Reputation and Cooperation in Voluntary
Exchanges: Comparing Local and Central
Institutions
T.K. Ahn Korea University
Justin Esarey Emory University
John T. Scholz Florida State University
Our experimental study compares the effectiveness of three reputation mechanisms believed to enhance cooperation. Groups of 14 subjects repeatedly select partners, play two-person prisoner’s dilemmas, and rely only on
individual experience to find trustworthy exchange partners in the baseline condition. The local condition represents emergent, bottom-up networks that allow partners to voluntarily share recommendations. The central
condition represents designed, top-down institutions that allow wide dissemination of recommendations provided
voluntarily. Surprisingly, the greater provision and use of information in the local condition supports the highest
level of cooperation, suggesting an unrecognized advantage of exchange networks over centralized institutions in
credibility and information provision.
E
very person, organization, agency, and state is
involved in voluntary relationships that present
the possibility for mutual gain at the risk of
betrayal. To cement these relationships, formal and
informal institutions can allow individuals to build
reputations and to select transaction partners carefully. These reputational mechanisms provide incentives for individuals to behave in a trustworthy
manner by making future opportunities contingent
on present choices. Thus, a well-functioning reputation mechanism can foster trust among the individuals affected by the reputation system by making trust
consistent with these individuals’ self-interest (Hardin
2002). As a consequence, reciprocal cooperators who
share information among themselves about other reliable partners can create networks of mutual reciprocity, a critical foundation for social capital (Coleman
1987; Granovetter 1985; Putnam 1993).
Central mechanisms that disseminate information beyond parochial relationships to a broader
audience can potentially magnify the effect of reputation. Institutions ranging in scope from the United
Nations and the European Union to local watershed
or economic development partnerships can potentially provide reputation-enhancing information
about members. The ability to share information
The Journal of Politics, Vol. 71, No. 2, April 2009, Pp. 1–16
Ó 2009 Southern Political Science Association
and hence expand reputations more broadly within
parties and caucuses in Congress, within regional
planning bodies or interdepartmental commissions
for governmental agencies and interest groups, or
within international organizations for nations may
dramatically increase the scope for mutually advantageous, self-enforcing agreements among individual
members of these bodies. This possibility has sparked
interest in institutions like medieval guilds (Greif,
Milgrom, and Weingast 1994), law merchants (Milgrom,
North, and Weingast 1990), international trade associations (Maggi 1999), and eBay’s feedback system
(Brown and Morgan 2006; Malaga 2001; Resnick
et al. 2006; Standifird 2001).
While formal models have demonstrated the
possibility that reputation can enhance cooperation
(e.g., Kandori 1992; Raub and Weesie 1990), less is
known about how different reputational institutions
condition the effect of reputation on cooperative outcomes. In particular, we do not know when central,
top-down information clearinghouses will outperform local, self-organizing, bottom-up systems. Can
the Better Business Bureau do a better job of enforcing consumer rights than word-of-mouth (Grinnell
1931)? Can central monitoring institutions like the
World Trade Organization or the International
doi:10.1017/S0022381609090355
ISSN 0022-3816
1
2
t.k. ahn, justin esarey, and john t. scholz
Atomic Energy Agency improve international cooperation beyond what can be achieved with purely
bilateral relationships (Greif 1989; Maggi 1999)?
Would information clearinghouses decrease rampant
business fraud better than reputations developed
by personal networks in newly emerging markets
(Radaev 2004)? Can blogs about the academic job market in political science (http://irrumormill.blogspot.
com/) provide a stronger reputation effect than traditional word-of-mouth?
Our analysis of central and local reputation
mechanisms focuses on the incentives to provide,
request, and use the information that is critical to any
reputational system. If providing information is
voluntary, individuals face the temptation to free
ride by using the information provided by others
without bothering to report their own experiences.
Thus, although reputation mechanisms can help
resolve strategic dilemmas that prevent cooperation
and broader exchanges, the solution itself involves an
additional information provision dilemma.
