Bajari Hortacsu JEL2004.PDF (IBPS Term-4)

7/27/2019 Bajari Hortacsu JEL2004.PDF (IBPS Term-4)

1/30

Journal of Economic LiteratureVol. XLII (June 2004) pp. 457486

Economic Insights from InternetAuctions

PATRICK BAJARI and ALI HORTASU1

457

1 Bajari: Duke University and NBER. Hortacsu:University of Chicago and NBER. We thank the editor,two anonymous referees, Ginger Jin, Axel Ockenfels,David Reiley, Paul Resnick, and Alvin Roth for their verydetailed and insightful comments on various drafts of thisdocument. We thank the National Science Foundation forpartial research support, grants SES-0012106 (Bajari) andSES-0242031 (Hortasu).

2 There are several interesting accounts of the earlydevelopment of internet auctions. Cohen (2002) chroni-cles the development of eBay through in-depth interviewsof eBay founders, executives, and users. David Lucking-Reiley (2000b) provides an insightful description of inter-net auction sites and auction mechanisms across differentmarket segments.

1. Introduction

Electronic commerce continues to growat an impressive pace despite widelypublicized failures by prominent onlineretailers. According to the Department ofCommerce, total retail e-commerce in theUnited States in 2002 exceeded $45 billion,a 27-percent increase over the previous year.

Online auctions are one of the most success-ful forms of electronic commerce. In 2002,more than 632 million items were listed forsale on the web behemoth eBay alone, a 51-percent increase over the previous year. Thisgenerated gross merchandise sales of morethan $15 billion.

The rapid development of these marketsis usually attributed to three factors.2

The first is that online auctions provide a

3 Beanie Babies are stuffed dolls that are popularamong collectors.

less-costly way for buyers and sellers onlocally thin markets, such as specialized col-lectibles, to meet. Adam Cohen (2002, p. 45)states, It would be an exaggeration to saythat eBay was built on Beanie Babies, butnot by much.3 In May 1997, nearly$500,000 worth of Beanie Babies was sold oneBay, totaling 6.6 percent of overall sales.While it may be difficult to find a particular

Beanie Baby locally, such as Splash theWhale or Chocolate the Moose, you have agood chance of finding it online. Collectiblessuch as Beanie Babies, first edition books,Golden Age comics, and Elvis paraphernaliaare among the thousands of categoriesactively traded in online auctions.

The second factor is that online auctionsites substitute for more traditional marketintermediaries such as specialty dealers inantiques, sports cards, and other collectibles.For instance, an antique dealer from Seattleexplained on an eBay message board why heclosed his store:

A couple of years or so ago my best buyers start-ed spending their money at eBay. Then mypickers started selling on eBay instead of sellingto me. Then when I went to the flea market andasked how much an item was, I got quoted what


2/30

458 Journal of Economic Literature, Vol. XLII (June 2004)

one sold for on eBay, not what the seller want-ed for the item. I have a toy show that sold outfor years, but nowadays all my vendors sell oneBay, and all the buyers are spending theirmoney on eBay. I used to buy and sell a lot inthe toy magazines before they got reduced to

mere pamphlet-sized rags Get my drift?(Cohen, p. 110)

Online auctions have extensive listingsand powerful search technologies that createliquid markets for specialized product cate-gories. The resulting reduction in transac-tion costs forced some intermediaries, likethe antique dealer above, to exit the market.

Finally, online auctions can be fun! Many

bidders clearly enjoy contemplating thesubtleties of strategic bidding and sharingtheir insights with others. Most online auc-tion sites have active message boardswhere one can learn the fine points of col-lecting. The boards also provide a sense ofcommunity for diehard collectors.

In this paper, we survey recent researchconcerning online auctions. First, wedescribe the mechanics of the auction rules

used on the most popular sites and someempirical regularities. An especially interest-ing regularity is that bidders frequentlysnipe; that is, they strategically submit theirbids at the last seconds of an auction thatlasts several days. Several authors haveempirically examined sniping and proposedexplanations.

We then survey a growing literature inwhich researchers have attempted to docu-

ment and quantify distortions from asym-metric information in online auctionmarkets. As pointed out by EichiroKazumori and John McMillan (2003), theinformation asymmetry problem consti-tutes perhaps the biggest limitation posed tothe impressive growth of online auctions. Inonline auctions, transactions take placebetween complete strangers who may notlive in the same state or the same country,

making it very difficult for buyers to directlyinspect the good or to make sure that thegood will be delivered at all. This creates

4 A number of high-profile incidents of fraud haveoccurred on online auctions. For instance, the FBIlaunched an investigation called Operation Bullpen thatled to the indictment of 25 persons for selling tens of mil-lions of dollars of forged collectibles such as forged signa-tures from Babe Ruth and Lou Gehrig (Cohen, p. 308).

5

Another survey article that provides a more in-depthsurvey of theoretical models of reputation is ChrysanthosDellarocas (2002). See also the survey by David Baron(2002).

opportunities for misrepresentation ofobjects and fraudulent behavior by sellers,which may limit trade in these markets.4

The informational asymmetry may mani-fest itself as a winners curse problem, in

which bidders recognize that winning anauction is conditional on being the mostoptimistic bidder about the items worth.Auction theory then predicts that bidderswill respond strategically to the winnerscurse by lowering their bids, thus leading tolower prices and volumes in these markets.In section 4, we will survey empirical workthat uses detailed data from online auctionsites to test whether bidders indeed act

strategically in the face of a possible winnerscurse, and whether this strategic response islarge enough to prevent fraudulent sellersfrom extracting (short-term) rents.

Given the above discussion, a very impor-tant component of the online auction busi-ness is to decrease the informationalasymmetries between market participants. Aparticularly popular method, pioneered byeBay, is the use of feedback mechanisms that

allow buyers and sellers to leave publiclyavailable comments about each other. Insection 5, we will survey a rapidly growingempirical literature that utilizes data fromthe feedback mechanisms of online auctionsites to quantify the market value of onlinereputations.5

Auction theory, with its sharp predictionsabout the optimal way to design andconduct auctions, has become one of the

most successful and applicable branches ofmicroeconomic theory. Internet auctionsites, with their abundance of detailed dataon bidding and selling behavior, provide


3/30

Bajari and Hortasu: Economic Insights from Internet Auctions 459

fertile ground to test some of these theories.On many internet auction sites, sellers areallowed to fine-tune their auctions byexperimenting with minimum bids or secretreserve prices. Several sites also allow the

sellers to make choices regarding whichauction format to use. Thus, empiricalresearchers can utilize these sources of vari-ation, along with a newfound ability to setup randomized field experiments to testpredictions of auction theory. In section 6,we discuss several strands of researchaddressing this purpose. We first discuss theuse of internet auction sites as a medium forrandomized field experiments, and survey

empirical work that has utilized thismethodology to compare bidder behavioracross alternative auction formats. Second,we discuss the use of both field experimen-tation and structural econometric modelingtechniques to investigate how sellers shouldset reserve prices for their auctions. In par-ticular, we focus on the question of whyreserve prices are kept secret in many inter-net auctions. Third, we discuss the preva-

lence of ascending auctions on the internet.We consider alternative theoretical explana-tions, along with a discussion of new exper-imental evidence motivated by thisempirical observation. Finally, we discussthe importance of taking endogenous par-ticipation decisions into account whenconducting theoretical and empirical com-parisons of different auction rules, sinceonline auction sites typically feature multi-

ple auctions that are taking place at thesame time (and the sites themselves can bethought of as competing with each otherthrough their site designs). We concludewith a brief discussion of open researchquestions that remain to be explored.

2. Buying and Selling in Online Auctions

On any given day, sellers list millions of

items in online auctions. These sites includebusiness-to-business (B2B) sites,6 and siteswhere the typical buyer is a consumer. We

6 See David Lucking-Reiley and Daniel Spulber (2001).Examples are MuniAuction.com, which conducts auctionsof municipal bonds; ChemConnect.com for trade in

chemicals; Wexch.com for electricity contracts.7 Lucking-Reiley (2000b) surveyed 142 online auctionsites that were in operation in autumn of 1998. Exact esti-mates of the number of auction listings are few and farbetween. Lucking-Reiley (2000b) reported that in summer1999, 340,000 auctions closed on eBay every day, asopposed to 88,000 on Yahoo!, and 10,000 on Amazon(most auctions last seven days). Sangin Park (2002)reports, based on Nielsen/NetRatings surveys that eBayhad 5.6 million listings, and Yahoo!, 4.0 million listings asof autumn 2000. When we counted the number of activelistings on January 25, 2004, we found 14.5 million listingson eBay, and only 193,800 listings on Yahoo!. Our counts

of Amazon listings were much more inexact, since thewebsite does not divulge this information as visibly as eBayor Yahoo!. However, our estimate is between 500,000 and1.5 million listings.

will focus on the latter category since this hasbeen the focus of most empirical research.On the largest sites, such as eBay, Amazon,and Yahoo!, bidders can choose from amind-boggling array of listings.7

These sites facilitate search in two ways.First, the sites have a carefully designed setof categories and subcategories to organizethe listings. For instance, eBays main pagelists general categories that include automo-biles, real estate, art, antiques, collectibles,books, and music. Within a subcategory,there are multiple layers of additional cate-gories. For instance, the subcategory ofpaintings includes antique American and

modern European. These categories arecarefully designed by the auction site tofacilitate buyers searches. Secondly, userscan search the millions of listings by key-words, category, price range, and completeditems. This allows users to gather consider-able information about similar products,which is useful in forming a bid.

