dollars Spent Online in Previous Six Months for IPIE Respondents (2001–2002) by Product Category.

Hardware Software Books Music Investments Travel Apparel Home Electronics

$0

Under $50

$50– $100

$100– $500

$500 or more

Don’t Know

38.7% 34.9% 30.0% 30.7% 62.1% 43.5% 40.6% 45.5%

5.4% 19.5% 30.4% 31.8% 1.7% 1.7% 9.7% 7.2%

8.0% 16.0% 13.7% 16.9% 2.0% 2.6% 13.6% 11.5%

13.9% 11.5% 10.9% 5.8% 3.4% 16.7% 15.4% 13.4%

18.5% 3.5% 2.9% 1.8% 11.9% 19.8% 6.0% 5.5%

15.5% 14.5% 12.1% 13.0% 18.9% 15.6% 14.8% 16.9%

consumers have made investment purchases online, but when they do they are most often for more than $500. To guard against generalizing from popular but low dollar volume categories, we analyze each product category separately. Both the GVU and the IPIE survey data used in this study have features that may limit the generalizability of our results. First, the sample could be biased toward experienced Internet users who may be more likely than the average consumer to use the Internet for product searching and purchasing. Second, the GVU survey was linked to high-exposure Internet sites such as Netscape and Yahoo! and the IPIE survey was linked to sweepstakes Internet sites. Therefore, even among Internet users the sample might be biased. Third, data were collected from those consumers that were willing to spend the time to fill out the questionnaire. We expect that they have lower opportunity cost of time than a representative consumer. Acknowledging these limitations, we estimate information search and product purchase behavior to uncover determinants of consumers’ search and purchase decisions across channels. To do so, we identify the theoretical constructs represented by these equations with responses to the survey questions. Some of the measures better reflect underlying theory than others. For example, while theory allows for the information obtained through a search channel to be a continuous variable, the surveys asked only whether or not the respondent searched for information in the channel. For our purposes, we identify this discrete search variable with seeking information. Similarly, the surveys do not contain direct measures of human capital specific to a channel. Instead, we develop proxy variables indicating channel specific human capital from answers regarding the respondent’s use of the channel for decisions in

Consumer Acquisition of Product Information

241

other product categories. That is, greater use of a channel for other product category choices is indicative of human capital specific to that channel. Certain variable definitions are required to estimate the product search relationship. Let information search by individual i in product category j via channel c be denoted as Iijc where c can take on values O, N, M, and T denoting online, newspaper or magazine, direct mail, and television and radio (the GVU survey does not ask about information gathered through the television or radio). Likewise, let product purchase by i in category j be denoted as Pijc where the subscript c can take on values O, R, and C denote online, traditional retail store, and catalog. From the survey data, Iijc and Pijc are dummy variables indicating information search and product purchase within the previous six months. As a proxy for purchase human capital specific to a channel, HPic, we construct TPijc, as the sum of Pijc over the other product categories or ⌺k ≠ jPikc. Omitting category j from the summation eliminates one obvious source of endogeneity. Likewise, we construct TIijc, a proxy for purchase human capital specific to an information source, HIic, to be the sum of Iijc over the other product categories or ⌺k ≠ jIikc. Individuals with more human capital specific to a purchase channel will tendto have larger values of TPijc and those with more human capital specific to an information source will tend to have larger values of TIijc. Our estimator for Eq. (1) is a logit regression based on:

冉 冘

Pr(Iijc = 1) = Pr ␪jc +

␪jcdTIjcd +

d = O, N, M, D

冘

冊

␾jcdTPjcd > 0

d = O, R, C

(3)

in which we will test for ␪jcc > 0 and ␾jcc > 0. Likewise, our estimator for Eq. (2) is a logit regression based on:

冉 冘

Pr(Pijc = 1) = Pr ␥jc +

␥jcd Ijcd +

d = O, N, M, D

冘

d = O, R, C

␦jcd TPjcd > 0

冊

(4)

in which we will test for ␦jcc > 0, ␥jcc > 0 and ␥jcd > 0. Since Eq. (2) was derived for a specific purchase decision, the channel choices were mutually exclusive. However, since the survey question refers to all decisions over the previous six months, we cannot impose an adding up constraint that one and only one channel will be chosen.

RESULTS We find strong general support for the importance of specific human capital in consumers’ information gathering and product purchasing strategies. We also find strong support for the importance of congruencies between information

242

MICHAEL R. WARD AND MICHELLE MORGANOSKY

sources and purchase channels. Finally, we find evidence of consumers’ use of online information to guide both online and off-line purchase decisions. Choice of Product Information Source The results of the estimation of Eq. (3) are reported in Table 5 for the GVU data (1998) and Table 6 for the IPIE data (2001–2002). For each of six product categories and three information sources in the GVU data (eight categories and four sources for IPIE), the results from Logit regressions are reported as columns. The three dependent variables are dummy variables indicating product information search online, through newspapers or magazines, and through direct mail. The separate rows represent the independent variables for these regressions. Rather than report Logit coefficients, we report the implied odds ratio for each independent variable computed at the sample mean. A value of 1.00 indicates that respondents with higher values of the variable are no more or less likely to gather information for this product and channel than those with lower values. In contrast, the 1.14 value in column one, row one of Table 5 indicates a 14% increase in the probability of searching for computer hardware online for every other product category that a consumer purchases other products online. The estimates provide strong support for search human capital increasing search behavior but weaker support for purchase human capital increasing search behavior. The first three rows of Table 5 are meant to capture human capital specific to a purchase channel while the next three are meant to capture human capital specific to an information source. Across all products and information sources, the probability that a respondent will make use of an information source tends to double for each additional source used for other products. The actual odds ratios indicate an increase of 45% (online investments) to 204% (direct mail software) and every estimate is significant at the 1% level. The off-diagonal terms indicate that the use of an information source is affected by human capital specific to a different information source. Only a few of these terms are significant and those that are have relatively smaller magnitudes (9–22% change in probability). In contrast, only a few measures of purchase channel human capital affect choice of information source. In Table 5, no non-congruent effect is significant at the 1% level and seven out of 18 congruent effects are significant. The magnitude of these effects range from 12% (online music) to 26% (direct mail music). The same pattern appears in the 2001–2002 data from IPIE. Table 6 indicates that the the probability that a respondent will use of an information source also increases substantially for each additional source used for other products. The

Consumer Acquisition of Product Information

Table 5.

Determinants of Product Information Search Behavior GVU (1999) Data.

