CORE Metadata, citation and similar papers at core.ac.

uk
Provided by Open Access LMU

Discussion Paper No. 465

Identity changes and the

efficiency of reputation systems

Matthias Wibral *

* University of Bonn and IZA Bonn

Financial support from the Deutsche Forschungsgemeinschaft through SFB/TR 15 is gratefully acknowledged.

Sonderforschungsbereich/Transregio 15 · www.sfbtr15.de
Universität Mannheim · Freie Universität Berlin · Humboldt-Universität zu Berlin · Ludwig-Maximilians-Universität München
Rheinische Friedrich-Wilhelms-Universität Bonn · Zentrum für Europäische Wirtschaftsforschung Mannheim

Speaker: Prof. Dr. Klaus M. Schmidt · Department of Economics · University of Munich · D-80539 Munich,
Phone: +49(89)2180 2250 · Fax: +49(89)2180 3510
Identity changes and the efficiency of reputation systems∗

Matthias Wibral∗∗
University of Bonn and IZA Bonn

JEL classification: C91, D02, L14

Keywords: trust; reputation; identity changes

Abstract: Reputation systems aim to induce honest behavior in online trade by pro-
viding information about past conduct of users. Online reputation, however, is not
directly connected to a person, but only to the virtual identity of that person. Users can
therefore shed a negative reputation by creating a new account. We study the effects of
such identity changes on the efficiency of reputation systems. We compare two markets
in which we exogenously vary whether sellers can erase their rating profile and start
over as new sellers. Buyer trust and seller trustworthiness decrease significantly when
sellers can erase their ratings. With identity changes, trust is particularly low towards
new sellers since buyers cannot discriminate between truly new sellers and opportunis-
tic sellers who changed their identity. Nevertheless, we observe positive returns on
buyer investment under the reputation system with identity changes, and our evidence
suggests that trustworthiness is higher than in the complete absence of a reputation
system.

∗
Financial support from the German Research Foundation (grant KR 2077/2-1 and SFB/TR 15) is
gratefully acknowledged. The author would like to thank Johannes Abeler, Steffen Altmann, Christine
Harbring and Armin Falk for many insightful discussions and Markus Antony, Holger Gerhardt, Alexander
Koch, Sebastian Kube, Rosemarie Nagel, Axel Ockenfels, Gert Pönitzsch, Mirko Seithe, Dirk Sliwka, Florian
Zimmermann and seminar and conference participants at Alicante, Amsterdam, Bonn, Cologne, Graz,
Milan and Zurich for helpful comments.
∗∗
Institute for Applied Microeconomics, University of Bonn, Adenauerallee 24–42, 53113 Bonn, Ger-
many. E-mail: [email protected], phone: +49 228 73-9480, fax: +49 228 73-9239.
1 Introduction
Online trade usually takes place between strangers, payment occurs before the good is
shipped, and legal enforcement of an agreement may be prohibitively costly. This partic-
ular constellation generates moral hazard and adverse selection problems, especially on
the seller side. Sellers have an incentive to ship a good of lower quality than promised
or not to ship at all. Reputation systems are the most important tool in e-business to
address these problems and to induce honest behavior among users. They store infor-
mation about past conduct of a user provided by other users in a reputation profile and
disseminate this information to the whole community. In principle, a reputation profile
thus allows buyers to distinguish honest sellers from dishonest ones and to interact only
with the former.
The successful use of reputation systems in traditional markets with similar moral
hazard problems can be traced back at least to the beginnings of long-distance trade in
the Middle Ages (Greif 1989, Milgrom et al. 1990). Online reputation in most settings,
however, differs from reputation in traditional markets in a very important way. Online
reputation is only connected to the virtual identity of a person, i.e., the user name, and
not the person itself. After a bad rating, a dishonest seller can comparatively easily
abandon his old virtual identity and create a new account under a new user name
with no reputation attached to it. The implications of this distinctive feature of online
reputation for the efficiency of reputation systems are not well understood.
At first sight, identity changes seem to severely weaken the disciplinary power of
a reputation system. Dishonest behavior can still be punished with a negative rating
but these ratings may lose their edge when they can easily be shed by creating a new
identity. Real newcomers and sellers who have changed identity after a bad rating
become indistinguishable. This argument suggests a high frequency of identity changes
accompanied by opportunistic behavior and lower buyer trust. Theory on the other
hand also provides some guidance on why reputation systems might still work effectively
(Friedman and Resnick 2001, Ockenfels 2003). If buyers anticipate that dishonest sellers
start over as new players they will not trust newcomers or only interact with them at very
unfavorable conditions, e.g., low sale prices. Starting as a new seller may then become
so costly that cheating and creating a new identity is not profitable anymore. In this
case, we would observe a high level of seller trustworthiness and no identity changes.
Buyer trust would be high in transactions with experienced sellers and low towards new
sellers. This creates a negative externality for honest new sellers.
In this paper, we examine empirically how the option to change one’s virtual identity
affects the efficiency of markets and the performance of reputation systems in inducing
trust and trustworthiness. We study two experimental markets in which buyers play
the trust game (Berg et al. 1995) with varying sellers. Buyers can rate sellers after each
transaction. Before deciding whether and how much to trust a seller, a buyer can see the
rating profile of this seller. New players enter the market over time in both treatments.
The only difference is that sellers in one market can change their identity (change
treatment), i.e., erase their rating profile and start over as new players, while in the other
market this is not possible (no-change treatment). In view of the discussion above we
aim at answering the following questions: Do sellers use the opportunity to change their
identity? If so, does this go along with an increase in opportunistic behavior? How is

1
buyer trust affected? In particular, are new members treated differently when identity
changes are possible?
These questions are of high practical relevance given widespread reports that dis-
honest behavior remains an important problem for online interaction.1 According to
a survey conducted by the European Commission, 49% of internet users in the EU 27
countries are concerned about becoming a victim of online fraud, where goods are
not delivered, counterfeit or not as advertised, and 19% have actually become a victim
(European Commission 2012). In an earlier survey (European Commission 2004), 21%
of the EU 15 population reported abstaining from buying online because they did not
trust the internet. Recent studies on eBay’s reputation system (Reichling 2004, Jin and
Kato 2006, Dellarocas and Wood 2008, Klein et al. 2009, Bolton et al. 2013) also indicate
that the low percentages of neutral and negative ratings (e.g., Resnick and Zeckhauser
2002) grossly understate the true extent of problematic transactions since users may
be reluctant to provide negative feedback for fear of retaliatory ratings (Masclet and
Peenard (2012)). Dellarocas and Wood (2008) estimate the true figure of mildly or very
dissatisfied buyers at up to 21%.
It is difficult to determine whether these problems (also) arise because sellers can
shed a negative reputation comparatively easily. So far, there is little evidence on the
effects of identity changes, mainly because identity changes are inherently difficult
to observe in field data. In addition, even if one could observe identity changes, com-
paring the same reputation system with and without the possibility to change one’s
identity would be extremely difficult in the field. We therefore use a controlled laboratory
experiment to address our research questions.
The main findings can be summarized as follows. First, buyer trust and seller trust-
worthiness are significantly lower when sellers can change their identities. Trust is
especially lower for new sellers. However, the reputation system in the change treatment
maintains trustworthiness at a level that is high enough to make investing profitable
for the buyers. The evidence is at least suggestive that trustworthiness is also higher
than in the complete absence of a reputation system. Second, the basic principles of
reputation systems function in both treatments. Lower trustworthiness translates into
lower ratings, and buyers trust sellers with lower ratings less. Third, in the change treat-
ment sellers with a sufficiently bad reputation circumvent lower buyer investments by
switching identity and starting over as a new seller. Incentives to be trustworthy are
therefore lower. Sellers seem to choose between two kinds of behavior—being trust-
worthy and maintaining a fixed identity, and behaving opportunistically and changing
identity multiple times. The latter group drives the overall lower trustworthiness in the
change treatment. Finally, opportunistic behavior does not pay off in the no-change
treatment. In the change treatment, however, opportunistic players do earn more than
their counterparts who share more equally with the buyer.
We believe that this work contributes on several fronts. The results for the change
treatment demonstrate that the efficiency of reputation systems in inducing trust and
trustworthiness is reduced if sellers have the opportunity to change their identity. While
the negative effects of identity changes have long been discussed theoretically in the
literature (e.g. Dellarocas and Wood 2003) this paper provides controlled evidence
1
See, for example, the reports prepared by the National Consumer League (www.fraud.org) or the
Internet Crime Complaint Center (www.ic3.gov).

