How Effective Are Online Reputation Mechanisms?

Gildas Marot- Cadet Thomas Master 2 e-Business

Shadow of the future or an analyze of

«How effective are online reputation mechanisms?» Gary E Bolton, Elena Katok and Axel Ockenfels May 2002

«On peut tromper une fois 1000 personnes, mais on ne peut pas tromper 1000 fois une personne.» Emile Gravier, La Cité de la peur (1994)

Introduction

3

Review of litterature

3

"Does a seller's e-commerce Reputation Matter?" Melnik, M.I and J. Alm

3

"The role of Institutions in the revival of trade: the law merchant, Private judges, and the champagne fairs" Milgrom, North and Weingast

3

" Experimental games for the design of reputation management system" C.Keser

4

"ERC: A Theory of Equity, Reciprocity and Competition." Bolton, G.E and A. Ockenfels (2000). 4 "Trust among Strangers in Internet Transactions: Empirical Analysis of eBay's Reputation System" Resnick, P. and Zeckhauser

4

“Evidence of the Effect of Trust Building Technology in Electronic Markets: Price Premiums and Buyer Behavior.” Ba, S. and P.A. Pavlou (forthcoming). 5 “The Coevolution of Trust and Institutions in Anonymous and Nonanonymous Communities.” Güth, W. and A. Ockenfels (forthcoming).

5

“Reputation in Auctions: Theory and Evidence from eBay.” Houser, D. and J. Wooders (2001). 6

Current paper : "How Effective are Online Reputation Mechanisms?"

6 Game presentation

6

The game design

6

Result of the  three market experimental design:

7

Analysis of the results:

8

Some Criticism about this study:

Conclusion

11

11

Addressed to Professor Masclet and Professor Denant-Boèmont

2

Introduction We try to analyze the article "How Effective are Online Reputation Mechanisms?" written by Gary E Bolton, Elena Katok and Axel Ockenfels in May 2002. This article explain in what extent the feedback mechanisms is mitigate the moral hazard problems associated with auction markets on internet. This problem come from the fact that seller could not send the product to his buyer when he receive the money if he is not honest, so their is a problem of trust for the buyer. It's a Prisoner's dilemma, a feedback mechanisms may solve it. First we analyze few references of the authors, then we present the game and his result, then we try to analyze this study and make few criticism. We realize this table in order to reveal some equilibrium on the stranger market.

Buyer/ Seller

Ship

no Ship

Buy

0,5 ; 0,5

0 ; 0,7

Don't Buy

0,35 ; 0,35

0,35 ; 0,35

We observe that it is a Prisoner's dilemma which conduct to no transaction. This is understable considering the one-shot effect. The reputation mechanism is here to solve this problem.

Review of litterature "Does a seller's e-commerce Reputation Matter?" Melnik, M.I and J. Alm Working paper, Georgia

They examine the willingness to pay for a mint condition US $5 Gold coin on eBay in 1999. They wanted to know if the seller's reputation have an impact on the price of the product. They finally managed after their studies to the conclusion that a good reputation have positive impact on the product price. It's mean that if you have a good reputation you can set a price higher. "The role of Institutions in the revival of trade: the law merchant, Private judges, and the champagne fairs" Milgrom, North and Weingast Economics and politics

In this paper, the authors works on the law merchant. This law was applied by gild of merchant in medieval time. This law solves the classic prisoner's Addressed to Professor Masclet and Professor Denant-Boèmont