We argue that the relative performance of reputation mechanisms depends on their ability to resolve
this information provision dilemma. Local mechanisms that provide word-of-mouth information
through endogenously created networks can create
strong incentives to provide honest reputation information by embedding information exchange in ongoing relationships of mutual cooperation. But by
limiting information exchange to close associates,
these mechanisms spread information slowly and
hence slow the evolution of cooperation and expansion
of exchange. Central mechanisms can make information available to much wider audiences, but may offer
weak incentives to motivate adequate and honest
reporting.
We test the performance of these reputation mechanisms with a laboratory experiment that compares the
level of cooperation achieved under central and local
information-sharing mechanisms against a baseline
condition that allows no sharing of information. The
control afforded by a laboratory environment allows us
to examine each institution in isolation to observe
information provision and its impact on cooperation.
Reputation, Cooperation, and
Information Provision
In two-person repeated prisoner’s dilemmas, direct
past experience allows a player to establish a reputation with the partner that can support mutual
cooperation. Cooperation provides an expensive signal to the other player that is rational to send in
a wide variety of environments (Friedman 1971;
Fudenberg and Maskin 1986; Rubinstein 1979; Taylor
1976), even including the finitely repeated dilemma
(Kreps et al. 1982), as long as the likelihood that the
partner will cooperate conditionally is sufficient. The
power of what we call experiential reputation has been
validated in a number of experiments (Andreoni and
Miller 1993; Dal Bo 2005) and evolutionary studies of
cooperation (Axelrod 1984; Bendor and Swistak 1997;
Boyd and Lorberbaum 1987; Lorberbaum 1994).
The reputation literature generally focuses on
informational reputation in which information from
one game influences others playing other games,
thereby constraining behavior in anticipation of
future reputational effects. For example, both Boyd
and Richerson (1989) and Nowak and Sigmund
(1998) have proposed models of indirect reciprocity
in an environment of randomly mixed pairwise encounters. In these models, cooperative outcomes are
sustained by making cooperation with a present partner contingent on that partner’s cooperation with
his/her previous partners. These indirectly reciprocal
strategies can be evolutionarily stable, even if individuals never meet more than once, as long as
information about a player’s past follows him into
future encounters with new partners. Related studies
by Kandori (1992) and Janssen (2008) model information sharing in terms of labels or symbols that are
imperfect but reliable and illustrate that shared information can support cooperative outcomes by triggering punishment for defection.
Most directly relevant to our experiment is the
analysis by Raub and Weesie (1990) of repeated twoperson dilemmas among ‘‘neighbors’’ embedded in a
fixed spatial grid. They demonstrate that the range of
payoffs capable of sustaining cooperative equilibria
increases with the number of the partner’s choices in
other games that are reported to the player and with
the accuracy of these reports. They suggest that further studies need to investigate alternative systems
that distribute information beyond a player’s local
partners and in particular the ‘‘incentive problems
associated with the supply of information’’ (Raub
and Wessie 1990, 648).
In order to consider alternative reputational systems and the resultant incentive problems, our
experiments move beyond the available theoretical
results to incorporate three relatively novel features
we believe to be critical for understanding the influence of reputation systems operating in all realms
of institutional life: voluntary exchange relationships,
comparing local and central institutions
voluntary information provision, and alternative
information exchange mechanisms.
Voluntary exchange relationships. Most people
are involved not in fixed games between prisoners
who cannot exit the relationship, but rather in a
range of ongoing voluntary interactions with selected
partners (Tullock 1985). We therefore move beyond
the one-shot or repeated prisoner’s dilemma framework of past models to consider individuals who may
repeatedly encounter each other in two-person dilemma-type exchanges, but are free to select their
partners and to terminate unsatisfactory relationships. As previous experiments have observed, the
addition of an exit option can substantially enhance
the benefits of cooperation and constrain defectors by
denying future payoffs to them (Boone and Macy
1999; Orbell and Dawes 1993). Ule (2006) found that
cooperation levels were much higher in all experimental conditions allowing selection than in comparable conditions with fixed partners, especially when
subjects had full information of outcomes and the
cost of refusing to play with defectors was low.
The power of reputation in voluntary exchange
relationships extends beyond the better-known punishment-oriented behavioral constraints because selection advantages further enhance the benefit of a good
reputation. If reliable new business partners can be
found via recommendations from trusted associates, a
good reputation creates new and beneficial exchanges.
