SELLER REPUTATION AS A DETERMINANT OF PRICE IN ONLINE AUCTIONS

SELLER REPUTATION AS A DETERMINANT OF PRICE IN ONLINE AUCTIONS: THEORY AND EVIDENCE FROM GIFT CARD SALES Jennifer Pate Department of Economics Loyola Marymount University 1 LMU Drive, Suite 4200 Los Angeles, CA 90045 jennifer.pate@lmu.edu October 2006 This study examines the impact of a seller’s reputation on sale price in online auctions. A model of bidding behavior demonstrates how a faulty reputation signal, in terms of its ability to convey information to potential bidders, may cause the bidder with the highest valuation of the item to lose the auction to a bidder with a lower valuation. Data from 2,002 auctions of gift cards provide evidence that the reputation component, as currently formulated by eBay, has a very small, though significant, impact on price. The lack of information contained in the reputation score may be responsible for its ineffectiveness, and several possibilities for improvement are presented. JEL Classification Codes: D44, L86, C51 Keywords: Internet Auctions, Reputation, Information I. Introduction The relatively recent proliferation of online auctions, namely the success of the largest internet auction website, eBay, has generated an abundance of economic research. Whether studying the act of bidding in the last few seconds before an auction closes, or “sniping” as it is referred to in Ockenfels and Roth (2002), or examining the issue of dishonest sellers bidding on their own auctions to increase the sale price, the availability of data and a genuine interest in understanding buyer and seller behavior has led to a wealth of new research. One extensively studied topic regarding internet auctions has been the concept of seller reputation. The study presented here empirically examines the role of reputation in sale prices using data from 2,002 completed eBay transactions. In online auctions, there exists a positive probability that the seller will default on the transaction after receiving payment by not sending the item. Other possible risks may be that the seller is misrepresenting the quality of the good, or fails to ship the item within a reasonable threshold of time. The anonymity of sellers, given that they use eBay “member” names which are generally not their own, creates a separation between the sellers actions and themselves that may increase the risk of default or other undesirable behavior. In an attempt to decrease these risks, eBay instituted a “feedback” system that allows buyers to send and receive information about a specific seller’s actions. After each transaction, both parties may leave positive, negative or neutral feedback, along with a brief explanation. For each eBay member, their feedbacks are totaled up according to a point system to create a member’s reputation score. Positive feedback is worth one point each, neutral feedback causes no change, and a negative feedback 2 reduces the reputation score of the member by one point. Therefore, a reputation score also allows for information about experience, as higher scores also indicate a greater number of completed transactions. In sum, this reputation measure should help buyers to properly account for the likelihood of seller default and allow them to adjust their bids accordingly. Previous research has analyzed the value of reputation in internet auctions by collecting information about sale characteristics of specific items, like coins in LuckingReiley et al (2000) or Harley-Davidson Barbie dolls in McDonald and Slawson (2002). The results have varied widely, as some studies found significant reputation effects on prices, where others found none. One explanation for the observed variance in results is that all research to date shares one fundamental flaw: object heterogeneity. Although most studies attempted to control for the differences in items across auctions, none were entirely successful. Heterogeneity across auctions can introduce a significant omitted variable bias in the empirical analysis. The study presented here is the first to control for the common value of the good to buyers, by examining auctions of transferable store credit; sales of gift cards. Since the value of the good is clearly indicated to buyers and is common knowledge across all bidders, it can be controlled for in the analysis. By including this term, the effect of seller reputation on sale price, along with many other sale characteristics, can be clearly measured. The results show that reputation has a positive and significant, although very minor effect on price, and that the commodity value of the good to the bidders explains over 99% of the sale price. 3 This article is organized as follows. Section II provides a more thorough discussion of the “feedback” component on eBay, while Section III discusses previous findings. Section IV describes a model to examine the importance of an effective reputation measure. Section V describes the data, and the empirical results are reported in Section VI. Section VII concludes. II. The eBay Feedback System In an online marketplace where buyers and sellers barely interact and often times only share one transaction in a lifetime, the role of establishing a reputation component is vital. eBay does this using a “feedback” system where buyers and sellers rate their trading partners regarding their eBay transaction. Members may leave positive (+1), negative (-1), or neutral feedback (0), along with a short explanation. Figure 1 shows an example of an eBay auction for a $50 GAP card, which displays the seller information for pinkync2003. Here, the seller has a reputation score of 37, consisting of 100% positive feedback. Potential bidders may click on a seller’s feedback