3

dilemma. In this dilemma, each merchant have incentive to cheat the other if there is only one transaction. In the case of repeated interaction, it would be honest dealing if they adopt the "tit-for-tat" strategy: "if you cheat me i will punish you". The "tit-for-tat" becomes "tit-for-tat" adjusted: it's mean that if you cheat a merchant, you will be punished by the next merchant you deal with. It is obvious, that kind of mechanism works only on community where everybody knows everybody. This mechanism works well because it is profitable to cheat the cheater and if you not punish the cheater, you will be punished. " Experimental games for the design of reputation management system" C.Keser Dr Keser realized this experiment in order to quantify the increase in trust produced by two versions of reputation management systems. He used the long run reputation management system ( seller history of notation) and a short run reputation system (buyers see only the last notation). The doctor used experimental economics which involved 320 students. Keser underline thank to his experiments that a reputation mechanisms increase the number of transaction but he also remark that the long run reputation is more efficient than the short run. The trust and truswothiness are both increased by more than 50% while they are only rose by more than 30% in the short run mechanism. "ERC: A Theory of Equity, Reciprocity and Competition." Bolton, G.E and A. Ockenfels (2000).  American Economic Review, 90, 166-193.

This study explain how and why people collaborate on a market where the relative standing is important. It explain that the strategic situation can be deduce from two of the most elementary games: ultimatum and dictator. Taken together, these games have a flash point where self-interest is subjugated to concern for relative standing. But what is this concern for relative standing? Is it altruism, equity, or reciprocity? And there is a second deeper question: Why should people care about relative standing? This study speculate that the answer to the first question is ‘reciprocity’ and that the answer to the second question has to do with biology. Several experimental studies cast doubt on the proposition that people care about distribution in a way that we would expect an altruist to care. The same evidence suggests that people are willing to sacrifice little to defend equity as a principle. The authors suppose that evolution has molded human for successful group living. People may then have a propensity to contribute to the group, because a successful group contributes to their own individual biological success. In our case, it means that sellers are fair-play. "Trust among Strangers in Internet Transactions: Empirical Analysis of e-Bay's Reputation System" Resnick, P. and Zeckhauser Draft paper prepared for NBER workshop.

The authors realized this study in order to answer to the question : why buyers trust unknown sellers in the vast electronic garage sale ? They used data from Addressed to Professor Masclet and Professor Denant-Boèmont

4

e-Bay on 1999 which represent million of items. Finally they underlined two facts. The first one was there is a hight rate of providing evaluate, that may explain why buyers trust sellers who have good reputation.The other fact they observed was the extreme rarely of neutral or negative evaluation. The last one was named the High Courtesy Equilibrium which involved that when buyers rate, they rate positively or they don't evaluate at all. More over this phenomenon is strengthen on platform like e-Bay which allow sellers to rate buyers. That means a seller will rate positively the buyer in order to be rated in the same way. “Evidence of the Effect of Trust Building Technology in Electronic Markets: Price Premiums and Buyer Behavior.” Ba, S. and P.A. Pavlou (forthcoming). MIS Quarterly.

This study show that the mechanism of feedback is equivalent to a reputation built between two partner but it certainly be weaker. This article suggest that seller's reputation, reflected through his feedback profile, plays a very important role in buyer trust building. Moreover this study suggest that people who have a better reputation could sell their goods to a higher price.

“The Coevolution of Trust and Institutions in Anonymous and Non-anonymous Communities.” Güth, W. and A. Ockenfels (forthcoming). Jahrbuch für Neue Politische Ökonomie, 20, Tübingen: Mohr Siebeck.

This article analyze how the institutional environment drives the evolution of trust and trustworthiness through the evolution of moral preferences. I explain that modern online communities that are cheaply connected via electronic communication channels may imitate small group detection technologies Addressed to Professor Masclet and Professor Denant-Boèmont

5

through sophisticated computerized feedback systems and thus partly crowd out legal institutions. “Reputation in Auctions: Theory and Evidence from eBay.” Houser, D. and J. Wooders (2001). Working Paper, University of Arizona.

This article examine the effect of reputation on price in a data set drawn from the online auction site eBay. Our main empirical result is that seller, but not bidder, reputation has an economically and statistically significant effect on price.