2
and establishes a causal link between the possibility of identity changes and a higher
incidence of dishonest behavior. This is the main contribution of our paper.
Our results for the no-change treatment complement the findings of previous labora-
tory experiments on reputation systems without identity changes. Keser (2003), Bolton
et al. (2004) and Masclet and Peenard (2012) find that a reputation system leads to high
levels of trust and trustworthiness. We extend this finding to an environment in which
new players enter the market over time: The mere fact that there are new players need
not affect the efficiency of a reputation system as long as buyers can be sure that a new
seller really is a person who has just entered the market.
In addition, the controlled environment of the laboratory allows us to take a closer
look at the mechanisms through which identity changes influence market outcomes.
Several of our specific findings, which we can causally attribute to potential identity
changes, are consistent with observations from online platforms. For example, Ockenfels
(2003) finds a very low percentage of sellers with more negative than positive ratings
in a sample from the platform half.com. In the same sample, new sellers also ask for
lower prices than experienced sellers. Resnick et al. (2006) conduct a randomized field
experiment on eBay, in which matched pairs of vintage postcards are sold via a high
reputation identity and accounts with little or no previous feedback. Buyers are willing
to pay 8% more to the high reputation seller. While the design of their study nicely
controls for many potential confounds that plagued earlier empirical studies, it is still
unclear to which degree the effect is driven by personal experience of repeat customers
of the high reputation identity. We can rule out this and other potential confounds in
our setup. Finally, our results also show that a substantial part of the negative effects
of a system which allows identity changes may arise from buyers simply refusing to
participate in market interaction, thus effectively dropping out of the market. In this
aspect, we complement survey evidence indicating that large effciency losses occur
because consumers do not trust online trade(rs).
The rest of the paper is structured as follows. Section 2 describes the experimental
design. Section 3 reports and discusses our results. Section 4 concludes.

2 Experimental Design
2.1 Trust Game and Reputation System
We use the trust game, originally introduced as investment game by Berg et al. (1995),
to mimic essential features of online interaction in our experiment.2 The trust game is
played by two players, which we refer to as buyer and seller in our setup. Both players
have an endowment of 10. In the first stage, the buyer can send an amount between 0
and 10—the investment—to the seller. On the way to the seller this amount is tripled by
the experimenter. In a second stage, the seller can then decide how much (if anything at
all) of the tripled amount he wants to send back to the buyer.
2
Two baseline paradigms have been used to study reputation systems. Keser (2003) and Masclet and
Peenard (2012) who also use the standard trust game, while Bolton et al. (2004) use a binary variant
(“buyer-seller game”), in which efficiency gains only arise when trust is honored. We use the standard
trust game for our main analysis because it allows studying certain types of strategic seller behavior.
For example, we can study whether sellers build up their reputation when it is relatively cheap to do so
and then “milk” this reputation later on. However, we also test the robustness of our results using the
buyer-seller game (see section 3.2.2).

3
 The trust game captures the basic moral hazard problem inherent in a broad class
of social interactions. There are efficiency gains to be realized but this requires that
the buyer sends money without a guarantee of receiving anything in return from the
seller. The buyer’s investment can therefore be taken as a measure of his trust while the
amount that the seller sends back (as a share of what he received) reflects the latter’s
trustworthiness. More precisely, our measure of trustworthiness is return on (buyer)
investment which we define as [(Amount Received/Amount Sent) − 1]. Trustworthiness
can, of course, only be measured when the seller was trusted in the first place, i.e., in
those instances when buyer investment is greater than zero.
In our experiment subjects play the trust game described above, half of them in the
role of buyers and the other half in the role of sellers. Every participant keeps his role
throughout the entire experiment which lasts 20 rounds. This is common knowledge.
Buyers and sellers are randomly rematched after every round. The matching mechanism
ensures that players never play with the same person in two consecutive rounds, which
is common knowledge as well.
Based on Keser (2003), we enhance the trust game with a reputation system in the
following way. At the end of each round, after the seller has decided how much to return
to the buyer, the latter has to rate this decision.3 He can give the seller a positive, neutral
or negative rating. If there was no interaction, i.e., the buyer did not send anything to
the seller, no rating can be made. The rating is stored in the seller’s rating profile.
At the beginning of each round, the seller sees his rating profile. It lists four categories:
the last rating and the total number of positive, neutral and negative ratings. In the
no-change treatment every seller keeps his rating profile for the whole experiment and
therefore is simply informed about his current profile. Then the buyer sees the seller’s
rating profile and decides how much money to send to the seller. In the change treatment
the seller can decide before the buyer sees his rating profile whether he wants to erase
the profile and start as a new player. This profile erasure is what we refer to as an identity
change. If the seller keeps his rating profile it is then shown to the buyer. If the seller
erases his profile, the buyer will only be given the information that he is matched with
a new player before making his investment decision. Identity changes are free to the
seller.
If buyers in real online markets face a new seller they are unable to tell whether this
person really is a newcomer or a (dishonest) seller who has abandoned his old account
and created a new one. To achieve this crucial feature in our laboratory experiment we let
new subjects enter the experiment over time. Only six subjects (3 buyers, 3 sellers) start
the experiment right from the beginning in every session. The others enter in the course
of the experiment. The timing of the entry of new players is the same for all sessions in
both treatments. One buyer and one seller each enter in rounds 4, 8 and 13, respectively.4
Players who entered the experiment at a later stage had to do a paid real effort task
3
We thus abstract from the public good problem of voluntary feedback provision. Gazzale and Khopkar
(2011) study a setup without identity changes, in which buyers can decide whether they want to leave
feedback or not. Since there is no scope for opportunistic buyer behavior in our setup, sellers cannot rate
buyers in our experiment.
4
We chose this timing of entry to minimize the number of repeat encounters. In the post-experimental
questionnaire, only 2 out of 192 subjects indicated that they thought they had recognized someone whom
they had played with before.

