Complete Research Paper from attached feedback

RESEARCH ARTICLE

STRATEGIC BEHAVIOR IN ONLINE REPUTATION SYSTEMS: EVIDENCE FROM REVOKING ON EBAY1

Shun Ye School of Business, George Mason University, 4400 University Drive, Fairfax, VA 22030 U.S.A. {sye2@gmu.edu}

Guodong (Gordon) Gao and Siva Viswanathan R. H. Smith School of Business, University of Maryland, College Park, MD 20742 U.S.A.

{ggao@rhsmith.umd.edu} {SViswana@rhsmith.umd.edu}

This study examines how sellers respond to changes in the design of reputation systems on eBay. Specifically, we focus on one particular strategic behavior on eBay’s reputation system: sellers’ explicit retaliation against negative feedback provided by buyers to coerce buyers into revoking their negative feedback. We examine how these strategic sellers respond to removal of their ability to retaliate against buyers. We utilize one key policy change of eBay’s reputation system, which provides a natural experimental setting that allows us to infer the causal impact of the reputation system on seller behavior. Our results show that coercing buyers to revoke their negative feedback through retaliation enables low-quality sellers to manipulate their reputations and masquerade as high-quality sellers. We find that these sellers reacted strongly to eBay’s announcement of a proposed ban on revoking. Interestingly, after the power of these strategic sellers is curtailed, we find evidence that they exert more efforts to improve their reputation scores. This study provides valuable insights about the relationship between reputation system and seller behavior, which have important implications for the design of online reputation mechanisms.

Keywords: Reputation mechanisms, online ratings, quality transparency, online auctions

Introduction1

Reputation systems play a critical role in electronic markets given the significant information asymmetry between sellers and buyers (Ba and Pavlou 2002; Dellarocas 2003). A wide variety of reputation systems have been designed and implemented to mitigate problems arising from information asymmetry, with eBay’s feedback mechanism being the most established and well-studied among them. In keeping with the importance of reputation systems for online markets, both practitioners and academic researchers have invested substan- tial efforts in examining the design of online reputation and

feedback mechanisms as evidenced by the growing number of studies in recent years (e.g., Aperjis and Johari 2010; Bolton et al. 2004; Cabral and Hortacsu 2010; Dellarocas et al. 2006; Melnik and Alm 2002; Resnick and Zeckhauser 2002; Resnick et al. 2006). The importance of such feedback and ratings for transaction partners has been well documented, and prior studies have shown that a seller’s reputation score has a significant impact on sales and price premiums (e.g., Houser and Wooders 2006; Resnick et al. 2006).

Clearly, the effectiveness of a reputation system critically depends on the behavior of the transacting partners (Dini and Spagnolo 2005). Given the importance of reputation, it is not surprising that opportunistic sellers try to “game” the system to boost their reputation scores. It has been inferred that a substantial percentage of buyers would rather remain silent

1Ravi Bapna was the accepting senior editor for this paper. Bin Gu served as the associate editor.

MIS Quarterly Vol. 38 No. 4, pp. 1033-1056/December 2014 1033

Ye et al./Strategic Behavior in Online Reputation Systems

than provide negative ratings to a seller due to fear of retalia- tion. Therefore, one critical mission for reputation system design is to promote desirable seller behavior. Up to this point, however, there have been few studies examining how sellers respond to changes in reputation system design.

Our study represents an effort to fill this gap in the literature. In particular, we focus on one strategic behavior within eBay’s reputation mechanism: sellers’ retaliation against buyers who left negative feedback to force buyers to revoke the feedback (hereafter, we call these sellers strategic sellers). Before May 19, 2008, eBay allowed “revoking”—the ability to withdraw negative feedback subsequent to mutual agree- ment between the buyer and the seller. While the ability to revoke negative feedback enables transacting partners to correct honest mistakes, it is also prone to abuse by strategic sellers. Specifically, after receiving a negative rating from a buyer, the seller could retaliate by giving a negative rating to the buyer, and then suggest that both transaction partners withdraw their negative ratings. Since negative ratings are very rare (typically less than 1 percent of total ratings) and carry significantly more weight than positive ratings (Resnick and Zeckhauser 2002; Standifird 2001), such revoking can be especially damaging to the effectiveness of the online reputa- tion system. Starting in May 2008, eBay banned the with- drawal of negative feedback, and disallowed sellers from leaving neutral and negative feedback for buyers, in essence eliminating the possibility of retaliation and revocation by opportunistic sellers. This key policy change provides a “natural experiment” setting that allows us to infer the causal effect of reputation system design on seller behavior with greater confidence.

Our study seeks to empirically examine how strategic sellers behave before as well as after eBay’s policy change, com- pared to nonstrategic sellers. Specifically, we examine two research questions: (1) How do strategic sellers differ from nonstrategic sellers in their reputation profiles before the policy change? (2) What are the impacts of the changes to eBay’s reputation system on the behavior of strategic sellers as compared to those of nonstrategic sellers? In answering the first question, we are able to assess the extent to which the availability of revoking benefits strategic sellers, but hurts the reputation system. We find that, prior to the policy change, strategic sellers had a superficially similar, but underneath significantly worse, reputation compared to a matched sample of nonstrategic sellers. Interestingly, we find that strategic sellers were also more likely to participate in the online strike initiated to protest against the change compared to nonstra- tegic sellers, perhaps indicative of their concern about the potential damage to their reputation from the proposed ban on

revoking. The second research question seeks to examine the differential impact of the policy change on strategic sellers and nonstrategic sellers. Using a difference-in-differences analysis, we find that while both types of sellers receive a higher percentage of negative ratings as expected, the magni- tude of the increase is much smaller for strategic sellers. Additional tests and robustness checks suggest that strategic sellers are more likely to improve their service quality after the policy change to compensate for their inability to use revoking as a strategic tool to fix their reputation. These findings not only provide one of the first pieces of empirical evidence of seller reactions to changes in the reputation mech- anism design, but generate valuable insights on the crucial behavioral assumption about how reputation systems should be modeled.

The remainder of the paper is structured as follows: In the next section, we provide an overview of existing literature. We then explain the data collection and the natural experi- mental setting. Next, we describe all of the major variables used in the econometric models, examine the differences between strategic sellers and nonstrategic sellers before the policy change, including their true reputation scores and their reactions to the announcement of the policy change, and focus on examining how they respond to the changes in the reputa- tion mechanism design (i.e., the elimination of revoking). Finally, we discuss the implications and present our conclusions.

Background

Online Reputation System

In online exchange markets like eBay, sellers and buyers are often geographically separated. The buyer has few means to verify the quality of the seller or hold the seller responsible. The potential of seller opportunism is even more significant when buyers and sellers have infrequent interactions. Reputa- tion systems, which disseminate information on the past behavior of individual traders, are designed to facilitate trustworthy transactions among strangers on the Internet. Numerous online markets, such as Elance.com, vWorker.com, Amazon.com, and eBay have adopted reputation mechanisms to promote honesty and better efforts in traders’ behavior.

Whereas an increasing number of studies have focused on designing different reputation mechanisms (e.g., Masclet and Pénard 2012; You and Sikora 2011), eBay’s reputation mechanism is arguably the most established and the most

1034 MIS Quarterly Vol. 38 No. 4/December 2014

Ye et al./Strategic Behavior in Online Reputation Systems

scrutinized by the popular press as well as by academics. On eBay, the primary source of information about the trust- worthiness of a seller is his/her feedback profile. Upon the completion of a transaction, both buyers and sellers have the opportunity to leave feedback within 90 days. Resnick and Zeckhauser (2002) find that buyers leave feedback 52.1 percent of the time and sellers leave feedback 60.6 percent of the time.

The feedback has three levels of valence: positive, neutral, and negative. In addition, buyers and sellers can each provide detailed comments about the other party regarding the transaction. The feedback a seller or a buyer receives is aggregated to calculate his/her feedback score, which is one key metric indicating the user’s reputation. A user’s reputa- tion score is calculated as the count of distinct users who gave positive feedback minus the count of those who left negative feedback, and it is displayed right next to the user’s ID wherever it appears on eBay. In addition, the percentage of positive feedback among all distinct positive and negative ratings for each seller is also reported. Since examining each individual feedback comment would entail a huge investment of time by the buyer, the reputation score, together with the percentage of positive feedback, is displayed to signal a seller’s quality. Given the importance of the feedback a user receives, eBay allowed buyers and sellers to negotiate to mutually revoke negative feedback ratings while unilateral attempts are disallowed. This policy has remained in place since eBay was founded in 1995, until the 2008 policy change that disallowed revoking.

Despite eBay’s popularity and success, there has been evi- dence of inefficiencies in its reputation mechanism. Some sellers continue to peddle fraudulent items with misleading descriptions without being caught. For instance, it is esti- mated that over 70 percent of the Tiffany jewelry sold on eBay is fake (Hafner 2007). Furthermore, one would expect an effective reputation mechanism to reward good sellers. However, researchers have failed to find consistent evidence for the impact of a seller’s reputation on auction price. Resnick et al. (2006), for example, find that negative feedback seems to have no impact on buyers’ willingness-to-pay. Cabral and Hortacsu (2010) examine sales of laptops, coins, and beanie babies on eBay and find that neither positive nor negative feedback influences the final auction price. Melnik and Alm (2002) find that even when a seller doubles his ratings, the consumer’s willingness-to-pay for gold coins in- creases by only 18 cents. Similarly, Lucking-Reiley et al. (2007) find that positive ratings have a negligible impact on price. This is echoed by Eaton (2005), who finds that a sel- ler’s reputation has little or no impact on the actual bid prices.

One critical issue that is detrimental to eBay’s reputation system is seller strategic behavior relating to feedback. On eBay, sellers and buyers may independently leave feedback within 90 days of the transaction and the feedback is available immediately to the other party. While the system is sym- metric (two-way), allowing both buyers and sellers to rate each other, buyers are at a disadvantage because they face product uncertainty before payment and seller opportunism after payment. While the reputation system intends to facili- tate buyers’ reporting of dishonest sellers to warn others, the symmetric nature of the previous reputation system makes it convenient and nearly costless for sellers to retaliate against any buyer providing them a negative rating. Thus it was apt to say that for buyers, “a negative first feedback can never be given without the fear of retaliation” (Klein et al. 2009, p. 315). This fear of retaliation reduces a buyer’s propensity to leave negative feedback on the seller (Dellarocas and Wood 2008). As a result, this creates an incentive for one party to strategically withhold its feedback as a means of retaliation (Dellarocas and Wood 2008; Yamagishi and Matsuda 2002). In addition to this direct feedback retaliation, a seller can also threaten to report buyers as scammers or abusers of the feedback system as a way to discourage negative feedback. This happens through private messaging and is not directly observable.