Current paper : "How Effective are Online Reputation Mechanisms?" Game presentation The Study on which we focus on is «How Effective are Online Reputation Mechanisms». This study was realized by Gary E Bolton, Elena Katok and Axel Ockenfels. The goal of this paper is to measure the effectiveness of online reputation mechanisms and to regard the behavior that they induce. To realize this study they have 144 participants who have incentive to participate by cash. In order to measure the effectiveness of reputations mechanisms, the authors focused on the shipping. They set 3 kind of game: • Stranger Market: it’s a market with no reputation system. It is use as reference • Reputation Market: it's market with a reputation mechanism which track seller's history of shipping. This history is available for prospective buyers. • Partner market: it's market with no reputation mechanisms but when the same people interact together all the time. The game design They set a game with 16 traders playing during 30 rounds. In each round, they matched 2 players together: one buyer and one seller. The role assignment (buyer or seller) is random but is restricted by the fact that each trader is buyer half time and seller half time.

Addressed to Professor Masclet and Professor Denant-Boèmont

6

to understanding differences in transaction behavior across markets (we find substantial differences) is how buyers react to information about the sellers they are paired with. If buyers Figure 1. The buyer-seller encounter

discriminate between sellers who have been trustworthy in the past and those who have not been, 1.

Buyer's Choice

The role assignment is pick randomly

then sellers will have an incentive to be trustworthy. For this reason, we look at buyer behavior buy

between buyer and seller

not buy

in some depth. Seller's Choice

2.

ship

not ship

3.1 Treatment effects

Buyer earns: $0.50 Seller earns: $0.50

Buyer choose to buy or not one item at a fixed price

Buyer earns $0.35 Seller earns $0.35

3.

$0 $0.70

The seller choose to ship the product

or topatterns. keep the These buyer’s money The major treatment effects have to do with trading can be measured 4.

The player are paid

three ways: efficiencyor the percentage of potential transaction completed (Figure 2), trust or

Figure 1 illustrates the buyer-seller encounter.

Game moves and payoffs can be

the percentage givenare(Figure trustworthiness nterpreted in the following way. Both of the buy seller orders and the buyer endowed3), withand $0.35, which is

or the percentage of shipped

he payoff when noitems, trade takes place. The seller offers an item(Figure for sale at4). a price $0.35three whichfigures, conditioned on buy orders In all

the treatment data has been

Result of the  three market experimental design: 

has a value of $0.50 to the buyer. The seller’s cost of providing the buyer with the item, costs

aggregated across sessions. This experimental investigation test three different market type. Let's view results on each market.

associated with executing the trade, shipping, handling etc., as well as production costs,7 is

$0.20. So each successfully completed trade increases efficiency by creating a consumer surplus

Figure 2. Efficiency measured as how often the gain from trade is realized, by round

of $0.15 and a $0.15 profit for the seller. If the buyer chooses to buy the item, he sends his

100% endowment of $0.35 to the seller, who then has to decide whether to ship the item, or whether to 90%

keep both the money and the item. If the seller does not ship the item he receives the price plus 80% 70%

Efficiency

his endowment of $0.35 for a total of $0.70. If he ships, he receives the price minus the costs 60%

Partners

50% whereas the buyer receives his value of the item. If the plus his endowment for a total of $0.50,

Reputation Strangers

40%

buyer chooses not to buy the item, both players keep their endowment. 30% 20% buyer-seller encounter is one-shot, the seller, once he Under the assumption that the 10%

eceives the money from the buyer, has 0% no pecuniary incentive to be trustworthy and to ship the 1

3

5

7

9

11

13

15

17

19

21

23

tem. Anticipating this, the buyer may not trust the seller, so that trade does not take place, even

25

27

29

Round

hough it would make everybody better off.

This is the basic dilemma online reputation

mechanisms are designed to solve.

Production costs • where either the seller only produces the item once he knows the demand, or the product is The strangers market: produced before the buyer’s decision is known but costs are not sunk (e.g., when the item can be resold at a price qual to production On costs). this market the realized trade is less than 50%

in first periods and fall to less than 10% at the middle of the game with a mean of 14,3%. 6

Trust, measured as the percentage of buying per round, start at about 70% on the beginning to fall at less than 20% at the end with a mean of 37,1%. Trustworthiness, measured as percentage of shipping per round, is very 11 variable during the game, with a mean of 35,7%. • The reputation market:

On this market the realized trade is between 30% and 60%. Trust decrease from 80% to 40% in the ten first period and become very variable after. Trustworthiness is higher than 50% except on the last two periods.