number to receive additional information, such as how long the seller has been registered on eBay, or to read the feedback comments. All feedback is a public service, meaning that participation by all members is entirely voluntary and not rewarded monetarily or otherwise. The feedback system allows potential bidders to examine the reputation of a seller before bidding on an item. Although a significant amount of negative feedback on a seller’s record should be cause for concern, it is not necessarily true that this seller would or should be entirely avoided. If the price of the item being sold is low enough to offset 4 the additional risk assumed by buying from the seller, then the transaction should still take place. Buyers should observe the auction characteristics including a seller’s reputation, then calculate the risks and bid accordingly. Do buyers actually account for the reputation score when bidding on an item? According to a survey of 1,000 eBay members, only 16% of users said that they would be less likely to do business with low-feedback sellers (Steiner, 2003). Given that only around 1% of all feedback on eBay is negative, this answer is not surprising.1 Besides the measures that eBay has in place to discourage non-positive feedback, there are other associated problems, such as the issue of “retaliatory” feedback, where users refuse to leave negative feedback in fear that the other user responds in a similar style. This threat helps to limit the amount of negative feedback on eBay, which in turn hurts the effectiveness of a reputation measure as a tool in guiding bidder behavior. The goal of this study is to measure the effect of seller reputation on sale price, and to use this information to test whether the current feedback system serves its purpose in relaying a seller’s risk of default to potential bidders. III. Related Literature A number of studies have empirically examined eBay’s feedback system.2 McDonald and Slawson (2002) collected information from completed auctions of HarleyDavidson collectible Barbie dolls and found that a 1-point increase in a seller’s reputation score increased sale price by $0.04. This finding of a small, positive, and significant The approximation that negative feedback occurs in only 1% of all transactions was calculated using the sample set, and may not be exact across all transactions of all items on eBay. 2 For a complete survey of the literature, see Resnick, Zeckhauser, Swanson, and Lockwood (2003). 1 5 coefficient on reputation is supported by many other studies, including Houser and Wooders (2006) and Dewan and Hsu (2004). Other research has examined panel data to test how negative feedback affects the number of future transactions, additional seller feedback reports, and seller exit. Cabral and Hortaçsu (2006) find that after a seller receives their first negative feedback, the probability of subsequent negative feedback increases, as does the probability of exit. They also show that sellers experience a decrease in their growth of sales after receiving negative feedback. Though many earlier studies have examined this topic, they have all faced the same difficulties in controlling for item heterogeneity. It is unclear whether the reputation coefficient is significant simply because other factors have been omitted, or whether the coefficient should be larger or smaller for the same reason. Studies that focus on sales of heterogeneous goods like Melnik and Alm (2003), who examine various conditions of seller-graded coins in an attempt to control for the quality of the item and its value to the bidders, still face issues with heterogeneity across the same grades of coins. Although the coins are graded by the sellers according to a “standardized and generally accepted” scale, this leaves room for seller interpretation that is unobservable. This situation itself may increase the value of a good reputation, however if the value of the item to the bidder is uncertain, then it will be difficult to know whether the estimate for the reputation coefficient is measuring the effect of reputation, or a combination of reputation and an overpriced book value of a coin. The unique contribution of this study is the data employed in analyzing the determinants of sale price in auctions. By examining items that are similar to collectible 6 coins in that there are different varieties, but where the quality of the item is always certain and known ex-ante to bidders, the obstacle of object heterogeneity is avoided by examining auctions for store credit.3 Although this study will focus primarily on the context of a seller fulfilling their agreement with the winning bidder, the possibility of a seller misrepresenting the value of the store credit in the item description could also be another interpretation of the risk of default. IV. The Model Modeling bidder behavior subject to auction characteristics in online auctions has been undertaken by previous researchers. However, the model presented here has several critical differences that allow for a more intuitive examination of the empirical data, producing more informative results and practical implications for the eBay feedback system. Following a model by Houser and Wooders (2006), the basic set-up here is similar to a buyer purchasing a contract. A single seller puts up for auction one indivisible unit of a good, with cost normalized to zero. A set of n potential bidders are interested in buying a contract from the seller to deliver the good being auctioned. In every auction there is a common value component equal to V>0, which is the commodity value of the good. Bidders also have a private value component of the good equal to their cash equivalent value, where each bidder i’s privately known value of the good is vi = αiV, αi i.i.d. є [0,1]. 