Once a buyer leaves a negative rating, the seller can retaliate and then try to “fix” the feedback using eBay’s revoking policy (Bolton et al 2009; Klein et al. 2009). In the vast majority of cases, revoking (the withdrawal of feedback based on mutual agreement) is preceded by a reciprocal negative feedback. When a seller responds to a negative rating with a negative rating, about 27 percent are later withdrawn through the revoking mechanism (Bolton et al. 2009).

In summary, the ability to retaliate and revoke feedback cre- ates an incentive for opportunistic sellers to manipulate their reputations by nullifying negative feedback. Whereas Bolton et al. (2009) and Klein et al. (2009) have pointed out the possibility of such strategic revoking, no study has thus far empirically and systematically examined this phenomenon.

eBay’s Policy Change: Ending Seller Coercion

Given the potential problems of eBay’s reputation system, scholars have suggested different ways to enhance the design of reputation systems. In a theoretical analysis, Ba et al. (2003) suggest that digital certificates issued by a trusted third party can motivate market participants to behave honestly. Others have also proposed that eBay should allow only the

MIS Quarterly Vol. 38 No. 4/December 2014 1035

Ye et al./Strategic Behavior in Online Reputation Systems

buyer to rate the seller (Chwelos and Dhar 2006) or that eBay should simultaneously reveal both partners’ ratings (Reichling 2004). Eventually, in January 2008, eBay announced dra- matic changes to its reputation mechanism, and starting on May 19, 2008, sellers were no longer allowed to provide negative or neutral feedback to buyers. A seller now has only two choices: not leaving any feedback, or leaving positive feedback to the buyer. Furthermore, revocation or mutual withdrawal of the feedback was disallowed. Any feedback that is left cannot be removed unless it is investigated and determined as a violation or abuse of eBay’s feedback policy after a dispute is filed. Bill Cobb, CEO of eBay, made the following comments in his public announcement on the reputation mechanism changes:

The original intent of eBay’s public feedback system was to provide an honest, accurate record of member experiences....But overall, the current feedback system isn’t where it should be. Today, the biggest issue with the system is that buyers are more afraid than ever to leave honest, accurate feedback because of the threat of retaliation. In fact, when buyers have a bad experience on eBay, the final straw for many of them is getting a negative feedback, espe- cially of a retaliatory nature.

Now, we realize that feedback has been a two-way street, but our data shows a disturbing trend, which is that sellers leave retaliatory feedback eight times more frequently than buyers do...and this figure is up dramatically from only a few years ago.

So we have to put a stop to this and put trust back into the system. (eBay 2008)

This change—from a symmetric to an asymmetric feedback system—removed a seller’s ability to retaliate against a buyer providing negative feedback. This change serves as an exoge- nous event that enables us to investigate how different sellers (both strategic and nonstrategic sellers) respond to the pro- posed as well as the actual changes in the design of eBay’s reputation system. The change in the reputation system shields buyers from retaliation by the sellers; hence they should be more willing to express their negative opinions about sellers. As for sellers, since the policy change mostly affects strategic sellers who have used retaliation and revoking to fix their reputations, they should be the most affected by the new policy. If so, these strategic sellers should be more likely to express their displeasure to the policy change. Further, if these sellers continue to under- perform, they could easily attract more negative feedback than

other sellers under the new reputation mechanism. Therefore, this policy change offers a valuable opportunity to examine how strategic sellers respond to reputation system design, which we examine in the remainder of this paper.

Research Context

eBay’s radical overhaul of its reputation mechanism, described above to be effective in May 2008, was announced on January 30, 2008. We examine the period before and after this policy change. To allow enough time for the new reputa- tion mechanism to take effect, we define a 3-month period— July 1, 2008, to September 30, 2008—as the post-change period.2 Correspondingly, we define July 1, 2007, to Sep- tember 30, 2007, as the pre-change period for two reasons. First, the pre- and post- periods cover the same months of a year, which alleviates potential seasonal effects on seller behavior. Second, because the pre-change period ends four months before eBay’s announcement, it is unlikely that buyers and sellers had changed their behavior in anticipation of the policy change. Comparing the pre- and post- periods allows us to examine the impact of the change in the reputation system design on seller behavior. Figure 1 shows the timeline of the events.

We draw a random sample of 2,890 sellers from the eBay marketplace (which we refer to as “general sellers”).3 To control for product categories, the sampling is based on the distribution of products listed on eBay. From this random sample of general sellers, we identify strategic sellers and nonstrategic sellers and examine how they respond to the policy change differently.

In addition to a comparison of strategic sellers and nonstra- tegic sellers during the pre- and post-change periods, we also

2eBay instituted additional changes in October 2008. For example, eBay stopped allowing users to send checks or money orders as payment for items purchased on the U.S. version of the site after October 20, 2008. Buyers would only be able to pay using PayPal, ProPay, credit or debit cards (if the seller had an Internet merchant account), or pay for the item upon pickup. These changes are beyond our study period, and thus they should not interfere with the effect of feedback policy change on seller behavior in our study.

3We restrict our sample to well-established sellers with total lifetime feed- back of 500 or more at the time of data collection in the year 2008. This reduces the noise from casual sellers and allows for a more accurate measurement of seller behavior based on transaction volume. These sellers account for 69.98 percent of all active listings on eBay at the time of our data collection.

1036 MIS Quarterly Vol. 38 No. 4/December 2014

Ye et al./Strategic Behavior in Online Reputation Systems

Figure 1. Timeline of eBay’s Reputation System Change

exploit an unusual event that provides further insights on seller reactions to the announcement of the policy change. The announcement of the policy change caused outrage among some sellers and culminated in a week-long strike, from February 18 to February 25, 2008, to protest the changes (Zouhali-Worrall 2008). In keeping with our primary objec- tive of understanding the differences in behaviors between strategic sellers and nonstrategic sellers before as well as after the policy change, we collect data relating to this strike to examine whether strategic sellers are more likely to partici- pate in the strike compared to other sellers.

We use eBay’s seller central forum to identify the sellers who participated in the strike. This forum is an online space for sellers to discuss a variety of issues related to eBay sellers, and it was established several years before the strike. Fol- lowing the announcement of the policy change in January 2008, a thread on eBay’s seller central forum was created with the title “Sign the pledge: No sales Feb 18-25!” From this thread we identify 398 unique IDs of sellers who signed

the pledge, whom we refer to as “strikers.” From this group of naturally disclosed sellers, we also identify strategic sellers and nonstrategic sellers.

For all of the sellers, we collect two sets of data: sellers’ feedback history and sellers’ listing records. The data covers all listings (including sold and unsold items) for the years 2007–2008, as well as the feedback if received. Based on sellers’ feedback history data, we calculate each seller’s pro- file, including their reputation scores and specific types of feedback ratings, which are dynamically updated at the time of each listing.

Our main analyses focus on the differences between strategic sellers and nonstrategic sellers, which are identified from the above sources.4 Since our focus is on the sellers’ revoking

4To increase the generalizability of our findings, we also conduct all of the analyses on general sellers and strikers separately and obtain consistent results.

MIS Quarterly Vol. 38 No. 4/December 2014 1037

Ye et al./Strategic Behavior in Online Reputation Systems

behavior, we differentiate between the cases of seller retali- ated and revoked (SRR), buyer retaliated and revoked (BRR), and non-retaliated and revoked (NRR). SRR feedback refers to the situation wherein the buyer leaves the seller a negative rating followed by the seller retaliating with a negative rating, and then both parties mutually agreeing to revoke their nega- tive feedback. BRR feedback refers to the situation wherein the seller leaves the buyer a negative rating followed by the buyer retaliating with a negative rating, and then both parties mutually withdrawing negative feedback. NRR feedback refers to the situation wherein the buyer gives the seller a negative rating and the seller directly asks for a withdrawal without any retaliation.

Because only SRR feedback reflects sellers’ strategic retali- ation behavior, we define strategic sellers as sellers who had SRR feedback in the pre-change period (before the announce- ment of the policy change). Nonstrategic sellers are sellers with zero SRR feedback (but they may have a small propor- tion of BRR or NRR feedback). This results in a sample of 387 strategic sellers (221 from general sellers and 166 from strikers) and 2,901 nonstrategic sellers.

Analyses

Data Description

We first examine how strategic sellers differ from nonstra- tegic sellers before the policy change and then examine how strategic sellers respond to the policy change differently from nonstrategic sellers. Table 1 provides a full list of all the dependent variables and explanatory variables that are used in different regression specifications.

Comparison of Strategic Sellers and Nonstra- tegic Sellers Before the Policy Change

Before the Announcement: The Effect of Revoking on Seller Reputation

Before examining how the policy change affects strategic sellers’ behavior, it is important to assess the extent of the benefit these sellers derive from revoking. If revoking plays a major role in affecting these sellers’ reputation, then it is more reasonable to assume that disallowing revoking should affect seller behavior in a substantial way. Therefore, we examine (1) the extent to which revoking contributes to boosting the displayed reputation scores of strategic sellers; and (2) how the displayed and real reputation scores of strategic sellers compare to the reputation of nonstrategic sellers.

Because SRR feedback is relatively rare, observing a higher percentage of SRR feedback for strategic sellers requires that they have a significantly higher number of feedback ratings than nonstrategic sellers. Therefore, it is not surprising that the average reputation score for strategic sellers (559.76) is higher than that of the nonstrategic sellers (149.71). This result is also consistent with the findings of Wood et al. (2002), which show that sellers with high reputation scores are more likely to engage in opportunistic behavior because buyers have a higher tolerance for them.

To confirm that our findings are not driven by the difference in the number of feedback ratings or other seller character- istics, we use the propensity score matching method to correct for potential sample selection bias due to the observable differences (Dehejia and Wahba 2002). We first predict propensity score based on a logit regression of the treatment (i.e., the status of being a strategic seller) on several key covariates, including the seller’s reputation score, the seller’s tenure on eBay, the average product price of the seller’s listings, and if the seller is a Powerseller5 or not. Then, for each strategic seller in the treatment group, we identify a matching seller in the control group (i.e., nonstrategic sellers) using nearest neighbor matching on the propensity score. Common support condition is imposed so that the treatment observations whose propensity scores are higher than the maximum or less than the minimum propensity score of the controls are dropped. This results in 354 strategic sellers and 354 nonstrategic sellers.

On eBay, a seller’s displayed reputation is reflected in his/her reputation score and in the percentage of positive feedback. Reputation score is defined as the number of unique positive feedback subtracted by the number of unique negative feed- back.6 The displayed percentage of positive feedback for a given seller is calculated by dividing the number of unique positive ratings by the total number of unique positive ratings and unique negative ratings. Once a feedback is revoked, it is not included in the calculation of reputation score and per- centage of positive feedback. Therefore, the displayed reputa- tion is subject to gaming. We further calculate a seller’s “true reputation” by taking into account neutral and revoked feedback.

5A Powerseller is an eBay seller who participates in the Powersellers program and maintains a high quality feedback profile and constant or growing trading volume. Powersellers enjoy a closer trading relationship with eBay, in- cluding increased attention, specialized tools, and discounts on final value fees.