Addressed to Professor Masclet and Professor Denant-Boèmont

7

• The partners market:

On this market the realized trade is very stable at about 70%, it only decrease on the last two periods. Trust is also very stable at more than 80%. Trustworthiness is higher than 80% except on the last two periods. For all those variable there is a end game effect, player know there is a high chance of deviation because they wouldn't have to trade with this players after the end of the game. Below all the means of different variables on different markets.

MEANS

trade

trust

trustworthiness

strangers

14,3 %

37,1 %

35,7 %

reputation 40,7 %

55,6 %

72,8 %

partners

83,3 %

88,5 %

73,9 %

Analysis of the results:  After they collected all the data from their experiments they made a probit model which explain the buyers trust and to know if the variable are significant. The main coefficients which interest us are the TOTALSHIPreputation, TOTALNOSHIPreputation, CBHS, CBNH, SHIPLASTreputation and NSHIPLASTreputation. With TOTALSHIPreputation and TOTALNOSHIPreputation, we watched that a "bad reputation" is more impacting than a good reputation. We noticed also with the table that the last ship is very important, in fact when the seller ship the last order it has a significant impact on the buyers trust ( O, 212 to 0,06 for SHIPLASTreputation against TOTALSHIPreputation).

Addressed to Professor Masclet and Professor Denant-Boèmont

8

Table 2. Random effects probit models, buyersa Maximum likelihood estimates (and two-sided p-values) for buyer behavior Dependent variable = “1” for buy Independent variable Model 1 Model 2

CONSTANT

0.533 (.0040) -0.020 (.9473) 0.963 (.0001)

REPUTATION = 1 if buyer is from reputation treatment, and 0 else.

PARTNERS = 1 if buyer is from partners treatment, and 0 else.

0.347 (.0185) 0.200 (.4347) 1.48 (.0000)

0.045 (.0180) -0.412 (.0000)

-0.390 (.1649) -0.944 (.0000) -1.200 (.0000) 0.456 (.0000)

-0.404 (.1671) -0.974 (.0000) -1.322 (.0000) .444 (.0000)

2160 -1056.57 .0000

2160 -988.67 .0000

= number of seller ships prior to last order.

TOTALNOSHIPreputation = number of seller no ships prior to last order.

SHIPLASTreputation = 1 if reputation seller shipped last order, and 0 else.

NSHIPLASTreputation = 1 if reputation seller did not ship last order, and 0 else.

SHIPLASTpartners = 1 if seller in partners shipped last order, and 0 else.

NSHIPLASTpartners = 1 if seller in partners did not ship last order, and 0 else. = number of past times item was shipped to buyer.

CBNH = number of past times buyer bought but not shipped.

ROUNDstrangers

-0.062 (.0000) -0.019 (.0006) 0.006 (.4806) -0.151 (.6414) -0.903 (.0000) -1.15 (.0000) 0.399 (.0000)

= round in strangers treatment, and 0 else.

ROUNDreputation = round in reputation treatment, and 0 else.

ROUNDpartners = round in partners treatment, and 0 else.

LAST2ROUNDstrangers = 1 if round 29 or 30 in strangers treatment, and 0 else.

LAST2ROUNDreputation = 1 if round 29 or 30 in reputation treatment, and 0 else.

LAST2ROUNDpartners = 1 if round 29 or 30 in partners treatment, and 0 else.

RHO

(random effects) Number of observations 2160 Log-likelihood -1087.77 .0000 !2 p-value a Analogous estimates for fixed effects linear models are given in Appendix B. b History for Partner’s buyers does not include last transaction.