In figure 1, we display a listing from eBay.The web page describes the item for sale,

which in this case is a collection of signa-tures of fourteen Nobel laureates in eco-nomics. The user can see the current highbid, the time left in the auction, the identityof the seller, and the highest bidder. Theweb page describes the item for sale and


4/30


Figure 1. Sample eBay Listing


5/30


8 Similar mechanisms are used on Yahoo! andAmazon. Yahoo!s feedback pages look almost exactly likeeBays. Amazon uses a slightly different five-star systemto summarize the comments.

9 The fees that eBay charges have been updated sev-eral times, with vocal dissent from eBay sellers in certaininstances, such as when eBay decided to charge a $1 fee touse the secret reserve option (Troy Wolverton 1999). Also,see Park (2002) for some evidence on the impact of net-

work effects on the competition (in listing-fees) betweeneBay and Yahoo!

10 Yahoo! also allows straight bidding, i.e., biddingprecisely the price you want to pay.

sometimes displays a seller-supplied pic-ture. The listing also displays the sellersfeedback to the right of the sellers identity.Users on eBay can leave each other feed-back in the form of positive, neutral, or neg-

ative comments. The total feedback is thesum of positive comments minus the num-ber of negative comments. eBay also com-putes the fraction of positive feedbackreceived by the seller.8

The listing displays the minimum bid andwhether the seller is using a secret reserve.The minimum bid is analogous to a reserveprice in auction theory; that is, the sellerwill not release the item for less than the

minimum bid. A secret reserve price func-tions similarly except that it is not publiclydisplayed to the bidders. Bidders can onlysee whether or not the secret reserve ismet. Sellers on eBay pay an insertion feeand a final value fee. The insertion fee isbased on the minimum bid set by the seller.For instance, if the minimum bid isbetween $25 and $50, the insertion fee iscurrently $1.10. The final sale fee is a non-

linear function of the final sale price. Foritems with a final sale price of less than $25,the fee is 5.25 percent of the final saleprice. Higher sale prices are discounted atthe margin. Also, eBay has additional fees ifa secret reserve is used or if more than onepicture is used.9

At the bottom of the listing, buyers cansubmit their bids. All three major onlineauction sites use some variant of proxy bid-

ding.10 Heres how it works. Suppose that aseller lists an Indian-head penny for sale

11eBay has a sliding scale of bid increments, based on

the current price level in an auction.12 There is a 10-cent charge for running a ten-day

auction on eBay.

with a minimum bid of $15.00. At this pricelevel, eBay requires a bid increment of $.50to outbid a competitor.11 If bidder A places aproxy bid of $20, the eBay computer willsubmit a bid of $15.00, just enough to make

bidder A the highest bidder. Suppose thatbidder B comes along and submits a bid of$18.00. Then the eBay computer will updateAs bid to $18.50 (the second-highest bidplus one bid increment). The proxy biddingsystem updates As bid automatically, until Ais outbid by another bidder. If this occurs,the bidder will be notified by email andgiven a chance to update their bid.

Although all three sites use the proxy-bid-

ding mechanism, they differ in how theyend the auction. eBay auctions have a fixedending time set by the seller, who chooses aduration between one, three, five, seven,and ten days, and the auction closes exactlyafter this duration has passed.12 OnAmazon, if a bid is submitted within the lastten minutes of the previously fixed endingtime, the auction is automatically extendedfor ten additional minutes. Furthermore,

for every subsequent bid, the ten-minuteextension still applies. Hence, the auctionends only if there is no bidding activity with-in the last ten minutes. Interestingly, Yahoo!takes the middle-ground between these twoending rules, and allows sellers to choosewhich one to apply to their auction. As wewill see in the next section, this seeminglyinnocuous difference in ending rules is asso-ciated with divergent bidding behaviors

across auction sites.

3.Last-Minute Bidding

Bids commonly arrive during the last sec-onds of an internet auction that lasts as longas several days. For instance, Alvin Rothand Axel Ockenfels (2002) and Ockenfelsand Roth (2003) find, in a sample of 240


6/30


antique auctions on eBay, that 89 had bidsin the last minute and 29 in the last ten sec-onds. Other researchers, including RonaldWilcox (2000), Bajari and Hortasu (2003),and Julia Schindler (2003), have document-

ed similar patterns. In this section, we sum-marize some of the proposed explanationsfor last-minute bidding and the relatedempirical work.

Last-minute bidding is difficult to explainusing standard auction theory. Proxy biddingbears a strong similarity to the second-pricesealed-bid auctions since, in both cases, thepayment by the winning bidder is equal tothe second highest bid. William Vickrey

(1961) observed that in a second-pricesealed-bid auction with private values, it is aweakly dominant strategy for a bidder to bidtheir reservation value. The intuition is sim-ple. If the bid is less than their private value,there is a probability that she will lose theauction. However, the payment the winningbidder makes only depends on the second-highest bid. As a result, bidding ones valua-tion weakly increases ones payoff. Thus, at

first glance, it appears that bidders can leavetheir proxy agents to do their bidding, anddo not need to wait until the last seconds ofthe auction.

However, this is clearly not what happensin practice, and therefore several explana-tions for late bidding have been proposed inthe literature. A first explanation, whichdeparts only slightly from the independentprivate values environment, is proposed by

Ockenfels and Roth (2003). They argue thatlate bidding may be a form of tacit collusionby the bidders against the seller. In theirmodel, bidders can choose to bid early or late.A late bid, however, might not be successful-ly transmitted due to network traffic. Thereare (at least) two possible equilibria in theirmodel. In the first, agents bid early in theauction, and in the second agents only bid atthe last second. Bidding late is a risk because

the bids may not be successfully transmitted.On the other hand, late bidding softens com-petition compared to the first equilibrium.

Ockenfels and Roth (2003) demonstratethat this tacit collusion explanation of last-minute bidding equilibrium hinges on theassumption that there is a hard deadline forsubmitting bids. As mentioned in the last

section, however, while eBay auctions have ahard deadline, Amazon auctions are auto-matically extended if a late bid arrives.Ockenfels and Roth (2003) then demon-strate that late bidding is no longer an equi-librium in Amazon-style auctions. Theyconclude that there are more powerfulincentives for late bidding in eBay auctionsthan in Amazon auctions.

As a test of this theory, Roth and Ockenfels

(2002) and Ockenfels and Roth (2003) com-pare the timing of bids for computers andantiques on Amazon and eBay. They find thatlate bidding appears to be more prevalent inthe eBay auctions. On eBay, bids are submit-ted within the last five minutes in 9 percentof the computer auctions and 16 percent ofthe antique auctions. On Amazon, about 1percent of the auctions in these categoriesreceive bids in the last five minutes. Bidder

surveys reveal that late bidding on eBay is adeliberate strategy meant to avoid a biddingwar. Similar evidence comes from Schindler(2003), who studies bidding in Yahoo auc-tions for computers, art, and cars. In theseauctions, the sellers can choose to have a hardclose or an automatic extension, similar toAmazon auctions, if late bids arrive. She findsthat, consistent with the Ockenfels and Roththeory, the winning bidder tends to arrive

later in the auctions, with a hard ending forall three product categories.

We should note that the empirical findingby Roth and Ockenfels (2002), Ockenfelsand Roth (2003), and Schindler (2003)thatthere is less sniping in flexible-ending ruleauctionsis not confirmed in all studies.Gillian Ku, Deepak Malhotra, and J. KeithMurnighan (2003) study bidding in onlineauctions in several U.S. cities for art that has

been publicly displayed. The proceeds of theauctions were donated in part to charitablecauses. The authors observe online auctions


7/30


13 However, this revenue ranking may be reversed in anaffiliated interdependent values environment.

with both hard endings and flexible endings.They observe substantially less late biddingthan the previously mentioned studies. Only1.6 percent of the bids arrive in the last fiveminutes of the auctions with hard deadlines

and 0.5 percent in the auctions with flexibleendings. Also, Ku et al. find that for auctionswith flexible endings, a greater percentageof the bids arrive in the last hour and the lastday than for auctions with hard endings. Thisis not qualitatively consistent with Ockenfelsand Roths prediction. However, it is impor-tant to note that the authors have limitedcontrols for heterogeneity across auctionsthat are held in different cities.