Online Purchases for Other Products Retail Purchases for Other Products Catalog Purchases for Other Products Online Information Searches for Other Products News/Mag Information Searches for Other Products Direct Mail Information Searches for Other Products Spring Pseudo R

2

Online Purchases for Other Products Retail Purchases for Other Products Catalog Purchases for Other Products Online Information Searches for Other Products News/Mag Information Searches for Other Products Direct Mail Information Searches for Other Products Spring Pseudo R

2

243

Online

Hardware News/ Mag

Direct Mail

Online

Software News/ Mag

Direct Mail

1.14 +

1.01

1.08

1.09

0.98

1.07

1.10 +

1.03

1.00

1.05

1.15*

1.10 +

0.89

0.97

0.88

0.98

1.03

0.94

2.35*

1.04

1.03

2.27*

0.97

1.01

0.99

2.36*

0.94

1.06

2.22*

0.90 +

1.01

1.18*

2.90*

1.00

1.12 +

3.04*

0.69 +

0.78 +

0.76 +

0.54*

0.77 +

0.75 +

0.221

0.230

0.229

0.214

0.216

0.246

Online

Music News/ Mag

Direct Mail

Online

Books News/ Mag

Direct Mail

1.12*

0.92 +

0.92 +

1.13 +

0.90 +

0.89 +

0.93

1.14*

0.91 +

0.99

1.03

0.90 +

0.94

1.06

1.24*

1.02

1.11

1.26*

1.71*

0.91 +

0.89 +

1.92*

0.99

0.89 +

1.00

2.19*

0.97

0.99

2.19*

1.05

1.01

0.94

2.08*

1.04

0.97

2.51 +

1.16

1.13

1.07

1.16

0.99

1.10

0.099

0.189

0.171

0.141

0.193

0.231

This table reports the Odds Ratios from Logit regressions of product information gathering for various product categories. Asterisks and plus signs indicate that the coefficient estimate is significant at the 1% and 10% levels. Each regression include 1,786 observations

244

MICHAEL R. WARD AND MICHELLE MORGANOSKY

Table 5.

Online Purchases for Other Products Retail Purchases for Other Products Catalog Purchases for Other Products Online Information Searches for Other Products News/Mag Information Searches for Other Products Direct Mail Information Searches for Other Products Spring Pseudo R

2

Continued.

Investments News/ Direct Online Mag Mail

Online

Travel News/ Mag

Direct Mail

1.18*

1.04

1.03

1.15*

1.03

0.89 +

0.93 +

1.00

1.00

0.98

0.99

1.02

1.00

1.03

0.97

1.00

1.02

1.03

1.45*

1.03

0.91

1.49*

0.98

1.05

1.11*

1.62*

1.22*

1.04

1.91*

1.25*

0.97

0.96

1.61*

0.94

0.99

1.80*

0.86

0.84

1.47 +

0.86

1.45*

1.39 +

0.077

0.095

0.099

0.078

0.140

0.146

This table reports the Odds Ratios from Logit regressions of product information gathering for various product categories. Asterisks and plus signs indicate that the coefficient estimate is significant at the 1% and 10% levels. Each regression include 1,786 observations

percentage changes in probability are slightly smaller with a low of 47% (online investments again) to 142% (TV/Radio home electronics) and every estimate is again significant at the 1% level. Again, significant off-diagonal effects are few and much smaller in magnitude (16–19%). Again, only a few measures of purchase channel human capital affect choice of information source, but they are no longer consistently congruent. For example, use of newspapers and magazines for music information is higher by 24% for respondents who purchase through catalogs and lower by 11% for those who purchase online. Choice of Product Purchase Channel Results of the estimation of Eq. (4) are reported in Tables 7 for the 1998 GVU data and Table 8 for the 2001–2002 IPIE data. These regressions provide estimates of the effect of information source and purchase channel human capital on the choice of product channel for purchase. In general, they provide strong support for both congruency between information source and purchase

Consumer Acquisition of Product Information

Table 6.

Determinants of Product Information Search Behavior IPIE (2001–2002) Data.

Online Purchases for Other Products Retail Purchases for Other Products Catalog Purchases for Other Products Online Information Searches for Other Products News/Mag Information Searches for Other Products Direct Mail Information Searches for Other Products TV/Radio Information Searches for Other Products Pseudo R2

Online Purchases for Other Products Retail Purchases for Other Products Catalog Purchases for Other Products Online Information Searches for Other Products News/Mag Information Searches for Other Products Direct Mail Information Searches for Other Products TV/Radio Information Searches for Other Products Pseudo R2

245

Online

Hardware News/ Dir Mag Mail

TV/ Radio

Online

Software News/ Dir Mag Mail

TV/ Radio

1.02

0.91 +

0.87 +

0.89 +

0.90

0.97

0.83*

0.83 +

1.00

0.90*

0.92 +

0.94

1.06

0.93 +

0.95

1.03

1.00

1.05

0.99

0.89

0.91

1.18*

1.05

2.24*

1.13*

1.14 +

1.05

2.39*

1.07

1.05

0.92

1.05

2.03*

1.05

0.99

1.02

2.04*

1.02

0.96

1.03

1.10 +

2.05*

1.18*

1.06

1.13 +

2.23*

1.03

1.02

0.96

0.93

1.97*

0.98

1.00

0.99

2.39*

0.390

0.344

0.266

0.308

0.399

0.355

0.323

0.370

Online

Music News/ Dir Mag Mail

TV/ Radio

Online

Books News/ Dir Mag Mail

TV/ Radio

1.03

0.89*

0.99

0.90 +

1.18*

0.97

0.90 +

0.98

1.00

1.03

0.95

1.07 +

1.02

1.08 +

1.01

1.04

1.15

1.24*

1.21 +

1.21*

0.93

1.15 +

1.38*

1.19 +

1.84*

1.01

0.99

1.11 +

1.80*

1.03

1.08

1.00

+

1.10*

1.69*

0.94

1.03

0.82

+

1.01

1.61*

0.97

1.14

0.95

0.95

1.95*

0.94

1.11 +

0.98

2.23*

0.92

1.01

1.81*

0.94

1.20*

1.07

2.00*

0.239

0.231

0.328

0.305

0.316

0.291

1.12

+

0.299

1.11

+

0.243

This table reports the Odds Ratios from Logit regressions of product information gathering for various product categories. Asterisks and plus signs indicate that the coefficient estimate is significant at the 1% and 10% levels. Each regression include 1,237 observations

246

MICHAEL R. WARD AND MICHELLE MORGANOSKY

Table 6.

Online Purchases for Other Products Retail Purchases for Other Products Catalog Purchases for Other Products Online Information Searches for Other Products News/Mag Information Searches for Other Products Direct Mail Information Searches for Other Products TV/Radio Information Searches for Other Products Pseudo R2

Online Purchases for Other Products Retail Purchases for Other Products Catalog Purchases for Other Products Online Information Searches for Other Products News/Mag Information Searches for Other Products Direct Mail Information Searches for Other Products TV/Radio Information Searches for Other Products Pseudo R2

Continued.