Exploiting current partners not only damages currently
ongoing games, but also limits future opportunities
with other players. In other words, the advantage for
conditional cooperators increases if reputational information not only increases the likelihood of continued cooperation in existing relationships, but also
reduces the risks involved in seeking additional partners. This first extension of the literature is particularly
important for reputation studies, because reputation
could operate primarily through selection and termination decisions rather than through behavioral modification within ongoing games.
Voluntary information provision. Unlike the
models referenced above, most real-world reputation
mechanisms are based on voluntary provision of
information about others’ behavior. Individuals can
often choose whether or not to complain about bad
behavior or report good behavior, as is the case with
the Better Business Bureau, eBay, product reviews
on the internet, informal recommendations from
friends, and so on. When information is provided
voluntarily, a successful reputation mechanism must
motivate agents to report accurate information despite the temptation to free ride on the reporting of
3
others. Thus reputation mechanisms must resolve an
information provision dilemma in order to mitigate
the original exchange dilemmas.
Alternative information exchange mechanisms.
Most previous models seek generality and tractability
by abstracting away the details of how information is
actually exchanged. Our goal of comparing central
and local reputational mechanisms, on the other
hand, requires a clear specification of the different
means of seeking and providing information in each.
Our third extension of the literature compares two
alternative informational reputation systems with a
baseline condition dependent only on experiential
reputation. One system represents local, self-organizing networks of information sharing associated
with the social capital perspective (Coleman 1987;
Granovetter 1985; Putnam 1993) and the bottom-up
approach to organization (e.g., Bianco and Bates
1990; Bohnet, Frey, and Huck 2001; Ostrom 1990;
Ostrom, Walker, and Gardner 1992). The other
represents the top-down, central information distribution institutions whose broader scope can potentially enhance performance by allowing information
to flow freely to anyone who seeks it, whether or not
they know the information provider.
Our research design replicates one critical distinction between these paradigmatic systems with
critical consequences for information provision.
When buyers use a local mechanism they generally
seek information from a friend who can recommend
a seller that the buyer may not even know; when
buyers use a central mechanism like the internet they
can generally request information about a particular
seller and see the comments posted by previous
buyers that the buyer is unlikely to know. The local
reputation mechanism we model allows an information seeker to choose the source (or the provider) of
information from among her group of associates,
which restricts information to the scope of the
source’s relationships. By contrast, the central reputation mechanism allows an information seeker to
choose any target of reputational information from a
common reputation repository and receive all available information about that target. This critical
difference translates into a two-person information
provision dilemma in the local condition and an
n-person dilemma in the central condition.
Information Provision Incentives
Determine Performance
We conjecture that the performance of any reputation mechanism relying on voluntary information
4
provision depends critically on the incentives for
providing information. Accurate information can
potentially be useful for providing credible commitments and for determining the trustworthiness of
current and prospective partners, but why should
anyone provide such information to others if it is
costly to do so? It takes time and effort to provide
accurate feedback on eBay, to report malfeasance to
the Better Business Bureau, or to introduce a reliable plumber or gardener to a neighbor. What if a
recommended lawyer turns out to be incompetent or
untrustworthy, thereby hurting the reputation of the
information provider? Vengeance may provide a
strong incentive for an aggrieved partner to report
bad behavior, but the risk of retaliation may offset
these psychological benefits. So why not free ride by
taking advantage of the information others provide
but not taking the time and risk of providing information? For a reputation mechanism to work,
individuals who can collectively benefit from the
mechanism must overcome this second-level collective action problem of information provision.
The local and central mechanisms differ substantially in the potential incentives they provide,
although both have means of reducing the costs of
information provision. Local interactions may reduce
costs by conveying reputational information through
gossip, which requires no organizing structure and is
valued in its own right. Yet information moves slowly
through such a system, constrained as it is to pass
through an incomplete and decentralized network.
Central interactions, on the other hand, may reduce
costs by simplifying the reporting process and communicating reports automatically to all interested inquirers. Internet exchanges, for example, provide
quick-access links, email reminders, and preset forms
with standardized reporting tools that make feedback
about the buyer’s experience relatively easy to post and
easy for other prospective buyers to find. Yet centralized systems rely on a costly infrastructure, and on the
willingness of strangers to share their experiences.