3 For a discussion of the market for gift cards and their welfare implications, see Offenberg (forthcoming). 7 In each auction, there is a positive probability that the seller will default on the contract and not deliver the good after payment has been received. Therefore, the probability of a successful sale, where the seller completes the transaction, pS є [0,1], is incorporated into the model.4 First, x = (x1, x2,…xM) is a positive real vector of common information available in the auction, restricted to values between zero and 1. Each x corresponds to the information signal from a particular observable characteristic, such as the seller’s reputation score, percentage of positive feedbacks, or promotional choices. An xl near zero can be interpreted as the lth observation having a negative impact on the probability of a successful sale. For example, having numerous negative feedbacks would make the value of x move closer to zero, while a perfect feedback rating may suggest to the bidder that the seller is more likely to complete the transaction, and x would be closer to 1. Bidders weight the available information according to the vector λ = (λ1, λ2,...λM). Each weighted term λk is related to the information content in xk. The probability of a successful sale is evaluated by bidders according to an equation of M observable auction characteristics, where p S = ∑ λl xl where 1 M ∑λ 1 M l =1. Auction characteristics providing higher quality information are weighted more heavily in the probability decision, whereas observations that provide weaker signals receive a smaller weight in the calculation of pS.5 It is initially assumed that the values of It would be straightforward to include other possible problems with auctions, such as the item arriving but either not as described in the auction or broken in transport. 5 The decision to model the informational content of the auction characteristics and use weighted terms, as well as the intuitive results due to this specification, is unique to the model presented here. 4 8 λ are common across all bidders, as are the observable auction characteristics.6 Since, at this time, all components of the probability of a successful sale are common across bidders, pS is also common across bidders. The value of the contract to a bidder i if he wins and pays b is U (ci ) = p S ⋅ u (vi − b) + (1 − p S ) ⋅ u (−b) . If we assume bidders are risk neutral, then utility takes the form U ( x) = x and the expected value of winning the contract is p S vi − b . All non-winning bidders receive zero utility.7 Given the value of the contract to bidder i, he chooses a series of proxy bids, and bidder i has the highest auction bid when he has the highest proxy bid bi ≥ b j ∀j ≠ i . If bidder i has the highest bid when the auction ends then he wins and pays the winning bid, equal to the 2nd highest proxy bid, max k ≠i bk . Thus, the winning bid paid is bi = max k ≠i bk . We also need a condition satisfied where no other bidder j, j ≠ i , wishes to bid above the current highest bid and fails. That is, the winning bid for bidder i must be more than any other bidder j’s expected value of winning, p S v j , requiring bi = max k ≠i bk ≥ p Sα j V ∀j ≠ i . It is important to consider two issues here. First, there is a positive probability that more than one bidder values the good at a certain v, in which case there would need to be considerations for the event of a tie. The eBay procedure for this circumstance is For empirical studies of internet auctions, it is important that bidders’ observations of auction characteristics are common across all bidders, as well as to the researcher. 7 The reputation of the bidder is ignored since bidders with bad reputations can still bid on and win any item, therefore the bidder reputation will not affect the sale price. 6 9 that the first person to place the bid has priority, that is, auctions bids are first come, first served. This alleviates the need to take into account tying outcomes in the model. Also, there is a small bid increment that must be satisfied in order to place a new high bid, where if bidder j wishes to outbid the current high bid of bidder i, they must bid b j = bi + ε . The bid increment is established by eBay according to the magnitude of the current high bid. In order to simplify the equilibrium conditions, the bid increment is ignored here, although it would be straightforward to model. Finally, we need to specify a no-default rule, where a bidder will not choose to bid above their expected value of winning. This establishes that any winning bidder will win and prefer to default, rather than complete the transaction. This requires that b j ≤ p S v j ∀j . Similar to Houser and Wooders, equilibrium is established where, if bidder i* wins the auction, the following conditions must hold: (1) bi* = max k ≠i bk (2) max k ≠i bk ≥ p Sα j V ∀j ≠ i (3) b j ≤ p S v j ∀j In equilibrium, it must be that no bidder wishing to place the highest bid, given their expected value of the auction, loses, and that no bidder valuing the item less than the highest bid wins. This establishes that every bidder is best off bidding their expected 10 value of winning the contract as their highest proxy bid.8 It is also important to note here that the value of the highest bid will not be equal to the commodity value of the item being auctioned unless pS = 1 and αi = 1. If all bidders bid their expected value of winning the contract, then it is straightforward to demonstrate that the bidder with the highest α draw wins the auction. For example, if 2 bidders exist with α1 > α2, then v1 > v2. If they bid their expected values, then b1 = p S v1 > p S v2 = b2 and bidder 1 wins and pays pSv2.9 The previous result is efficient in that the bidder with the highest cash equivalence of the commodity good V always wins. However, it may be more reasonable to relax the assumption that λi = λ ∀i ∈1,..., n and instead allow the weights of the observed auction characteristics to vary across bidders.10 The probability of a successful sale now becomes piS = ∑ λil xl where 1 M ∑λ 1 M il = 1. It is not necessarily the case that bidders always make a single proxy bid equal to their expected value of winning the auction. Although eBay recommends this as the “best” bidding method, bidders may instead place multiple bids over a period of time as the auction continues and may not need to bid as high as their expected value of winning in order to win. This equilibrium simply establishes the highest proxy bid that bidders would be willing to place. 9 The possibility of “sniping” or placing last second bids has been discussed by Roth and Ockenfels (2002) as a method for bidders to pay less than their and possibly other bidders’ expected values of winning the auction and still win. This is the reasoning behind eBay’s “bid your highest valuation” advice, so that a bidder with a lower valuation of the item will be unable to “snipe” the auction and win. Sniping is appropriately not predicted by this model for two reasons. First, last second bidding was not observed in the data for store credit auctions and, second, Roth and Ockenfels found that sniping was more likely “in eBay auctions in which expertise plays a role in appraising values” (2002, p. 1101), which is clearly not the case here. 10 Evidence from surveys of eBay members supports relaxing this assumption. For example, when asked about the quality of the eBay reputation score, 29% of respondents claimed it was fair or poor, 35% believed it to be just adequate, while 36% rated it above average. The variance across answers suggests that different bidders value the reputation score differently (Steiner, 2003). 8 11 Allowing bidders to value information from observable auction characteristics differently directly affects pS by making the probability of a successful sale a private value, equal to piS . Given that the equilibrium conditions established earlier still hold, allowing pS to vary across bidders introduces an interesting and likely possibility; that the bidder with the highest cash value of the good may not always win the auction. For instance, take the following numerical example: The auction is for a store credit of $100 and two bidders have values α1 = 0.9 and α2 = 0.85, which correspond to private values v1 = 90 and v2 = 85. Now, if bidder 1 is more critical of a certain component of pS than bidder 2, all else S equal, then the probabilities may take the form p1S = 0.8 and p2 = 0.86 . If both bidders bid their expected value of winning the auction, then b1 = 72 and b2 = 73.10. In this case, bidder 2 wins the auction and pays $72, even though bidder 1 had the higher valuation of the good. The occurrence of this outcome will be more frequent the greater the variance of piS across bidders, which is directly related to the quality of the common information contained in the auction. If auction characteristics are less informative, then the weights may vary greatly, as will piS . However, if the informational content was improved, then the weights would become more similar across bidders, decreasing the variance of the informative quality of the signal and reducing the likelihood of a bidder with the secondhighest cash valuation or lower of winning the item. The quality of information in different auction characteristics is examined empirically in later sections to assess the validity of this issue. 12 V. Data The data consist of information from 2,002 completed eBay transactions of transferable store credit collected from February 18, 2004 to June 11, 2004. The benefit to examining auctions selling store credit, commonly referred to as gift cards, is that the commodity value of the item to buyers is common knowledge. For example, take the sale of a $50 Gap card, as seen in Figure 1. The value of the item being sold is clearly $50, without any uncertainty regarding the quality of the good, which is a common problem in prior research as discussed earlier.11 Also notice that the shipping for this particular item is free, which is another benefit to examining store credit auctions. Prior studies have had problems with sellers not specifying the cost of shipping in their auction listings and were forced to either enter approximations or drop observations. With the median shipping cost across the 2,002 auctions for store credit equal to zero, the effect of shipping and handling on sale price is almost completely controlled for prior to regression analysis and therefore avoids this issue. Figure 1 displays the source of most of the auction characteristics collected for this study as it appears on eBay. The information recorded from each auction appears in Table 1, along with summary statistics for each term.12 ‘Promotions’ refers to the use of bold or subtitles in the listing of the auction, which is the component that buyers see and click on to arrive at the auction page like that of Figure 1. Sellers must incur an This study does not assume that the cash value of a $50 GAP card is equal across all bidders, and therefore this is not a perfect measure of individual bidder value, while the common component necessary for the test is the amount of store credit stored on the card. Also, since the cash equivalence of store credit may vary across types of stores, it is controlled for in the analysis. 