6Consistent with eBay’s approach to calculate reputations, we only consider unique feedback: multiple positive feedback ratings from the same buyer are counted as only one positive feedback rating. Other types of feedback ratings are treated similarly.

1038 MIS Quarterly Vol. 38 No. 4/December 2014

Ye et al./Strategic Behavior in Online Reputation Systems

Table 1. Description of Variable Variable Name Variable Description

ifStrike 1 if the seller participated in the strike; 0 otherwise NumberofListingsLog The seller’s number of listings in one month prior to the strike (logrithamized) ifPowerSeller 1 if the seller is a PowerSeller on eBay; 0 otherwise SellerTenure How many months the seller had stayed on eBay FeeDifference The financial loss a seller would suffer under the new policy for one month’s listing ReputationScoreLog The seller’s reputation score (logrithamized) TotalNegativePct The seller’s percentage of initial negative feedback (with revoked feedback counted

as negative feedback) RemainingNegativePct The seller’s percentage of negative feedback RevokedFeedbackPct The seller’s percentage of revoked feedback SRRFeedbackPct The seller’s percentage of seller retaliated and revoked feedback BRRFeedbackPct The seller’s percentage of buyer retaliated and revoked feedback NRRFeedbackPct The seller’s percentage of non-retaliated and revoked feedback ifFeedbackNegativeit 1 if the feedback received by seller i at time t is negative; 0 otherwise AfterPolicyChanget 1 if the time t is after the policy change; 0 otherwise StrategicSelleri 1 if the seller i is a strategic seller; 0 otherwise MonthDummyt Dummy variables for months MonthIsAugt 1 if the time t is in August; 0 otherwise MonthIsSeptt 1 if the time t is in September; 0 otherwise TransDurationit How long the transaction for seller i at time t took to complete TransPriceLogit The final price of the transaction for seller i at time t (logrithamized) SellerTenureit The seller i’s number of months on eBay at time t MonthlyNegativePctim The seller i’s aggregated monthly percentage of negative feedback at month m ifFeedbackPositiveit 1 if the feedback received by seller i at time t is positive; 0 otherwise HonestRevokeri 1 if the seller i is a honest revoker; 0 otherwise

Table 2. Pre-Change Overall Reputation Profile Comparison: Strategic Sellers Versus Nonstrategic Sellers

Displayed Reputation† True Reputation

Score Positive Negative Positive Negative Neutral Revoked eBay-Withdrawn Strategic Sellers 564.99 99.58% 0.42% 97.85% 0.41% 0.64% 0.99% 0.11%

Nonstrategic Sellers

521.91 99.65% 0.35% 98.93% 0.34% 0.51% 0.13% 0.09%

T-value 0.57 -1.09 1.09 -7.98*** 1.08 2.10* 17.41*** 1.04 †eBay displays the percentage of positive feedback as the key metric of a seller’s reputation. Percentage of negative feedback is simply 1 minus the percentage of positive feedback. To be consistent with eBay’s practice, we only report the percentage of positive feedback and the percentage of negative feedback for “Displayed Reputation,” which is what a buyer observes. *p < 0.05; **p < 0.01; ***p < 0.001.

MIS Quarterly Vol. 38 No. 4/December 2014 1039

Ye et al./Strategic Behavior in Online Reputation Systems

Table 2 provides the comparison of both displayed reputation and true reputation profiles for strategic sellers and nonstrategic sellers. For the displayed reputation, the average percentage of positive feedback for strategic sellers and nonstrategic sellers is 99.58% and 99.65%, respectively. The difference is not significant at the 5% level, suggesting that the displayed reputation is similar between strategic sellers and nonstrategic sellers.

We next compare the true reputations of strategic sellers and nonstrategic sellers. Note that the revoked feedback was originally a negative feedback that had been withdrawn upon the mutual agreement of both the seller and the buyer. After adding in the original negative value of revoked feedback, we find that strategic sellers actually have a much higher true negative feedback percentage than nonstrategic sellers (0.99% + 0.41% = 1.40% for strategic sellers, and 0.13% + 0.34% = 0.47% for nonstrategic sellers, t-value = 10.51, p < 0.001). Combined with the comparison of the displayed reputations, our results indicate that while strategic sellers have a much higher percentage of true negative feedback, the revoking mechanism helps these lower-reputation sellers masquerade as sellers with higher reputations.

Are Strategic Sellers More Likely to Participate in the Strike Against the New Policy?

Given the evidence above that revoking can be used as a tool to strategically nullify negative feedback, in this section we seek to examine if there was a significant difference in the propensity of strategic sellers to participate in the strike, compared to nonstrategic sellers.

Since the strike was initiated in the eBay forum, one may argue that sellers active in the forum were more likely to strike merely because they knew about it. To control for this potential confounding factor and ensure the robustness of our results, we introduce a control group in our analysis on the strike propensity: forum sellers who were active in the forum but did not participate in the strike. We create a random sample of 2,280 such sellers (which we refer to as forum sellers), and analyze them together with general sellers and strikers in predicting strike propensity.

To confirm that the sellers who pledged to join the strike actually participated in the strike, we check their listing activities during the strike week. We do find that strikers reduced their listings significantly during the one-week period whereas we observe no such trend for general sellers and forum sellers.

In addition to the changes in the reputation system, there are other factors that might drive participation in the strike. Spe-

cifically, at the same time eBay announced changes to its fee structure, with lower listings fees (the price charged for each item listed to be sold on eBay) and higher final value fees (a percentage of the closing price extracted by eBay). Based on their listing and sales patterns, some sellers believed that they would have to pay more because of these changes. Thus, potential financial loss under the new fee structure could have also motivated some sellers to join the strike.

To control for the potential impact of changes in the fee structure, we collect detailed listings of sellers in all three groups one month prior to the strike (from January 18, 2008, to February 17, 2008). We collect detailed information about each listing, including product category, auction style, starting price, final price, and usage of features such as gallery pic- tures and subtitles. This allows us to calculate the exact fee charged by eBay. To measure potential financial loss, we calculate, for each listing, the difference between fees actually charged by eBay under the old fee structure and fees that would be charged by eBay under the new fee structure. We then aggregate the differences at the seller level.

In addition to changes to the fee structure, several other factors could potentially influence participation in the strike as well. Sellers with a larger number of listings (logarith- mically) would suffer more financially if they joined the strike and hence may have been less likely to participate. Powersellers would also be less likely to join the strike because they would enjoy significant final value fee discounts under the new fee structure. The longer a seller has used eBay, the higher his/her switching cost due to the accumu- lated loyal customer base on eBay. These sellers should have a stronger reaction to the reduction of seller power under the new reputation mechanism. Therefore, we included number of months on eBay as another control variable. Seller repu- tation is measured by both reputation score (log-transformed) and total negative feedback percentage (i.e., the sum of revoked negative feedback percentage and remaining negative feedback percentage). The full specification of the model is

Logit (ifStrike) = α + β1 × NumberofListingLog + β2 × ifPowerSeller + β3 ×SellerTenure + β4 × FeeDifference + β5 × ReputationScoreLog + β6 × TotalNegativePct + g

The descriptive statistics and correlation matrix of the variables in the regression are provided in Tables 3 and 4. The maximum VIF is 1.59, well below the threshold of 10, indicating that there is no multicollinearity among the independent variables.

The results of the logit regression model are shown in Table 5. Model 1 is the baseline model. The coefficient of fee

1040 MIS Quarterly Vol. 38 No. 4/December 2014

Ye et al./Strategic Behavior in Online Reputation Systems

Table 3. Summary Statistics for the Strike Analysis

Variable Number of

Observations Mean Std. Dev. Min Max (1) NumberofListingLog 5568 4.00 2.15 0.00 9.19

(2) ifPowerSeller 5568 0.43 0.50 0.00 1.00

(3) SellerTenure 5567 76.77 30.58 5.73 145.83

(4) FeeDifference 5568 -21.38 89.66 -699.02 1715.68

(5) ReputationScoreLog 5568 4.46 1.19 0.00 9.61

(6) TotalNegativePct 5568 0.27% 0.10% 0.00% 28.57%

(7) RemainingNegativePct 5568 0.26% 0.96% 0.00% 25.00%

(8) RevokedFeedbackPct 5568 0.14% 0.61% 0.00% 25.00%

(9) SRRFeedbackPct 5568 0.08% 0.39% 0.00% 10.00%

(10) BRRFeedbackPct 5568 0.02% 0.20% 0.00% 6.67%

(11) NRRFeedbackPct 5568 0.04% 0.42% 0.00% 25.00% (12) ifStrike 5568 0.07 0.26 0.00 1.00

Note: There is one missing value for number of months on eBay.

Table 4. Correlation Matrix Variable VIF (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12) (1) 1.39 1.00 (2) 1.29 0.36* 1.00 (3) 1.04 0.02* -0.07* 1.00 (4) 1.16 -0.28* -0.16* 0.07* 1.00 (5) 1.59 0.45* 0.45* -0.18* -0.19* 1.00 (6) 1.01 0.01* 0.04* -0.05* -0.01 0.00 1.00 (7) 1.01 0.03* 0.01 -0.03* -0.00 -0.02 0.86* 1.00 (8) 1.01 0.06* 0.06* -0.04* -0.02 0.04* 0.58* 0.07* 1.00 (9) 1.01 -0.02 0.05* -0.03* -0.03* 0.08* 0.38* 0.07* 0.63* 1.00 (10) 1.00 -0.00 -0.01 -0.02 0.01 -0.01 0.19* 0.02 0.33* -0.00 1.00 (11) 1.00 0.01 0.03* -0.02 -0.01 -0.02 0.38* 0.03* 0.69* 0.00 -0.00 1.00 (12) 0.03* -0.02 0.05* 0.04* -0.04* 0.07* -0.01 0.16* 0.27* 0.01 -0.01 1.00

Note: Pair-wise Spearman correlation is reported. *Indicates p < 0.05.

difference is significantly positive, suggesting that sellers who stand to lose more (or save less) under the new fee structure are more likely to strike. Consistent with our prediction, sellers with a longer tenure on eBay are more likely to strike. Powerseller status and the volume of listings do not have a significant effect on a seller’s propensity to strike. The coefficient of total negative feedback percentage is significantly positive, suggesting that sellers with more negative feedback before revoking are more likely to strike. In Model 2, we divide total negative feedback percentage into remaining negative feedback percentage and revoked feed-

back percentage in the regression. We find that the pseudo R2 increases by almost 160 percent, supporting the assertion that a seller’s revoking behavior has significant explanatory power on his/her participation in the strike. The coefficient of the percentage of revoked feedback is positive and is significant at the p < 0.001 level. This suggests that sellers with a history of revoking negative feedback are more likely to strike.