The main differences in incentives arise from the
more intimate relationship between information
seekers and providers in the local system and the
corresponding two-person provision dilemma. Local
reputation systems can use this more intimate relationship to punish free riders who do not provide
recommendations. One party to an ongoing exchange
can ask the partner for recommendations about that
partner’s partners and will know if the partner does
not respond. Nonresponding partners can be punished by noncooperation or termination of the
exchange, which is not really a costly punishment if
t.k. ahn, justin esarey, and john t. scholz
the nonresponse correctly indicates that the partner is
not trustworthy. Providing misleading information
may also lead to future punishment or termination,
although the potential for punishment here depends
on the ability to detect misinformation. One plausible
strategy capable of supporting a cooperative equilibrium under the local reputation mechanism would
be an extension of the well-known nice but vengeful
strategy that would always ask for recommendations,
always provide truthful information when asked, and
cease to play with anyone who either defects, fails to
provide a recommendation, or gets a bad recommendation from another partner.
The link between information provision and cooperation is weaker in the central condition, where the
request and provision of recommendations is mediated
by a passive central information distribution system.
Central systems tend to rely on relatively anonymous
reports, in part to minimize concerns about retaliation
in response to negative reports.1 Information seekers
and providers both act anonymously, so providers do
not know whether others will seek their information,
and anonymous free riders face little threat of retaliation for providing insufficient or incorrect information. Because reporting requires effort but creates no
direct benefit to the subject herself and no clear
benefits even to cooperating partners, a self-interested
person will not contribute. Vengeful subjects may
provide negative recommendations to punish defectors, but can negative recommendations alone support
reputation-induced cooperation (Standifir 2001)?
In theory at least, free riders could be deterred in
the central condition if all cooperation were conditional on some minimal level of information reporting
by everyone. However, the need to punish everyone in
this n-person provision game for the lapses of a few
free riders makes such a strategy much less plausible
than strategies in the local condition that can directly
punish individual free riders in the two-person game,
particularly in uncertain and noisy environments in
which a plausible minimum level of reporting would
be difficult to set. Furthermore, such trigger strategies
may actually reduce cooperation to lower levels than
in the baseline condition if the observed initial
reporting falls below the expectations of subjects
following such strategies. Free riding might more
plausibly be reduced if some critical proportion of
the population had an internalized incentive to
1
E-Bay identifies recommenders and provides a system to rate
them, which allows merchants to see who has given them bad
ratings and to retaliate for negative feedback (Brown and Morgan
2006). Brown and Morgan criticize even E-Bay’s rating system as
ineffective in preventing bogus recommendations.
comparing local and central institutions
report—after all, successful internet rating sites do
provide reports from substantial numbers of buyers
willing to assess sellers and products. For example,
the critical population might support a reciprocating
norm that would require honest reporting from those
who use and identify with the system. Some proportion of free riders could be tolerated in such a system
as long as the proportion of normatively motivated
participants could provide sufficient information to
establish credible reputations.
Our experiments are designed to test the conjecture that the ability of each reputation mechanism
to enhance cooperation depends on its ability to
overcome its characteristic information provision
problem. The central mechanism has a natural
advantage in that information once provided can be
disseminated widely in comparison with the local
mechanism, where only existing partners have access
to recommendations. The local mechanism, on the
other hand, has the advantage of more intimate
relationships between information seekers and providers that may more effectively limit the critical freerider problem of information provision.
Experimental Design
To capture the critical dimension of voluntary exchange, our experiment represents the basic exchange
dilemma with a two-person prisoner’s dilemma (PD)
game; like the real-world interactions that we hope to
represent, the PD allows for mutual gain at the risk
of exploitation. Payoffs in the PD are set at 0, 25,
75, and 100 in Experimental Currency Units (400
ECUs5$1.00). Mutual cooperation (payoff 5 75)
provides a substantial advantage over mutual defection (payoff 5 25), although the choices were
presented to subjects as alternatives A and B rather
than as ‘‘cooperate’’ and ‘‘defect.’’