4
which consisted of counting certain letters in a short text. We introduced the real effort
task to keep subjects busy and to avoid income effects. Payments were calibrated such
that earnings from the real effort experiment were not significantly different from those
of the participants who played the trust game right from the beginning. Subjects know
that new players enter the experiment but not when and how many. When faced with a
new seller buyers in the change treatment thus cannot discern whether the seller really
is a new player or has just erased his rating profile.
In addition to our two main treatments, we conducted a control treatment without
a reputation system to establish baseline levels of trust and trustworthiness. Apart from
the reputation system the control treatment is completely identical to the two other
treatments in all aspects (matching, players per session, entry of new players, etc.).
Subjects thus play a series of one-shot trust games with changing partners.
The experiment was programmed using z-Tree (Fischbacher 2007) and conducted
at the BonnEconLab at the University of Bonn. Eight sessions of each treatment with
12 subjects each were played, yielding a total of 288 participants. At the beginning
of each session participants were seated in separate cubicles. Before the experiment
started the experimenter read out the instructions and answered all remaining questions
privately.5 After the experiment, subjects answered a short questionnaire with free form
questions regarding the motivation for their decisions. All subjects were undergraduates
from different fields recruited via announcements posted on campus. The experiment
lasted approximately one hour and subjects earned 12.35 Euro on average, including a
show-up fee of 2 Euro.

2.2 Related literature and behavioral predictions

In the trust game presented above, the buyer does not send anything to the seller in the
first place when agents only care about their own material payoffs. This follows from the
standard unravelling argument. The introduction of a reputation mechanism does not
change this result as long as the game is finitely repeated, independent from whether
identity changes are possible or not. Based on this standard model we would therefore
expect no differences between our treatments.
Previous experiments with finite repetitions of a trust game (Keser 2003, Bolton
et al. 2004, Bohnet et al. 2005, Masclet and Peenard 2012) suggest that this is not a
good description of actual behavior. In these experiments, a reputation system without
identity changes substantially reduces seller moral hazard and improves efficiency. This
finding can be rationalized by models which assume different types of players (Kreps
et al. 1982). Brown et al. (2004) provide an example in which some players are inequity
averse. If some sellers are intrinsically trustworthy in our setup it can be profitable
for opportunistic sellers to camouflage as trustworthy at least until the final rounds of
play.6 Another potential explanation for the success of reputation systems in previous
experiments is that subjects are myopic in the sense that they fail to fully account for
5
A translated version of the instructions can be found in the appendix.
6
The behavioral reputation literature starting with Camerer and Weigelt (1988) tests the predictions
of this type of reputation models by exogenously introducing uncertainty about sellers’ preferences via
the experimental design. The results of this literature also suggest that other-regarding preferences or an
intrinsic concern for appropriate behavior play an important role in understanding reputation (Grosskopf
and Sarin 2010).

5
the finite nature of the experiments.
The only models which explicitly consider the effects of identity changes (Friedman
and Resnick 2001, Ockenfels 2003) use a setup with an infinite time-horizon. They stand
in the tradition of earlier work (Kandori 1992, Okuno-Fujiwara and Postlewaite 1995)
demonstrating that the seller moral hazard problem can be completely solved with a
reputation system when identity changes are not possible, even in large communities,
in which agents only interact infrequently and enter and exit over time. The simple
intuition is that sellers, who behave opportunistically, can be punished later on by
other members of the community if past conduct is common knowledge. With identity
changes, direct punishment of opportunistic players is not feasible anymore since
players with a bad reputation can—in anticipation of punishment—simply reenter
the game as new players with no reputation at all. However, even though the direct
link “opportunistic behavior–bad reputation–future punishment” is severed, this need
not lead to a complete breakdown of cooperation. Both models of identity changes
(Friedman and Resnick 2001, Ockenfels 2003) show that an equilibrium characterized
by high levels of cooperation can be sustained.
The idea behind this result is straightforward. Since there is no way to distinguish
between truly new players and opportunistic players, who have reentered the game,
new players will not be trusted and receive unfavorable treatment from experienced
players. This is referred to as making new players “pay their dues". In an online auction
setup, for example, new sellers could receive a lower sales price for their goods. If this
unfavorable treatment of new players is sufficiently severe, i.e., the dues for newcomers
are high enough, opportunistic behavior and subsequent identity changes do not pay
off anymore. These “pay your dues” strategies lead to a welfare maximizing (second-
best) equilibrium, in which no one behaves opportunistically and changes identities.
Efficiency losses compared to a situation without identity changes can arise because
truly new players “undeservedly” also receive unfavorable treatment from experienced
players. The insights in Friedman and Resnick (2001) and Ockenfels (2003) are derived
for an infinite time horizon, but the key mechanisms may also be informative for our
finite setup, in particular, if some agents are myopic or intrinsically cooperative.
Summarizing the theoretical arguments and previous experimental evidence, the
following hypotheses can be formulated. The two basic principles, upon which repu-
tation systems are based, should work in both reputation treatments: Opportunistic
behavior will translate into lower ratings and buyers will condition their investment
decision on the rating profile in the sense that lower ratings will lead to lower buyer
investments. As a consequence, we expect high levels of trust and trustworthiness in
the no-change treatment. The degree to which these carry over to the change treatment
crucially depends on the treatment of new players. Based on the models, we hypothesize
that buyer investment towards new sellers (and therefore investment overall) will be
lower in the change treatment compared to the no-change treatment. If this discount is
sufficiently big, the lower buyer investment could make identity changes unprofitable
and prevent sellers from behaving opportunistically and changing their reputation.

3 Results
In this section we first present the basic treatment effects. We then explore whether the
mechanisms described in the previous section can explain the treatment differences

6
 10
9
8
7 6
Investment
4 5
3
2
1

Control Change No−Change

0

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
Period

Figure 1: Investment over time

observed between the no-change and the change treatment.

3.1 Overall Treatment Effects

A reputation system should ultimately promote seller trustworthiness and, as a conse-
quence, buyer trust. As described above, we use buyer investment to measure trust and
return on (buyer) investment as a measure of sellers’ trustworthiness. Figure 1 shows
average buyer investment over time for the two reputation treatments and the control
treatment without a reputation system, including those instances in which the buyer
did not invest anything at all. Figure 2 displays return on buyer investment over time for
the different treatments.
We first examine the effect of identity changes by comparing our two main treat-
ments. Buyers trust more when sellers cannot change their identity. While buyers in the
no-change treatment invest 7.04 on average, they send only 6.00 to the sellers in the
change treatment. The difference is (weakly) significant using a non-parametric U-test
(p = 0.09) and when comparing the change to the no-change treatment using a random-
effects estimation of buyer investment on treatment dummies (Table 1, columns 1
and 2).7 Regarding the relative frequency of the different investment levels the biggest
treatment difference occurs for investments of zero, i.e., those instances in which buyers
withdraw from the market altogether and which are difficult to observe with field data.
Buyers in the no-change invest zero in 64 cases or 8% of all rounds.8 Strikingly, buyers
in the change treatment do not invest anything almost twice as often, namely in 117 or
15% of all rounds (U-test p=0.06).
A similar pattern can be found for trustworthiness. Return on investment is highest
in the no-change treatment where buyers on average make a return of 61% on their
investment. For every period including the last one, average return on investment is
7
All reported p-values are two-sided. Unless otherwise noted all non-parametric tests use session
averages as independent observations, i.e., N = 8 for each treatment. Table 4 in the appendix provides
detailed information for each session.
8
In total 98 rounds were played in each session (3 players in each role played 20 rounds, 1 player each
played 17, 13, or 8 rounds respectively). For each role there are thus (60 + 17 + 13 + 8) × 8 = 784 rounds per
treatment.

