«THE DYNAMICS OF SELLER REPUTATION: THEORY AND EVIDENCE FROM eBAY Luís Cabral Ali Hortaçsu Working Paper 10363 ...»
NBER WORKING PAPER SERIES
THE DYNAMICS OF SELLER REPUTATION:
THEORY AND EVIDENCE FROM eBAY
Working Paper 10363
NATIONAL BUREAU OF ECONOMIC RESEARCH
1050 Massachusetts Avenue Cambridge, MA 02138 March 2004 We thank Kenny Ballendir, Tim Miller and Jeremy Shapiro for truly outstanding research assistance.
We also thank Damien DeWalque, Svetlozar Nestorov, Mike Riordan, Anne Rogers, Steve Tadelis, and seminar participants at Chicago, Essex, Copenhagen, SED meetings, NBER Summer Institute, Georgetown, Tufts, Rochester, NYU and Columbia for helpful comments and suggestions. Hortaçsu acknowledges financial support from the National Science Foundation (Grant SES-0242031). The views expressed herein are those of the authors and not necessarily those of the National Bureau of Economic Research.
©2004 by Luís Cabral and Ali Hortaçsu. All rights reserved. Short sections of text, not to exceed two paragraphs, may be quoted without explicit permission provided that full credit, including © notice, is given to the source.
The Dynamics of Seller Reputation: Theory and Evidence from eBay Luís Cabral and Ali Hortaçsu NBER Working Paper No. 10363 March 2004 JEL No. D44, L15, L86
ABSTRACTWe propose a basic theoretical model of eBay's reputation mechanism, derive a series of implications and empirically test their validity. Our theoretical model features both adverse selection and moral hazard. We show that when a seller receives a negative rating for the first time his reputation decreases and so does his effort level. This implies a decline in sales and price; and an increase in the rate of arrival of subsequent negative feedback. Our model also suggests that sellers with worse records are more likely to exit (and possibly re-enter under a new identity), whereas better sellers have more to gain from "buying a reputation" by building up a record of favorable feedback through purchases rather than sales. Our empirical evidence, based on a panel data set of seller feedback histories and cross-sectional data on transaction prices collected from eBay is broadly consistent with all of these predictions. An important conclusion of our results is that eBay's reputation system gives way to strategic responses from both buyers and sellers.
Luís Cabral Ali Hortaçsu New York University Department of Economics Henry Kaufman Mgmt Ctr. University of Chicago 44 W 4th Street, 7-70 1126 East 59th Street New York, NY 10012 Chicago, IL 60637 and CEPR and NBER firstname.lastname@example.org Hortacsu@uchicago.edu 1 Introduction Electronic commerce presents the theoretical and the empirical economist with a number of interesting research questions. Traditional markets rely signiﬁcantly on the trust created by repeated interaction and personal relationships.
Electronic markets, by contrast, tend to be rather more anonymous. Can the same level of trust and eﬃciency be obtained in these markets?
One possible solution, exempliﬁed by eBay auctions, is to create reputation mechanisms that allow traders to identify and monitor each other. In this paper, we study eBay-type reputation mechanisms, both from a theoretical and from an empirical point of view. Speciﬁcally, we propose a basic theoretical model of eBay’s reputation mechanism, derive a series of implications and empirically test their validity.
Our focus on eBay’s reputation mechanism is justiﬁed for two reasons.
First, electronic commerce in general and eBay in particular are a signiﬁcant economic phenomenon: in 2003, more than $21bn were transacted on eBay by 69 million users. Second, with its well deﬁned rules and available information, eBay presents the researcher with a fairly controlled environment for theory testing. Speciﬁcally, a reasonable assumption on eBay is that the information one trader has about other traders is the same as the researcher’s. Essentially, this information consists of a series of positive and negative feedback comments given by past trading partners. In this context, we can make sharper predictions about agent behavior than in other markets, in particular in markets where buyers and sellers share information that is not observed by the researcher.
Our theoretical model features both adverse selection and moral hazard on the seller’s side. In the spirit of Diamond (1989), we show that in equilibrium there is a positive correlation between seller reputation and seller eﬀort.