Several other empirical researchers haveexamined the tacit collusion hypothesis ofOckenfels and Roth further. Kevin Hasker,Raul Gonzalez, and Robin Sickles (2003) testthe Ockenfels and Roth theory by examiningbids for computer monitors on eBay. If latebids soften competition and lower the proba-bility of price wars, then the distribution ofthe winning bids conditional upon a snipeshould not equal the distribution of the win-

ning bids if no snipe occurs. Hasker et al. findthat in most of the specifications they exam-ine, they are unable to reject the equality ofthese two distributions. They argue that thisis inconsistent with the tacit collusion theo-ry. Similarly, in a data set of bidding for eBaycoin auctions, Bajari and Hortasu (2003)find that reduced form regressions suggestthat early bidding activity is not correlatedwith increased final sales prices. Schindlers

(2003) study also casts some doubt on thetacit collusion hypothesis on the revenuefront. Under private values, the Ockenfelsand Roth model predicts that a hard closewill decrease revenues for the seller.13

Schindler finds sellers more frequently useauctions with automatic extensions for art(82 percent vs. 18 percent), and cars (73 per-cent vs. 27 percent), but not for computers(91 percent vs. 9 percent). Therefore, under

14 Based on survey data, Ku et al. (2003) also find thatsome bidders may be driven by emotional factors, whichthey label as competitive arousal. For instance, one sur-vey respondent from Cincinnati who purchased a ceramic

pig explained that she really wanted the pig, and probablyalso got caught up in the competitive nature of the auc-tion. Another respondent explained her behavior by stat-ing, Auction fever took over.

the Ockenfels and Roth theory, the first twocategories are consistent with seller revenuemaximization while the last category is not.

A second theoretical explanation, alsooffered by Ockenfels and Roth (2003) is the

presence of nave bidders on eBay who donot understand the proxy-bidding mecha-nism, and hence bid incrementally inresponse to competitors bids.14 Ockenfelsand Roth (2003) demonstrate that last-minute bidding is a best-response by ration-al bidders against such nave bidders. Theypresent empirical evidence that more expe-rienced bidders are less likely to place mul-tiple, incremental bids on eBay. This

evidence is complemented by survey evi-dence reported in Roth and Ockenfels(2002). Furthermore, Dan Ariely, Ockenfels,and Roth (2003) conduct a controlled labo-ratory experiment in which one of the exper-imental treatments is an eBay-typefixed-deadline auction, where the probabili-ty of losing a bid due to transmission erroris zero. Notice that this feature rules out thetacit collusion explanation; however, it is still

a best-response against nave incrementalbidders to bid at the last second.Accordingly, Ariely et al. (2003) find a signif-icant amount of late-bidding activity in thisexperimental treatment.

A third explanation is based on a commonvalue. In a common-value auction, such asRobert Wilsons (1977) mineral-rights model,the item up for sale has a true value Vthat isnot directly observed by the bidders. For

example, the common value Vcould be theresale value of a collectible. Each bidderreceives an imperfect signal x ofVwhich isprivate information. By bidding early, a bid-der may signal information about x to otherbidders and cause them to update their


8/30


beliefs about V. Conditional on winning, thismay increase the price that a bidder has topay for the item. Bajari and Hortasu (2003)formalize this intuition and demonstrate thatlast-minute bidding occurs in models of

online auctions with a common valueinfact, they show that in a symmetric common-value environment, the eBay auction can bemodeled as a sealed-bid second-price auc-tion. A similar result is also provided inOckenfels and Roth (2003) in a simpler set-ting. This explanation also finds some sup-port in the data. Roth and Ockenfels (2002)and Ockenfels and Roth (2003) report thatthere is more last-minute bidding on eBay

antiques auctions than in eBay computerauctions. They argue that antiques auctionsare more likely to possess a common valueelement than computer auctions, and hencethe observed pattern is consistent with thetheoretical prediction.

A fourth explanation for late bidding is pro-posed by Joseph Wang (2003), who studies amodel in which identical items are simultane-ously listed, as opposed to the usual assump-

tion that only a single unit of the item is up forsale. Last-minute bidding is part of the uniqueequilibrium to his model. Michael Peters andSergei Severinov (2001) argue that theirmodel of bidding for simultaneous listings ofidentical objects is also qualitatively consistentwith late bidding. Taken together, these twopapers suggest that the multiplicity of listingsis another explanation for late bidding.However, it is not clear whether the Wang

(2003) and Peters and Severinov (2001) mod-els predict that last-minute bidding should beless prevalent on Amazon as opposed to eBay.

A fifth explanation is given by EricRasmusen (2001), who considers a model inwhich bidders have uncertainty about theirprivate valuation for an item. As in themodel of Dan Levin and James Smith(1994), he assumes that some bidders mustpay a fixed fee to learn their private infor-

mation. This is not an unreasonable assump-tion in online auctions. In order to learntheir private valuations, bidders will inspect

the item and may search for sales prices ofpreviously listed items. The fixed fee can bethought of as the opportunity cost of timerequired to do this research. Rasmusendemonstrates that late bidding can occur

because bidders wish to economize on thecosts of acquiring information.The multiplicity of explanations provided

in the literature regarding the causes of aseemingly innocuous phenomenon like last-minute bidding is a great example of how theanalysis of online auctions enables us toappreciate the richness and complexity ofstrategic interaction in markets. The preced-ing discussion also illustrates how a seeming-

ly simple empirical regularity can havemultiple explanations, and that it is not easyto discern between these explanations with-out creative exploitation of sources of varia-tion in the data. In this regard, theexperimental study of Ariely, Ockenfels, andRoth (2003) is a good demonstration of howlaboratory experiments can complementdata obtained from real markets to helpexplain complex strategic interactions.

4. The Winners Curse

In online markets, buyers are not able toperfectly observe the characteristics of thegoods for sale. In the market for collectiblesand other used goods, buyers value objectsthat appear new. Scratches, blemishes, andother damage will lower collectors valua-tions. Since a buyer cannot directly touch or

see the object over the internet, it may behard to assess its condition. This introducesa common-value component into the auc-tion, and therefore bidders should accountfor the winners curse.

The winners curse can be illustrated bythe following experiment discussed in PeterCoy (2000):

Paul Klemperer illustrates the winners

curse to his students by auctioning off a jar withan undisclosed number of pennies. The stu-dents bid a little below their estimate of thejars contents to leave a profit. Every time,


9/30


though, the hapless winner of the jar is the stu-dent who overestimates the number of penniesby the greatest amount, and therefore overpaysby the most.

The winners curse occurs when bidders

do not condition on the fact that they willonly win the auction when they have thehighest estimate. If there is a large numberof bidders, the highest estimate may bemuch larger than the average estimate.Therefore, if the winner bids naively, hemay overpay for the item. In addition to theclassroom experiment above, a number ofexperimental studies find that inexperi-enced bidders frequently are subject to the

winners curse (see John Kagel and AlvinRoth 1995 for a survey of the experimentalliterature).

Several authors empirically examinewhether bidders are subject to the winnerscurse and, more generally, measure distortionsfrom asymmetric information in online mar-kets. Two main methodologies have been uti-lized to answer this question so far. The firstmethod, utilized by Ginger Jin and Andrew

Kato (2002), is to sample goods in an actualmarket, and determine whether the pricesreflect the ex-post quality of the goods theybuy. Online auctions present perhaps a uniqueopportunity in the utilization of this methodol-ogy, since items sold in these markets are inex-pensive enough to allow researchers to shopout of their own pockets or research grants.

In particular, Jin and Kato (2002) studyfraudulent seller behavior in an internet

market for baseball cards. Fraud on theinternet is a problem. The Internet FraudCenter states that 48 percent of the 16,775complaints lodged in 2001 were from onlineauctions. In this market, Jin and Kato findthat sellers frequently misrepresent howbaseball cards will be graded.

Professional grading services are com-monly used for baseball cards. Gradingservices produce a ranking from 1 to 10

based on the condition of the card. Cardsthat are scratched or have bent corners areassigned lower grades than cards that are in

15 They attempted to bid just enough to win the auc-tion and to have a minimal effect on the actions of otherplayers.

mint condition. Jin and Kato bid on eBayauctions for ungraded cards and then sub-mitted the cards to a professional gradingservice.15 Their sample contained 100ungraded baseball cards. Of these, sellers

of nineteen claimed that they had a rankingof 10 (gem mint), 47 claimed mint (9 or9.5), sixteen near-mint to mint (8 or 8.5),seven near-mint (7 or 7,5); eleven cards hadno claim.

Many sellers of ungraded cards misrepre-sented the card quality. Among sellers whoclaimed that their cards were grade 9 to 10,the average grade was 6.34. In comparison,the average grade of cards claiming 8.5 or

below was 6.87. The price differencebetween a grade 10 and a grade 6.5 for somecards could be hundreds or thousands ofdollars. Jin and Kato find evidence that somebuyers were misled by these claims. Buyerswere willing to pay 27 percent more forcards that were seller-reported as 9 to 9.5and 47 percent more for cards that werereported as 10. They conclude that somebuyers have underestimated the probability

of fraud and therefore have fallen prey to thewinners curse.

To establish whether the winners curseproblem is more of an issue in online asopposed to offline markets, the authorsemployed male agents between 2535 yearsof age to purchase the same types of cardsfrom retail collectibles stores in eleven metro-politan districts, and found that the fraud ratethere (3.2 percent) was much lower than

online (11 percent). Moreover, in their retailpurchases, the authors found that retail sellerswere more reluctant to quote the likely gradeof the cards when one was not available.