12 The counts for each variable are not equal, since information regarding Starting Bid and Reputation Percentage was not collected until later in the sample. 11 13 additional fee to use these promotions, usually a flat fee $1.00 for bold and $0.50 for listing a subtitle.13 The ‘Commodity Value’ is the amount of commodity credit being held on the card, and ‘Sale Price’ is the amount that the highest bidder must pay to the seller to receive the item, in addition to the shipping fees if applicable. Number of bids, Starting bid, and Shipping cost are all straightforward variables. Ending time is the time the auction closes, recorded in a 24-hour format. The ‘Buy-It-Now’ option is another feature of eBay auctions, allowing sellers to set a price that they would be willing to accept to end the auction prior to the pre-designated closing time. If the transaction ended early because a buyer chose to ‘Buy-it-now,’ then this variable takes a value of 1. Reputation Score is the total number of positive feedbacks a seller has received, after subtracting the number of negative feedbacks. As can be seen in Figure 1, the exact amount of each type of feedback is unknown, unless buyers take an additional step in examining the seller’s Feedback Profile, by clicking on the reputation score. However, buyers can see the Reputation percentage, or percentage of positive feedbacks in the seller’s reputation. If sellers accepted only one form of payment, generally Paypal, then the ‘Payment Restrictions’ variable takes the value of 1. Paypal is an eBay company that assists transactions by allowing for buyers to pay by credit card or direct bank transfer and for sellers to accept these funds. Though paying by Paypal is faster than mailing a money order or waiting for a personal check to clear, the registration process can be inconvenient and time consuming. Paypal requires new members to allow a small charge Occasionally eBay has temporary subtitle sales where any seller may list their item with a subtitle for only $0.01. 13 14 to their credit card and then new members have to wait for their next credit card bill to receive the confirmation code, which must be entered into the Paypal account in order for it to become active. A similar process must be followed to set up bank account transfers. If the auction closed on a weekend, defined as the time between midnight on Friday and midnight on Sunday, then that information was also recorded, and the variable takes the value of 1. A dummy variable for auctions ending between midnight and 4am, called “Late hour” and monthly dummies were also included in the analysis (the dummy for June was dropped).14 VI. Empirical Results Results from several different specifications of the aforementioned variables appear in Tables 2 and 3, all following the general regression formula ln( saleprice) = c + β (ln(truevalue)) + γ (reputation) + φ (other demand factors ) + ε .15 The “other demand factors” affect either the interpretation coefficients (λ) or the cash equivalence distribution of α, or both, depending on the particular variable. For example, choosing to use a promotional tool draws attention to the auction, which would affect the distribution of αi by increasing the bidding pool. This may also affect how bidders interpret the auction characteristics, by implying that the seller is willing to put in extra effort and therefore may be less likely to default. The ‘Late hour’ variable was calculated according to Indiana Central Standard time zone as a benchmark. This is not specified exactly according to the model, due to the equilibrium solution where the sale price is actually the second highest proxy bid. There is not full disclosure of information regarding the highest proxy bid in each auction, making this specification a next best alternative. 15 14 15 All standard errors reported are robust. The reputation measure was not found to be highly correlated with any of the other demand factors.16 The variable for starting bid was found to be highly correlated with the variable for number of bids, and was excluded from the regression here, as well as in previous studies (i.e. McDonald and Slawson, 2002).17 The regressions also include dummy variables to control for the type of store credit being auctioned, but the discussion of this secondary market for gift cards is left to Offenberg (forthcoming). Table 2 presents results from two regressions. Although each regression examines a slightly different set of variables, there are a few important general findings that occur throughout. First, the true value of the item, equal to the commodity value adjusted for the cost of shipping and handling, accounts for over 99% of the sale price, and is consistently significant at the 1% level regardless of the specification. This discovery is entirely unique to this study. This finding, along with the very small, though positive and significant coefficient for the seller’s reputation score and percent of positive feedback, indicate that either nearly 100% of all sellers on eBay are trustworthy, or that bidders do not place a high value on the reputation score, suggesting that the eBay feedback system does not provide quality information to bidders.18 If approximately 2.2% of all feedback is neutral or negative, with the median seller having at least one negative feedback comment as found by Cabral and Hortaçsu (2004), the data suggest that far less than 100% of eBay sellers are trustworthy. 16 The highest correlation of reputation to any of the demand factors was to the buy-it-now option at 0.107. Sellers, who are aware of their feedback score, do not tend to alter their behavior regarding other sellerspecified auction characteristics depending on their reputation. 17 The remaining independent variables were not found to be correlated. 