To represent ongoing voluntary exchanges in our
experimental sessions, the 14 subjects in each session
know that they can simultaneously propose to play
with as many specific partners as they wish in each of
20 periods. However, the PD game only takes place if
both partners propose each other at the beginning of
a given period (cf. Ule 2006).2 Thus, either player can
2
To minimize ordering effects in the proposal process, each
subject receives a different list of subject ID numbers in a
different order; this prevents (for example) a disproportionate
number of subjects from proposing to play with Subject #1
simply because s/he is at the top of every person’s list of subjects.
The experiment was conducted in a computer lab setting using
z-Tree (Fischbacher 2007).
5
terminate an existing relationship unilaterally in the
next period simply by not proposing to play, so any
player can ensure no further games will be played if
a partner turns out to be undesirable. To reflect
bounded rationality limitations on the number of
exchanges individuals are able to manage, subjects
pay a cost per game that increases with the number of
games played in a given period.3 Our payoff structure
sets two important benchmarks that illustrate the
potential cooperators’ advantage. A rational player
expecting all others to always defect should optimally
play four games in each period and always defect,
earning a net payoff of 60.4 ECU’s per period after
subtracting maintenance costs for four games totaling
39.6 from the 100 ECU payoffs for four games with
mutual defection. On the other hand, for the conditional cooperator who expects others to cooperate,
the mutual cooperation payoff of 75 per game implies
an optimum of 10 games, providing the socially
optimal benchmark net payoff of 393.6 ECU’s per
period (750 in total payoffs minus 356.4 in maintenance costs). The social optimum net payoff is
more than six times as large as the defectors’ optimum benchmark prediction, leaving much room
for variation between the two. By comparing the
proportion of cooperation and mutual cooperation,
the number of exchanges, and the average payoffs
achieved under different reputation mechanisms, we
can compare the social welfare achieved under each
reputation mechanism.
In the Baseline Condition, subjects gain information about other subjects only from their own experience. To reinforce the experiential reputation
effect, at the beginning of each period (after the first)
each subject sees only her own experiences (her proposals, proposals from others to her and PD outcomes when games take place with her) with each of
the 13 other subjects for up to the past four periods
of play.4 After viewing this information, all players
make proposals of whom to play in the current
3
This is modeled as a cost function c 5 0 if m 5 0, and c 5 4.4
(m – 1)2 for m . 0, where m, a nonnegative integer, is the
number of games the player has in a period. Subjects were
presented with the list of total costs associated with each number
of exchanges, not with this formula. Note that the minimal sucker’s
payoff is zero, so there would be no limit on the preferred number
of games except for the increasing cost function.
4
The four-period limit was imposed due to space constraints on
the computer screens used in the experiment, but it also reflects a
bounded capacity to recall the past and process the information.
Subjects are provided access to past outcomes without overwhelming them with the full sequence of play.
6
period without knowing whether others will propose
them in return or not. Each player plays the resultant
PD games with all the proposed players who had also
proposed to that player, and the experiment proceeds
to the next period once all games are finished.
The other two conditions are the same as the
baseline condition except that subjects may provide a
positive or negative recommendation about any partner after each PD game. To represent the voluntary
feature of reputation mechanisms discussed above,
subjects are not required to give recommendations,
and the recommendations can be misleading (i.e.,
subjects can recommend defectors positively and
cooperators negatively). Recommendations are presented after each period on the information screen
described for the baseline condition. As noted above,
the process of seeking and making recommendations
differs significantly in the remaining conditions.
In the Local Information Condition, a player can
ask any current partner for recommendations before
choosing new partners. The partner making recommendations is told who if anyone has asked for
recommendations and can then make positive or
negative recommendations about any number of his
or her current partners that will be shared with the
requesting partners. The requesting player then sees
the recommendations and proceeds to the choice of
partners with this information available. In this
condition, the flow of information depends on the
endogenous pattern of PD relationships created by
the subjects.
In the Central Information Condition, after finishing all games in a given period players can post
positive or negative recommendations to an information board about any number of their partners.
These anonymous recommendations are then available to any other player who requests them. Players
can request information for as many other players as
they want, and they see the number of positive and
negative recommendations posted for each requested
player during each of the past four periods. The
number of posted recommendations is not known
before the request is made, and no information is
available about who made the recommendations.