In Model 3, we split revoked feedback into SRR feedback, BRR feedback, and NRR feedback. The Pseudo R2 further increases by about 100 percent. The coefficient of SRR feedback percentage is significant and positive, but the coeffi-

MIS Quarterly Vol. 38 No. 4/December 2014 1041

Ye et al./Strategic Behavior in Online Reputation Systems

Table 5. Logit Regression Analyses of Strike Propensity Dependent Variable: ifStrike

Model 1 Model 2 Model 3

Independent Variable coefficient (std. err.)

coefficient (std. err.)

coefficient (std. err.)

Intercept -2.768*** (0.273)

-2.767*** (0.278)

-2.607*** (0.286)

NumberofListingsLog 0.005 (0.029)

0.009 (0.030)

0.000 (0.031)

ifPowerSeller -0.006 (0.121)

-0.011 (0.122)

0.017 (0.126)

SellerTenure 0.007*** (0.002)

0.007*** (0.002)

0.007*** (0.002)

FeeDifference 0.002* (0.001)

0.002** (0.001)

0.002** (0.001)

ReputationScoreLog -0.088 (0.053)

-0.108* (0.055)

-0.158* (0.057)

TotalNegativePct 13.958** (3.029)

RemainingNegativePct -15.667* (7.791)

-24.764* (9.469)

RevokedFeedbackPct 69.149*** (7.098)

SRRFeedbackPct 166.126*** (11.073)

BRRFeedbackPct 15.029 (19.976)

NRRFeedbackPct -52.169 (33.837)

Pseudo R2 0.017 0.045 0.098

*p < 0.05, **p < 0.01, ***p < 0.001

cients of BRR feedback percentage and NRR feedback percentage are insignificant. This indicates that sellers who strategically retaliate and then revoke negative feedback are indeed more likely to strike. A 0.1 percent increase in SRR feedback percentage would lead to 18.07 percent increase in the odds of joining the strike.

The logit regression analyses on the strike provide empirical evidence that revoking after retaliation is a significant factor that motivates the participation in the one-week strike: since strategic sellers will lose a strategic tool to deceptively “boost” their reputation after the policy change, they are more likely to protest against the ban on revoking.

Comparison of Strategic Sellers and Non- strategic Sellers after the Policy Change

Given the initial evidence that strategic sellers are more likely to demonstrate their displeasure by joining the strike, in this

section we focus on the impact of reputation system change on strategic sellers compared to nonstrategic sellers by utilizing the random sample of eBay sellers as well as the sample of strikers.7 Our major analyses, as detailed below, focus on how strategic sellers differ from nonstrategic sellers in the efforts they exert to reduce the chance of receiving negative feedback8 when moving from the pre-change period to the post-change period.

7We also conduct analysis using the strikers as the convenient sample of strategic sellers and the other sellers as the control group of nonstrategic sellers and get similar findings.

8Revoked feedback in the pre-change period is converted to their original values and count as negative feedback.

1042 MIS Quarterly Vol. 38 No. 4/December 2014

Ye et al./Strategic Behavior in Online Reputation Systems

The Difference-in-Differences Approach

To measure the impact of the 2008 policy change on seller behavior, we adopt a difference-in-differences model, which is commonly used to examine the causal effect of an interven- tion. One major advantage of the difference-in-differences model is that it circumvents many of the endogeneity issues that can arise when comparing heterogeneous individuals (Bertrand et al. 2004; Meyer 1995).

Denote Δs as the change in a strategic seller’s propensity to receive negative feedback from buyers after the new policy, and Δn as that of nonstrategic sellers. A negative value of Δ indicates a decrease in the propensity of receiving negative feedback. For example, if the propensity of receiving nega- tive feedback is reduced from 3 percent to 2 percent, then Δ equals -1 percent. Some unobserved factors can contribute to the change in seller reputation scores (e.g., changes in eBay’s buyer population, competition from other market places, etc.). However, since these factors are common to both the strategic sellers and nonstrategic sellers, we can difference out their effects, and identify the extra impact of policy change on strategic sellers by Δs – Δn (that is, beyond the impact received by nonstrategic sellers). A negative value of the above differencing term means that strategic sellers have a greater decrease in the propensity of receiving negative feedback.

We delve deeper to identify these impacts. In our case, two major factors contribute to the change in feedback by buyers. First, because of the removal of seller retaliation, there is a change in buyers’ propensity to leave negative feedback, which we term as δb. If buyers are more likely to leave negative feedback, then δb will be positive, indicating a higher chance to receive negative feedback in the post period. Second, with less power in the reputation system, sellers might change their behavior as well, which affects their chances of receiving negative feedback. We term the change in propensity of receiving negative feedback due to seller behavior change as δs. If a seller exerts more effort (which can be in the form of a more accurate description of items, faster delivery, better packaging, among others), this will lower the chance of receiving negative feedback, leading to a negative δs. For strategic sellers, the net change in the pro- pensity of receiving negative feedback from buyers, Δs, can be expressed as

Δs = δbs + δss, where the second s for each δ denotes strategic sellers

Similarly, the change for nonstrategic sellers can be expressed as

Δn = δbn + δsn, where n denotes nonstrategic sellers

Δ can be positive or negative depending on the relative magnitude of δb compared to δs. The difference-in-differences in the propensity of receiving negative feedback between revokers and non-revokers is

Δs – Δn = (δbs + δss) – (δbn + δsn) = (δbs – δbn) + (δss – δsn)

In the above equation, if strategic sellers exert more efforts in improving service than nonstrategic sellers, they should expect a greater drop in the propensity of receiving negative feedback, which leads to a negative value of δss – δsn. δbs – δbn reflects the difference across strategic sellers and nonstrategic sellers over the buyer’s propensity to leave them negative feedback when holding seller service quality constant. Since the displayed reputation profiles of strategic sellers and nonstrategic sellers are very similar (as shown earlier), we should not expect the buyers of strategic sellers to be system- atically different from buyers of nonstrategic sellers. Indeed we find that both strategic sellers and nonstrategic sellers face similar groups of buyers. Table 6 shows that the buyers of strategic sellers are not statistically different from the buyers of nonstrategic sellers, in terms of how long they have been on eBay and their reputation scores both before the policy change and after the policy change. Furthermore, buyers of the strategic sellers group and buyers of the nonstrategic sellers group have a similar propensity to leave negative feedback to sellers. Therefore, it is reasonable to assume that δbs and δbn have the same magnitude.

Rearranging the above equation, we have

(δss – δsn) = (Δs – Δn) – (δbs – δbn)

Since (δbs – δbn) is expected to be 0, δss – δsn can be proxied by Δs – Δn. A negative Δs – Δn would indicate that strategic sellers exert more effort (compared to that of the pre-period) in reducing negative feedback in the post-change period compared to nonstrategic sellers.

Still, one might argue that buyers of strategic sellers respond differently to the policy change than buyers of nonstrategic sellers. If this were the case, given the historical higher retaliation rate of strategic sellers, their buyers should exhibit a marked increase in their propensity to leave negative feed- back, therefore δbs should be greater than δbn. Since (δss – δsn) = (Δs – Δn) – (δbs – δbn), this means that strategic sellers’ behavioral improvement could be even greater than what Δs – Δn reflects (since seller effects are partly negated by buyer effects), and our analysis provides a more conservative estimate.

MIS Quarterly Vol. 38 No. 4/December 2014 1043

Ye et al./Strategic Behavior in Online Reputation Systems

Table 6. Buyers of Strategic Sellers Versus Buyers of Nonstrategic Sellers Pre-Change Post-Change

Feedback Score

Tenure on eBay in month by May 8, 2008

Propensity of Leaving Negative Feedback

Feedback Score

Tenure on eBay in month by May 8, 2008

Propensity of Leaving Negative Feedback

Buyers of Strategic Sellers

165.64 46.95 1.50% 174.08 47.72 1.68%

Buyer of Nonstrategic Sellers

188.72 47.03 1.38% 192.46 49.33 1.70%

T-test -0.54 -0.32 0.47 -0.44 -0.28 -0.12

*p < 0.05, **p < 0.01, ***p < 0.001

We estimate the following specification:

Logit (ifFeedbackNegativeit) = α + β1 × AfterPolicyChanget + β2 × StrategicSelleri + β3 × AfterPolicyChanget × StrategicSelleri + β4 × MonthDummyt + β5 × TransDurationit + β6 × TransPriceLogit + β7 × SellerTenureit + git

where i indexes the sellers and t indexes the timestamp for each feedback. The coefficient of AfterPolicyChanget reflects the general change in the possibility of receiving negative feedback by sellers. As discussed above, the coeffi- cient of AfterPolicyChanget × StrategicSelleri captures Δs – Δn. Therefore, a negative coefficient of AfterPolicyChanget × StrategicSelleri implies that strategic users exert more efforts compared to nonstrategic users in the post-change period as explained earlier. We also include the MonthDummyt vari- ables to control for possible fixed seasonality effects. As detailed in the “Research Context” section, our pre-change period is defined as July–September 2007, and the post- change period is July–September 2008. Because there are three different months (i.e., July, August, and September), two dummy variables, MonthIsAugt and MonthIsSeptt are included in the regression models. Finally, three control variables are added to control for transaction heterogeneity: duration of the transaction, final price, and the seller’s tenure at the time of transaction.

Empirical Findings

For the 354 strategic sellers and 354 nonstrategic sellers, we collect 294,586 feedback ratings in the pre-change period and

264,086 feedback ratings in the post-change period.9

The estimates are reported in Table 7. In the random effects logit model, the coefficient of AfterPolicyChanget is significantly positive, indicating that sellers generally are more likely to receive negative feedback after the change in the reputation system. This is consistent with our prediction: since sellers are no longer able to use retaliation to prevent buyers from providing negative feedback, or to eliminate negative feedback by revoking, they are expected to have more negative feedback displayed in their profiles under the new reputation mechanism. The significantly positive coeffi- cient of StrategicSelleri suggests that strategic sellers are more likely to receive negative feedback than nonstrategic sellers in the pre-change period, as expected.

Interestingly, the coefficient of AfterPolicyChanget × StrategicSelleri is significantly negative across various speci- fications. This indicates that the increase in negative feed- back percentage is much smaller for strategic sellers than for nonstrategic sellers after revoking is disallowed. This result still holds after we control for possible seller fixed effects in the fixed effects logit model.

Because negative feedback is extremely rare on eBay, we also estimate a rare-event logit model to correct for the potential underestimation bias, as suggested by King and Zeng (2001). We once again observe a significantly negative coefficient of the interaction term AfterPolicyChanget × StrategicSelleri as shown in Model 2.

9For simplicity, we report the findings from the matched sample, based on reviewers’ recommendations. All of our results are consistent when using the full sample, or the random sample from eBay sellers.