Online

Hardware Investments News/ Dir Mag Mail

TV/ Radio

1.16*

1.17*

1.07

1.23*

Online

Software Travel News/ Dir Mag Mail

TV/ Radio

1.10 +

0.97

1.01

0.91 +

+

0.97

1.05

0.93

1.02

1.03

1.09

0.93

0.96

0.83 +

1.08

0.96

0.94

0.86 +

0.76 +

1.47*

1.02

0.86 +

1.08

1.48*

0.99

1.03

1.02

1.06 +

1.50*

1.15 +

1.17*

1.05

1.63*

1.04

1.19*

0.92 +

0.91 +

1.69*

0.89

0.97

1.04

1.72*

1.11

1.02

1.04

1.02

1.52*

1.05

1.03

1.07

1.69*

0.181

0.204

0.192

0.201

0.195

0.239

0.207

0.296

Home Electronics News/ Dir TV/ Online Mag Mail Radio

Online

Apparel News/ Dir Mag Mail

TV/ Radio

1.16*

1.08

0.93

0.92

1.09 +

0.95

1.12 +

0.99

1.00

1.06

1.08

1.05

0.96

1.16*

1.03

1.08

1.05

1.02

1.15 +

1.18 +

1.05

0.91

1.09

1.00

2.02*

1.00

1.12 +

1.04

1.49*

1.05

0.99

1.09

1.01

2.17*

1.05

1.07

0.99

1.59*

1.07 +

1.05

0.89 +

0.95

2.16*

1.12 +

1.12 +

1.11 +

1.90*

1.07

1.14 +

1.16*

1.18*

2.42*

1.05

1.03

1.03

2.09*

0.376

0.414

0.333

0.380

0.179

0.256

0.227

0.343

1.11

1.08

This table reports the Odds Ratios from Logit regressions of product information gathering for various product categories. Asterisks and plus signs indicate that the coefficient estimate is significant at the 1% and 10% levels. Each regression include 1,237 observations

Consumer Acquisition of Product Information

Table 7.

Determinants of Product Purchase Behavior GVU (1999) Data. Online

Online Product Information Search Newspaper/Magazine Product Information Search Direct Mail Product Information Search Online Purchases for Other Products Retail Purchases for Other Products Catalog Purchases for Other Products Spring Pseudo R

2

Online Product Information Search Newspaper/Magazine Product Information Search Direct Mail Product Information Search Online Purchases for Other Products Retail Purchases for Other Products Catalog Purchases for Other Products Spring Pseudo R

2

247

Hardware Retail Catalog

Online

Software Retail

Catalog

15.71*

2.08*

2.55*

9.55*

1.65*

2.06*

0.71 +

1.81*

1.37 +

0.94

2.21*

1.54*

1.26 +

1.07

1.80*

1.61*

1.04

1.97*

2.18*

1.07

1.24*

2.23*

0.96

1.18*

0.98

2.07*

1.14*

1.04

2.05*

1.13 +

1.13 +

1.16 +

2.52*

1.03

1.10

3.08*

0.76 +

1.08

0.95

1.21

1.26 +

1.09

0.277

0.183

0.212

0.276

0.176

0.250

Online

Music Retail

Catalog

Online

Books Retail

Catalog

20.64*

2.41*

1.70*

9.13*

1.40 +

1.61*

0.66*

2.66*

1.22

0.89

2.51*

1.68*

1.31 +

1.22

7.97*

1.01

1.01

5.06*

1.79*

0.90 +

0.95

1.89*

0.98

0.93

0.99

2.26*

1.13 +

0.93 +

2.09*

1.09 +

1.06

1.13 +

2.04*

0.98

1.13 +

2.04*

0.94

0.95

1.18

1.00

0.95

0.86

0.327

0.226

0.258

0.260

0.188

0.225

This table reports the Odds Ratios from Logit regressions of channel for product purchase for various product categories. Asterisks and plus signs indicate that the coefficient estimate is significant at the 1% and 10% levels. Each regression include 1,786 observations

248

MICHAEL R. WARD AND MICHELLE MORGANOSKY

Table 7. Online Online Product Information Search Newspaper/Magazine Product Information Search Direct Mail Product Information Search Online Purchases for Other Products Retail Purchases for Other Products Catalog Purchases for Other Products Spring Pseudo R

2

Continued.

Investments Retail Catalog

Online

Travel Retail

Catalog

16.92*

1.28

0.41 +

29.45*

1.44*

0.91

1.07

3.26*

10.23*

0.72*

2.48*

3.97*

0.98

2.08*

2.85*

0.93

1.25

2.46*

1.56*

0.94

1.00

1.62*

0.95

0.92

1.11 +

1.82*

0.84

1.01

1.74*

1.12

1.04

1.00

2.13*

1.11 +

1.15*

1.89*

0.83

1.09

1.75

0.88

0.83

0.70

0.299

0.149

0.230

0.281

0.134

0.193

This table reports the Odds Ratios from Logit regressions of channel for product purchase for various product categories. Asterisks and plus signs indicate that the coefficient estimate is significant at the 1% and 10% levels. Each regression include 1,786 observations

channel and for the importance of purchase channel human capital. Moreover, they provide evidence of interesting non-congruent patterns in consumers’ behaviors. Specifically, they indicate that information obtained online tends to have larger affects on retail and catalog purchasing than other information sources have on non-congruent channels. In Table 7, our measure of channel specific purchase human capital always has a large positive and significant effect on channel usage. The likelihood that one will select the a channel for other purchases increases about 100% for each additional purchase a respondent makes with the a channel for other product categories. The actual range in Table 7 varies from a 56% (online investments) to 208% (catalog software). The effects seem to be slightly higher for the catalog channel than the others, suggesting that human capital developed with catalog shopping tends to be more specific and less transferable. The effects also appear to be larger for computer hardware and software and smaller for investments and travel. Most “off-diagonal” elements are not significant and all are smaller than diagonal elements. The magnitudes of the significant effects, for computer hardware catalog purchases from online and retail human capital,

Consumer Acquisition of Product Information

Table 8.

Determinants of Product Purchase Behavior IPIE (2001–2002) Data. Online

Online Product Information Search Newspaper/Magazine Product Information Search Direct Mail Product Information Search Television/ Radio Product Information Search Online Purchases for Other Products Retail Purchases for Other Products Catalog Purchases for Other Products Pseudo R2

Online Product Information Search Newspaper/Magazine Product Information Search Direct Mail Product Information Search Television/ Radio Product Information Search Online Purchases for Other Products Retail Purchases for Other Products Catalog Purchases for Other Products Pseudo R2

249

Hardware Retail Catalog

Online

Software Retail

Catalog

7.19*

1.83*

3.59*

9.45*

1.49 +

0.86

1.21

1.42 +

2.16 +

0.86

1.27

1.67 +

1.48 +

1.74 +

1.48

1.43

1.88 +

2.13 +

0.87

1.69 +

1.63

0.99

1.27

2.24 +

1.56*

1.21*

0.98

1.85*

1.04

1.12

0.99

1.79*

1.13 +

1.00

1.84*

1.27*

1.13 +

0.53*

3.02*

0.94

1.09

2.40*

0.276

0.309

0.418

0.343

0.277

0.355

Online

Music Retail

Catalog

Online

Books Retail

Catalog

5.43*

1.06

1.20

7.96*

1.37 +

0.51 +

1.44 +

2.36*

1.74 +

1.29

2.11*

1.73 +

1.10

1.00

5.95*

1.07

1.17

4.51*

1.12

1.93*

1.26

0.69

2.87*

1.23

1.70*

0.86*

0.94

1.59*

0.99

0.95

1.01

1.97*

1.12 +

1.03

1.89*

1.06

1.09

1.05

2.11*

1.20

1.14

2.20*

0.313

0.324

0.304

0.344

0.332

0.284

This table reports the Odds Ratios from Logit regressions of channel for product purchase for various product categories. Asterisks and plus signs indicate that the coefficient estimate is significant at the 1% and 10% levels. Each regression include 1,237 observations

250

MICHAEL R. WARD AND MICHELLE MORGANOSKY

Table 8. Online Online Product Information Search Newspaper/Magazine Product Information Search Direct Mail Product Information Search Television/ Radio Product Information Search Online Purchases for Other Products Retail Purchases for Other Products Catalog Purchases for Other Products Pseudo R2

Continued.