After this chance to see recommendations, subjects
proceed to the next period in which they again propose partners, play the PD games, and make and
request recommendations.
We later added a variant Central Truthful Condition that eliminates the information provision
dilemma by simply reporting truthfully all information about the PD choices for each specific player to
anyone who requests this information. Players still
t.k. ahn, justin esarey, and john t. scholz
need to request information for specific players, but
they know that they will receive full information
about the number of A or B choices made by each of
the specified players in each of the previous four
periods.
To encourage the subjects to think about their
decision to provide and obtain information rather
than always providing and obtaining everything, both
requesting and providing recommendations impose
a very minor cost of 1 ECU per request or recommendation.5 In the local condition, requests are charged
for each partner from whom requests are solicited,
while in both central conditions requests are charged
for each subject whose recommendations from others
is sought.
A total of 14 sessions were run, four sessions for
each of the three initial conditions plus two additional sessions for the central truthful condition. The
196 subjects recruited from social science courses at
Florida State University earned performance-based
payoffs averaging $20–$25 for sessions that lasted
from 60 to 90 minutes. A sample of computer screens,
instructions, and other details of the experimental
procedures are provided along with the data and
replication package in the online appendix at http://
journalofpolitics.org/.
Results: Local Reputation
Condition Outperforms Others
The four graphs in Figure 1 report average performance per subject on four related measures of social
welfare for all 20 periods in each of the three initial
experimental conditions; the key in the upper left
graph indicates that the solid, dashed, and dotted
lines, respectively, represent the baseline, central, and
local conditions. Each measure captures a slightly
different aspect of performance, but all tell the same
story. As expected, informational reputation does
outperform experiential reputation, but only for the
local condition. More surprisingly, the central condition if anything does worse than the baseline
condition. We will first present these results in greater
5
Our experimental design tries to make the costs of requesting
and obtaining information as equal as possible across conditions.
Given the intrinsic value of gossip in informal relationships, costs
may be lower or even negative in naturally occurring local
settings, so the equality of costs in our design may actually favor
the central condition.
comparing local and central institutions
7
F IGURE 1 Comparison of Outcomes by Experimental Condition and Period
0.4
50
0.2
0
5
10
15
20
5
10
15
Average # CC Links
Average # of Links
20
6
4
2
0
1
2
3
Average # of Links
4
8
Period
5
Period
0
Average # CC Links
100 150 200 250 300
Average Earnings
Baseline
Central
Local
0.6
0.8
Average Earnings
0.0
Average % of C Plays
1.0
Average % of C Plays
5
10
15
20
Period
detail and then analyze information provision patterns
to explain them.
The percentages of choices that were cooperative
(upper-left graph) in all conditions considerably
exceed the one-period Nash equilibrium of zero
cooperation and remains relatively stable until the
expected decline in the last periods. The observed
level of cooperation is on the upper end of the range
reported in other experiments with dilemmas; as
anticipated (Ule 2006), the ability of subjects to select
partners and exit from unsatisfactory relationships
apparently sustains high levels of cooperation even in
the baseline condition with no information provided
by other players.
Comparing the results across treatments, the
local treatment produces considerably higher levels
of cooperation than the baseline in every period,
while the central treatment produces lower results.
Averaged across all periods, local provision produces
15 percentage points more cooperation than the
baseline condition, a difference that is significant
5
10
15
20
Period
even with a conservative test6 (t52.09). Central provision, on the other hand, produces on average 10 percentage points less cooperation than the baseline
across all periods, although the difference is not
significant (t 5 21.0).
The lower left graph in Figure 1 reports the
average number of games with mutually cooperative
6
When periods are analyzed individually, the local condition
produces significantly higher percentages of cooperation in
periods 11–19. Due to the possibility of correlated errors among
observations within a given session, we calculate robust standard
errors, clustered by session. Given the distribution of the dependent variable, we report the results based on OLS estimation for
percent cooperation and earnings and Poisson regression for the
number of mutual cooperation (CC) and the number of games.
The analyses reporting significance by period are done in the
following manner. For the regressions reporting the number of
CC links, the subject’s proportion of cooperative choices in the
first period is included as a control in later periods for any
differences across treatments due to potential differences in the
frequency of more cooperative subjects in any session. This proxy
for the initial cooperative tendency of the subject does not alter
the significant impact of the treatment.