1044 MIS Quarterly Vol. 38 No. 4/December 2014

Ye et al./Strategic Behavior in Online Reputation Systems

Table 7. The Impacts of Removal of Revoking: Strategic Sellers Versus Nonstrategic Sellers Model 1

(Random Effects Logit)

Model 2 (Fixed Effects Logit)

Model 3 (Rare Event Logit)

Model 4 (Random Effects Logit)

Model 5 (Fixed Effects Logit)

Model 6 (Generalized

Linear Model)

Dependent Variable ifFeedback Negative

ifFeedback Negative

ifFeedback Negative

ifFeedback Negative

ifFeedback Negative

Monthly NegativePct

coefficient (std. err.)

coefficient (std. err.)

coefficient (std. err.)

coefficient (std. err.)

coefficient (std. err.)

coefficient (std. err.)

Intercept -5.992*** (0.136)

-5.013*** (0.060)

-5.983*** (0.290)

-5.417*** (0.198)

AfterPolicyChanget 0.323*** (0.085)

0.273** (0.104)

0.158* (0.071)

0.423* (0.197)

0.550* (0.246)

0.500* (0.199)

StrategicSelleri 1.173*** (0.121)

0.880*** (0.054)

0.593* (0.253)

1.277*** (0.189)

AfterPolicyChanget *StrategicSelleri

-0.278** (0.089)

-0.304*** (0.091)

-0.309*** (0.078)

-0.367*** (0.107)

-0.438* (0.215)

-0.385* (0.173)

SellerTenureit -0.007*** (0.002)

0.000 (0.019)

-0.006*** (0.000)

-0.006* (0.003)

-0.022 (0.033)

-0.005+

(0.003) MonthIsAugt 0.137***

(0.035) 0.133***

(0.040) 0.049

(0.034) 0.172*

(0.077) 0.190*

(0.084) -0.058 (0.169)

MonthIsSeptt 0.109** (0.039)

0.104* (0.051)

0.050 (0.037)

0.200* (0.078)

0.233* (0.099)

-0.181 (0.157)

TransDurationit -0.000 (0.001)

-0.000 (0.001)

TransPriceLogit 0.228*** (0.029)

0.231*** (0.030)

# of obs. 558672 558672 558672 153678a) 153678(a) 4248(b)

+p < 0.10, *p < 0.05, **p < 0.01, ***p < 0.001 Note: (a) 404994 feedback instances are removed from the regression because of missing data on duration and price, either due to non-US transactions or missing transaction ID in the feedback history data; (b) The total number of data points is 354 × 2 × 6 = 4248.

As a further robustness check, we add transaction-related variables such as the duration of the transaction (i.e., the interval between the listing date and the transaction closing date) and the final price into the conventional logit regression analyses. The estimates are reported in Model 4 and Model 5. While the sample size is reduced due to missing values in these two variables in our raw data, we find that all major results hold.

In the formal difference-in-differences model, we assume that the error term εit follows an independent and identically distributed (i.i.d.) standard logistic distribution. However, as Bertrand et al. (2004) suggest, serial correlation between εit and εit+1 can lead to an underestimation of the standard error and an overestimation of t-statistics and significance levels. One way to circumvent this issue without making any specific assumption about the autocorrelation form is to aggregate the

time series information. Therefore, as a robustness check, we aggregate our data at two levels for further robustness checks.

First, we aggregate each seller’s performance by month. The dependent variable now is the seller’s monthly aggregated percentage of negative feedback, a ratio whose predicted value should also fall between 0 and 1, requiring the use of a generalized linear model (Papke and Wooldridge 1996). As shown in Model 6 of Table 7, the coefficient of AfterPolicyChanget × StrategicSelleri is consistently significantly negative.

Second, we aggregate each seller’s overall performance in the pre-change and the post-change periods. As shown in Figure 2, the displayed negative feedback percentage in- creases for all sellers (from 1.40 percent to 1.42 percent for strategic sellers, and from 0.47 percent to 0.80 percent for

MIS Quarterly Vol. 38 No. 4/December 2014 1045

Ye et al./Strategic Behavior in Online Reputation Systems

Figure 2. Comparison: Change in Negative Feedback Percentage

nonstrategic sellers). However, the 0.02 percent increase in actual negative feedback percentage for strategic sellers is much lower than the 0.33 percent increase for nonstrategic sellers, and the difference is statistically significant at the 5 percent level (t-value = -2.08).

Overall, we find that while both strategic sellers and nonstra- tegic sellers experience a higher percentage of negative ratings in the post-change period, the magnitude of the increase is much smaller for strategic sellers. Prior literature (e.g., Pavlou and Dimoka 2006) has suggested that feedback can be a proxy for the effort exerted by a seller in a trans- action. Our difference-in-differences estimate, therefore, provides supporting evidence that strategic sellers exert extra efforts (compared to nonstrategic sellers) to improve service quality. This indicates that strategic sellers changed their behavior in a positive way to mitigate the increase in negative feedback caused by the change in the reputation mechanism.

While in the above analyses strategic sellers are defined as sellers who had successfully convinced a buyer to revoke negative feedback, there are also 70 sellers in our sample who retaliated but with no success in revoking. These cases might be caused by the inadequate effort the seller made to negotiate with the buyer. Despite the fact that a revocation outcome was not reached in these cases, retaliation nonetheless reflects a seller’s endeavor to game the reputation system under the

old policy. Therefore, it is reasonable to assume that not only sellers with success of revoking but also sellers who retaliated (even though they did not successfully revoke) would be affected by the change in the feedback policy. Accordingly, we broaden the definition of strategic sellers by also including sellers who retaliated but did not succeed in revoking. In total, we have 424 strategic sellers. Using the propensity score matching method based on reputation score, tenure on eBay, average product price, and Powerseller status, we con- struct a matched sample of 424 nonstrategic sellers who had never retaliated against any negative feedback and who were not involved in revoking. As shown in Table 8, we consis- tently find that strategic sellers improve their efforts more compared to nonstrategic sellers after the policy change. This suggests that our findings are applicable to a broader range of strategic sellers in addition to pure sellers with success of revoking.

Falsification Test and Additional Robustness Checks

Falsification Test: As discussed earlier, the focus of the above analysis is how the behavior of strategic sellers is affected by the changes to the reputation system. We find evidence that there are sellers who previously attempted to fix their reputations by retaliating against buyers and revoking

1046 MIS Quarterly Vol. 38 No. 4/December 2014

Ye et al./Strategic Behavior in Online Reputation Systems

Table 8. The Impacts of Removal of Revoking: Broadened Strategic Sellers Versus Nonstrategic Sellers

Model 1 (Random Effects Logit)

Model 2 (Fixed Effects Logit)

Model 3 (Rare Event Logit)

Model 4 (Random Effects Logit)

Model 5 (Fixed Effects Logit)

Model 6 (Generalized

Linear Model)

Dependent Variable ifFeedback Negative

ifFeedback Negative

ifFeedback Negative

ifFeedback Negative

ifFeedback Negative

Monthly NegativePct

coefficient (std. err.)

coefficient (std. err.)

coefficient (std. err.)

coefficient (std. err.)

coefficient (std. err.)

coefficient (std. err.)

Intercept -6.056*** (0.133)

-5.319*** (0.064)

-6.664*** (0.353)

-5.528*** (0.222)

AfterPolicyChanget 0.271** (0.086)

0.244** (0.082)

0.159* (0.079)

0.252* (0.120)

0.301* (0.147)

1.130* (0.275)

StrategicSelleri 1.324*** (0.114)

1.183*** (0.059)

1.323*** (0.310)

1.637*** (0.166)

AfterPolicyChanget *StrategicSelleri

-0.233* (0.090)

-0.256** (0.092)

-0.247** (0.085)

-0.308** (0.105)

-0.309** (0.115)

-0.321*** (0.047)

SellerTenureit -0.007*** (0.001)

0.004 (0.018)

-0.006*** (0.000)

-0.003 (0.002)

-0.008 (0.033)

-0.008 (0.003)

MonthIsAugt 0.126*** (0.035)

0.118** (0.039)

0.048 (0.034)

0.110 (0.071)

0.110 (0.079)

-0.047 (0.178)

MonthIsSeptt 0.102** (0.037)

0.088+ (0.053)

-0.039 (0.064)

0.141+ (0.072)

0.156 (0.095)

-0.142 (0.164)

TransDurationit 0.000 (0.000)

0.000 (0.000)

TransPriceLogit 0.203*** (0.027)

0.208*** (0.028)

# of obs. 592408 592408 592408 210546(a) 210546(a) 5088(b)

+p < 0.10, *p < 0.05, **p < 0.01, ***p < 0.001 Note: (a) 381862 feedback instances are removed from the regression because of missing data on duration and price, either due to non-US transactions or missing transaction ID in the feedback history data; (b) The total number of data points is 424*2*6=5088.

Table 9. Pre-Change Overall Reputation Profile Comparison: Honest Revokers Versus Nonstrategic Sellers

Displayed Reputation True Reputation

Score Positive Negative Positive Negative Neutral Revoked eBay-

Withdrawn Honest Revokers 447.52 99.47% 0.53% 96.31% 0.51% 0.88% 1.25% 1.05%

Nonstrategi c Sellers 453.09 99.73% 0.27% 99.11% 0.27% 0.54% 0.00% 0.08%

T-value 0.05 -1.77+ 1.77+ -2.25* 1.71+ 2.38* 3.91*** 1.06 +p < 0.10, *p < 0.05, **p < 0.01, ***p < 0.001

MIS Quarterly Vol. 38 No. 4/December 2014 1047

Ye et al./Strategic Behavior in Online Reputation Systems

negative feedback. After revoking was banned, these sellers began expending more effort to improve their services. To further verify this inference, we also conduct a falsification test. Specifically, there are sellers who, upon receiving nega- tive feedback, did not retaliate against the buyer. Rather, they admitted their mistakes and took remedial action, and then asked the buyers to withdraw the negative feedback. To further corroborate that these sellers were behaving honestly, we examine the communications between buyers and sellers through text replies to negative feedback. We find that these sellers typically did not retaliate because they committed a genuine error and were attempting to fix it (for instance, replies include “Would be happy to give a full refund” and “Sorry for the confusion, we guarantee quality and delivery, will get enough back!”). We call these sellers honest revokers.10 Since this group of sellers do not strategically retaliate against buyers with negative feedback, they should be less affected by the policy change. In other words, this group of honest revokers who do not retaliate should behave differently from strategic sellers who retaliate and revoke.

We identify a total of 100 honest revokers in our sample. Using a similar propensity score matching method based on reputation score, tenure on eBay, the average product price of the seller’s listings, and the seller’s Powerseller status, 98 sellers who had neither retaliated against any negative feedback nor participated in revoking are matched as the control group for 98 honest revokers. The comparison of reputation profiles between these two groups is shown in Table 9.

Even though honest revokers initially receive more negative feedback than nonstrategic sellers, they look similar to non- strategic sellers after correcting their mistakes and removing the negative feedback. This suggests that revoking is a useful tool for honest sellers to remedy their mistakes, perhaps the primary reason why eBay introduced this policy initially. However, the existence of strategic sellers who abuse the policy dampens its effectiveness.