Does this mean that online auctions forbaseball cards are fraught with fraudulentsellers and nave buyersa market that coulduse further regulation and intervention bythird parties to operate more efficiently?


10/30


Before we jump to this conclusion, webelieve that a cautious interpretation of Jinand Katos findings is in order. In our opinion,the main finding of this paper is that whileonline bidders account for the possibility of

misrepresentation, they might make mistakeswhen assessing the probability of fraudulentclaims. It should be pointed out that themistakes that buyers make are not likely tolead to very large monetary losses, since mostitems in this market are mundanely priced(between $60 and $150). Even for more valu-able items, the estimated losses from buyernaivete are not very large. For instance, Jinand Kato report that the average eBay price

for a grade-10 Ken Griffey Jr.s 1989 UpperDeck Card (the most actively traded card oneBay) is $1,450. An ungraded Griffey with aself-claim of 10 only sold for an average of$94.26. Jin and Katos estimates imply thatthis self claim generated an extra $30 in rev-enue for a fraudulent seller. If bidders naive-ly took the claims at face value, this wouldhave generated a mistake of over $1,300hence the Ken Griffey example shows that

bidders by and large do correct for thewinners curse. The contention is whetherthe bidders correct enough for this.Furthermore, the authors did manage to pur-chase a card of mint quality (grade 9) fromtheir ungraded group. Apparently, taking achance with ungraded cards sometimes doesin fact pay off.

Regardless of how one may interpret Jinand Katos (2002) findings, their methodolo-

gy to assess the prevalence of fraud and themagnitude of the corresponding winnerscurse suffered by buyers is a very direct andeffective way to get precise measurements.Unfortunately it would be difficult, if onlydue to budget constraints, to replicate thismethodology to examine markets for moreexpensive items like computers or automo-biles (which are actively traded on eBay).Hence, the second set of papers we will sur-

vey utilizes more indirect methods thatrely on testing the implications of strategicbidding in common-value auction models (in

16We should note, however, that Paarsch (1992) imple-mented his idea in the context of first-price auction, wherethe comparative static with respect to the number of bid-ders in the auction is ambiguous. See Susan Athey and

Philip Haile (2002) and Haile, Han Hong, and MatthewShum (2003) for more rigorous derivations of this compar-ative static result in the context of second-price andascending (button) auctions.

which bidders account for the presence of awinners curse.)

The first study we will summarize in thisregard is Bajari and Hortasu (2003), whoexamine the effect of a common value on

bidding for collectible coins in eBay auc-tions. The authors first argue that bidding inan eBay auction in the presence of a com-mon-value element can be modeled as bid-ding in a second-price sealed-bid auctionwith common values (and an uncertain num-ber of opponents). The reasoning is basedon the observation of sniping in these auc-tions: the presence of a common-value ele-ment in an eBay auction suggests that

bidders should wait until the last minute inorder to avoid revealing their private infor-mation. If all bids arrive at the last minute, abidder will not be able to update his beliefsabout the common value Vusing the bids ofothershence his bidding decision will beequivalent to that of a bidder in a sealed-bidsecond-price auction.

Given this argument, the authors then testfor the presence of a common-value element

using an idea first proposed by Harry Paarsch (1992).16 If the auction environmentis one with purely private values, it is a dom-inant strategy for bidders to bid their privatevalues, independent of the number of com-petitors that they face. If there is a common-value element, however, the bids will dependon the number of bidders present. This isbecause when there are more bidders, thepossibility of suffering a winners curse con-

ditional on winning the auction is greater,further cautioning the bidders to tempertheir bids to avoid the curse.

Bajari and Hortasu (2003) implementthis using cross-sectional regressions of bidson the number of bidders, and apply various


11/30


instruments to account for the endogeneityof the number of bidders in the auction.Consistent with the presence of a common-value element, they find that bids declinewith the number of competing bidders.

They test whether bidders with little eBayexperience tend to bid systematically higher,by regressing individual bids on total eBayfeedback. While they find a statistically sig-nificant effect on bids, it is very small inmagnitude. This is not consistent with exper-imental evidence that suggests inexperi-enced participants overbid, and hence aremore likely to suffer from a winners curse.

Next, the authors attempt to directly

measure the distortions from asymmetricinformation by estimating a structuralmodel. In their structural model, the datagenerating process is a Bayes-Nash equilib-rium where: 1) there is a common value, 2)bidders have to pay a fixed cost to learn theirsignalx, and 3) entry is endogenously deter-mined by a zero profit condition. Given thisstructural model, they estimate the parame-ters for the distribution of the common value

V, the distribution of bidders private infor-mation, x and the parameters that governentry decisions.

The authors find that bidders prior infor-mation about V is quite diffuse. In theirmodel, ex ante beliefs about Vare normallydistributed. The standard deviation is 0.56times the book value of the coin if there is noblemish and 0.59 times the book value if thereis a blemish. For instance, the standard devia-

tion ofVfor a coin with a book value of $100with no blemish is $56. However, the distri-bution ofx has much lower variance than V.For instance, the distribution forx for the coindescribed above would be normally distrib-uted with a mean equal to Vand a standarddeviation of $28. Apparently, after a buyer hasspent the time and effort to research a coin,his estimate x is considerably more precisethan his prior information about V.

Given their parameter estimates, theauthors simulate their structural model toquantify the distortions from asymmetric

information. First, they compute theequilibrium bidding strategies for an auctionwhere all of the covariates are set equal totheir sample averages (most importantly, thebook value is equal to $47). The bid func-

tions are roughly linear. In this auction, bid-ders shade their bids to be $5.50 less thantheir signalsx. Given their uncertainty aboutV, the buyers bid 10-percent less than theirsignals to compensate for the winners curse.Second, they simulate the impact of addingan extra bidder on expectation (since biddersarrive at random). When the expected num-ber of bidders increases, the winners curseis more severe. Bajari and Hortasu find that

for an auction with all characteristics setequal to their sample averages, adding anadditional bidder reduces equilibrium bidsby 3.2 percent as a function of the bidderssignalx. Finally, they study the effect of low-ering the variance ofVfrom 0.52 times thebook value to 0.1. They find that the bidfunction shifts upwards by $2.50.

There are two main limitations to theiranalysis. First, the results depend on cor-

rectly specifying both the game and the para-metric assumptions used in the structuralmodel. The common value model that theyuse is fairly strict. All players are perfectlyrational, symmetric, and have normally dis-tributed private information. However, theynote that it is not computationally tractable,with current methods, to generalize themodel since computing the equilibriuminvolves high dimensional numerical integra-

tion. Second, unlike Pai-Ling Yin (2003) or Jinand Kato (2002), the authors do not have richex-post information about the realization ofthe common value or for bidder uncertainty.

In a related study, Yin (2003) utilizes com-parative static predictions of the common-value second-price auction model to test forthe presence of common values in eBay auc-tions for used computers. Based on numeri-cal simulations, Yin establishes that bidding

strategies in the common-value second-priceauction model respond strongly to changesin the variance ofx, a bidders private signal


12/30


0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

.

0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 1.1 1.2 1.3 1.4 1.5 1.6 1.7 1.8

Normalized Dispersion SD (=sd/V)

NormalizedPrices=P/V

Figure 2. Bidder Uncertainty and Winning Bids

about V. Holding Vfixed, if the variance ofxincreases, the bidder is more likely to be

subject to the winners curse. Conditional onwinning, the value ofx should be largerholding V fixed. Therefore, bidders shouldbehave more conservatively.

Yin supplements her data set with surveyinformation in order to understand howchanging the variance of the signal x influ-enced bidding. To do this, she downloadedthe web pages for 223 completed auctionsand asked survey respondents from the

internet to reveal their estimates,x, purginginformation about the bids and the reputa-tion of the seller from the web page so thatthe survey responses would primarily reflectbidders ex-ante beliefs about the item forsale and not the seller characteristics. Bycomputing the variance of an average of 46such responses per auction, she constructeda proxy for the variance ofx. The highestvariance occurred in the auctions where the

seller poorly designed the web page orwhere the object for sale had inherentlyambiguous characteristics.

Yin finds that the winning bid is negative-ly correlated with the normalized variance of

the survey responses. This is illustrated infigure 2. The vertical axis is the winning biddivided by the average survey response. Thehorizontal axis is the variance of the respons-es, divided by the average response. For anauction where the normalized variance was0.4, the expected value of the normalizedwinning bid was 0.7. As the varianceincreased to 1.3, the expected value of thenormalized winning bid was less than 0.2.

Given that the auctions were for computersworth hundreds of dollars, these are fairlylarge magnitudes.

One interpretation of these results is thatbidders act as if they understand the win-ners curse. The bidders are reluctant to sub-mit high bids when they are uncertain aboutthe condition of the used computer, asreflected by a high variance in x. On theother hand, if they are more certain about

what they are purchasing, they will bid withmore confidence. These findings suggestthat asymmetric information can generate


13/30


significant distortions and therefore sellershave a strong incentive to communicatewhat they are selling by designing a clearweb page to reduce bidder uncertainty.