Investments Retail Catalog

Pseudo R2

Travel Retail

Catalog

14.71*

1.18

0.65

18.94*

0.87

0.42

1.16

2.47*

2.63 +

1.07

2.24*

2.28

0.90

1.547.01*

0.70

1.62 +

8.02*

1.35

0.81

0.38

0.80

1.13

0.81

1.40*

0.95

0.94

1.41*

1.00

0.92

0.99

1.54*

1.12

1.06

1.87*

1.47 +

0.89

0.98

1.34 +

0.99

1.08

1.71*

0.335

0.141

0.133

0.335

0.222

0.260

Online

Apparel Retail

Catalog

Home Electronics Online Retail Catalog Online Product Information Search Newspaper/Magazine Product Information Search Direct Mail Product Information Search Television/ Radio Product Information Search Online Purchases for Other Products Retail Purchases for Other Products Catalog Purchases for Other Products

Online

4.80*

1.38 +

1.46

9.79*

1.93*

2.50*

1.04

0.90

1.68 +

1.10

2.23*

1.50 +

0.72 +

1.24

3.15*

1.52 +

1.75 +

3.37*

0.86

1.19

0.55 +

0.97

1.22

1.19

1.81*

0.94

1.10

1.46*

1.02

1.08 +

1.03

2.31*

1.22 +

0.97

1.90*

1.03

1.21 +

1.16

2.70*

1.08

1.00

1.94*

0.327

0.368

0.370

0.330

0.337

0.244

This table reports the Odds Ratios from Logit regressions of channel for product purchase for various product categories. Asterisks and plus signs indicate that the coefficient estimate is significant at the 1% and 10% levels. Each regression include 1,237 observations

Consumer Acquisition of Product Information

251

computer software catalog purchases from online human capital, and retail travel from catalog human capital, range from 14% to 24% increases. Table 7 also reveals strong congruent effects. Every hypothesized information source to purchase channel linkage was found to be large and statistically significant. These diagonal terms in the table tend to suggest an 8–28 fold increase for online information to online purchases, a 74%–226% increase for newspaper and magazine information to retail purchases, and an 80%–697% increase for direct mail information to catalog purchases. The larger percentage increase for online and catalog could reflect that, since fewer consumers engage in these purchases, the same absolute change will be a larger percentage change. Nevertheless, these are large effects and they suggest strong congruent linkages for all three source-channel pairs. Product categories that benefit more from database search, such as music, travel, and investments, tend to have even larger online-online linkages. Besides the congruent effects, many “off-diagonal” or non-congruent effects emerge. Most of these involve online information affecting retail or catalog purchasing. For example, the probability of retail computer software and hardware and music purchases increases as much from non-congruent online information as it does from congruent newspaper and magazine information. Similar, albeit smaller, effects are found for many online information to catalog purchase pairs. Interestingly, use of newspaper and magazine information substantially reduces the probability of an online music or travel purchase. This is consistent with consumers who purchase music and travel online substituting online information for printed material. Most of the patterns from the 1998 GVU data are also found in Table 8 for the 2001–2002 IPIE data, with a few notable differences. Channel specific human capital still appears to be important. The likelihood that one will select the a channel for other purchases still increases about 100% for each additional purchase a respondent makes with the a channel for other product categories. The high and the low are the same as before: 40% for online investments and 202% for catalog software. The effects again seem to be slightly higher for the catalog channel than the others, suggesting that human capital developed with catalog shopping tends to be more specific and less transferable. The effects also appear to be smaller for investments and travel, however those for computer hardware and software are no longer appreciably larger. A few “offdiagonal” elements are significant at the 1% level. Again, the magnitudes of the significant effects tend to be smaller. Interestingly, experience in purchasing computer hardware online increases retail purchases but experience purchasing through catalogs decreases retail purchases. This is consistent with complementarities in online and retail human capital but retail and catalog

252

MICHAEL R. WARD AND MICHELLE MORGANOSKY

human capital crowding each other out. This crowding out effect also appears between human capital specific to online and retail music purchasing. We continue to find strong congruent effects in Tables 8. Almost every hypothesized information source to purchase channel linkage was found to be large and statistically significant. The IPIE data allow for the inclusion of information gathered through television and radio. This measure only has a significant effect for music and books, but these are the hypothesized congruent effects. These diagonal terms in the table tend to suggest a 4.4 to 18 fold increase for online information to online purchases, a 84%–147% increase for newspaper and magazine information to retail purchases, and a 140%–702% increase for direct mail information to catalog purchases. Again, the larger percentage increase for online and catalog could reflect that, since fewer consumers engage in these purchases, the same absolute change will be a larger percentage change. However, the online magnitudes are noticeably smaller than in Table 7. This could simply reflect that a larger fraction of consumers purchase online in 2002 than in 1998. Again, product categories that benefit more from database search, such as music, travel, and investments, tend to have even larger online-online linkages. The pattern of “off-diagonal” or non-congruent effects retains the same general pattern as with the 1998 data, but somewhat more muted. Online information still appears to increase retail or catalog purchasing, but the only strongly significant effects are for computer hardware and apparel. Many other online information to retail or catalog purchasing effects are large and positive (books, software, home electronics), but are significant only at the 10% level. A possible explanation for weaker linkages between online information and off-line purchases is that consumers have become more willing to use this information to complete their purchases online. This could be evidence of growing consumer acceptance of online shopping.

CONCLUSIONS The Internet is expected to dramatically increase the amount of information consumers can use to make purchase decisions. Increased consumer information can be beneficial to consumers by reducing asymmetric information problems (Ward & Lee, 1999) and by lowering search costs (Brynjolfson & Smith, 2000). However, the information obtained online need not be restricted to online purchases. For many products, the transactions costs of consummating a transaction online may be prohibitive. In these cases, increased online information may instead lead to more efficient transactions through other purchase channels. In this way, the traditional “congruent” linkages between

Consumer Acquisition of Product Information

253

information sources and purchase channels may be augmented by “noncongruent” linkages involving online information. This chapter finds empirical support for product information obtained on the Internet affecting both congruent and non-congruent channel for product purchase. McGoldrick (2002) claims that the future is “hybrid” where an integration of bricks and mortar retailing and e-tailing is an inevitable consequence of consumer demand. Using the Internet to transmit product information is efficient even when consummating transactions online is cumbersome. Our evidence on non-congruent online channel linkages suggests that a hybrid approach may emerge as a successful marketing strategy. In addition, our findings point to some general implications and concerns about such a hybrid marketing strategy. First, the volume of online transactions is likely to greatly underestimate the effect of the Internet on retailing activity. The dollar amount of online retailing transactions continues to grow at a significant rate. Our evidence suggests that consumers are researching product choices online before they visit traditional retail outlets. Second, the Internet is likely to be making consumers more informed about their product selection decisions. If consumers are researching online before purchasing through other channels, then revealed preference suggests that more information is being obtained at lower transactions cost. Thus, we find some support for this critical assumption underlying many studies of Internet pricing. Third, because consumers are more informed and are using this information for off-line purchase decisions, consumer markets should be less “imperfectly competitive.” Greater consumer information likely makes retailers closer substitutes and thus reduces market power. It also is makes these markets less susceptible to asymmetric information problems implying that costly quality assurance mechanisms are less important. Alleviating either source of imperfection should make markets less imperfect. Moreover, these effects should apply to both online retailers and, to some degree, off-line retailers. Fourth, and related to above, markets are likely to be broader than previously assumed. When economists define markets, say for antitrust purposes, they attempt to include products that are close substitutes from a demand perspective. Often, there is debate as to whether products marketed through different channels actually compete with each other. However, our finding of non-congruent online-to-retail linkages suggests that products sold by traditional retail outlets do in fact compete with products sold online. The market for a product may encompass both online and off-line outlets.