7
 1
.8 .6
Return on investment
0 .2 −.2.4

Control Change No−Change

−.4

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
Period

Figure 2: Return on investment over time

positive. The reputation system is less successful in inducing trustworthy behavior in

the market in which participants can erase their reputation. Return on investment is
significantly lower in the change treatment (U-test, p = 0.06; random-effects estimation
p = 0.02, columns 3 and 4 in Table 1). For example, the most extreme case in which
a buyer makes an investment but the seller does not return anything occurs 63 times
(8.8% of all investments made) in the no-change treatment but 102 times (15.3%) in
the change treatment. However, buyers in the change treatment on average still make a
substantial profit of 36% on their investment.9

Result 1: Trust and trustworthiness are significantly lower when sellers can change
their identities.

In the control treatment without reputation system buyers invest 5.16, and return
on investment is 16%.10 A main finding of previous experiments on reputation systems
without identity changes (Bohnet et al. 2005, Bolton et al. 2004 and Keser 2003) is
that a reputation system increases trust and trustworthiness compared to a situation
9
Note that, in principle, lower average trustworthiness in the change treatment could simply be driven
by lower buyer investments if sellers reciprocate lower investments with lower returns on investment. Two
analyses indicate that the treatment difference in our setup is not just due to the different distribution of
investments, but persists conditional on investment. First, we calculate a counterfactual average return on
investment by using the investment distribution of the change treatment and the returns on investment
observed for each investment level from 1 to 10 in the no-change treatment. The counterfactual return
on investment is 56% and thus substantially higher than the 36% actually observed. Second, we regress
return on investment in the two reputation treatments on a treatment dummy controlling for investment
in an random-effects estimation with standard errors adjusted for clustering at session level. While the
coefficient on investment is significant and positive, the treatment dummy remains economically and
statistically significant in this regression. Detailed results of both analyses are available upon request.
10
While some previous studies observe a negative or zero return on investment in trust games in a
stranger environment (e.g., Keser 2003), the results of our control treatment are in line with those of
a recent meta-study of 162 trust games regarding investment and return on investment (Johnson and
Mislin 2011), which finds an average investment of 5.00 and a small, but significant positive return on
investment of 0.11.

8
without reputation system. Our no-change treatment replicates this finding for an
environment in which new players enter over time (U-test control vs. no-change: trust
p=0.01; trustworthiness p < 0.01).

Table 1: Random-effects panel regressions of investment ((1) and (2)) and return on investment
((3) and (4)) on treatment dummies (“Control” and “Change”). “Last” is a dummy equal to 1 in
period 20. ∗∗∗ and ∗∗ indicate significance on the 1-percent and 5-percent level respectively.
Reported standard errors (in parentheses) are adjusted for clustering at session level.

Investment Return on investment

(1) (2) (3) (4)
Control -1.940** -2.068*** -0.461*** -0.478***
(0.76) (0.73) (0.09) (0.09)
Change -0.959* -0.998** -0.253** -0.246**
(0.53) (0.51) (0.11) (0.11)
Last -2.115*** -0.457***
(0.54) (0.14)
Control × Last 1.902*** 0.335*
(0.66) (0.18)
Change × Last 0.581 -0.156
(0.66) (0.19)
Constant 6.951*** 7.094*** 0.606*** 0.629***
(0.36) (0.34) (0.04) (0.05)
R2 0.046 0.057 0.056 0.073
N 2352 2352 2019 2019

In view of result 1 the question arises whether a reputation system under which
identity changes are possible still represents an improvement over a situation without
any reputation system. The evidence on this is mixed. While average investment is
higher in the change treatment compared to the control in all but two periods, the
difference is not significant (U-test p = 0.43; Wald test for equality of coefficients in
Table 1: column 1, p = 0.20; column 2, p = 0.16). This finding is unexpected in view of
the models discussed above. Compared to the control treatment, return on investment
in the change treatment is more than twice as high (36% vs. 16%). The difference is
not statistically significant using a U-test (p = 0.14), but is (weakly) significant in the
random-effects regressions (Wald test for equality of coefficients in Table 1: column 3,
p = 0.09; column 4, p = 0.05). The evidence thus is at least suggestive that the reputation
system with identity changes increases trustworthiness compared to an environment
without any reputation system.

3.2 The impact of identity changes

The results presented above suggest that the ability of reputation systems to foster
trust and trustworthiness seems to be negatively affected if sellers have the opportu-
nity to change their identity. In the following, we analyze how reputation information

9
influences behavior and also discuss whether the treatment differences between the
treatments with a reputation system can be traced back to the mechanisms discussed
in section 2.
3.2.1 Do the basic principles of reputation systems work?
A first prerequisite for the functioning of a reputation system is that buyer ratings
reflect seller behavior. In particular, opportunistic behavior should translate into a bad
reputation. We use ordered probit regressions of return on investment on the rating
given by the buyer for this return to investigate this question. We code a good rating as
1, a neutral rating as 0, and a negative rating as −1. Table 2 reports the regression results.
The coefficient for return on investment is positive and significant for both treatments,
also if we control for the level of investment and a variable indicating how long a buyer
has been in the experiment (“personal period”). In both treatments, sellers thus receive
lower ratings for lower returns.
To gain an impression of potential differences in rating standards across treatments
we calculate the return on investment thresholds which explain the highest percentage
of ratings for each treatment assuming that buyers rate a return on investment below a
threshold b neg as negative, between b neg and b neu as neutral, and above b neu as positive.
The optimal b neg is 0.19 in the change and 0.59 in the no-change treatment. The optimal
threshold between a neutral and a good rating, b neu , is 0.65 in the change and 0.88 in
the no-change treatment. These thresholds predict around 75% of choices in both
treatments.11 Overall, rating standards seem to be a bit lower in the change treatment.
Nevertheless, we observe a higher average rating in the no-change compared to the
change treatment. While in the no-change treatment positive (negative) ratings account
for 47% (30%) of all ratings, this figure is 41% (37%) in the change treatment. Using the
coding of ratings described above, the average rating in the no-change treatment is 0.17,
and 0.04 in the change treatment.
A second prerequisite for the success of a reputation system is that buyers condition
their actions on the reputation information they receive, i.e., reward a good reputation
with higher and punish a bad reputation with lower investments. Table 3 shows how
the reputation profile a buyer sees at the beginning of a period influences his invest-
ment decision in the two treatments. The table reports the results of random-effects
estimations with standard errors adjusted for clustering at session level. The results
indicate that buyers take a seller’s reputation into account when deciding how much
to invest. The coefficients for the number of good and bad ratings are significant and
have the expected sign. Bad ratings seem to have a stronger impact. On average, one
additional good rating increases investment by 0.17 in the no-change treatment whereas
the coefficient for an additional negative rating is -0.52 (0.21 and -0.75 in the change
treatment, column 1). A Wald-test confirms that the absolute values of the coefficients
are significantly different (no-change: p < 0.01; change: p < 0.01). A stronger influence
of negative ratings is in line with findings for laboratory data and field data from eBay.12
11
Since buyers did not make a decision for every potential return on investment only intervals can be
determined for the b neg and b neu , which explain the highest percentage of choices. All b neg ²[0.58; 0.6] and
all b neu ²[0.88; 0.88] are optimal in the no-change treatment (change: b neg ²[0.17; 0.2]; b neg ²[0.63; 0.66]).
The values reported represent the mid points of these intervals.
12
See Bajari and Hortaçsu (2004), Resnick et al. (2006) or Bolton et al. (2013) for references on the

10
Table 2: Buyer rating depending on seller return on investment, ordered probit estimates. “Per-
sonal Period” indicates how many periods a buyer has been in the experiment. ∗∗∗ and ∗∗
indicate significance on the 1-percent and 5-percent level respectively. Reported standard errors
(in parentheses) are adjusted for clustering at session level.