8
outcomes (CC) per person, which reflects the ability
of cooperators to find reciprocally cooperative relationships. Mutual cooperation increases in the first
half of the periods for all conditions even though the
percentage of cooperation appears to be stable or
decreasing, suggesting that cooperators become better matched over time under all conditions. The local
condition produces more CC outcomes in every
period compared with the baseline condition, with
significant differences (t.1.96) in periods 10–17.
Again, central provision falls below the baseline both
overall and in every period, although the differences
are not significant.
In all conditions, the average number of games
played per person (lower right graph) is closer to the
benchmark of four mutual defection games than to
the socially optimal benchmark of 10 mutual cooperation games. The local condition has a higher
number of games in every period and a significantly
(t.1.96) higher number in nine of the periods, but
the average across all periods of 4.9 is still closer to
the lower than the upper benchmark. The factor that
limits the number of games is evident in the previous
graph, which indicates that on average only three
games with mutual cooperation could be achieved
even in the best-performing local condition.
The upper-right graph reports the primary welfare measure of average earnings in experimental
units (ECUs), subtracting the maintenance cost from
the total earnings in games played. These results
confirm the enhanced overall efficiency of the local
condition in extracting resources in comparison to
the other conditions. Earnings were higher than the
baseline for local (140 vs. 198, t52.30) and lower
than the baseline for central (140 vs. 115, t50.38)
when averaged across all periods. Yet again, even the
highest earnings of 210 in later periods of the local
condition were far below the socially optimal level of
393.6 ECU’s per subject.
In sum, the local outperforms the central and
baseline conditions on all measures, while the central
condition unexpectedly performs no better than the
baseline. The improved performance of the local
condition demonstrates that reputational information
can clearly enhance outcomes in our social dilemma
and suggests that this implementation of the local
reputation mechanism at least partially resolves the
information provision dilemma. At the same time,
the poor performance of the central condition suggests that our implementation completely failed to
resolve the information dilemma. We next analyze
differences in information provision and its relationship to overall performance of each mechanism.
t.k. ahn, justin esarey, and john t. scholz
Subjects in Local Condition Provide
More Information
Table 1 compares the number of negative and positive
recommendations conditional on the game outcomes
for the local and central conditions and supports in
several ways the conjecture that subjects have greater
incentives to provide accurate information in the
local condition. Most prominently, the central condition provides only 30% of the total recommendations provided by the local condition (1,364 to 407).
Subjects only provide information after 9% of the PD
games played in the central condition compared with
27% in the local condition, where more games are
played, but where recommendations could only be
made when requested. Over time, more recommendations are reported in the earlier periods in both
cases, with the frequency dropping off gradually
throughout the 20 periods.
Second, recommendations in the local condition
are strongly tied to mutual cooperation: a positive
recommendation is provided after 37% of mutual
cooperation outcomes (CC) in the local condition,
but only after 15% in the central condition. Over
1,000 positive reports followed mutual cooperation
in the local condition, or 75% of all reports, compared with only 161 positive reports (40% of all
reports) in the central condition.
Third, cooperators (CC and CD outcomes) are
considerably more likely to provide recommendations in the local condition, providing recommendations 35% of the time compared to only 7% provided
by defectors. In the central condition cooperators
provide recommendations 18% of the time compared
to only 4% by defectors. Again, the link between
cooperation and information provision is strongest in
the local condition.
Vengeance, on the other hand, appears to motivate subjects equally in both conditions: cooperative
subjects give negative recommendations about a
defector after 28% of games with CD outcomes in
the local condition and after 24% in the central
condition. Thus evidence does not support a greater
frequency of the cooperators’ revenge strategies in the
central condition, where it might have played a greater
role in motivating provision. In sum, subjects in the
local condition supply considerably more information than those in the central condition, with the
greatest difference in the category of supporting a
partner after a mutually cooperative outcome.