254

MICHAEL R. WARD AND MICHELLE MORGANOSKY

Fifth, the management of “channel conflict” by multi-channel retailers is likely to become more complicated. Many multi-channel retailers run parallel operations in, say retail and mail order. In many applications, actions in one channel rarely affect responses in other channels because consumers rarely switch between the two. Our results suggest that retailers can benefit from exploiting consumers’ use of multiple channels for product and retailer selection. However, we suspect that doing so is likely to require new management and marketing expertise. Lastly, some retailers are concerned about online marketing “cannibalizing” retail sales. That is, online sales could increase at the expense of traditional retail sales as loyal consumers switch from one channel to another. In fact, we find evidence consistent with an Internet presence augmenting retail sales.

ACKNOWLEDGMENTS This research was partially supported by C-FAR Project No. 99I-027–2. We would like to thank the participants of the Third Berlin Internet Economics Conference, 2000 and the Third International Conference on Uses and Services in Telecommunications, 2001 for helpful comments.

REFERENCES Akhter, S. H. (1989). Schematic Information Processing: Direct Marketing and Purchase Decisions. Journal of Direct Marketing, 3, 31–38. Ainslie, A., & Rossi, P. E. (1998). Similarities in Choice Behavior Across Product Categories. Marketing Science, 17(2), 91–106. Bagozzi, R. (1975). Marketing as Exchange. Journal of Marketing, 39 (October), 32–39. Bagwell, K., & Garey, R. (1994). Coordination Economies, Advertising, and Search Behavior in Retail Markets. American Economic Review, 84(3), 498–517. Balasubramian, S. (1998). Mail vs. Mall: A Strategic Analysis of Competition Between Direct Marketers and Conventional Retailers. Marketing Science, 17(3), 181–195. Benabou, R. (1990). Search Market Equilibrium, Bilateral Heterogeneity, and Repeat Purchases. Mimeo, MIT. Benabou, R. (1993). Search Market Equilibrium, Bilateral Heterogeneity, and Repeat Purchases. Journal of Economic Theory, 60(1), 140–158. Brynjolfsson, E., & Smith, M. (2000). Frictionless Commerce? A Comparison of Internet and Conventional Retailers. Management Science, 46(4), 563–585. Carlson, J. A., & McAfee, R. P. (1983). Discrete Equilibrium Price Dispersion. Journal of Political Economy, 91(3), 480–493. DMA (1999). Statistical Fact Book (pp. 150–162). New York: Direct Marketing Association. Fishbein, M., & Ajzen, I. (1975). Theories of Attitude. In: M. Fishbein & I. Ajzen (Eds), Beliefs, Attitudes, Intention, and Behavior: An Introduction to Theory and Research (pp. 335–383). New York: Addison Wesley.

Consumer Acquisition of Product Information

255

Graphics, Visualization, and Utilization Center (1998). GVU’s WWW User Surveys. < http://www.gvu.gatech.edu > . Lancaster, K. J. (1966). A New Approach to Consumer Theory. Journal of Political Economy, 74 (April), 132–157. McGoldrick, P. J. (2002). Retail Marketing (p. 633). New York: McGraw-Hill. Messinger, P. R., & Narasimhan, C. (1997). A Model of Retail Formats Based on Consumers’ Economizing on Shopping Time. Marketing Science, 16(1), 1–23. Pashigian, P., & Bowen, B. (1994). The Rising Cost of Time of Females, the Growth of National Brands, and the Supply of Retail Services. Economic Inquiry, 32, 33–65. Rossi, P. E., McCulloch, R., & Allenby, G. (1996). On the Value of Household Purchase History Information in Target Marketing. Marketing Science, 15(4), 321–340. Schiffman, L. G., & Kanuk, L. L. (1997). Consumer Behavior (pp. 560–564). Upper Saddle River, NJ: Prentice Hall. Szymanski, D. M., & Hise, R. J. (2000). E-satisfaction: An Initial Examination. Journal of Retailing, 76(3), 309–322. Woodruff, R. B. (1997). Customer Value: The Next Source of Competitive Advantage. Journal of the Academy of Marketing Science, 25(2), 139–153.

This Page Intentionally Left Blank

AN ECONOMIC ANALYSIS OF MULTIPLE INTERNET QoS CHANNELS Dale O. Stahl, Rui Dai and Andrew B. Whinston ABSTRACT With the astounding growth of traffic carried by the Internet in recent years, congestion is becoming an increasing problem. Real-time and mission-critical traffic require levels of service quality exceeding the besteffort level currently provided. Intserv and Diffserv service models have been developed to provide multiple levels of quality of service (QoS) throughout the Internet. We develop a generic economic model of network QoS to address how these levels of QoS should be specified and how they should be allocated. The pricing of marketable tokens for QoS access and its impact on efficiency are analyzed.

1. INTRODUCTION With the rapid growth and commercialization of the Internet and the appearance of more bandwidth-hungry real-time multimedia applications, quality of service (QoS) is becoming a hotly discussed issue. The current Internet only provides best-effort service to users. This best-effort service model treats all data packets exactly alike: no matter whether a packet is in a mission-critical traffic flow or part of a non-urgent email. However, with the fast development of the Internet, this service model cannot satisfy the requirements for increasing business usage (e.g. to finish a transaction within a specified time interval) and real-time applications (e.g. to keep the transmission rate of an Internet phone call at 64 kb/s). The Economics of the Internet and E-Commerce, Volume 11, pages 257–268. © 2002 Published by Elsevier Science Ltd. ISBN: 0-7623-0971-7