Dependent variable: Rating

no-change treatment change treatment
(1) (2) (3) (4) (5) (6)
Return on investment 2.271*** 2.282*** 2.363*** 1.513*** 1.510*** 1.514***
(0.409) (0.344) (0.371) (0.186) (0.190) (0.189)
Investment - 0.099** 0.105** - 0.053** 0.0524*
- (0.050) (0.050) - (0.026) (0.027)
Personal Period - - -0.028*** - - 0.007
- - (0.006) - - (0.014)
Log-likelihood -510.19 -499.26 -495.05 -462.79 -458.62 -458.35
N 720 720 720 667 667 667

Neutral ratings seem to be perceived as a positive signal. The coefficient for neutral
ratings is positive in both treatments, but only significant for the change treatment
(column 1, Wald test, p < 0.01). Overall, the coefficients of the interaction terms indicate
that ratings tend to have a slightly stronger (i.e., more positive for positive and neutral
ratings, and more negative for negative ratings) influence in the change treatment. How-
ever, the difference is only significant for neutral ratings and this result is not robust to
including a time trend (column 2). Note that the rating profile of a seller also listed last
period’s rating separately. Including dummies for a positive, neutral or negative rating
in the last period and the respective interaction terms does not qualitatively change the
results reported above (column 3). Regarding the influence of the last rating we find that
only a positive last rating has a significant additional impact on investment. However,
this impact is substantial, as a positive last rating increases investment by almost 1.3.

Result 2: The basic principles of reputation systems function in both treatments.

Buyers give lower ratings for lower returns on investment. They also base their
investment decision on the reputation profile of the seller. A worse reputation leads
to lower buyer investments.

Finally, we study whether a buyer’s personal experience matters for his investment
decision in addition to the reputation information he receives. To this end we include
“personal period” and a variable indicating how often a buyer made a loss on his invest-
ment in the past (“no. loss”) in the regression (column 4). The results indicate that own
negative experiences in previous rounds lead to a significant and sizable reduction in
buyer investment, especially in the no-change treatment. Opportunistic sellers thus

growing number of field studies on the influence of reputation information on outcomes in online trade.

11
create a negative externality. A similar “own history effect” has been found in Bolton
et al. (2004).

Table 3: Buyer investment depending on seller reputation. “Change” is a dummy variable equal
to 1 for the change treatment. “Good” is a variable containing the number of positive ratings
in the rating profile of the seller a buyer was matched with, “Neutral”, and “Bad” are defined
analogously. Variables with the prefix “Last” are dummy variables for a good, neutral, or bad
rating in the previous period. “No. Loss” contains the number of times a buyer has made a loss
so far. ∗∗∗ , ∗∗ , and ∗ indicate significance on the 1-, 5- and 10% level respectively. Reported
standard errors (in parentheses) are adjusted for clustering at session level.

Dependent variable: Investment

(1) (2) (3) (4)
Change -1.844*** -1.285*** -1.943*** -1.327**
(0.512) (0.489) (0.588) (0.513)
Good 0.165*** 0.242*** 0.109*** 0.252***
(0.027) (0.035) (0.027) (0.044)
Good × Change 0.049 0.023 -0.018 -0.005
(0.061) (0.060) (0.047) (0.072)
Neutral 0.088 0.176** 0.052 0.171**
(0.067) (0.075) (0.075) (0.085)
Neutral × Change 0.191* 0.119 0.119 0.128
(0.113) (0.114) (0.091) (0.120)
Bad -0.525*** -0.406*** -0.419*** -0.396***
(0.065) (0.066) (0.069) (0.075)
Bad × Change -0.226 -0.248 -0.043 -0.262
(0.161) (0.170) (0.132) (0.170)
Period - -0.115*** - -0.102*
- (0.030) - (0.053)
Last_Good - - 1.282*** -
- - (0.315) -
Last_Good × Change - - 0.889 -
- - (0.572) -
Last_Neutral - - 0.705 -
- - (0.542) -
Last_Neutral × Change - - 0.601 -
- - (0.638) -
Last_Bad - - -0.415 -
- - (0.069) -
Last_Bad × Change - - 0.490 -
- - (0.462) -
No. Loss - - - -0.831***
- - - (0.211)
No. Loss × Change - - - 0.579**
- - - (0.251)
Personal Period - - - 0.046
- - - (0.045)
Constant 7.355*** 8.021*** 6.826*** 7.982***
(0.280) (0.317) (0.393) (0.321)
N 1568 1568 1568 1520
R2 0.187 0.206 0.223 0.222

12
3.2.2 Identity changes, the quality of new sellers, and consequences for buyer trust
We hypothesized above that investment towards new sellers would be lower in the
change treatment. Note that the treatment dummy in the regression in column 1 of
Table 3 represents the difference in investment towards sellers without any ratings
between the no-change and the change treatment. As predicted players without a
reputation receive a substantially lower investment in the change treatment.13

Result 3: Investment towards new players is substantially lower in the change

treatment.

In the models of reputation systems with identity changes described in section 2.2,
new players receive the same treatment as a player with the worst possible reputation.
Changing identity is not a profitable strategy and no identity changes occur in equi-
librium. We therefore look at how new sellers are treated compared to those with a
bad reputation in the change treatment. We calculate the aggregate reputation of a
seller by subtracting the number of bad ratings from the number of good ratings. In
the change treatment, sellers with an aggregate reputation of −1, i.e., with exactly one
more negative than positive rating, received 3.29, those who had an equal number of
positive and negative ratings received 4.89, and those with an aggregate reputation of
1 received 6.79. As a comparison, buyers invested 4.88 when they were facing a new
player. According to this, admittedly coarse, analysis strategic sellers should thus switch
their identity when their aggregate reputation reaches -1. While identity changes are not
readily observable in the field our data allow an analysis of whether and when sellers
actually change their identity. On average, sellers erase their reputation profile every 3.5
rounds, i.e., in 28.4% of the rounds.14 An analysis of when sellers erase their reputation
indicates that they understand that there is scope for a strategic use of identity changes.
63% of identity changes occur when the aggregate reputation is −1, another 25% when
it is 0.15
There is also tentative evidence for strategic behavior in terms of reputation dy-
namics. In view of the difference in investment towards new players and those with
an aggregate reputation of 1 shown above, opportunistic players have an incentive to
strategically build up a reputation in order to ‘milk’ this reputation later on. This holds
for those instances, in which they can buy a good rating cheaply, i.e., when they receive
a small investment. Two observations suggest that at least some sellers in the change
treatment employ this strategy. First, there is a negative (though not significant) corre-
13
Players without reputation include those players who entered as a new seller in the previous round
but did not receive any investment. If we only consider players with the label “new” the difference in
investment is even larger (see below).
14
For the analysis of identity change behavior we exclude the first round of each seller since identity
could not be changed in this round. 736 rounds are thus included in the analysis.
15
The conditional probability of changing identity is also highest for an aggregate reputation of −1.
86% of the sellers with an aggregate reputation of −1 change their identity. Note also that the substantial
positive impact of a positive last rating on investment found above could make a strategy profitable where
sellers split equally in one period to receive a positive rating, behave opportunistically in the next, and
change their identity after having received a negative rating. This could explain why 25% of sellers change
their identity at an aggregate reputation of zero.