Speciﬁcally, when a seller receives a negative rating for the ﬁrst time, his reputation decreases and so does his eﬀort level. This implies a decline in sales and sale price; and, moreover, an increase in the rate of arrival of subsequent negative feedback. Our empirical evidence is broadly consistent with these predictions. Speciﬁcally, we ﬁnd that the growth rate of a seller’s transactions drops from about 7% per week to about -7% following the ﬁrst negative feedback. We also ﬁnd that the rate of negative feedback arrival increases twofold following this event. Both ﬁndings are strongly statistically signiﬁcant across a variety of empirical speciﬁcations. We also ﬁnd that the sale price for identical goods varies across sellers with diﬀering feedback records: a 1% level increase in the fraction of negative feedback is correlated with a 9% decrease in price A natural experiment based on a change in eBay’s reporting format suggests there is indeed a causal relation between seller reputation and sale price.
We consider two extensions of our basic model. First we allow for the possibility of seller “exit”, which we assume corresponds to a secret change in identity. We show theoretically that exit is more likely the worse the seller’s record is. Our empirical ﬁndings are once again consistent with this result. We ﬁnd that a tenfold increase in a seller’s transaction record length is correlated with a 18 to 27% lower probability of exit within the observation period.
Moreover, a 1% level increase in the fraction of negative feedback is correlated with a 1 to 2% increase in probability of exit (however, this coeﬃcient is not statistically signiﬁcant).
Second, we consider the possibility of sellers building up a record (“buying a reputation”) by starting oﬀ as buyers and then switching to selling (anecdotal evidence suggests that it is easier and cheaper to accumulate positive feedback as a buyer than as a seller). Our theoretical model suggests that better sellers have more to gain from building such a record, a prediction that is borne out by the data. Speciﬁcally, we deﬁne a seller as a “switcher” if more than 50% of the ﬁrst 20 transactions were purchases whereas more than 70% of the last 20 transactions were sales. About 30% of all sellers fall in this category; sellers with 1% lower percentage of negative feedback are 6% more likely to have started out “switchers”.
A number of authors have conducted empirical studies of eBay’s reputation mechanism. Almost all of these prior studies focus on the buyer response to published feedback aggregates. In particular, a large number of studies estimate cross-sectional regressions of sale prices on seller feedback characteristics: Dewan and Hsu (2001), Eaton (2002), Ederington and Dewally (2003), Houser and Wooders (2003), Kalyanam and McIntyre (2003), Livingston (2002), Lucking-Reiley, Bryan, Prasad and Reeves (2000), McDonald and Slawson (2002), Melnik and Alm (2002), Resnick and Zeckhauser (2001).1 Resnick, Zeckhauser, Swanson and Lockwood (2003) point out the potential for a signiﬁcant omitted variable bias in these cross-sectional regressions, and conduct a controlled ﬁeld experiment in which a seasoned seller sells identical postcards using his real name and an assumed name. They ﬁnd an 8% premium to having 2000 positive feedbacks and 1 negative over a feedback proﬁle with 10 positive comments and no negatives. Ba and Pavlou (2002) conduct a laboratory experiment in which subjects are asked to declare their valuations for experimenter generated proﬁles, and ﬁnd a positive See Dellarocas (2002), Resnick, Zeckhauser, Swanson and Lockwood (2003), and Bajari and Horta¸su (2004) for surveys of these results.
c response to better proﬁles. Jin and Kato (2004) assess whether the reputation mechanism is able to combat fraud by purchasing ungraded baseball cards with seller-reported grades, and having them evaluted by the oﬃcial grading agency. They report that while having a better seller reputation is a positive indicator of honesty, reputation premia or discounts in the market do not fully compensate for expected losses due to seller dishonesty.
Our main contribution to the study of online reputation mechanisms is to devise a number of theory-driven empirical tests to investigate the incentives created by eBay’s feedback system. Our focus is on the empirical implications of sellers’ equilibrium behavior. By contrast, with the exception of Jin and Kato (2004), previous work has studied buyers’ reaction to seller’s feedback record. Moreover, our empirical tests are primarily based on panel data, whereas most of the previous work is primarily based on cross-section data.