257

258

DALE O. STAHL, RUI DAI AND ANDREW B. WHINSTON

The Internet Engineering Task Force (IETF) has proposed several new service models to offer better network service than best-effort, including integrated service (Intserv) (e.g. Braden et al., 1994) and differentiated service (Diffserv) (e.g. Blake et al., 1998). In these frameworks, a user has a budget of tokens that can be used to tag outgoing packets, and core routers provide differentiated treatment of tagged and untagged packets when congestion occurs. Thus, a token can be viewed as an admission ticket for higher QoS. Guaranteeing a higher QoS for tagged packets requires load control (or traffic shaping) by the user’s ISP. A common method of load control is to reserve a portion of the network capacity for “premium” services, and to set the supply of tokens so the reserved capacity is sufficient to provide the higher QoS for all tagged packets (e.g. Nichols et al., 1998, 1999). More generally, the network capacity would be divided into a number of channels, each providing a designated QoS and requiring an associated token to be admitted to that channel. The idea of providing QoS on the Internet by partitioning the network into several logically independent channels has been proposed and studied extensively in computer networking (e.g. Odlyzko, 1999). This approach is called PMP (Paris Metro Pricing), which was initially introduced in the Paris Metro System. With this method, service providers will post different prices for the access rights to different channels and expect that the more expensive channels will be less congested. Therefore, users who care more about service quality can pay more for access to the more expensive channels and enjoy better service. Although it is widely known that channeling could introduce inefficiencies in bandwidth usage, experiments have shown that many real-time applications that cannot run properly in today’s Internet work well in a PMP system. Most researches in the literature are focused on the implementation of PMP system. However, few have studied the fundamental economic and system design issues concerning multiple QoS channels: (1) How are the tokens to be allocated among the users for these QoS levels? If a service provider offers several levels of service quality, it is natural to use pricing to differentiate network users. A network user will choose the optimal service level so that her utility is maximized. The question is from the social planner’s point view, what are the optimal prices for different service levels. (2) Are the tokens transferable among users? If so, what would be the market price for tokens, and how would token markets affect the overall efficiency of the network? In order to ensure a certain level of service quality and

An Economic Analysis of Multiple Internet QoS Channels

259

hedge the risk of high token prices during congestion, a network user need buy a certain amount of tokens. However, usually the real network capacity usage for a user will be fluctuating and subjected to change from time to time. So creating a token market and allowing users to exchange tokens (people who find they do not need all the reserved capacity can sell tokens to those people who need more capacity) could improve the efficiency of the network system. Then the question is how to organize this token market and whether it will be a more efficient approach. (3) How should the QoS levels be determined, and how should the network capacity be partitioned into QoS channels given a diverse population of users? From a pure economics point of view, providing more service quality options to network users would improve the social welfare of network service. However, the standard economics results on vertical (quality) differentiation cannot be directly applied to the Internet service because network service quality is endogenous in our model: QoS depends on intensity of usage of each channel. In this specific application, offering too many service quality levels could lower network utilization level and hence affect social welfare negatively. Moreover, when a service provider designs the service plans, the distribution of the users must also be considered. For example, if there are a lot of users who require high service quality, the service provider may want to allocate more capacity for higher service levels. Therefore, the service provider has to tradeoff between network utilization level and number of options for network users conditioning on the characteristics of the user distribution, and then decide the optimal service plan, including the number of service levels and the capacity to be allocated for each service level. In this paper we develop a generic network model of this problem that can be analyzed economically to answer each of these questions. We characterize the optimal allocation of tokens, the price that would support the optimal allocation in a token market. One important finding is that creating a market for tokens itself does not necessarily improve efficiency; and the information available from a token market is insufficient to determine the optimal aggregate allocation of tokens. For any fixed partitioning of capacity into channels, we characterize the optimal specification of QoS in each channel. Finally, we prove the surprising result that the optimal number of channels is exactly one, no matter what the diversity of users. In other words, creating multiple QoS levels via channelization is an inefficient use of resources. An important implication of this research is that: although PMP can improve the usability for some real-

260

DALE O. STAHL, RUI DAI AND ANDREW B. WHINSTON

time applications, the overall costs of abandoning statistical multiplexing will actually offset the benefits. Therefore, we believe economics analysis should be more closely involved in network engineering design in the future.

2. THE MODEL OF MULTIPLE QoS LEVELS VIA CHANNELS A simple way to provide multiple QoS levels is to divide the bandwidth capacity into non-interacting channels. For example, standard multiplexing technology naturally divides the transmission capacity of a cable into multiple channels. Each such channel can carry a proportion of the entire cable capacity. By regulating the flow into each channel, the QoS can be controlled (at least on average). For example, if the flow is regulated to provide a low channel capacity utilization rate, then the waiting times and delays of user of that channel will be less than for a comparable channel with a higher average capacity utilization rate. We will develop a model of this problem that can be analyzed economically. First, let ␮j denote the capacity of channel j in terms of work per unit time. We will number the channels from 1 to J. Here we assume there are only one link and one router in the network. This might be considered over-simplified, because usually a network consists of many routers and links, and network users are concerned about end-to-end QoS. However, the single-link model is enough to present the problems introduced by channelization, and it is not very difficult to be extended to more complex models. Let xj denote the average flow of work per unit time into channel j, and let ␶j denote the average throughput time per unit of work in channel j. We assume that this average throughput time can be represented by a function of capacity and flow: T(xj, ␮j), which is strictly increasing in xj, approaches infinity as xj approaches ␮j from below, and is strictly decreasing in ␮j for all ␮j > xj. We also assume that the locus of constant throughput times (i.e. {(xj, ␮j) | T(xj, ␮j) = ␶} has a slope d␮j/dxj ≤ 1 for all ␮j > xj.1 To keep the model simple, we assume that the user’s preferences are quasilinear, so the net benefits per unit of work are V ⫺ ␦␶j, where V and ␦ are private parameters of each user. Let G(V, ␦) denote the joint distribution of these parameters in the population of users. Each user must decide whether or not to submit work given her V and ␦, her expectation of ␶j, and any monetary charge per unit work, say rj. Given these, her expected net value is V ⫺ ␦␶j ⫺ rj.2 Clearly, she will prefer the channel with the maximum expected net value, or equivalently the least net cost ␦␶j + rj. Further, she will submit work to the least cost channel only if that expected cost does not exceed the value V. This linear

An Economic Analysis of Multiple Internet QoS Channels

261

Fig. 1.

programming problem leads to a partition of (V, ␦)-space into regions for which specific channels are best for her, and a region for which it is best not to submit work. Accordingly, let Aj(␶, r) denote the set of (V, ␦) values for which channel j is best, where (␶, r) denotes the array of (␶j, rj) for all channels. Then, G[Aj(␶, r)] is the proportion of users who submit work to channel j. Letting X0 denote the ex ante potential flow of work totaled over all users, the average flow of work into channel j is xj = X0G[Aj(␶, r)].

(1)

But the throughput time depends on the flow, so Eq. (1) is not in reduced form, as substituting ⍀(xj, ␮j) for ␶j reveals: xj = X0G[Aj({T(xj, ␮j), j = 1 . . . J}, r)].