13
 1
Return on investment
.5
Life length
1
2
3
4
5
0 −.5

1 2 3 4 5
Period under current identity

Figure 3: Return on investment for identities of different length

lation between investment and return on investment for new sellers (Spearman rank
order correlation, ρ = −0.11, p = 0.15, N = 181), whereas this correlation is positive and
significant overall (ρ = 0.11, p = 0.01, N = 644). Second, we study return on investment
dynamics over the life span of a seller identity. For example, a seller who participates in
the experiment right from the beginning and changes identity every fifth period has four
identities of five periods each. Figure 3 shows average return on investment per period
for identities of different length (1-5 periods). Return on investment in the first period
of an identity is negative for identities which only last one period but starts out above
zero for those who stay with their current identity for more than one period. In addition,
we see positive returns on investment for the early periods of multi-period identities
and a negative return for the last at about the level observed for one-period-identities.
These findings are in line with anecdotal evidence from eBay.16
Finally, we confirm that sellers who behave opportunistically and then change
their identity are driving the lower trustworthiness in the change treatment. Return
on investment and the frequency of identity changes are negatively correlated on an
individual seller basis (Spearman rank order correlation ρ = −0.67; p < 0.01; N = 48).17
Sellers can roughly be divided into two groups—those who change their identity at most
once and exhibit comparatively high returns on investment, and those who change their
identity more frequently and return little to the buyer.

Result 4: Sellers in the change treatment use the opportunity to strategically change
identities. Frequent identity changes are associated with low returns on investment.

The previous analyses suggest that the composition of the pool of new players differs
drastically between the treatments. In contrast to the no-change treatment, sellers with
16
Jin and Kato (2006), for example, observe several sellers who build up their reputation first and then
fail to deliver on a large number of parallel auctions in the eBay submarket for unrated baseball cards.
See Liu (2011) for a model which produces similar reputation dynamics.
17
The correlation is also significant if we take each session as an independent observation (Spearman
rank order correlation ρ = −0.90; p < 0.01; N = 8).

14
the label “new” in the change treatment are a mix of truly new sellers (10%) and sellers
who erased their rating profile (90%). Average return on investment by real new sellers
is 40%.18 In contrast, the remaining 90% of the sellers labeled as “new”, i.e., those who
changed their identity, display an average return on investment of −10%. Overall, we
observe a significant difference in the behavior of sellers labeled as “new”, and those
with a reputation profile (two-sided Wilcoxon test p = 0.02). While buyers on average
make a loss when they interact with new sellers in the change treatment, interactions
with sellers with a reputation profile yield a return on investment of 52%.
Taking the returns on investment presented in the previous paragraph as given
it would be optimal for buyers in the change treatment not to invest anything when
facing a new seller. However, while buyers in the change treatment do indeed invest less
towards new sellers compared to the no-change treatment (4.88 vs 7.38; two-sided U-
test, p = 0.01), the investment is still substantial.19 As shown above, sellers can therefore
avoid the negative consequences of opportunistic behavior to some extent when identity
changes are possible: new sellers on average receive a higher investment than those with
a negative aggregate reputation and about the same as sellers with a neutral aggregate
reputation. Compared to the no-change treatment, incentives to behave trustworthy are
therefore lower. An important question in this respect is whether a reputation system
with identity changes can still prevent that opportunistic sellers make a higher profit
than trustworthy sellers. We therefore turn to the analysis of payoffs in the next section.
3.2.3 Payoffs
Buyers generate a substantial surplus in both treatments by investing more than the
Nash prediction of 0, but the surplus generated is not distributed equally. In line with the
trust game literature, buyers on average earn less than sellers in both treatments. Both,
buyers and sellers, are nevertheless better off compared to their initial endowment. In
the no-change treatment buyers earn 14.49 per round while sellers earn 19.58 (Wilcoxon
test p = 0.01). The difference is even larger for the change treatment where the respective
figures are 12.33 for buyers and 19.67 for sellers (Wilcoxon test p = 0.01). There is no
significant difference between seller profits in the two treatments (U-test p = 0.92),
but buyers earn significantly less when sellers can change identity (U-test p = 0.04).
The efficiency losses arising from lower investments in the change treatment are thus
entirely borne by the buyers.20
For judging the performance of a reputation system it is important to consider
the payoff consequences of different types of behavior. Ideally, a reputation system
prevents that opportunistic agents make a higher profit than honest ones. In the models
described above this desirable property is achieved in equilibrium. Our final analysis is
therefore a simple test of whether bad behavior is profitable. We classify a seller as “bad”
(“good”) if he has collected more bad than good ratings (more good than bad) in the
18
Given that all players were new players at the start of the experiment and the label could thus not
carry any information we exclude the first period from the analyses in this paragraph. The qualitative
results are robust to including these observations.
19
One reason why buyers might not treat new sellers in the worst possible way could be that they are
guided by “preferences for appropriate choices" (Grosskopf and Sarin 2010).
20
This result also holds for the control treatment where buyers earn 11.07 and sellers earn 19.25.

15
course of the experiment.21 In the no-change treatment, good sellers earn 19.22 and
bad sellers earn 19.75 on average, but the difference is not significant (Wilcoxon test
p = 0.21). In the change treatment, good sellers earn 18.39 on average, bad sellers make
20.74. In contrast to the no-change treatment, this difference is significant (Wilcoxon
test p = 0.02). The reputation system in the change treatment thus does not succeed in
making good behavior at least as profitable as bad behavior. Since opportunistic players
camouflage as new players, all new players receive lower investments. But the reduction
for new sellers is not large enough to offset the lower returns by opportunistic sellers.22

Result 5: Opportunistic sellers do not earn more than sellers with high returns
in the no-change treatment. However, opportunistic behavior does pay off in the
change treatment.

3.2.4 Robustness Check

Two different designs have been used to study reputation systems, the standard trust
game (Keser 2003, and Masclet and Peenard 2012) and a binary variant called the
buyer-seller game (Bolton et al. 2004). The main difference between the two games
is that efficiency gains in the standard trust game arise once the buyer trusts while
in the buyer-seller game they only arise when trust is honored. In addition, Bolton
et al. (2004) abstract from a reputation system based on subjective ratings and instead
provide the history of past actions of a seller. To check wether our results also hold in
this environment we ran two additional buyer-seller control treatments (N = 48 for
each treatment). Subjects either participated in the no-change or the change treatment,
but, instead of the trust game, played the buyer-seller game from Bolton et al. (2004).
Payoffs and move structure in this game are as follows. Buyers can choose buy/not buy,
and sellers can choose ship/not ship if the buyer has decided to buy. If the buyer does
not buy, both players earn 35. If the buyer buys and the seller ships, both earn 50. If
the buyer buys and the seller does not ship, the buyer earns 0 and the seller 70. The
feedback mechanism showed the history of a seller’s actions.
To keep the setup in our buyer-seller control treatments as similar as possible to
our original treatments, everything else was kept equal, i.e., number of participants per
session, entry and timing of new players, number of rounds, and possibility to erase
one’s history in the change treatment. For the same reason buyers and sellers did not
switch roles, in contrast to Bolton et al. (2004).
Figure 4 and Figure 5 in the appendix show the results for buyer and seller decisions.
The main results from the trust game treatments are replicated in the buyer-seller game
setup—trust and trustworthiness are lower in the change treatment. The average buy
21
There are 20 (23) bad and 27 (24) good sellers in the no-change (change) treatment. For our analysis,
we ignore the remaining seller in each treatment who ends the session with an equal number of good and
bad ratings. Including them either with the good or the bad sellers does not change the results.
22
In view of this result one might ask why not more subjects choose to behave opportunistically and
change their identity afterwards. Analyzing behavior in the last round suggests that a substantial fraction
of sellers does not act purely out of strategic concerns. 11 out of the 34 sellers who receive an investment
in the last period provide a return on investment for the buyer of 50% or more, and thus seem to have an
intrinsic preference for trustworthy behavior.