0.524 (.0001) 0.852 (.0011) 0.0616 (.0014) -0.124 (.0144) 0.212 (.2111) -0.646 (.0005) 1.330 (.0000) -.697 (.0100) -0.005b (.8386) -0.386b (.0000)

TOTALSHIPreputation

CBSH

Model 3

In the and same way as weindicated can also notice what we named a "loss confident reputation strangers, by the treatment dummy PARTNERS, but controlling foreffect" with the CBHN. It's mean that when you were disappointed in the past by one, end-game the trust shownfrom by partners remarkably stable over time: The sellers two or effects even more cheating sellers, isyou have some difficult to trust after. ROUNDpartners coefficient is small and not significant. ROUNDreputation is also small, but

14

Addressed to Professor Masclet and Professor Denant-Boèmont

9

condition their buying decision on shipping history. (Models in Table 2, to be discussed in a moment, show this more formally.) Figure 5. Marginal trust conditioned on last feedback across treatments*

marginal effect on probability of trust

20%

Strangers

Reputation

Partners

10% 0% -10% -20%

untrustworthy trustworthy

-30% -40% -50%

* The base rate (the zero line) is the average buy over all encounters for each treatment separately (37.08 percent in strangers, 55.56 percent in reputation and 83.22 percent in partners).

This kind of conditional buying is rational since the seller’s history has predictive power In order to measure the improvement in trust, the author compared Stranger for his future 3 presents a random effect probit We can see that Market andperformance. ReputationTable market. They remarked that for thesellers. last feedback is very important market the stranger market. shipping the on last the time reputation both a reputation and than partneron market seller received a buy Buyers order is atrust with a probability of 33 % if seller no ship the last time and rose to 65% if the significant predictor of whether will do so this time. (The coefficient for the seller ship the last time. the So seller the reputation mechanism may improve number of transaction. LASTSHIPstrangers is significant as well but with a negative sign.) Further, a last decision to ship is more highly predictive of shipping this time in partners than in reputation markets (two-

They also the Partner Market and the Reputation Market in order to tailed pcompared = 0.0121, Wald test). see what is good and what is wrong. Comparing to the partner market, the feedback do not work perfectly on the reputation market. They explained that by the fact that when you send a feedback on the reputation market it profit to everybody. When you send a signal on a partner market, it will benefit only to the one who sent it. In fact we can16consider that the signal is a public good and people , so people have to deal with free-riding problem. We can say that feedback in reputation market is costly than in the partner one. An other thing which can explain the less efficiency of reputation market is the fact that seller history is diluted by the own buyers history. Concerning the pay-off, there is a positive correlation between the trust and the payoffs. That mean the payoffs are higher in reputation market than in the stranger market.

Addressed to Professor Masclet and Professor Denant-Boèmont

10

Some Criticism about this study: 

Figure A1. Buyer screen

We have few criticism to do about this study. First of all, players see a feedback of all previous transaction of their temporary partners. Results might be different if they see a global grade which is more common in reality even if people can see all feedback by a second click. Not all buyer go to see the historical of their seller which is very different because when you build your trust on a grade you don't know if the seller had a good behavior in his last transaction, maybe because you don't care or you don't think about this... An other criticism could be that this study don't care about the price of the exchanged product, a buyer would spend more time to inspect his seller feedback history if he want to buy to him a 1000$ good whereas for a 10$ good the opportunity cost is to high. If a product is very rare, if only one seller sell it and if a buyer really want it he may risk his money even if the seller haven't got a good feedback history. 30 The study venture the hypothesis that post a feedback is not beneficial to the buyer. It neglect the social aspect of the Internet. It's socially enhancive for people to produce information, it's one of the major motor of web 2.0.

Conclusion After this study the authors take away two main fact. They conclude first that a market works better with a reputation system. This observation is strengthen by the strong end game effect which involved that their some strategies played during the game. The second point which they keep in mind is the public characteristic of the feedback that explain why the partner market works better than the reputation one. Authors also noticed that there may be newbies problems due to the non-cost of change your identity on internet. There is also the emotional side with revenge effect and people generally stopped their choice when they are satisfied and not when they reach optimum choice.

Addressed to Professor Masclet and Professor Denant-Boèmont

11