(2)

It is straightforward to show that there exist functions xj(r) that satisfy this flow equation. In other words, given a list of monetary charges for the channels (r), the realized demand for channel j is given by xj(r), which leads to expected throughput times ␶j = T(xj, ␮j) consistent with Eq. (1). But what determines these monetary charges? In the current Internet, there are no usage charges, so r = 0. Hence, xj(0) is the average flow demand in channel j. Note that xj(0) < ␮j, since otherwise throughput times would be infinite. Demand increases until rational expectations of throughput times rise sufficiently to discourage further increases. 2.1. Using Tradable Tokens to Control the Demand Both Diffserv and Intserv incorporate a token system for controlling or shaping traffic (e.g. Tang et al., 1999; Dovrolis et al., 1997). One application of this tagging scheme would be to require token tags for every job (packet) that enters

262

DALE O. STAHL, RUI DAI AND ANDREW B. WHINSTON

the network, blocking jobs without token tags. A prospective user i would contract with the network access provider for a supply rate of tokens for each channel, say ␪ij␮j, where ␪ij denotes the supply of tokens as a fraction of the capacity of channel j. Obviously, feasibility requires that the sum of ␪ij over all users [denote this by ␪j] cannot exceed 1, and prohibiting short positions implies ␪ij ≥ 0. To submit a flow of xij user i must have a token flow of at least this amount. However, the stochastic nature of user needs will typically generate instances when the user’s demand exceeds the contracted rate and other times when demand is less than the contracted rate. Therefore, it might be desirable to allow users to trade tokens in a spot market. Given a spot market price for tokens for channel j of rj, the out-of-pocket monetary cost of submitting a flow of xij is rj(xij ⫺ ␪ij), which can be positive or negative. With this interpretation of rj, we ask whether the market for tokens can be cleared, at what price (rj), and at what impact on network efficiency. By virtue of our assumption of quasi-linear preferences, the potential revenue from selling all one’s tokens, rj␪ij, has no effect on the user’s demand for network services. Thus, xj(r), as derived from Eq. (2) still fully characterizes the aggregate demand for channel j. The market clearing condition for channel j is then (3) xj(r) ≤ ␪j, and xj(r) < ␪j implies rj = 0. In other words, demand cannot exceed supply in any channel, and if there is excess supply in any channel the spot price for that channel falls to 0. The latter result holds because any user who holds excess tokens would be willing to undercut any positive price to make a sale and spend the resulting revenue to consume a desirable scarce good. This simple observation has an important implication. Suppose the entire capacity of channel j is sold (␪j = ␮j). Since xj(r) is always strictly less than ␮j, there will be excess supply, so it must be that rj = 0. In other words, if the entire capacity of channel j is allocated contractually, and the tokens are tradable, tokens will be worthless. Further, if xj(0) ≤ ␪j for each channel, then rj = 0 is a market-clearing price. Hence, for positive prices to exist generally, the contractually allotted tokens must be less than the demand would be at zero usage prices. What mechanism would bring this about, and how should the supply levels (␪j) be determined? 2.2. Optimal Token Supplies The optimal token supply levels are those values of ␪j that maximize the total flow of net benefits through the network. To determine this, we take the point

An Economic Analysis of Multiple Internet QoS Channels

263

of view of a planner who can make submission decisions for each user based on the user characteristics (V, ␦) subject to technological feasibility. Accordingly, let ␲j(V, ␦) denote the probability of admitting this user’s job to the network. These functions are the control variables of the planner. Then, the total net flow of value through channel j is Wj = X0

冕

(V ⫺ ␦␶j)␲j(V, ␦) dG(V, ␦).

The Kuhn-Tucker conditions for an optimal allocation are: V ⫺ ␦␶j ⫺ (⭸⍀/⭸xj)xj␦¯ j < ␯ implies ␲j(V, ␦) = 0,

(4)

(5)

where ␯ = max{0, maxk{V ⫺ ␦␶k ⫺ (⭸T/⭸xk)xk␦¯ k}}, and ␦¯ j =

冕

冕

␦␲j(V, ␦) dG(V, ␦) /

␲j(V, ␦) dG(V, ␦),

the average value of ␦ over those users who submit to channel j. For notational convenience, we define (6) r*j = (⭸T/⭸xj)xj␦¯ j. This is the marginal social cost of congestion in channel j; (⭸T/⭸xj) is the marginal increase in throughput time caused by increased traffic flow, ␦¯ j is the average sensitivity of delay per user affected, and xj the number of users affected per unit time. Thus, the Kuhn-Tucker conditions tell us that a user (or job) with characteristic (V, ␦) should submit to channel j only if channel j would give the maximum and non-negative net social benefits V ⫺ ␦␶j ⫺ r*j . Clearly, this optimal decision would be internalized by the user if she faced monetary charges equal to the marginal social costs (r*j ).3 Further, the optimal flows that solve Eq. (5), x*j = X0 兰 ␲j(V, ␦) dG(V, ␦), also satisfy users facing these charges, so xj(r*j ) = x*j . It follows that if total supply of tokens were ␪j = x*j , then r*j would be a market-clearing price for tokens that would support the optimal allocation. Conversely, if ␪j ≠ x*j , then the market outcome will be suboptimal. In other words, the desirability of marketable tokens crucially depends on getting the aggregate supply right. Can the behavior of market prices as a function of supply, rj(␪), help us discover the right supply level? For example, one might hope that rj(␪) would

264

DALE O. STAHL, RUI DAI AND ANDREW B. WHINSTON

exhibit a discontinuity at the optimal supply level. Unfortunately, the token economy with optimal token supplies is “regular”, and hence the price correspondence is generically locally continuous. In general, rj(␪) itself provides no obvious clues about the optimal supply of tokens. Alternatively, the owners of channel j could post a price rj and dynamically adjust that price towards r*j based on observed demand and channel status ␶j. This approach has been extensively studied by Gupta et al. (e.g. Gupta et al., 1997 and Gupta et al., 1999), and the practical problems of stable price adjustment and measurement of ␦¯ j were solved. However, since r*j does not necessarily correspond to the profit-maximizing price, the owners have no incentive to do this. On the other hand, in a VPN (Virtual Private Network) (e.g. Lin et al., 1999), the network administrator does have the same objective as our fictitious planner, and so would want to post the optimal price r*j . 2.3. Optimal Provision of QoS Levels Up to this point we have taken the channel capacities, ␮j, as exogenous. In that framework, setting the supply of tokens is equivalent to setting the QoS. To see this, recall that the throughput time (which is the only QoS variable in this simple model) is determined by the average demand and the capacity: ␶j = T(xj, ␮j). If aggregate supply ␪j ≥ xj(0), then tokens are free, and so ␶j = T(xj(0), ␮j). On the other hand, if ␪j < xj(0), then there will be a positive market-clearing price, so ␶j = T(␪j, ␮j). By restricting supply, throughput decreases (QoS increases). An important implication of the previous section is that there is an optimal QoS level for channel j; namely that obtained with the optimal token supply: ␶*j = T(x*j , ␮j). The question we address now is: given the distribution of user characteristics G(V, ␦), how should the total bandwidth be partitioned into channels? Equivalently, what is the optimal provision of QoS levels? To answer this question, we maximize the net flow of value through all channels (⌺Jj= 1 Wj), where Wj is given by Eq. (4)), with respect to ␮j, subject to the capacity constraint that ⌺Jj= 1 ␮j = ␮. The Kuhn-Tucker conditions for this problem are that there exists a positive scalar ␣, such that ⫺ (⭸T/⭸␮j)xj ␦¯ j ≤ ␣, and ⫺ (⭸T/⭸␮j)xj ␦¯ j < ␣ implies ␮j = 0.

(7)

In other words, allocate capacity to the channel(s) for which the social value of increasing QoS is highest. By Eqs (5) and (6), r*j = (⭸T/⭸xj)xj ␦¯ j. Dividing this by Eq. (7): r*j /␣ = ⫺ (⭸T/⭸xj)/(⭸T/⭸␮j) ⬅ d␮j/dxj holding ␶j constant. In other words, r*j is proportional to the marginal increase in capacity per unit increase in flow required to maintain a fixed QoS.