As shown in Model 1 and Model 2 of Table 10, the inter- action term AfterPolicyChanget × HonestRevokeri (HonestRevokeri is a dummy variable for being an honest revoker) is not significant. This result strengthens our finding that only strategic sellers who previously used retaliation are incentivized to perform better compared to nonstrategic sellers after the policy change.

Examining Changes in Positive Feedback: In the above analyses, we focus on the negative feedback received by

sellers. As another robustness test, we examine the positive feedback received by sellers after the policy change. If our finding that strategic sellers exert more efforts than nonstra- tegic sellers after the policy change is correct, this should be reflected in the positive feedback they receive as well. In other words, because of their improved service quality in the post-period, these strategic sellers should also experience a greater increase in the likelihood of receiving positive feedback than nonstrategic sellers when compared to the pre- change period. Our empirical finding confirms this conjec- ture. As shown in Model 3 and Model 4 of Table 10, the coefficient of AfterPolicyChanget × StrategicSelleri is consistently positive and significant in both the random effects logit model and the fixed effects logit model. These tests give us greater confidence that the reputation system change does motivate strategic sellers to improve their efforts to serve buyers.

We further carefully examine and rule out alternative explana- tions as to why strategic sellers are less likely to receive negative feedback in the post-change period, other than improving efforts,11 as detailed below.

Switching Product Categories. One alternative explanation for strategic sellers’ “improved” feedback scores compared with nonstrategic sellers is that strategic sellers simply switch to safer product categories instead of improving their services. To rule out this possible alternative explanation, we first calculate the distribution of listings among product categories for strategic sellers and nonstrategic sellers in the pre-change period and the post-change period. We then compare the change in each product category for strategic sellers and non- strategic sellers. As shown in Figure 3, no significant differ- ence is detected between strategic sellers and nonstrategic sellers in terms of changes in the total number of product categories in which they are selling and the percentage of listings in the top five product categories sold.12

Researchers have argued that different product categories might inherently have different potential for receiving nega- tive feedback. For example, MacInnes et al. (2005) find that in eBay online auctions, transactions in services are more likely to result in disputes than transactions in physical goods. Scott and Gregg (2004) propose that, when purchased online, high sensory products such as clothing and furniture are more likely to generate negative feedback compared with low sensory products. Product categories may differ in their inherent riskiness and, consequently, in the number of com-

10We thank one anonymous reviewer for this point.

11We conduct all of these checks for strategic sellers and nonstrategic sellers and obtain consistent results.

12This result also holds for the other 26 product categories.

1048 MIS Quarterly Vol. 38 No. 4/December 2014

Ye et al./Strategic Behavior in Online Reputation Systems

Table 10. Falsification Test and Robustness Test Model 1

(Random Effects Logit)

Model 2 (Fixed Effects Logit)

Model 3 (Random Effects Logit)

Model 4 (Fixed Effects Logit)

Dependent Variable ifFeedback Negative

ifFeedback Negative

ifFeedback Positive

ifFeedback Positive

Coefficient (std. err.)

coefficient (std. err.)

coefficient (std. err.)

coefficient (std. err.)

Intercept -5.840*** (0.234)

4.811*** (0.101)

AfterPolicyChanget 0.266* (0.130)

0.385* (0.188)

0.145* (0.056)

0.105* (0.053)

HonestRevokeri 1.149*** (0.210)

AfterPolicyChanget*HonestRevokeri -0.243 (0.155)

-0.254 (0.157)

StrategicSelleri -0.848*** (0.086)

AfterPolicyChanget*StrategicSelleri 0.124* (0.061)

0.132* (0.061)

SellerTenureit -0.005 (0.003)

-0.070 (0.113)

0.006*** (0.001)

0.026* (0.012)

MonthIsAugt 0.052 (0.081)

0.122 (0.143)

-0.087*** (0.025)

-0.110*** (0.027)

MonthIsSeptt 0.035 (0.085)

0.171 (0.247)

-0.089*** (0.026)

-0.130*** (0.034)

# of obs. 131419 131419 558672 558672 +p < 0.10, *p < 0.05, **p < 0.01, ***p < 0.001

Figure 3. Comparison: Change in Distribution of Categories

MIS Quarterly Vol. 38 No. 4/December 2014 1049

Ye et al./Strategic Behavior in Online Reputation Systems

Figure 4. Comparison: Change in the Distribution of Product Categories by Risk

plaints received by their sellers (MacInnes et al. 2005). This product category risk is aligned with the consumers’ beliefs regarding whether the products will perform according to their expectations (Bhatnagar et al. 2000). Product category risk increases with greater technical complexity, price, and needs of feel and touch (Bhatnagar et al. 2000; Chang et al. 2006; Finch 2007). We then examine whether retaliators have switched to low-risk product categories more than non- retaliators after the change in the reputation mechanism. We consider only the top five product categories: clothing, collectibles, books, jewelry, and electronics. These top five categories account for about half of all listings. Clothing is considered a high-risk product category because of the sensory nature of the product and the difficulty in describing its features accurately (Bhatnagar et al. 2000). Collectibles are considered a high-risk product category because they have many attributes and a complex description is required (Scott and Gregg 2004). Books, which are typically lower priced items, are considered a low-risk product category. Jewelry is considered a high-risk product category as sellers who cheat stand to benefit more from higher priced items. Electronics are considered to be a high-risk product category because, in

general, electronic items are technically and descriptively complex. According to Bhatnagar et al.’s (2000) rank of product category risk, electronics are much riskier than clothing and books.

Figure 4 presents the percentage of listings in high-risk and low-risk product categories for strategic sellers and nonstra- tegic sellers. Strategic sellers and nonstrategic sellers show a similar proportion of listings in high-risk products and low- risk products respectively in both the pre-change and the post- change periods. Also, the magnitude of change for strategic sellers and nonstrategic sellers is not significantly different.

Buying Reputation. Another potential alternative explanation for strategic sellers’ smaller increase in negative feedback is that strategic sellers intentionally buy more positive feedback through selling very low-value items to buyers and engaging in reciprocally positive feedback exchange (Dini and Spag- nolo 2009). Typically the title of such listings clearly states “100% positive feedback.” However, an examination of the product listings by both strategic seller and nonstrategic sellers suggest that no such feedback-buying behavior exists.

1050 MIS Quarterly Vol. 38 No. 4/December 2014

Ye et al./Strategic Behavior in Online Reputation Systems

Figure 5. Comparison: Sell-Through Rate

Sell-Through Rates. In the above analysis, we focus on seller reputation profiles. Another important measure of seller performance is the sell-through rate, which has important implications for market liquidity and efficiency. As shown in Figure 5, strategic sellers and nonstrategic sellers do not differ in their sell-through rates for either high-risk or low-risk products. Also, there is no difference in the magnitude of change in sell-through rates between the two groups of sellers (t-value = -0.12). This indicates that the smaller increase in negative feedback for strategic sellers is not driven by successfully selling in more low-risk product categories but is instead largely due to their quality improvement in selling.

Product Price. It is possible that strategic sellers might be likely to intentionally reduce the product price to lower buyers’ expectations of service quality in order to get less negative feedback in the post-change period. Therefore, we compare the average change in product price from the pre- change period to the post-change period for both strategic sellers and nonstrategic sellers. As shown in Figure 6, we do not find any significant difference between strategic sellers and nonstrategic sellers for either high-risk products or low- risk products. This helps rule out the alternative explanation.

To summarize, our results consistently show that the repu- tation system design has a meaningful and significant impact on seller behavior. After the power balance shifts in favor of buyers, strategic sellers improve their efforts more than nonstrategic sellers in the post-change period, and therefore have a smaller increase in negative feedback.

Discussions and Implications

Reputation mechanisms are vital to the success of online marketplaces such as eBay. However, the efficacy of these mechanisms depends crucially on how robust they are to potential gaming by participants. Ours is among the first studies to examine strategic gaming behavior in the context of online reputation systems. We utilize eBay’s policy change banning revoking to examine the impact of reputation system design on seller behavior. Our analysis of the protest/strike following the new policy announcement provides supporting evidence that strategic sellers do react strongly to the repu- tation system change: those who have revoked before are much more likely to participate in the online strike. After the new policy is implemented, we find that, in general, buyers are more likely to leave negative feedback after the seller loses the power to retaliate. More interestingly, we find that those strategic sellers have indeed acted opportunistically as they exert more efforts to improve the quality of their transactions.

This study makes several important contributions to the litera- ture on reputation systems (see Dellarocas 2005; Fan et al. 2005; Qu et al. 2008; Zhou et al.2008). To the best of our knowledge, there have been few studies on the explicit stra- tegic behavior of sellers in the context of online reputation systems. Previous studies (e.g., Dellarocas and Wood 2008) have inferred the threat of retaliation using statistical models. We build on these studies, and obtain direct and detailed measures of retaliation and revoking behavior, which allow us

MIS Quarterly Vol. 38 No. 4/December 2014 1051

Ye et al./Strategic Behavior in Online Reputation Systems

Figure 6. Comparison: Change in Product Selling Price

to generate deeper insights into the operational details of the gaming behavior within a reputation system. A reputation mechanism should facilitate market transactions by separating good players (either sellers or buyers) from bad ones and inducing honest behavior. We advance the existing literature that reputation matters in eBay auctions (Dewan and Hsu 2004; Lucking-Reiley et al. 2007) by providing one of the first pieces of empirical evidence that sellers do respond to the design of the reputation mechanism. Allowing revoking of feedback facilitates sellers’ strategic gaming behavior. After revoking is disabled, the more opportunistic sellers “behave better.” Empirically, the natural experiment setting, as well as the use of a difference-in-differences approach, allows us to infer the causal effect more rigorously.

Our findings also provide important theoretical insights into the development of online reputation systems (Dellarocas 2005). In recent years, theoretical work on the design of reputation systems has highlighted the significance of modeling how sellers respond to reputation mechanism design. There are three different ways to model a reputation system in a market wherein long-lived sellers interact with short-lived buyers: pure hidden information, pure hidden action, and mixed model (Bar-Issac and Tadelis 2008). In the pure hidden information model, sellers vary in their innate ability (or type) to deliver a product/service, and the reputation system’s goal is to reveal the seller’s type (Cripps et al. 2004; Mailath and Samuelson 2006). On the other hand, the pure hidden action model assumes that sellers have control over the outcome of a transaction by deciding how much effort to put into it. In such a case, the reputation system is designed to motivate the effort the seller exerts (Dellarocas

2005; Fan et al. 2005). The mixed model assumes that sellers differ in their innate abilities or qualities, but low quality sellers can increase the probability of a satisfactory trans- action by exerting more effort (Aperjis and Johari 2010; Cabral and Hortacsu 2010; Li 2010). While various theo- retical papers have adopted different models of reputation systems, there is little empirical evidence to verify these competing assumptions. Our study examines the extent to which sellers change their behavior in response to changes in the reputation system design, and generates valuable insights on the crucial behavioral assumption in these models. We find support for both hidden information and hidden efforts: strategic sellers improve their services after the policy change, but the reputation scores are now revealed as worse than average, as reflected in Figure 2 (with 1.42 percent negative feedback for strategic sellers and 0.80 percent negative feedback for nonstrategic sellers, t-value = 2.26). Therefore, the mixed model is likely closer to reality.