18 It’s also important to note here that the reputation coefficient estimated while including the effect of the value of the item is much smaller than that found in prior research not controlling for the value of the good. 16 The current measure of seller reputation according to eBay’s feedback system has a small and decreasing effect on price, according to the negative and significant coefficient on the variable reputation squared, implying that the system itself may be flawed. Across all observations, an increase the reputation score from the 25th percentile (38 feedbacks) to the 75th percentile (365 feedbacks) increases sale price by just $0.08. With problems like retaliatory feedback and voluntary reporting, it is not surprising that the absence of negative feedback has reduced the quality of the reputation measure. This finding also lends credibility to the model by suggesting that the informational content of the observed characteristics should be taken into account when examining bidding behavior. The second set of regressions in Table 3 segment the data by commodity value in order to avoid forcing the coefficients to be equal across different ranges of values. This is a unique ability of this study, since previous data has not contained the value of the item and could therefore not examine the effect of the independent variables at different price levels. The first two regressions in columns 1 and 2 divide the data into commodity values of less than $100 and values of $100 and above. Interestingly enough, the coefficient for seller reputation score is not much larger for greater values of commodity money. For auctions of less than $100 of store credit, an increase in seller feedback from 10 to 100 increases sale price by $0.07 (approximately 0.14% of the average gift card value of $50.46), while for auctions of over $100, the same increase in reputation increases sale price by $0.12 (just 0.04% relative to the average gift card value of 17 $310.17), all else equal.19 The final two specifications appearing in columns 3 and 4 of Table 3 present results from the two extremes of the data, for auctions of less than $50 of store credit and for over $200 of store credit, to further test for any effects of magnitude. In auctions for store credit valued at over $200, where reputation should become increasingly important as the stakes rise, a 1000% increase in seller feedback increases the sale price by only $0.30. If the measure was informative, then we would expect the importance of a positive seller reputation to increase as the amount of money on the line increases, but that does not seem to be the case here. Although many of the coefficients across the outermost values of store credit appear to be equivalent, such as the importance of the true value of the store credit, buyit-now option, and payment restrictions, there was a large difference in the importance of number of bids. At higher monetary levels, the number of bids is irrelevant, while the coefficient for the number of bids is positive and significant at the 1% level for auctions of less than $50 of store credit. This may be related to the finding of a decreasing value of store credit. In particular, if at high store credit values (like $300), many bidders have very little cash value for the credit as a percentage of its commodity value, they may decide not bid, whereas almost all bidders could find a use for a $10 store credit and they all place their proxy bids accordingly. This translates to reasoning that there may be more bidders in the market for a $10 store credit than a $300 store credit, and hence more bids. 19 Previous research by McDonald and Slawson (2002) reported that a 1-point increase in seller reputation would increase the sale price by $0.04, which is significantly larger than the estimate found in this study. After controlling for the value of the item, the estimate falls to a $0.003 increase in price for the same change in seller reputation under similar conditions. 18 There are also a few notable general findings that have a negative effect on sale price. If the auction ended in the month of February, payment was restricted to Paypal only, or if the auction was ended early using the buy-it-now option, the sale price was negatively impacted. The explanation for a decrease in sale price due to a restriction in the type of payment accepted may be as simple as a decrease in demand. Although Paypal may be more convenient for sellers than accepting other forms of payment, it is likely that some potential bidders choose not to endure the Paypal certification process, which may decrease the bidding pool, possibly excluding the bidder with the highest cash equivalent valuation of the item. The negative effect attributed to the buy-it-now option may be interpreted several different ways. It could be that sellers of store credit underestimate the cash value of their item to the bidding pool. The support for this reason is simply the fact that the seller put the item up for auction in the first place, suggesting that the seller may have a low cash equivalence for the store credit, and therefore may not be able to properly deduce the appropriate buy-it-now price to set. The negative effect may also be attributed to a time value of money argument, where the seller benefits from setting a low buy-it-now price in order to have the auction end sooner and receive payment faster. Of course, the seller could also initially establish the auction as a one day auction to accomplish the same goal. For whatever reason, it is interesting to document this finding and also to recall that only 18% of the observed auctions ended with buy-it-now, possibly indicating