An Economic Analysis of Multiple Internet QoS Channels

265

Theorem: The optimal provision of QoS levels entails ␮j = 0 for all j except one. Proof: Suppose to the contrary that ␮j > 0 for j = 1, 2. Without loss of generality suppose that r*1 ≥ r*2 ; hence, ␶*1 ≤ ␶*2 . This situation is depicted in the figure below (with ␶*1 < ␶*2 ). Suppose x1 and x2 are the corresponding optimal allocations, and without loss of generality that ␮1 + ␮2 = 1. We construct a line from (x1, ␮1) with a slope of 1 which crosses ␮ = 1 at the point labeled b. Since d␮1/dx1 along the T(x1, ␮1) = ␶*1 locus has a slope that never exceeds 1, the intersection of this locus with the ␮ = 1 line (point c) does not lie to the left of b. In other words, consolidating the capacity of channels 1 and 2 could accommodate traffic up to x* with QoS at least as good as the current channel 1 (␶*1 ). Point a in the figure is constructed by drawing a line from (x1, ␮1) with a slope equal to ␮2/x2 > 1. Since this slope is greater than one, and since ␮2 = 1 ⫺ ␮1, a lies strictly to the left of b. But by construction total traffic at point a is x1 + x2. In other words, the current total traffic x1 + x2 < x*. Therefore, T(x1 + x2, 1) < ␶*1 ; i.e. consolidating the two channels could accommodate the same total traffic at higher QoS for everyone, contradicting the optimality of having two active channels. Q.E.D. We can, therefore, conclude that partitioning bandwidth into separated channels is not an efficient or useful way to create multiple QoS levels. It is well-known that a single queue for multiple servers dominates separate queues for each server, and with M/M/s systems, if we fix the aggregated service rate of all the servers, s = 1 will dominate s ≥ 2 (e.g. Prabhu, 1997). However, it has

Fig. 2.

266

DALE O. STAHL, RUI DAI AND ANDREW B. WHINSTON

not been studied systematically in previous literature (to our knowledge) whether offering different service quality to different queues and imposing different tariffs for different service classes can really improve the system efficiency. What is new here is that with tradable tokens for access, more flexible design choices (instead of using discrete servers, we can divide the capacity arbitrarily), more general queueing distributions, and an arbitrary diverse distribution of users, it is also optimal to consolidate service into one channel with a single queue. Therefore, the effort to implement a more complex PMP system will not lead to better social welfare as some engineers have conjectured.

3. CONCLUSIONS We have presented a simple model of multiple QoS levels via the partitioning of bandwidth capacity into channels and using a token tagging system for load control. Our surprising conclusion is that this is not an efficient or useful way to create multiple QoS levels. What does this mean for Intserv and Diffserv? While we have analyzed a model with only a single congestible resource, rather than a network of resources, there is no reason to believe that our results would not carry over to a network model. Intserv creates virtual circuits which are conceptually similar to the channels in our model. Diffserv conducts its load control functions at edge routers, so markets for tokens issued at these edge routers cannot possibly support even nearly optimal allocations of the network resources beyond the local domain. Of course, one could envision a more complex tagging system with node-specific tokens, so to obtain a specific QoS on an entire route would require the purchase of the required tokens for each node and link in the route. Even if the core routers and lines have substantial excess capacity (so token prices there would be negligible), the edge router and line at the destination of a traffic flow would be as important as the edge router and line at the origin. Hence, there would need to be markets for token “bundles” corresponding to each possible origin-destination pair in the network, before one could begin to hope that token markets would support nearly-optimal resource allocations. An alternative approach is real-time pricing of network resources coupled with “wholesale” billing methods as proposed by Stahl et al. (e.g. Stahl et al., 1994). Optimal resource prices are analogous to those derived in section 2 of this paper. QoS levels were modeled as priority classes sharing the same resource. Further, GSW developed and simulation-tested real-time decentralized algorithms for computing nearly-optimal prices based on local traffic information (e.g. Gupta et al., 1997).

An Economic Analysis of Multiple Internet QoS Channels

267

NOTES 1. This condition is satisfied, for example, by M/G/1 queues provided that the relative standard deviation of service times is a constant not exceeding 1. 2. In reality, a network user will submit work on a session level, such as downloading a file, instead of on the packet level. Each session consists of many packets. In contrast, our model might appear to suggest that users need to make submission decisions for each packet, which could be very annoying. Actually, the session-level decision problem can be mapped to the decision problem presented in our model. Suppose that the total value of a whole session to a user is V ⬘, which consists of n packets. ISP will charge a liner price to the user based on the traffic volume of the session, ri * n, where ri is the unit price for service class i. Again, we assume users are concerned about the average delay of the packets. Therefore, the user’s decision problem is: Max V ⬘ ⫺ ␦⬘␶i ⫺ ri * n i

(*)

Where ␦⬘ is the delay cost on session level and ␶i is the average packet delay in class i. We can rewrite (*) as: Max (V ⬘/n) ⫺ (␦⬘/n) * ␶i ⫺ ri i

This is exactly the decision problem formulated in our model if we substitute V for V ⬘/n and ␦ for ␦⬘/n. 3. Provided the distribution G is sufficiently smooth, cases of indifference have zero probability and thus can be safely ignored.

REFERENCES Braden, R., Clark, D., & Shenker, S. (1994). Integrated Services in the Internet Architecture: an Overview. Internet RFC 1633, IETF. Blake, S., Black, D., Carlson, M., Davies, E., Wang, Z., & Weiss, W. (1998). An Architecture for Differentiated Service. Internet RFC 2475, IETF. Dovrolis, K., Vedam, M. P., & Ramanathan, P. (1997). The Selection of the Token Bucket Parameters in the IETF Guaranteed Service Class. University of Wisconsin-Madison, Madison, Wisconsin. Gupta, A., Stahl, D., & Whinston, A. (1997). A Stochastic Equilibrium Model of Internet Pricing. Journal of Economic Dynamics and Control, 21, 697–722. Gupta, A., Stahl, D., & Whinston, A. (1999). The Economics of Network Management. Communications of the ACM, 42(9), 57–63. Lin, Z., Ow, P. S., Stahl, D., & Whinston, A. (1999). Improving the Performance of Virtual Private Network by Pricing Traffic. Proceedings of the Workshop on Information Technology and Systems. Nichols, N., Blake, S., Baker, F., & Black, D. (1998). Definition of the Differentiated Services Field (DS Field) in the IPv4 and IPv6 Headers. Internet RFC 2474, IETF. Nichols, N., Jacobson, V., & Zhang, L. (1999). A Two-bit Differentiated Services Architecture for the Internet. Internet RFC 2474, IETF. Odlyzko, A. (1999). Paris Metro Pricing for the Internet. Proceedings of ACM conference on Electronic Commerce, pp. 140–147.

268

DALE O. STAHL, RUI DAI AND ANDREW B. WHINSTON

Prabhu, N. U. (1997). Foundations of Queueing Theory. Kluwer Academic. Stahl, D., & Whinston, A. (1994). A General Economic Equilibrium Model of Distributed Computing. In: W. W. Cooper & A. B. Whinston (Eds), New Decisions in Computational Economics (pp. 175–189). Tang, P., & Tai, C. (1999). Network Traffic Characterization Using Token Bucket Model. Proceedings of the Conference on Computer Communications (IEEE Infocom).