Furthermore, by examining the buyer–seller interactions before and after a fundamental change our study contributes to the understanding of reputation system design by shedding light on the importance of the power balance between the buyer and the seller on the effectiveness of a reputation mechanism. Our study also contributes to the understanding of the emerging influence of users’ actual interactions with feedback systems and other information systems on market mechanism design, as noted by Bapna et al. (2004) and Adomavicius et al. (2012).

The paper also contributes to the growing literature on the ways in which retailers can strategically influence or respond

1052 MIS Quarterly Vol. 38 No. 4/December 2014

Ye et al./Strategic Behavior in Online Reputation Systems

to buyer reviews. Whereas Stephen et al. (2012) show that monetary incentives offered by sellers can lead buyers to leave more helpful reviews, Cabral and Li (2012) find that monetary rewards can only increase the likelihood of buyers leaving unbiased ratings but not the values of the ratings. Abeler et al. (2010) examine sellers’ response to negative buyer reviews by comparing private apology to monetary compensation and find the former more effective in moti- vating buyers to withdraw negative ratings. Similarly, Gu and Ye (2014) find that a public management response can increase the future satisfaction of buyers who leave negative ratings. Jiang and Guo (2012) argue that retailers should allow more rating scales for popular products and fewer rating scales for niche products in order to induce more positive ratings. In another theoretical paper, Chen and Xie (2008) show that sellers should strategically respond to buyer reviews based on the type of the products. Our study makes contributions to this line of literature by studying sellers’ strategic gaming behavior with buyer feedback in the context of reputation systems.

This paper is also part of the growing literature on gaming behavior in online marketplaces. Kauffman and Wood (2005) examine the shilling behavior of sellers to artificially raise bidding prices. Cabral and Hortacsu (2010) find that roughly one-third of sellers built their reputations by acting as a buyer first. Jin and Kato (2006) find that some eBay sellers make non-credible claims of quality and mislead buyers. Stephen and Toubia (2010) find that sellers can strategically increase revenues by creating incoming links from other sellers who are dispersed. We contribute to the above literature by intro- ducing a new way to study seller strategic behavior. Our work is also related to the question of how consumers should interpret sellers’ online reputations. Zhang (2006) finds that reputation as a seller and as a buyer has a different impact on closing price. Our findings imply that the reputation system should make consumers aware of seller strategic behavior to better differentiate qualities.

Managerially, this study has two implications. First, the finding that revoking elicits strategic behavior in sellers suggests that, when revoking is available to sellers, online market makers should adopt other measures to reveal more quality information to buyers. One potential way to do this is to take revoked feedback into account when calculating over- all reputation and to display the percentage of revoked feedback to buyers. Currently there is no easy or straight- forward way of getting this information from eBay or other similar markets. Second, while banning revoking and the possibility of retaliation by sellers might help mitigate the retaliation problem, such a change could unduly shift the balance of power in favor of buyers. Providing more detailed

and granular feedback and reputation scores (for instance, their reputation in their role as a buyer versus as a seller) could help alleviate such a power imbalance, making market participants less vulnerable to strategic transaction partners.

We acknowledge several limitations in this study. First, in order to ensure that we examine active sellers with substantial numbers of transactions, who account for the majority of the transactions on eBay, we restrict our investigation to sellers with lifetime total feedback of 500 or more. Second, eBay made some other changes in October 2008 (e.g., no checks or money order as payment methods, as detailed in footnote 5). While we have limited our sample period to July–September, which is before these changes, the announcement effect may potentially influence selling behavior. Since these changes are not related to the reputation mechanism, we believe the confounding effects of these other changes should be trivial or nonexistent. Third, one alternative explanation for the observed difference between strategic sellers and nonstrategic sellers in negative feedback is that they may face significantly different buyers. In this study, even though we have con- trolled for the change in buyer behavior in the difference-in- differences model, direct investigation of buyer feedback- leaving behavior using detailed buyer-side data would further corroborate our findings. Also, several studies have shown that buyers’ opinions might be affected by previous ratings13 (e.g., Godes and Silva 2012; Li and Hitt 2008; Moe and Schweidel 2012; Moe and Trusov 2011). Given that strategic sellers and nonstrategic sellers look very similar to each other from the buyer’s perspective, we should not expect this social adjustment effect of seller ratings to be different between strategic sellers and nonstrategic sellers. A more detailed analysis of the evolution of seller feedback might help further support our findings. Finally, we infer the seller behavioral change using buyer feedback. Future research could strengthen our findings by seeking more direct measures of seller efforts and service quality.

The study can be extended in a number of interesting ways. First, one could conduct a more detailed analysis of how the process of revoking unfolds by looking at both sellers’ and buyers’ detailed feedback behavior. It is also important to understand how the changes in reputation mechanism influ- ence market efficiency. A detailed comparison of final auction prices between strategic sellers and nonstrategic sellers could shed light on this. In addition, it would be interesting to examine whether banning revoking in the new system benefits eBay or not. Prior to the policy change, eBay’s reputation mechanism was more symmetric with both buyers and sellers being allowed to post positive, neutral, or

13We thank one of the anonymous reviewers for this point.

MIS Quarterly Vol. 38 No. 4/December 2014 1053

Ye et al./Strategic Behavior in Online Reputation Systems

negative feedback about their transaction partners. However, the inherent asymmetry in the value of reputations (i.e., a good reputation is more valuable to a seller than it is to a buyer) made revoking more attractive to sellers. The change in the design of the reputation mechanism from a symmetric to an asymmetric one is likely to be in line with the asym- metric value of reputation to sellers and buyers, and therefore optimal. On the other hand, it is possible that these changes make buyers more powerful and induce them to behave opportunistically. Further research is needed on the costs and benefits of a symmetric versus an asymmetric feedback mechanism. Finally, we find supporting evidence that sellers improve their services as reflected in buyer feedback. Future research could examine more direct measures of seller efforts for a deeper understanding of how these efforts lead to better reputation portfolios.

References

Abeler, J., Calaki, J., Andree, K., and Basek, C. 2010. “The Power of Apology,” Economics Letters (107:2), pp. 233-235.

Adomavicius, G., Curly, S. P., Gupta, A., and Sanyal, P. 2012. “Effect of Information Feedback on Bidder Behavior in Con- tinuous Combinatorial Auctions,” Management Science (58:4), pp. 811-830.

Aperjis, C., and Johari, R. 2010. “Optimal Window for Aggre- gating Ratings in Electronic Marketplaces,” Management Science (56:5), pp. 864-880.

Ba, S., and Pavlou, P. A. 2002. “Evidence of the Effect of Trust Building Technology in Electronic Markets: Price Premiums and Buyer Behavior,” MIS Quarterly (26:3), pp. 243-268.

Ba, S., Whinston, A. B., and Zhang, H. 2003. “Building Trust in Online Auction Markets Through an Economic Incentive Mech- anism,” Decision Support Systems (35:3), pp. 273-286.

Bapna, R., Goes, P., Gupta, A., and Jin, Y. 2004. “User Hetero- geneity and Its Impact on Electronic Auction Market Design: An Empirical Exploration,” MIS Quarterly (28:1), pp. 21-43.

Bar-Isaac, H., and Tadelis, S. 2008. “Seller Reputation,” Founda- tions and Trends® in Microeconomics (4:4), pp. 273-351.

Bertrand, M., Duflo, E., and Mullainathan, S. 2004. “How Much Should We Trust Differences-in-Differences Estimates?,” Quar- terly Journal of Economics (119:1), pp. 249-275.

Bhatnagar, A., Misra, S., and Rao, H. R. 2000. “On Risk, Con- venience, and Internet Shopping Behavior,” Communications of the ACM (43:11), pp. 98-105.

Bolton, G., Greiner, B., and Ockenfels, A. 2009. “Engineering Trust: Reciprocity in the Production of Reputation Information,” Working Paper, Harvard Business School.

Bolton, G., Katok, E., and Ockenfels, A. 2004. “How Effective Are Electronic Reputation Mechanisms? An Experimental Investi- gation,” Management Science (50:11), pp. 1587-1602.

Cabral, L. M. B., and Hortacsu, A. 2010. “The Dynamics of Seller Reputation: Theory and Evidence from eBay,” Journal of Industrial Economics (58:1), pp. 54-78.

Cabral, L. M. B., and Li, L. 2012. “A Dollar for Your Thoughts: Feedback-Conditional Rebates on eBay,” Working Paper, New York University.

Chang, Q., Van, C., and Han, S. 2006. “Break the Trust Threshold: Customer Ratings and Trust Building on eBay Auctions,” in Proceedings of the 12th Americas Conference on Information Systems, Acapulco, Mexico, pp. 1805-1812.

Chen, Y., and Xie, J. 2008. “Online Consumer Review: Word-of- Mouth as a New Element of Marketing Communication Mix,” Management Science (54:3), pp. 477-491.

Chwelos, P., and Dhar, T. 2006. “Caveat Emptor: Differences in Online Reputation Mechanisms,” Working Paper, Sauder School of Business, University of British Columbia.

Cripps, M. W., Mailath, G. J., and Samuelson, L. 2004. “Imperfect Monitoring and Impermanent Reputations,” Econometrica (72:2), pp. 407-432.

Dehejia, R., and Wahba, S. 2002. “Propensity Score-Matching Methods for Nonexperimental Causal Studies,” Review of Economics and Statistics (84:1), pp. 151-161.

Dellarocas, C. 2003. “The Digitization of Word of Mouth: Pro- mise and Challenges of Online Feedback Mechanisms,” Manage- ment Science (49:10), pp. 1404-1424.

Dellarocas, C. 2005. “Reputation Mechanism Design in Online Trading Environments with Pure Moral Hazard,” Information Systems Research (16:2), pp. 209-230.

Dellarocas, C., Dini, F., and Spagnolo, G. 2006. “Designing Reputation (Feedback) Mechanisms,” in Handbook of Procure- ment, N. Dimitri, G. Piga, and G. Spagnolo (eds.), New York: Cambridge University Press.

Dellarocas, C., and Wood, C. A. 2008. “The Sound of Silence in Online Feedback: Estimating Trading Risks in the Presence of Reporting Bias,” Management Science (54:3), pp. 460-476.

Dewan, S., and Hsu, V. 2004. “Adverse Selection in Reputations- Based Electronic Markets: Evidence from Online Stamp Auctions,” Journal of Industrial Economics (52:4), pp. 497-516.