A potential limitation to Yins analysis is

that the variance ofx for survey respondentsmay be an imperfect measure of the varianceofx for the real bidders. The survey respon-dents reservation values are roughly twicethe final sales price on average. Perhaps theyare not well informed about used-computerprices. On the other hand, if the winnerscurse is a strong possibility, it may be anequilibrium for a bidder to shade his bid tobe one half of his estimatex.

Conditional on their caveats, all threepapers surveyed in this section suggest thatthe winners curse is an important concern inthe eBay markets analyzed (baseball cards,collectible coins, used computers). Two ofthe three papers, Bajari and Hortasu (2003)and Yin (2003), find evidence that biddersstrategically respond to the presence of awinners curse. Jin and Kato (2002) also pres-ent similar evidence of strategic bid-shad-

ing, but, through their ex-post appraisal ofcards sold on eBay, they argue that theamount of bid-shading by bidders in thebaseball cards market is not large enough.

As pointed out in the introduction, animportant implication of these results, notedby Kazumori and McMillan (2003), is thatdepressed prices due to the winners cursemay limit the use of online auctions by sell-ers to items for which informational asym-

metries do not play a very large role.Kazumori and McMillan (2003) report thatafter several years of experimentation withrunning art auctions on the internet in part-nership with eBay, Sothebys decided to dis-continue selling art on the internet as of May2003. As the authors note, there may havebeen several confounding factors that led tothe failure of Sothebys.com. Fortunately,this still leaves the investigation of the ques-

tion which goods can be sold using onlineauctions open for further empirical andtheoretical analysis.

17 Sellers are also required to list a valid credit-cardnumber.

18 Resnick and Zeckhauser (2001) had data on transac-tions conducted on eBay in a five-month period, hencethey cannot track transactions preceding this period.Within this time period, however, repeat transactions, ifthey occurred at all, happened in a very short time period.

19 Other auction sites, such as Amazon, allow for more

nuanced feedback. Amazon, for example, utilizes a scaleof 1 to 5.

20 In March 2003, eBay also began displaying the per-centage of positive feedbacks along with the total score.

5. Reputation Mechanisms

Perhaps the most important source ofinformation asymmetry on online auctions,aside from the inability to physically inspectgoods, is the anonymity of the sellers. For

instance, eBay does not require its users todivulge their actual names or addresses; allthat is revealed is an eBay ID.17 There arealso very few repeat transactions betweenbuyers and sellers. For instance, Paul Resnickand Richard Zeckhauser (2001), using a largedata set from eBay, report that fewer than 20percent of transactions are between repeatedbuyer-seller pairs within a five-month peri-od.18 This obviously limits a buyers informa-tion about the sellers reliability and honesty.

To ensure honest behavior, online auctionsites rely on voluntary feedback mecha-nisms, in which buyers and sellers alike canpost reviews of each others performance.On eBay, a buyer can rate a seller (and viceversa) by giving them a positive (+1), neutral(0), or negative (-1) score, along with a textcomment.19 eBay records and displays all ofthese comments, including the ID of theperson making the comment. eBay also dis-plays some summary statistics of users feed-back. The most prominently displayedsummary statistic, which accompanies everymention of a user ID on eBays web pages, isthe number of positives that a particular userhas received from other unique users, minusthe number of negatives.20 eBay alsocomputes and reports the number of posi-

tives/neutral/negatives that a seller hasreceived in their lifetime, along with the lastweek, last month, and last six months.


14/30


21 Resnick and Zeckhauser (2001) also report in a small-er sample that in 18 out of 87 cases (20 percent) where abuyer left negative feedback about a seller, the sellerresponded with a negative for the buyer. There is also thepossibility that some buyers are more critical than othersalthough eBay allows users to observe the feedback that auser left about others, it does not provide summary statis-tics of left feedbackhence, making it quite costly for a

prospective buyer to gauge the attitude of a particularcommentor (other sites that provide user reviews, such asAmazon, allow readers of a comment to rate that commenton the basis of its usefulness).

A prospective buyer on eBay therefore hasaccess to a considerable amount of informa-tion about the reputation of a seller. Forinstance, almost all of the feedback on eBayis positive. Resnick and Zeckhauser (2001)

report that only 0.6 percent of feedbackcomments left on eBay by buyers about sell-ers was negative or neutral. One interpreta-tion of this result is that most users arecompletely satisfied with their transaction.Another interpretation, however, is thatusers are hesitant to leave negative feedbackfor fear of retaliation. Luis Cabral and AliHortasu (2003) report that a buyer wholeaves a negative comment about a seller has

a 40-percent chance of getting a negativeback from the seller (whereas a neutral com-ment has a 10-percent chance of being retal-iated against).21 Another factor limiting thepotential usefulness of feedback, reportedby Resnick and Zeckhauser, is that feedbackprovision is an arguably costly activity that iscompletely voluntary, and that not all buyers(52.1 percent) actually provide reviewsabout their sellers.

5.1 Empirical Assessment of FeedbackMechanisms

Do online reputation mechanisms work?The fact that we observe a large volume oftrade on sites like eBay, Yahoo! and Amazonmay suggest that the answer to this questionis affirmative. However, there have also beena number of attempts to answer this ques-tion in a more direct manner, by estimating

the market price of reputation in online auc-tions through hedonic regressions.

Table 1, which is adapted from Resnick etal. (2003), summarizes these empirical stud-ies. The first ten studies on this list used dataon completed auctions on eBay to run cross-sectional hedonic regressions of the sale

price or sale probabilities of similar objectson sellers observable feedback characteris-tics. The (log) number of positive and nega-tive comments is used as a right-hand sidevariable in almost all of these studies. Someof these studies also attempted to accountfor nonlinearities in the functional relation-ship by putting in dummy variables for theexistence of negatives, or dummy variablesfor different ranges of feedback. Some stud-

ies also explicitly recognize the truncationproblem caused by auctions that did not endin a sale, and estimated a probability of sale equation either separately or jointlywith the (log) price regression.

As can be seen in table 1, although thesigns and statistical significance of theregression coefficients are mostly of theexpected kind, the empirical results fromthese studies are not easily comparable,

since some of them are reported in absoluteterms and some are reported in percentageterms. We should also note that the absolutenumber of feedback comments received bysellers follows a highly skewed distribution,reflecting a Gibrats Law type effect. Giventhis, however, the starkest differencesappear to be between sellers who have nofeedback records, and those with a very large number of positive comments. For

example, Mikhail Melnik and James Alm(2002) report that the difference between452 and 1 positive comment is $1.59 for$33 items, implying a price premium of 5percent. Jeffrey Livingstons (2002) esti-mates imply a more than 10-percent pricepremium for sellers with more than 675positive comments, as opposed to those withno feedback. Similarly, Kirthi Kalyanamand Shelby McIntyre (2001) report that a

seller with 3000 total feedbacks and nonegatives is estimated to get a 12-per-cent higher price than a seller with ten total


15/30


22Standard models of dynamic industry equilibrium,

such as Hugo Hopenhayn (1992) and Richard Ericson andAriel Pakes (1998), suggest that survival is positively corre-lated with measures of seller productivity.

comments and four negatives. The presenceof an up to 1012 percent price premiumbetween an established eBay seller (withhundreds or thousands of feedback com-ments) and a seller with no track record

appears to be the most robust result amongthese studies, especially in light of the 8.1-percent price premium uncovered by arecent field experiment conducted byResnick et al. (2003) to account for some ofthe more obvious confounding factorsinherent in the hedonic regressions, whichwe shall discuss now.

The first confounding factor is that the esti-mate of reputation may be subject to an omit-

ted variable bias. Since there is very littlenegative feedback, the number of positives isessentially equal to the number of transactionsthat a seller has completed on eBay. Moreexperienced eBayers are on average moreadept at constructing a well-designed webpage and responding to buyer questions. Thework of Yin (2003) suggests that a well-designed web page has a sizeable effect onfinal sale prices by reducing buyer uncertain-

ty about the object for sale. As a result, moreexperienced sellers (as measured by positivefeedback) will have higher sales prices for rea-sons that have nothing to do with reputation.22

Obviously, these factors could lead to biasedestimates that overstate the market value ofreputation scores obtained by sellers on eBay.