16
rate is 81% in the no-change and 67% in the change treatment. The average shipping
rate conditional on a buy decision is 88% in the no-change and 69% in the change
treatment. The treatment difference for both measures is significant in a random-effects
probit regression of the respective measure on a treatment dummy (Buy: p = 0.01; Ship:
p = 0.02) or using a non-parametric U-test (Buy: p = 0.06; Ship: p = 0.02). Regarding
the treatment of new sellers we also observe a qualitatively similar pattern in the buyer-
seller control treatments. Buyers facing a new seller decide to buy in 96% of the cases in
the no-change treatment, but only in 52% of the cases in the change treatment. Overall,
the results from the control treatment clearly indicate that our main results hold for
both the trust game and the buyer-seller game.

4 Conclusions
In this paper, we studied the influence of identity changes on the efficiency of reputation
systems by comparing two markets. In both markets, new buyers and sellers enter over
time, but in one market sellers cannot erase their reputation while in the other market
they can. We find reduced seller trustworthiness and buyer trust in the latter market.
A substantial fraction of sellers uses the opportunity to behave opportunistically and
change their identity afterwards. Buyers cannot distinguish between these players and
those who are really new and react with lower investments towards players with the
label “new player” as predicted by (infinite-horizon) models of reputation systems
with identity changes. However, this reduction in investment is not large enough to
make opportunistic behavior with frequent identity changes an unprofitable strategy in
the change treatment. Nevertheless, return on buyer investment is positive under the
reputation system with identity changes and somewhat higher than in an environment
without reputation system at all.
An important implication of our findings for online market design is that any mea-
sure which reduces the uncertainty connected with the label “new player” and thus
increases the ability to distinguish between good and bad sellers can increase effi-
ciency.23 An obvious suggestion is to introduce fixed identifiers, another less invasive
measure is to give sellers the opportunity to have their identity verified by a third party.
However, users may value a certain degree of anonymity (Friedman and Resnick 2001).
The results in this paper thus need not be interpreted as arguments against allowing
identity changes in general. They rather suggest that doing so comes at a cost which has
to be weighed against the potential benefits of increased anonymity. These are likely to
depend on the environment in which the reputation system is to be used.
Our tentative finding of an improvement in trustworthiness between the control and
the change treatment hints at the potential reputation systems offer even in environ-
ments in which identities can be changed. An interesting question for future research is
whether introducing competition in the form of partner choice increases the efficiency
of reputation systems in these environments. On the seller side competition is likely
to create stronger incentives for trustworthy behavior (Bolton et al. 2008, Cassar and
Rigdon 2011). Competition on the buyer side (Keck and Karelaia 2011) could lead to a
greater buyer willingness to take risks with sellers whose reputation is not entirely posi-
23
Friedman and Resnick (2001) draw the same conclusion from their model and provide a discussion of
different measures to achieve this.

17
tive. This seems to hold in particular for new buyers (Jin and Kato 2006). Which of these
effects dominates is ultimately an empirical question. Our simple and parsimonious
framework can successively be enriched to study partner choice and other measures
such as verified IDs in future research.

18
References
Bajari, Patrick and Ali Hortaçsu (2004): “Economic Insights from Internet Auctions.”
Journal of Economic Literature, 42:457–486.

Berg, Joyce E., John Dickhaut and Kevin McCabe (1995): “Trust, Reciprocity, and Social
History.” Games and Economic Behavior, 10:122–142.

Bohnet, Iris, Heike Harmgart, Steffen Huck and Jean-Robert Tyran (2005): “Learning
Trust.” Journal of the European Economic Association, 3:322–329.

Bolton, Gary E., Ben Greiner and Axel Ockenfels (2013): “Engineering Trust: Reciprocity
in the Production of Reputation Information.” Management Science, 59:265–285.

Bolton, Gary E., Elena Katok and Axel Ockenfels (2004): “How Effective are Online
Reputation Mechanisms? An Experimental Investigation.” Management Science,
50:1587–1602.

Bolton, Gary E., Claudia Loebbecke and Axel Ockenfels (2008): “Does Competition Pro-
mote Trust and Trustworthiness in Online Trading? An Experimental Study.” Journal
of Management Information Systems, 25:145–170.

Brown, Martin, Armin Falk and Ernst Fehr (2004): “Relational Contracts and the Nature
of Market Interactions.” Econometrica, 72:747–780.

Camerer, Colin and Keith Weigelt (1988): “Experimental Tests of a Sequential Equilib-
rium Reputation Model.” Econometrica, 56:1–36.

Cassar, Alessandra and Mary Rigdon (2011): “Trust and Trustworthiness in Networked
Exchange.” Games and Economic Behavior, 71:282–303.

Dellarocas, Chrysanthos and Charles A. Wood (2003): “The Digitization of Word of

Mouth: Promise and Challenges of Online Feedback Mechanisms.” Management
Science, 49:1407–1424.

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

European Commission (2004): Issues Relating to Business and Consumer E-commerce.

Special Eurobarometer 60.0/Wave 201, European Opinion Research Group, Brussels.

European Commission (2012): Cyber Security. Special Eurobarometer 390/Wave EB

77.2, TNS Opinion & Social, Brussels.

Fischbacher, Urs (2007): “z-Tree: Zurich Toolbox for Ready-made Economic Experi-
ments.” Experimental Economics, 10:171–178.

Friedman, Eric J. and Paul Resnick (2001): “The Social Cost of Cheap Pseudonyms.”
Journal of Economics & Management Strategy, 10:173–199.

19
Gazzale, Robert S. and Tapan Khopkar (2011): “Remain silent and ye shall suffer: seller
exploitation of reticient buyers in an experimental reputation system.” Experimental
Economics, 14:273–285.

Greif, Avner (1989): “Reputation and Coalitions in Medieval Trade: Evidence on the
Maghribi Traders.” Journal of Economic History, 49:857–882.

Grosskopf, Brit and Rajiv Sarin (2010): “Is Reputation Good or Bad? An Experiment.”
American Economic Review, 100:2187–2204.

Jin, Ginger Zhe and Andrew Kato (2006): “Price, quality, and reputation: evidence from
an online field experiment.” The RAND Journal of Economics, 37:983–1005.

Johnson, Noel D. and Alexandra A. Mislin (2011): “Trust games: A meta-analysis.” Journal
of Economic Psychology, 32:865–889.