Dini, F., and Spagnolo, G. 2005. “Reputation Mechanisms and Electronic Markets: Economic Issues and Proposals for Public Procurement,” in Challenges in Public Procurement: An Inter- national Perspective, K. V. Thai, A. Araujo, R. Y. Carter, G. Callender, D. Drabkin, R. Grimm, K. Jensen, R. E. Lloyd, C. McCue, and J. Telgen (eds.), New York: Academic Press, pp. 227-247.

Dini, F., and Spagnolo, G. 2009. “Buying Reputation on eBay: Do Recent Changes Help?,” International Journal of Electronic Business (7:6), pp. 581-597.

Eaton, D. H. 2005. “Reputation Effects in Online Auction Markets,” Working Paper, Murray State University, Murray, KS.

eBay 2008. “A Message from Bill Cobb – New Pricing and Other News,” eBay General Announcements, January 29 (http://www2. ebay.com/aw/core/200801.shtml#2008-01-29054823).

Fan, M., Tan, Y., and Whinston, A. 2005. “Evaluation and Design of Online Cooperative Feedback Mechanisms for Reputation Management,” IEEE Transactions on Knowledge and Data Engineering (17:2), pp. 244-254.

Finch, B. J. 2007. “Customer Expectations in Online Auction Environments: An Exploratory Study of Customer Feedback and Risk,” Journal of Operations Management (25:5), pp. 985-997.

1054 MIS Quarterly Vol. 38 No. 4/December 2014

Ye et al./Strategic Behavior in Online Reputation Systems

Godes, D., and Silva, J. C. 2012. “Sequential and Temporal Dynamics of Online Opinion.” Marketing Science (31:3), pp. 448-473.

Gu, B., and Ye, Q. 2014. “First Step in Social Media-Measuring the Influence of Online Management Responses on Customer Satisfaction,” Production and Operations Management (23:4), pp. 570-582.

Hafner, K. 2007. “Tiffany and eBay in Fight Over Fakes,” New York Times, Technology Section, November 27 (http://www. nytimes.com/2007/11/27/technology/27ebay.html?_r=2&oref= slogin).

Houser, D., and Wooders, J. 2006. “Reputation in Auctions: Theory and Evidence from eBay,” Journal of Economics and Management Strategy (15:2), pp. 353-369.

Jiang, Y., and Guo, H. 2012. “Design of Consumer Review Sys- tems and Product Pricing,” NET Institute Working Paper No. 12-10.

Jin, G. Z., and Kato, A. 2006. “Price, Quality, and Reputation: Evidence from an Online Field Experiment,” RAND Journal of Economics (37:4), pp. 983-1004.

Kauffman, R. J., and Wood, C. A. 2005. “The Effects of Shilling on Final Bid Prices in Online Auctions,” Electronic Commerce Research and Applications (4:2), pp. 21-34.

King, G., and Zeng, L. 2001. “Logistic Regression in Rare Events Data,” Political Analysis (9:2), pp. 137-163.

Klein, T. J., Lambertz, C., Spagnolo, G., and Stahl, K. O. 2009. “The Actual Structure of eBay’s Feedback Mechanism and Early Evidence on the Effects of Recent Changes,” International Journal of Electronic Business (7:3), pp. 301-320.

Li, L. 2010. “Reputation, Trust, and Rebate: How Online Auction Markets Can Improve Their Feedback Mechanisms,” Journal of Economics & Management Strategy (19:2), pp. 303-331.

Li, X., and Hitt, L. M. 2008 “Self-Selection and Information Role of Online Product Reviews,” Information Systems Research (19:4), pp. 456-474.

Liu, Q. 2006. “Information Acquisition and Reputation Dynamics,” SIEPR Discussion Paper, Stanford University.

Lucking-Reiley, D., Bryan, D., Prasad, N., and Reeves, D. 2007. “Pennies from eBay: The Determinants of Price in Online Auctions,” Journal of Industrial Economics (55:2), pp. 223-233.

MacInnes, I., Li, Y., and Yurcik, W. 2005. “Reputation and Dis- pute in eBay Transactions,” International Journal of Electronic Commerce (10:1), pp. 27-54.

Mailath, G. J., and Samuelson, L. 2006. Repeated Games and Reputations, Oxford, UK: Oxford University Press.

Masclet, D., and Pénard, T. 2012. “Do Reputation Feedback Systems Really Improve Truth among Anonymous Traders? An Experimental Study,” Applied Economics (44:35), pp. 4553-4573.

Melnik, M. I., and Alm, J. 2002. “Does a Seller’s eCommerce Reputation Matter? Evidence from eBay Auctions,” Journal of Industrial Economics (50:3), pp. 337-349.

Meyer, B. 1995. “Natural and Quasi-Natural Experiments in Economics,” Journal of Business and Economic Statistics (13:2), pp. 151-162.

Moe, W. W., and Schweidel, D. A. 2012. “Online Product Opinions: Incidence, Evaluation, and Evolution,” Marketing Science (31:3), pp. 372-386.

Moe, W. W., and Trusov, M. 2011. “The Value of Social Dynamics in Online Product Ratings Forums,” Journal of Marketing Research (48:3), pp. 444-456.

Papke, L. E., and Wooldridge, J. M. 1996. “Econometric Methods for Fractional Response Variables with an Application to 401(k) Plan Participation Rates,” Journal of Applied Econometrics (11:6), pp. 619-632.

Pavlou, P. A., and Dimoka, A. 2006. “The Nature and Role of Feedback Text Comments in Online Marketplaces: Implications for Trust Building, Price Premiums, and Seller Differentiation,” Information Systems Research (17:4), pp. 392-414.

Qu, Z., Zhang, H., and Li, H. 2008. “Determinants of Online Merchant Rating: Content Analysis of Consumer Comments about Yahoo Merchants,” Decision Support Systems (46:1), pp. 440-449.

Reichling, F. 2004. “Effects of Reputation Mechanisms on Fraud Prevention in eBay Auction,” Working Paper, Stanford University.

Resnick, P., and Zeckhauser, R. 2002. “Trust among Strangers in Internet Transactions: Empirical Analysis of eBay’s Reputation System,” in The Economics of the Internet and E-Commerce, M. R. Baye (ed.), Greenwich, CT: JAI Press.

Resnick, P., Zeckhauser, R., Swanson, J., and Lockwood, K. 2006. “The Value of Reputation on eBay: A Controlled Experiment,” Experimental Economics (9:2), pp. 79-101.

Scott, J. E., and Gregg, D. G. 2004. “The Impact of Product Classification for Online Auctions,” in Proceedings of the10th Americas Conference on Information Systems, New York, pp. 2376-2380.

Standifird, S. S. 2001. “Reputation and E-Commerce: eBay Auctions and the Asymmetrical Impact of Positive and Negative ratings,” Journal of Management (27:3), pp. 279-295.

Stephen, A., Bart, Y., Du Plessis, C., and Goncalves, D. 2012. “Does Paying for Online Product Reviews Pay off? The Effects of Monetary Incentives on Consumers’ Product Evaluations,” INSEAD Working Paper No. 2012/96/MKT, Fontainebleau, France.

Stephen, A. T., and Toubia, O. 2010. “Deriving Value from Social Commerce Networks,” Journal of Marketing Research (47:2), pp. 215-228.

Wood, C., Fan, M., and Tan, Y. 2002. “An Examination of Reputation Systems for Online Auctions,” Working Paper, University of Washington.

Yamagishi, T., and Matsuda, M. 2002. “Improving the Lemons Market with a Reputation System: An Experimental Study of Internet Auctioning,” Working Paper, Hokkaido University.

You, L., and Sikora, R. T. 2011. “An Adaptive Reputation Mech- anism for Online Traders,” European Journal of Operational Research (214:3), pp. 739-748.

Zhang, J. 2006. “The Roles of Players and Reputation: Evidence from eBay Online Auctions,” Decision Support Systems (42:3), pp. 1800-1818.

Zhou, M., Dresner, M., and Windle, R. J. 2008. “Online Reputation Systems: Design and Strategic Practices,” Decision Support Systems (44:4), pp. 785-797.

Zouhali-Worrall, M. 2008. “Outraged eBay Sellers Plot Strike Week,” CNNMoney, February 10 (http://money.cnn.com/2008/ 02/07/smbusiness/ebay_boycott.fsb/).

MIS Quarterly Vol. 38 No. 4/December 2014 1055

Ye et al./Strategic Behavior in Online Reputation Systems

About the Authors

Shun Ye is an assistant professor in the Information Systems and Operations Management area at the School of Business, George Mason University. His research focuses on electronic market design and the strategic impact of emerging technologies, and has been published or presented at leading conferences such as the International Conference on Information Systems, Conference on Information Systems and Technology, and Workshop on Informa- tion Systems and Economics, among others. Prior to receiving his Ph.D. in Information Systems at the University of Maryland, College Park, he earned his B.S. in Management, B.E. in Computer Science and Technology, and M.S. in Management from the University of Science and Technology of China.

Guodong (Gordon) Gao is an associate professor in the Decision, Operations, and Information Technology Department at the Robert H. Smith School, University of Maryland, College Park. His research interests focus on the role of IT in transforming healthcare and improving quality transparency. His research has been published in leading journals including Management Sciences, Information Systems Research, Manufacturing and Service Operations Management, Journal of Management Information Systems, and Journal of Medical Internet Research, and presented at the International Health Economics Association Conference and the American Society of Health Economists Conference. He has

taught undergraduate, MBA, and doctoral courses. Gordon received his B.S. in Electrical Engineering and B.A. in Economics from Tsinghua University, his MBA from the Tsinghua–MIT Sloan Joint Program, and his Ph.D. from the Wharton School of the University of Pennsylvania.

Siva Viswanathan is an associate professor at the Robert H. Smith School of Business, and a codirector of DIGITS (the Center for Digital Innovation, Technology, and Strategy) at the University of Maryland. Siva’s research focuses on emerging issues related to online firms and markets, and on analyzing the competitive and strategic implications of digital innovations. His research has examined the growth of online information intermediaries as well as social media platforms, and their potential to disrupt traditional business models and transform the competitive landscape in a variety of sectors. His current work analyzes the sharing economy, focusing primarily on technology-enabled mechanisms that help alleviate information asymmetries in these nascent markets. Siva’s research has been published in top academic journals including Management Science, Information Systems Research, Journal of Marketing, and Decision Support Systems. He is on the editorial board of MIS Quarterly, and is also an active participant in inter- national conferences and industry forums. As a codirector of DIGITS, he also organizes the D. C. Forum on Digital Innovation, which brings together leaders from government, industry, and academia.

1056 MIS Quarterly Vol. 38 No. 4/December 2014

Copyright of MIS Quarterly is the property of MIS Quarterly and its content may not be copied or emailed to multiple sites or posted to a listserv without the copyright holder's express written permission. However, users may print, download, or email articles for individual use.