Sulin Ba and Paul Pavlou (2002), Cabraland Hortasu (2003), and Resnick et al.(2003) attempt to reduce this confounding

factor by manipulating feedback indicatorsindependently of other characteristics ofauction listings. For example, Resnick et al.(2003) conduct a field experiment with thehelp of an established eBay seller (2000 pos-itives, 1 negative), by selling matched pairsof postcards both under the sellers realname and under newly created identities.They report that the established seller

23 There may have been some other confounding fac-tors that Resnick et al. were not able to remove with theirexperimental design. In particular, some buyers may havebeen repeat customers of the established seller ID, and

some buyers may even have searched (or even had auto-matic watches) for sales by that seller, and not even lookedat the matched listings from the new buyer. We thank PaulResnick for pointing this out.

received, after correcting for non-sales, 8.1percent higher prices than his newly formedidentity. Ba and Pavlou (2002), on the otherhand, asked a sample of eBay bidders tostate their willingness to pay for eBay auc-

tion listings obtained from the web site, withthe twist that seller feedback characteristicswere manipulated by the experimenters.23

Cabral and Hortasu (2003) exploited achange in eBay feedback reporting policy: asmentioned before, eBay began displayingthe percentage of positives starting in March2003. Cabral and Hortasu (2003) run ahedonic regression where they interact dif-ferent feedback summary statistics with a

dummy variable for the policy change, andshow that the negative correlation betweenprices and the percentage of negatives islarger after the policy change, whereas thecorrelation between prices and total feed-back score, and the sellers age in days, issmaller.

A second limitation of hedonic studies isthat there is very little variation in certainindependent variables, particularly negative

feedback. For instance, in Dan Houser andJohn Wooders (2003), the maximum numberof negative feedbacks in their data set istwelve. Similarly in Melnik and Alm (2002),the maximum number of negatives is thirteen.In the data set analyzed in Bajari and Hortasu(2003), we found that only very large sellerswith hundreds, if not thousands, of positivefeedbacks, had more than a handful of nega-tive feedbacks. Obviously, it is hard to learn

about the value of negative feedback whenthere is not much variation in this variable.

A third limitation is that the objects ana-lyzed in the hedonic studies mentionedabove are fairly inexpensive and standard-ized. Researchers prefer using standardized


16/30

TABLE 1EMPIRICAL STUDIES ON THEVALUE OF REPUTATION ON EBAY

(ADAPTED FROM RESNICK ET AL. 2003)

Source Items Sold Mean Type of Study Covariates ("?" indicates dummy price variable)

Dewan and Hsu,2001

Eaton, 2002

Houser andWooders, 2000

Jin and Kato,2002

Kalyanam andMcIntyre, 2001

Livingston, 2002

Lucking-Reileyet al. 2000

McDonald andSlawson, 2000

Melnik and Alm,2002

Cabral andHortacsu, 2003

Ba and Pavlou,2002

Resnick et al.2003

Cabral andHortacsu, 2003

Collectible stamps

Electric guitars

Pentium chips

Sports tradingcards

Palm Pilot PDAs

Golf clubs

coins

collectible Dolls;new

gold coins in mintcondition

IBM Thinkpadnotebooks, goldcoins, mint coin

sets, Beanie Babies

music CD; modem;Windows Serversoftware CD;digital camcorder

vintage postcards

IBM Thinkpadnotebooks, goldcoins, mint coinsets, Beanie Babies

$37

$1,621

$244

$166

$238

$409

$173

$208

$33

$15-$900

$15- $1200

$13

$15-$900

hedonic regression, compareauction prices on eBay withspecialty site (MichaelRogers,Inc)

hedonic regression

hedonic regression

hedonic regression

hedonic regression

hedonic regression

hedonic regression

hedonic regression

hedonic regression

hedonic regression

lab experiment in the field:subjects responded with trustlevel and willingness to pay forauction listings with differentfeedback profiles spliced in.

field experiment

panel regression usingbackward-looking feedbackprofiles of a cross-section ofseller

book value; internationaltransaction?; number of bids;full set?; close on weekend?;buyer's net rating

guitar type; credit card?; escrow?;pictures?; interaction terms

auction length, credit card, bookvalue, used, processor type

same as (Melnik and Alm, 2002)

product type; picture?number of bids

product type; minimum bid; bookvalue; secret reserve?; credit card?;day and time of auction close;length; inexperienced bidder?

book value; minimum bid;auction length; end on weekend?;

Hidden reserve?month

gold price; closing time and day;auction length; credit cardacceptance; images; shipping charges

dummy variable for eBay formatchange date;closing time and day;auction length; credit card

acceptance; PayPal; images;whether item is refurbished

showed identical listings withdifferent feedback profiles

showed matched information formatched items, in different displayformat

control for seller fixed effects,age in days, reviewer profile


17/30

TABLE 1 (cont.)

Modeling Specification Results regarding market value of reputation

OLS for sold items only;ln(price) vs. ln(positive-negative)

reduced form, logit on probability of sale, OLS onitems sold: positives (number); negatives (dummyor number)

model of bidder distributions based on privatevalues motivates reduced-form GLS (on sold itemsonly), with expectation of more bidders for longerauctions; ln(high-bid) vs. ln(pos), ln(non-positive)

probit predicts probability of sale; OLS for solditems, ln(price/book) vs. ln(net score), has negatives?

OLS for sold items only; price against positives,negatives, interaction terms

probit predicts probability of sale; simultaneousML estimation of price if observed, probability ofsale; reputation specifed by dummies for quartilesof positive feedback, fraction of negative

reduced form, censored normal regression;ln(high-bid) vs. ln(pos), ln(neg)

Simultaneous regression for sold items only: priceagainst bids, reputation; bids against minimum bid,secret reserve, reputation. Various specifications ofreputation

reduced form, censored normal regression;ln(high-bid) vs. ln(pos), ln(neg), ln(neutral)

OLS; ln(highbid) vs. % of negatives, total numberof feedbacks, age of seller in days

OLS, ANOVA of trust, price premium, againstln(positive), ln(negative)

censored normal: ln(ratio of prices) dep variablewith no independent variables

look at the impact of first vs. subsequent negativeson sales growth, where sales levels are proxied bythe number of feedbacks received; look at thetiming between negatives

20 additional seller rating points translate into a 5 cent increase inauction price on eBay. Average bids higher on specialty site, butseller revenues on specialty site (after 15% commission) are samewith revenues on eBay.

No robust statistically significant relation between negativefeedback and probability of sale and price of sold items

10% increase in positive feedback increases price 0.17%; 10%increase in negative feedback reduces it by 0.24%. Increasingpositive comments from 0 to 15 increase price by 5%, or $12.

Positive feedback increases probability of sale; negative decreasesprobability of sale unless card is professionally grade; no significanteffects on price

Seller with 3000 total feedback and 0 negatives gets 12% higherprice than a seller with 10 total feedback and 4 negatives

Sellers with 1 to 25 positive comments receive $21 more thansellers with no reports. Sellers with more than 675 positivecomments receives $46 more. The first 11 good reports increaseprobability of receiving a bid 4%, subsequent reports do notappear to have a statistically significant effect on sale probability.

No statistically significant effect from positive feedback; 1%increase in negative feedback reduces price by 0.11%.

High reputation seller (90th percentile of eBay rating) gets 5%higher price than low reputation (10th percentile) and getsmore bids

Decline of positive ratings from 452 to 1 decreases price by $1.59for $32 items; halving negative feedback from 0.96 comments to0.48 comments increases price by 28 cents.

Effect of % negatives increases after eBay begins to reportpercentages in March 2003 (8% price premium if % negativesdeclines 1 point). Effect of total no. of feedbacks and age of seller

declines after format change. Effect most visible for Thinkpads.

Buyers willing to pay 0.36% more if feedback profile has 1% morepositives, -0.63% less with 1% more negatives (Table 8). Effect islarger for higher priced items.

Seller with 2000 positive comments and 1 negative fetched 8% higherprices for matched items sold by newly created seller identities with10 positives on average. Sale probability of large seller 7% higherthan new sellers. One or two negatives for a relative newcomer hadno statistically detectable effect over other newcomers.

In 4 week window after first negative, sales growth rate is 30% lessthan in 4 week window before negative; second negatives arrivefaster than first negatives


18/30


objects because it is fairly easy to collect bookvalues, which are an important independentvariable in hedonic regressions, for suchitems. Fairly inexpensive items are typicallyused because they are representative of the

objects bought and sold in online markets.It is arguable, however, that the role ofreputation is most important when a buyer isconsidering the purchase of a very expensiveitem of potentially dubious quality. When abuyer is purchasing a new, branded, andstandardized object (e.g., a new Palm Pilot)there is probably little uncertainty about theitem up for sale. When the item is expensive,used, and has uncertain quality, such as a

hard-to-appraise antique, reputation mightplay a more important role. A buyer in suchan auction might be reluctant to bid thou-sands of dollars when the seller has only lim-ited feedback. Such items, however, havetypically not been included in previous stud-ies since it is difficult to construct appropri-ate controls for the value of such an item. Inparticular, book values for such items areprobably not very meaningful.

A final limitation, which applies equally toboth the hedonic regressions and the field-experiment studies, is that it is difficult tointerpret the implicit prices as buyer valua-tions or some other primitive economicobject. In the applied econometric literatureon hedonics, such as Sherwin Rosen (1974)and Dennis Epple (1987), there is a fairly sim-ple mapping from implicit prices to buyerpreferences. The implicit price can typically

be interpreted as the market price for a char-acteristic. Hence the implicit price is equal tothe marginal rate of substitution between thischaracteristic and a composite commodity.This straightforward mapping breaks downwhen there is asymmetric information.