Using panel data allows us to account for seller-level heterogeneity in most of our empirical tests.2 In addition to the literature on eBay and its reputation mechanism, our paper also relates to the empirical study of models with adverse selection and moral hazard. In particular, one of the most striking and robust results in our paper is that, once a seller receives negative feedback from buyers, the frequency of such feedback increases dramatically. We show that this is consistent the presence of moral hazard and rejects a pure adverse selection model.
Abbring, Chiappori, and Pinquet (2003) suggest a related test for the presence of moral hazard in auto insurance by looking at interarrival times of reported accidents in panel data on claims histories. Their test exploits discontinuous changes in driver incentives created by an exogenously speciﬁed experience rating scheme determining insurance premia. They fail to ﬁnd evidence for moral hazard in a sample of French drivers. In our setting, seller incentives are created endogenously through buyers’ expectations of what the seller will do in the future, and hence a discontinuity in incentives in response to an “accident” (i.e. a negative comment) is more diﬃcult to establish. Nevertheless, we succeed in deriving a robust empirical implication that is strongly veriﬁed We believe the diﬀerence between panel and cross-section data is important. In fact, our results from panel data are typically very signiﬁcant, whereas our results from cross-section data, consistently with much of the previous literature, have weak statistical signiﬁcance.
in the data.3 The paper is structured as follows. In Section 2, we brieﬂy describe the institutional setup of eBay, in particular the mechanics of its reputation mechanism. In Section 3, we present our basic model of buyer and seller behavior, as well as a number of extensions. Section 5 tests the implications from our basic model regarding sales rate (Section 5.1), price (Section 5.2), frequency of negative feedback arrival (Section 5.3), exit (Section 5.4), and reputation building (Section 5.5). Section 6 concludes the paper.
2 The eBay reputation mechanism Since its launch in 1995, eBay has become the dominant online auction site, with millions of items changing hands every day. We will not attempt a detailed account of how eBay has evolved and what its trading rules are; the interested reader may ﬁnd this in a number of survey articles and in the popular press.4 Thus we are going to largely ignore the intricacies of the price formation process on eBay in what follows; however, from our modelling purposes it will not be too inaccurate to characterize the auction mechanism as a variant of the second-price auction.5 eBay does not deliver goods: it acts purely as an intermediary through which sellers can post auctions and buyers bid. eBay obtains its revenue from seller fees collected upon successfully completed auctions.6 Most importantly, to enable reputation mechanisms to regulate trade, eBay uses an innovative feedback system.7 After an auction is completed, both the buyer and the seller can give the other party a grade of +1 (positive), 0 (neutral), or −1 (negative), Abbring, Chiappori, and Pinquet (2003) show that, in the French auto insurance market, an accident increases the cost of future accidents. An implication of moral hazard is that the arrival rate of accidents decreases when an accident takes place. By contrast, our model predicts that the marginal beneﬁt of eﬀort decreases when an “accident” (negative feedback) happens. Therefore, the arrival rate of “accidents” should go up when the ﬁrst “accident” happens.
See Cohen (2002) for an entertaining historical account of eBay. Survey articles on Internet auctions include Lucking-Reiley (1999), Dellarocas (2003), and Bajari and Horta¸su c (2004).
In reality eBay auctions are dynamic auctions in which bidders place (possibly multiple) “proxy bids” indicating their maximum willingness-to-pay. See Roth and Ockenfels (2002), Ockenfels and Roth (2003), and Bajari and Horta¸su (2003) for detailed analyzes of dynamic c bidding behavior on eBay.
Success is deﬁned as a bid above the minimum bid or a secret reserve price set by the seller. eBay collects its fee even if the physical transaction does not take place.
eBay does oﬀer an escrow service for use with especially valuable goods, though this service is used for only a small fraction of the transactions.
along with any textual comments.8 eBay then displays several aggregates of the grades received by each seller
and buyer. These are:
1. Overall rating: this is the sum of positives minus negatives received by a seller from unique buyers throughout her entire history. Until March 1, 2003, this was the most prominently displayed feedback aggregate on eBay — it appeared next to the sellers’ user ID on the auction listing page, as can be seen in the sample eBay page in Figure 1. (Here, seller wsb5 is shown to have 127 net positive ratings from unique buyers.)