Kandori, Michihiro (1992): “Social Norms and Community Enforcement.” Review of

Economic Studies, 59:63–80.

Keck, Steffen and Natalia Karelaia (2011): “Does competition foster trust? The role of
tournament incentives.” Experimental Economics, 14:273–285.

Keser, Claudia (2003): “Experimental games for the design of reputation management
systems.” IBM Systems Journal, 43:498–506.

Klein, Tobias J., Christian Lambertz, Giancarlo Spagnolo and Konrad O. Stahl (2009):
“The actual structure of eBay’s feedback mechanism and early evidence on the effects
of recent changes.” International Journal of Electronic Business, 7:301–320.

Kreps, David, Paul Milgrom, John Roberts and Robert Wilson (1982): “Rational coop-
eration in the finitely repeated prisoners’ dilemma.” Journal of Economic Theory,
27:245–52.

Liu, Qingmin (2011): “Information Acquisition and Reputation Dynamics.” Review of

Economic Studies, 78:1400–1425.

Masclet, David and Thierry Peenard (2012): “Do reputation feedback systems really im-
prove trust among anonymous traders? An experimental study.” Applied Economics,
44:4553–4573.

Milgrom, Paul R., Douglass C. North and Barry R. Weingast (1990): “The Role of Institu-
tions in the Revival of Trade: The Law Merchant, Private Judges, and the Champagne
Fairs.” Economics & Politics, 2:1–23.

Ockenfels, Axel (2003): “Reputationsmechanismen auf Internet-Marktplattformen: The-

orie und Empirie.” Zeitschrift für Betriebswirtschaft, 73:295–315.

Okuno-Fujiwara, Masahiro and Andrew Postlewaite (1995): “Social Norms and Random
Matching Games.” Games and Economic Behavior, 9:79–109.

Reichling, Florian (2004): “Effects of Reputation Mechanisms on Fraud Prevention in

eBay Auctions.” Stanford University Working Paper.

20
Resnick, Paul and Richard Zeckhauser (2002): “Trust Among Strangers in Internet Trans-
actions: Empirical Analysis of eBay’s Reputation System.” The Economics of the
Internet and E-Commerce, 11:127–157.

Resnick, Paul, Richard Zeckhauser, John Swanson and Kate Lockwood (2006): “The value
of reputation on eBay: A controlled experiment.” Experimental Economics, 9:79–101.

21
Appendix A

1
.8
.6
Buy
.4
.2

Change No−Change
0

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
Period

Figure 4: Relative frequency of buy decisions in the buyer-seller treatments

1
.8
.6
Ship
.4
.2

Change No−Change
0

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
Period

Figure 5: Frequency of ship decisions in the buyer-seller treatments conditional on buying

22
 Treatment Session Inv Inv_Half1 Inv_Half2 ROI No_Inv Inv_10 Good Neutr Bad
Control 1 8.22 7.45 8.76 .45 2 1 - - -
Control 2 1.24 1.9 .79 -.19 60 7 - - -
Control 3 5.93 7.1 5.12 .35 15 5 - - -
Control 4 5.21 5.72 4.86 .28 14 7 - - -
Control 5 6.12 6.1 6.14 .15 1 2 - - -
Control 6 4.64 5.03 4.38 .1 13 4 - - -
Control 7 4.86 6.05 4.03 .05 19 0 - - -
Control 8 5.07 5.2 4.98 -.19 28 4 - - -
No-Change 9 7.27 8.25 6.59 .47 9 0 31 21 37
No-Change 10 8.27 8.45 8.14 .58 6 0 45 27 20
No-Change 11 8.47 8.25 8.62 .74 1 0 54 23 20
No-Change 12 5.59 6.47 4.98 .49 22 4 34 13 29
No-Change 13 6.71 7.88 5.91 .55 8 5 36 26 28

23
No-Change 14 7.46 6.6 8.05 .81 4 1 48 21 25
No-Change 15 6.13 6.8 5.67 .5 12 1 30 22 34
No-Change 16 6.4 6.8 6.12 .7 2 7 60 12 24
Change 17 4.98 5.3 4.76 .08 23 0 26 12 37
Change 18 5.82 5.9 5.76 .11 13 3 26 17 42
Change 19 4.82 5.5 4.34 .55 15 7 33 23 27
Change 20 5.07 6.07 4.38 .26 23 4 18 19 38
Change 21 7.51 7.7 7.38 .81 3 1 63 20 12
Change 22 7.16 7.25 7.1 .46 10 0 39 18 31
Change 23 5.28 5.53 5.1 -.06 23 2 23 17 35
Change 24 7.38 8.75 6.43 .55 7 2 48 16 27

Table 4: Descriptive statistics by session. Inv_Half1 and Inv_Half2 refer to investment in periods 1-10 and
11-20, respectively. No_Inv and Inv_10 state how often an investment of 0 or 10, respectively was made.
Good, Neutr and Bad describe how often the respective rating was given in a certain session.
Appendix B: Instructions for the change treatment
In what follows, we present a translation of the instructions for players in the change
treatment. The instructions for the no-change treatment differed only in the paragraph
regarding the possibility to erase one’s rating profile.

Instructions

There are two types of players in this experiment, player A and player B. You will be
assigned the role of either player A or player B, and you will keep your assigned role
throughout the entire experiment.
The experiment lasts 20 rounds. At the beginning of each round the number in the upper
part of the screen indicates the current round. In each round a player A is matched with
a player B. This matching process is random. However, you will never be matched with
the same player two rounds in a row.
Structure of each round:
At the beginning of each round each player is endowed with 10 Taler. Player A then
decides how many Taler of his endowment he wants to send to player B. Player A can
send any (integer) amount between 0 and 10 Taler.
The experimenter triples this amount, so that player B receives the tripled amount of
what player A decided to send.
Player B then decides how many Taler he wants to send back to player A. He can send any
(integer) amount between 0 Taler and the tripled amount transferred by player A. The
amount of Taler that player B sends back to player A is not tripled by the experimenter.
Rating:
After both players made their decisions, player A can rate player B’s decision as "good",
"neutral" or "bad". The rating is stored in the rating profile. There is no rating stage if
player A sends zero Taler to player B. In that case the phrase "No rating, because no
exchange took place" is displayed on the screen.
Before player A decides on how much to send to player B, the rating profile of the player
B he is matched with is displayed on the screen. The rating profile includes the rating
in the previous round as well as the total number of good, neutral and bad ratings that
player B has received so far.
Deleting the rating profile:
At the beginning of each round player B can delete his existing rating profile, and start
over as a new player.
New players:
During the course of the experiment new players enter the experiment. Players which
enter the experiment in later rounds find paper and pen at their desks and have to work
on a different task before starting as a player in the computerized experiment. These
players will also be paid for working on the non-computerized task at the end of the
experiment. They will find a more detailed description of the task at their desks. These
players will start with their decisions in the computerized experiment as soon as the
screen in their cubicle displays an entry mask.
The label "New player" is used for players who deleted their rating profile, as well as for
players who have just entered the experiment.

24
Questionnaire:
After the last round of the experiment a short questionnaire is displayed on the screen.
Please answer the questionnaire as precisely as possible. When you are done with the
questionnaire please wait in your cubicle until we ask you to collect your payment.
Payment: After the experiment the amount of Taler you earned is exchanged into Euros
at an exchange rate of 3,5 Cent/Taler.

25