The winning bid in an auction is a compli-cated function of the underlying privateinformation of all of the bidders. Only fairlystylized models of auctions would let the

economist directly interpret the coefficienton feedback as a marginal rate of substitu-tion between reputation and a composite

commodity. In the framework of L. Rezende(2003), if the economist makes the followingassumptions, then the standard hedonicinterpretation of implicit prices is valid:1. There are private values and there is no

asymmetric information among the bid-ders about the marginal value of theobserved product characteristics.

2. There are no minimum bids or reserveprices.

3. All bidders are ex ante symmetric.4. There are no product characteristics

observed by the bidders but not theeconomist.

5. Entry is exogenous and a dummy variable

for the number of bidders is included inthe regression.Clearly, these may be strong assumptions

in many applications. In particular, a largefraction of online auctions uses minimumbids or secret reserve prices. Also, it couldbe argued that uncertainty about quality nat-urally induces a common-value componentinto the auction.

Many of the papers in the literature do not

articulate a primitive set of assumptionsunder which the regression coefficients canbe interpreted as a measure of buyers will-ingness to pay for characteristics (such asreputation) or some other primitive econom-ic parameter. While this limits the generalityof the conclusions, it is nonetheless interest-ing to know the conditional mean of the saleprice as a function of characteristics.

5.2 Other Tests of the TheoryIn addition to measuring the market

price of a reputation, several other empiri-cal observations have been made aboutfeedback mechanisms. For example, Baand Pavlou (2002) report that the impact ofvariation in feedback statistics is largerwhen the value of the object being sold ishigher. This is consistent with economicintuitionthe value of a reputation is more

important for big ticket items. Jin andKato (2002) and Louis Ederington andMichael Dewally (2003) find that for


19/30


collectible objects for which professionalgrading is an option, sale prices of ungrad-ed objects respond more to eBays feedbackstatistics than graded objects. This is con-sistent with our conjecture abovereputa-

tion is more important the less certain thebuyer is about the quality of the item that isfor sale.

Cabral and Hortasu (2003) investigatewhether sellers respond to the feedbackmechanism. They use the feedback profilesof a cross-section of active sellers to con-struct a backward-looking panel data set thattracks the comments received by the sellers.They find that, on average, the number of

positive comments received by a seller untiltheir first negative is much larger than thenumber of positive comments receivedbetween their first and second negatives.They investigate a number of alternativehypotheses for this phenomenon; in particu-lar the possibility that buyers may be reluc-tant (possibly due to altruistic reasons) to bethe first one to tarnish a sellers reputation.They find that buyers who place the first

negative are not, on average, more likely togive negative comments than the buyerswho gave the subsequent negatives.Moreover, they do not find an observabledifference between the textual content offirst vs. subsequent comments.

5.3 Do Reputation Mechanisms Work?

Given the various results in the literature,it is natural to attempt to reach a judgment

as to whether reputation mechanismsachieve their purpose of reducing tradingfrictions on the internet. The robust growthin the number of users and transactions oneBay could be regarded as a testament to thefact that fraud is not perceived as a hugedeterrent in these markets. However, a num-ber of authors express skepticism about theeffectiveness of the feedback systems usedin online auctions. For example, the study by

Jin and Kato (2002), which we discussed insection 4, takes the stance that the observedpatterns of trade cannot be explained by

24 A pioneering attempt in this direction is a laboratoryexperiment by Gary Bolton, Elena Katok, and AxelOckenfels (2003), which shows that while reputationmechanisms induce a substantial improvement in transac-

tional efficiency, they also exhibit a kind of public goodsproblem in that the reporting of honest or dishonestbehavior have external benefits to the community that cannot be internalized by the individual making the report.

rational buyers. The conclusion they reach intheir study is that the prices ungraded cardswere fetching on eBay were higher thancould be rationalized by the frequency offraudulent claims in their graded sample.

They also found that although reputable sell-ers were less likely to make fraudulentclaims and were also less likely to default ordeliver counterfeits, the premium that buy-ers pay for reputation (after correcting forsales probability) is much lower than thepremium that buyers pay for self claims.Hence they conclude that In the currentonline market, at least some buyers drasti-cally underestimate the risk of trading

online and that some buyers have diffi-culty interpreting the signals from seller rep-utation. The study by Resnick et al. (2003)also concludes with the statement that:Nevertheless, it is hardly obvious that thisreputation system would work sufficientlywell to induce reliable seller behavior.

We believe that the jury is still out on theeffectiveness of the reputation systemsimplemented by eBay and other online

auction sites. There is still plenty of work tobe done to understand how market partici-pants utilize the information contained inthe feedback forum system, and whethersome of the seemingly obvious deficienciesof these systems, such as the free-ridingproblem inherent in the harvesting of userreviews and the presence of seller retalia-tion, are large enough to hamper the effec-tiveness of these systems. As in the analysis

of the late-bidding phenomenon, per-haps controlled laboratory experiments canhelp shed more light into how differentcomponents of this complex problem workin isolation from each other.24


20/30


25 See the survey articles by R. Preston McAfee, and

John McMillan (1987), Paul Klemperer (1999), and therecent textbook by Vijay Krishna (2002).

26 For a comprehensive survey of experimental work onauctions, see Kagel and Roth (1995, ch. 7).

6.Auction Design Insights fromInternet Auctions

Perhaps the most central question thatauction theorists try to answer is What kindof an auction should I use to sell my goods?

Since Vickreys (1961) seminal paper, a largebody of theoretical literature has investigat-ed how various informational and strategicfactors in the auction environment affectthe decision of how one should design auc-tions.25 Economic theorists studying auc-tions have also been influential in thedesign of important, high-profile auctionssuch as wireless spectrum auctions, and thedesign of deregulated energy marketsacross the world.

The dramatic development of our theoret-ical understanding of auction theory and theincreasing scope for real-world applicationsalso drive a natural need for empiricalassessment of how these theories perform inthe real world. To this date, most empiricalresearch on auctions has been conducted inlaboratory experiments. Laboratory experi-ments are well-suited for this purpose, sincethe empirical researcher has control over thedesign of auction rules and the flow of infor-mation to participants.26 However, most lab-oratory experiments employ college studentsas their subjects, who are typically not expe-rienced bidders, and are given relativelymodest incentives (by the yardstick of corpo-rate wages). Hence, empirical researchershailing from nonexperimental branches of

economics have frequently questioned theextent to which laboratory experiments canreplicate real-world incentives.

There are important difficulties, however,with the use of data from nonexperimental,real-world auctions for empirical testing oftheories. First, in most field settings of inter-est, it is very costly to conduct controlled, 27 According to Krishna (2002), page 4, footnote 2:

Two games are strategically equivalent if they have thesame normal form except for duplicate strategies. Roughlythis means that for each strategy in one game, a player has

a strategy in the other game, which results in the same out-comes. In this context, strategic equivalence means thatfixing the valuation constant, bids should be the sameacross the two auctions.

randomized trials; hence, statistical infer-ences based on nonexperimental data aretypically subject to much stronger assump-tions than inferences from experimentaldata. Second, detailed transactions data

from real markets is typically not easy to get,since such data is sensitive and confidentialinformation in many industries.

Internet auctions provide a very interest-ing and promising middle ground betweencontrolled laboratory experiments conduct-ed with student subject pools and large-scalefield applications such as wireless spectrumauctions or energy auctions. First, data onbids are readily available from online auction

sites, lowering this important barrier-to-entry for many empirical researchers.Second, as discussed in section 3, there aredifferences across the auction rules utilizedacross different auction sites or even withinan auction site, and these differences can beexploited to empirically assess whether onemechanism works better than another.Third, empirical researchers can actuallyconduct their own auctions on these sites,

which allows them to run randomized fieldexperiments with experienced, real-worldsubject pools.

We now survey several examples of howthe study of internet auctions has providednovel and potentially useful insightsregarding economic theories of auctiondesign.

6.1 Field Experiments on the Internet

In a set of field experiments, DavidLucking-Reiley (1999a) tests two basichypotheses derived from auction theory: 1)the strategic equivalence27 between theopen descending price (Dutch) and sealed-bid first-price auction, and 2) the strategic


21/30


28 Briefly, the auction formats are the following: in thesealed-bid first-price auction, the highest bidder wins theauction and pays the price bid. In the sealed-bid second-price auction, the highest bidder wins, but pays the sec-ond-highest bid. In the Dutch auction, price decreasescontinuously from a high starting value, and the first bid-der to claim the object wins the auction and pays the priceat the instant she called in. In the English auction consid-ered here, price rises continuously, and all bidders indicate(by pressing a button, for example) whether they are still

in the auction or not. The auction ends when the second-to-last bidder drops out, and the winner pays the price atwhich this final drop-out occurs. Bidders cannot rejoin theauction once they have dropped out.

equivalence between second-price andEnglish auctions.28 As discussed in PaulMilgrom and Robert Weber (1982), thestrategic equivalence between Dutch andsealed-bid first-price auctions follows quite

generally, since the only piece of informa-tion available to bidders in a Dutch auctionthat is not available to bidders in a sealed-bid first-price auction is whether a bidderhas claimed the object. But this informa-tion cannot be incorporated into biddingstrategies, since its revelation causes theauction to end.