that sellers may realize this negative effect on price and limit its use. The negative effect on sale price of an auction ending in February is interesting since no other month dummies were significant, while the February dummy was 19 significant at the 5% level. This result may signal the effect of level of search on sale price in online auctions, although it is unclear whether the amount of search is higher or lower during the month of February relative to June, and that information is not publicly available.20 Another explanation for the negative February effect may be that consumers as a group are less likely to shop in February, whether due to the post-holiday season or simply cold weather factors. Monthly retail sales reports from 1992 to 2004 collected by the U.S. Census Bureau confirm this, as February is historically one of the slowest retail months of the year and is also consistently worse than the month of June, even after controlling for the number of days in a month. This decline in retail shopping may cause a decrease in the demand for gift cards, resulting in lower sale prices.21 Also, as the commodity value of the item increases, the willingness to pay also increases but at a slower rate, meaning that, all else equal, a store credit of $50 sold for about 95% of its value, while store credits of $100 and $200 sold for around 90% and 80% of their value, respectively. Although the model examines only one general-valued auction and therefore does not take this issue into account, it is interesting to consider that for a given V, bidder i may value the item at vi = αiV, but that increasing the value of V may not correspond directly to the cash value, that is 2vi may be less than αi(2V). The results from studying data containing the common value of the item sheds light on many different questions left unanswered by previous studies of eBay auctions. 20 Occasionally sellers will place “counters” on their auctions that display the number of visits to the webpage, which could be used to assess the level of search. However, this may not count the number of distinct visitors, since people who block cookies may increase the counter every time they visit. 21 The dummy for an auction ending in February is not significant for auctions of commodity value $200 and greater, providing evidence that slower retail periods do not affect sales of high-valued store credit. This store credit may be more likely to be consumed over time relative to lower-valued store credit, which may make it less susceptible to retail cycles. 20 Specifically, measuring the importance of the value of the item to consumers and its effect on sale price has introduced a new estimate and another component to understanding online auctions. VII. Conclusion This study used data from auctions of store credit, or gift cards, to test the value of reputation measures in determining sale price in online auctions, and found a positive and significant, though very small effect. This data avoids the problem of object heterogeneity by analyzing auctions that disclose the commodity value of the item to consumers, resulting in a significantly smaller coefficient for seller feedback than in prior research. A model of bidding behavior was presented to validate the importance of having quality information available to bidders, in an attempt to demonstrate how weak signals could cause a bidder with the highest value of the item to lose the auction. Implications of the model, combined with the empirical findings, suggest that bidders in internet auctions would benefit from improved reputation measures. There are many different changes that eBay could institute in order to improve their feedback system. Paying a small “feedback bonus” to members once they have left feedback after a transaction would increase the amount of feedback given.22 Also, evaluating sellers according to specific performance issues would eliminate the problem of retaliatory feedback. For example, a seller should be rated across a 5 point scale on communication, delivery of the item, arrival of the item in a timely manner, and item as 22 Only around 50% of all transactions are currently rated by bidders (according to Resnick and Zeckhauser, 2001). 21 described. This information could then be converted to a reputation measure, by category, displayed directly on the auction page. Members should not receive feedback as buyers in a transaction, since sellers have the power to report non-paying bidder alerts to eBay, which removes “bad” bidders from the system. Also, the feedback received as a buyer should not have an effect on a member’s reputation as a seller, since these actions are unrelated. Improving the reputation measure would help to increase the quality of information available to bidders, which would reduce both the likelihood of a bidder getting “burned” by a bad seller, and the chance that a bidder with the highest valuation of an item loses the auction. In addition to examining the eBay seller reputation measure, this study also provided results and discussion regarding the negative effect on price of using the Buy-itnow option, offering restricted options for payment, and selling items in the month of February, relative to June. Future research will examine how the amount of bidder search affects sale price in online auctions, possibly by employing experimental methods. 22 References Cabral, L. and A. Hortaçsu, 2006, The Dynamics of Seller Reputation: Evidence from eBay, working paper. Dewan, S. and V. Hsu, 2004, Adverse Selection in Electronic Markets: Evidence from Online Stamp Auctions, Journal of Industrial Economics 52, 497-516. Houser, D. and J. Wooders, 2006, Reputation in Auctions: Theory and Evidence from eBay, Journal of Economics and Management Strategy 15, 353-369. Lucking-Reiley, D., D. Bryan, N. Prasad, and D. Reeves, 2000, Pennies from eBay: The Determinants of Price in Online Auctions, Working paper, Vanderbilt University #003. McDonald, C. and V. Slawson, 2002, Reputation in an Internet Auction Market, Economic Inquiry 40, 633-650. Melnik, M. and J. Alm, 2003, Reputation, Information Signals, and Willingness to Pay for Heterogeneous Goods in Online Auctions, SSRN Working paper #452820. Offenberg, J., The Market for Gift Cards, forthcoming, Journal of Economic Perspectives. Resnick, P. and R. Zeckhauser, 2001, Trust Among Strangers in Internet Transactions: Empirical Analysis of eBay’s Reputation System, The Economics of the Internet and ECommerce. Michael R. Baye, editor. Volume 11 of Advances in Applied Microeconomics. Amsterdam, Elsevier Science. Roth, A. and A. Ockenfels, 2002, Last-Minute Bidding and the Rules for Ending SecondPrice Auctions: Evidence from eBay and Amazon Auctions on the Internet, American Economic Review 92, 1093-1103. Steiner, D., 2003, Survey: How do Users Feel About eBay’s Feedback System? Auctionbytes-Update 8, www.auctionbytes.com [April 3, 2005]. 23 Back to list of items Everything Listed in Else > Gifts & category: Occasions > Gift Certificates $50 Gap Gift Card You are signed in Item number: 2993565095 Bidding has ended for this item (happypat35 is the winner) Winning bid: Ended: US $42.15 Mar-18-04 15:03:58 PST This is a $50 Gap Gift Card Seller information Never Expires USE IT AT ANY GAP STORE, GAP KIDS, BABY GAP, GAP BODY, GAP OUTLET, AND EVEN AT GAP.COM Great Gift Free shipping!!!! Go to larger picture pinkync2003( 37 ) Feedback Score: 37 Positive Feedback: Mar-1104 15:03:58 100% Start time: Member since Oct-06-03 PST in United States 9 bids (US Read feedback comments History: $10.00 starting bid) Ask seller a question View seller's other items Winning bidder: happypat35 ( 11 ) Item location: Morrisville, NC United States Ships to: United States only Shipping and payment details Figure 1: Example of an eBay Gift Card Auction 24 Table 1: Summary Statistics of the Independent Variables Mean Promotions Commodity Value Sale Price Number of Bids Starting Bid Shipping Cost Ending Time Buy-It-Now Reputation Score Reputation Percentage Payment Restrictions Ended on Weekend 0.25 192.16 164.99 9.33 83.85 1.48 1520 0.18 523 98.92 0.31 0.29 Median 0 100 87.13 8 9.99 0 1631 0 99 100 0 0 Mode 0 100 90 1 0.99 0 1200 0 87 100 0 0 Std Dev 0.44 411.42 339.26 6.99 342.45 2.64 499.65 0.39 1346 2.39 0.46 0.45 Min 0 3.3 2.26 1 0.01 0 0 0 -1 80 0 0 Max 1 11000 8800 43 8800 40 3321 1 16150 100 1 1 Count 2002 2002 2002 2002 1755 2002 2002 2002 2002 490 2002 2002 25 Table 2: Regression Estimates Dependent variable: Ln(Sale Price) (1) Ln(True Value) Reputation Score (in thousands) (2) 0.991*** (0.007) 0.996*** (0.003) 0.01*** (0.000) Reputation Squared Reputation Percentage Ended in February Ended on a Weekend Buy-It-Now Late Hour Number of Bids Payment Restrictions Promotions Dummy-Apparel Dummy-Coffee Shop Dummy-Discount Store Dummy-Electronics Dummy-Home Improvement Dummy-Office Supplies Dummy-Pet Store Dummy-Department Store Dummy-Toy Store Constant Observations, R-squared -0.000009** (0.000) 0.004** (0.002) -0.02** (0.008) -0.007 (0.01) 0.01 (0.02) -0.017*** (0.005) -0.019 (0.02) -0.007 (0.006) -0.006 (0.006) 0.0006** (0.000) 0.0007* (0.0004) -0.017*** (0.005) -0.003 (0.019) 0.016* (0.008) 0.035 (0.024) -0.008 (0.007) -0.059 (0.014) 0.085*** (0.011) 0.083*** (0.012) 0.092*** (0.010) 0.088*** (0.010) 0.077*** (0.008) 0.074** (0.01) 0.11*** (0.009) 0.11*** (0.003) 0.085*** (0.01) 0.081** (0.016) 0.005 (0.014) -0.035 (0.022) 0.031*** (0.009) 0.031* (0.009) 0.051*** (0.009) 0.053*** (0.009) -0.191*** (0.017) -0.542*** (0.166) 2002, 0.993 490, 0.986 ***Significant at 1%, ** 5%, and * 10% levels Robust Standard Errors are in parentheses. 26 Table 3: Estimates According to Data Segmented by Commodity Value Dependent variable: Ln(Sale Price) (1) under $100 Ln(True Value) Reputation Score (in thousands) (2) $100 and up 1.003*** (0.002) (3) under $50 1.015*** (0.014) (4) over $200 1.014*** (0.005) 1.002*** (0.010) 0.02* (0.000) 0.007*** (0.000) 0.071** (0.000) 0.01*** (0.000) Reputation Squared Ended in February Buy-It-Now Number of Bids Payment Restrictions Promotions Dummy-Apparel Dummy-Coffee Shop Dummy-Discount Store Dummy-Electronics Dummy-Home Improvement Dummy-Office Supplies Dummy-Pet Store Dummy-Department Store Dummy-Toy Store Constant Observations, R-squared -0.000001* (0.000) -0.0000001** (0.000) -0.0000001** (0.000) -0.0000001** (0.000) -0.013 (0.013) -0.021** (0.010) 0.0016 (0.012) -0.005 (0.015) -0.0175* (0.010) -0.014*** (0.005) -0.021* (0.012) -0.025*** (0.007) 0.0009 (0.001) 0.0006** (0.000) 0.002*** (0.000) 0.0002 (0.0003) -0.021* (0.012) -0.012*** (0.004) -0.031*** (0.008) -0.023*** (0.005) 0.039* (0.022) 0.002 (0.004) 0.033 (0.037) 0.002 (0.005) 0.017 (0.017) -0.007 (0.010) -0.08*** (0.016) -0.025 (0.044) 0.125*** (0.018) 0.046** (0.018) 0.032** (0.016) 0.035 (0.052) 0.115*** (0.025) 0.094*** (0.010) 0.009 (0.029) 0.115*** (0.043) 0.092*** (0.017) 0.089*** (0.009) -0.014 (0.015) 0.112** (0.043) 0.134*** (0.035) 0.111*** (0.008) 0.008 (0.016) 0.132*** (0.042) 0.104*** (0.022) 0.089*** (0.013) 0.005 (0.031) 0.095** (0.045) 0.06** (0.026) -0.008 (0.016) -0.009 (0.029) -0.009 (0.044) 0.044** (0.017) 0.043*** (0.012) -0.06*** (0.018) 0.065 (0.045) 0.069*** (0.02) 0.059*** (0.012) -0.064*** (0.019) 0.097** (0.045) -0.25*** (0.043) -0.224*** (0.022) -0.215*** (0.035) -0.314*** (0.053) 922, 0.971 1082, 0.994 462, 0.962 540, 0.991 ***Significant at 1% level, ** 5%, and * 10% level, Robust Standard Errors are in parentheses. 27