On the other hand, the strategic equiva-lence between second-price and English

auctions only applies in a private-valuesenvironment, where bidders would notmodify their personal willingness to pay forthe object even if they knew how muchother bidders were willing to pay. It can beshown that with private values, biddingones value, or staying in the auction untilthe price reaches ones valuation is a (weak-ly) dominant strategy in both the second-price auction and the English auction. In

an interdependent-values or common-value environment, however, seeing anoth-er bidders willingness to pay may affect theinference one makes regarding the value ofthe object, and hence changes their willing-ness to pay. As shown by Milgrom andWeber (1982), bidding strategies may bevery different in an English auction inwhich drop-out points of rival bidders arerevealed, as opposed to sealed-bid second-

price auctions, where bidders do not

29 This finding does have an alternative explanationin that bidders have decreasing marginal valuations forthe cards.

observe each others actions, and hencecannot update how much they are willing-to-pay for the object during the course ofthe auction.

Lucking-Reileys experiment used the

above auction formats to sell trading cardsfor the role-playing game, Magic: TheGathering, on an internet newsgroup thatwas organized as an online marketplace forenthusiasts in pre-eBay days. Lucking-Reiley invited participants to his auctionsusing e-mail invitations and postings on thenewsgroup. To minimize differences acrossparticipants distribution of values for theauctioned cards, Lucking-Reiley used a

matched-pair design. For example, in thecomparison of first-price and Dutch auc-tions, he first auctioned a set of cards usingthe first-price auction; a few days after hisfirst set of auctions ended, he sold an identi-cal set of cards using a Dutch auction. Toaccount for temporal differences in biddersdemand for these cards, Lucking-Reileyrepeated the paired experiment about fourmonths later, but this time selling the first

set using Dutch auctions, and the second setusing first-price auctions.

Lucking-Reiley found that Dutch auc-tions yielded 30-percent higher average rev-enue than the first-price auctions, andrejected the strategic equivalence of the twoauction formats. On the other hand, in thesecond-price vs. English auction experi-ment, he found that the auction formatsyielded statistically similar revenues.

However, he found less convincing evidencefor the hypothesis that the two auctionswere strategically equivalent. In particular,he found that in 81 out of 231 cases wherethe same bidder placed a bid for the samecard on both an English auction and a sec-ond-price auction, the bidders highestEnglish-auction bid exceeded their bid onthe second-price auction.29


22/30


30 Lucking-Reiley also discusses whether the speed atwhich prices decline in Dutch auctions may have led to adifference between his results and laboratory results; inhis field experiment, prices declined much slower than inthe laboratory. Interestingly, Elena Katok and AnthonyKwasnica (2000) find, in a laboratory setting, that slower

price declines in Dutch auctions led to increased rev-enues as compared to sealed-bid auctions, suggesting thatbidder impatience may have played a role in the fieldexperiment setting.

We should note that previous studiesusing laboratory experiments, by VickiCoppinger, Vernon Smith, and Jon Titus(1980) and James Cox, Bruce Roberson,andVernon Smith (1982) also rejected the

strategic equivalence of first-price auctionswith descending auctions. However, theseexperiments reported higher prices in first-price auctions than Dutch auctions, withDutch auctions yielding 5-percent lowerrevenues on average. Similarly, John Kagel,Ronald Harstad, and Dan Levin (1987)report the failure of strategic equivalenceof the second-price and English auctions inthe laboratory setting. In particular, they

report a tendency for subjects to bid abovetheir valuations in the second-price auc-tion, whereas they rapidly converge to bid-ding their values in the English auction,yielding 11 percent higher revenue for thesecond-price auction.

What explains the differences acrossLucking-Reileys results and results fromlaboratory experiments? First of all, asLucking-Reiley notes, his field experiment

cannot control for the entry decisions of thebidders; he reports that Dutch auctionsattracted almost double the number of bid-ders as his first-price auctions, and it is nothard to see that an increase in the number ofbidders would increase revenues. The caus-es of the higher participation in the Dutchauction remains a mystery: Lucking-Reileyargues that this is not solely due to the nov-elty of the Dutch auction mechanism, since

market participants had been exposed toDutch auctions before. One wonders, how-ever, whether a taste for novelty effect maypersevere over longer periods of time.30

Second, Lucking-Reileys experiment can-not control for the informational structure ofthe auction. In the laboratory, the researchercan choose whether to run a common/inter-dependent value or a private-value auction;

however, Lucking-Reiley cannot predeter-mine whether bidders will regard his tradingcards as private-value or common-valueobjects. This may affect the interpretation ofhis second-price vs. ascending auctionresults: in a common-value environment, thesecond-price auction is predicted to yieldlower average revenues. This factor, com-pounded with the experimentally reportedbias of the bidders to overbid, may result in

the observed revenue equivalence result.On the other hand, one may think that theonline marketplace utilized by Lucking-Reiley is populated by veterans of previousauctions, who are experienced enough not tooverbid in a second-price auctionhencethe theoretical prediction of revenue equiv-alence is more likely to be borne out in thefield than in the laboratory. Lucking-Reileybriefly mentions this possibility; though

more direct evidence comes from else-where. In a recent paper, Rod Garratt, MarkWalker, and John Wooders (2002) invitedexperienced bidders on eBay to take part insecond-price auction experiments conduct-ed in a laboratory setup. In contrast to previ-ous experimental findings, Garratt, Walker,and Wooders found that bidders experi-enced on eBay do not overbid in second-price private-value auctions, and, very often,

give the correct reasoning to their action.Hence, field-experiments conducted ononline auction sites may indeed provide amore qualified subject pool for auctionexperiments.

Another argument for the use of fieldexperiment methodology, as pointed out byLucking-Reiley (1999b), is that in practicalsettings where a seller is trying to assess theproper auction mechanism to use, controls

for endogenous entry decisions and on theinformational environment will typically notbe present. For example, the fact that a


23/30


31 Lucking-Reiley quotes Robert Hansens (1986) find-ing of systematic differences in timber lots sold usingsealed-bid vs. ascending auction by the U.S. Forest

Service.32 Lucking-Reiley notes that the marketplace was activeenough that his 80 auctions would not have a market-wide impact.

Dutch auction may lead to higher bidderparticipation may be an outcome variable ofinterest to a seller considering switching tothis auction format; especially since it leadsto higher revenues. Hence, Lucking-Reiley

argues that for practical or policy applica-tions, results of field experiments may pro-vide more insights regarding the outcome ofa possible policy change.

The field experiment methodology alsohas advantages over nonexperimentalanalyses of outcomes across different auc-tion formats. As Lucking-Reiley notes, anonexperimental analysis of whether a first-price or a second-price auction yields higher

revenues may have to worry about why someauctions were conducted using a first-priceauction, and others using a second-priceauction.31 On the other hand, we shouldnote that the field experimentation method-ology, as used by Lucking-Reiley, can onlyshed light onto partial-equilibrium respons-es to changes in mechanism design. Forexample, his finding that Dutch auctionsyield 30-percent higher revenues than first-

price auctions may not apply if every selleron an auction site decides to use a Dutchauction as opposed to a first-price auction.32

Whatever its pros and cons, Lucking-Reileys paper is a good example of howonline auction sites can be used as a labora-tory for creative field experiments. Last, butnot least, we should point out that a veryimportant advantage of this field labora-tory is that it was relatively inexpensive to

use: Lucking-Reileys experiment had an ini-tial cost of $2,000 to buy about 400 tradingcards, which he claims to have recoupedwith a profit (in addition to a published the-sis chapter) after selling them. Moreover, atthe cost of some lack of customizability,

online auction sites make it easy for theresearcher to set up and track experiments;instead of requiring them to spend a signifi-cant amount of programming time to writesoftware for their own experiments.

6.2 Should Reserve Prices Be Kept Secret?

Auction design is not only about choosingbetween formats such as first-price, second-price, English or Dutch. As first shown byRoger Myerson (1981), a seller may signifi-cantly increase their revenues by optimallysetting a publicly observable reserve price,or, equivalently, a minimum bid.

A look at sellers practices of setting reserve

prices on eBay reveals an interesting regular-ity: as noted by Bajari and Hortasu (2003),many sellers on eBay, especially those sellinghigher-value items, choose to keep theirreserve prices secret, as opposed to publiclyannouncing it. This immediately brings upthe question: When, if ever, should one use asecret reserve price, as opposed to a publiclyobservable minimum bid?

Auction theory has been relatively silent

regarding this question, with two notableexceptions. Li and Tan (2000) show that withrisk-aversion, secret reserve prices mayincrease the auctioneers revenue in an inde-pendent private value first-price auction, butin second-price and English auctions, regard-less of the bidders risk preferences, the auc-tioneer should be indifferent between settinga secret vs. observable reserve price. DanielVincent (1995) provides an example in which

setting a secret reserve price in an interde-pendent value second-price auction canincrease the auctioneers

Documents

Bajari Hortacsu JEL2004.PDF (IBPS Term-4)