Finally, the table shows that both information
conditions prompt relatively truthful reporting: positive recommendations are given for cooperators and
comparing local and central institutions
T ABLE 1
Prior
Outcome
9
Information Provided by Subject about Partners, By Prior Outcome
Negative
Report
Central Information Treatment
DD
63
CD
119
DC
12
CC
5
Total
199
Local Information Treatment
DD
60
CD
143
DC
10
CC
2
Total
215
Positive
Report
Number of
Occurrences
of Outcome
Reports as
Proportion of
Occurrences
Reports as Proportion
of All Reports
14
5
28
161
208
2,352
518
518
1,136
4524
0.03
0.24
0.08
0.15
0.09
0.19
0.30
0.10
0.41
1.00
43
13
71
1,022
1149
1,214
560
560
2,794
5128
0.08
0.28
0.14
0.37
0.27
0.08
0.11
0.06
0.75
1.00
Notes: Recommendations are provided immediately after each game outcome. In the ‘‘Prior Outcome’’ column, C indicates
cooperation, D indicates defection, and the recommender’s choice is presented first—that is, CD indicates that the recommender
cooperated and the recommended subject defected in the prior period, the situation in which vengeful negative recommendations
would be most expected.
negative recommendations for defectors 95% of the
time in the local condition and 92% in the central
condition. Apparently, the incentives for strategic
misrepresentation are equally low in both conditions,
despite the greater incentive for truthful reporting
because of the more intimate relationship between
seekers and providers in the local condition.
Subjects in Local Condition Request
More Information
Requests for information are more difficult to compare across conditions, because subjects in the local
condition request recommendations from a partner
about unknown others, while in the central condition
they retrieve recommendations about a specific target, whether a current or potential future partner,
from unknown others. A single request in the local
condition may produce recommendations from the
selected source about several potential partners, while
in the central condition the request may produce
recommendations from multiple recommenders about
the single selected target. With this caveat in mind,
Table 2 reports the number of requests following each
possible outcome in the previous period. For the
central condition, the outcome is reported between
the information requestor and the target of the
requested information, so it reveals who the requestor is interested in. For the local condition the
outcome is with the partner from whom information
is being requested, so it reveals who the subject thinks
has the best information.
Subjects in the local condition request information 1,162 times compared with 1,028 in the central
condition. Since the opportunity to request is limited
by the subject’s number of partners in the local
conditions, requests are made in 23% of all opportunities in the local condition, compared with only
7% of the unrestricted opportunities in the central
condition. Since the number of recommendations
provided is considerably greater in the local condition, as noted above, the average number of recommendations provided per request is also greater.
Perhaps the most striking result in Table 2 is that
78% of all requests by local subjects are made to
mutually cooperative partners, again confirming the
strong relationship between cooperation in the initial
and informational dilemmas that was already evident
in the provision of information. Requests are made
after one third of all mutual cooperation outcomes,
with a higher request frequency in earlier periods.
Equally noteworthy is the fact that 67% of all requests
by central subjects inquire about potential partners
who had not played with the subject in the previous
period. Only 6% of all requests in the central condition
are inquiries from a cooperator about a partner who
defected, where the cooperator might seek confirmation from others about the trustworthiness of the
defector. That is, information requests in the central
condition appear predominantly to be used to search
10
t.k. ahn, justin esarey, and john t. scholz
T ABLE 2
Information Requested by Subject about Partners and Potential Partners, By Prior Outcome
Number of
Occurrences
of Outcome
Requests as
Proportion of
Occurrences
Requests as Proportion
of All Requests
Central Information Treatment
No contact
247
Partner Try
256
Subject Try
181
DD
128
CD
60
DC
44
CC
112
Total
1,028
4,914
2,683
2,683
2,882
502
502
1,122
14,560
0.06
0.10
0.07
0.06
0.12
0.09
0.10
0.07
0.24
0.25
0.18
0.12
0.06
0.04
0.11
1.00
Local Information Treatment
No game
[0]
DD
97
CD
94
DC
67
CC
904
Total
1,162
[8,704]
1,214
560
560
2,794
5,128
NA
0.08
0.17
0.12
0.32
0.23
0.08
0.08
0.06
0.78
1.00
Prior
outcome
Information
Requested
Notes: * indicates that the occurrences include only situations with games, since no request can be made when there is no game.
for new partners rather than to check up on existing
partners, consistent with an extended trigger strategy
that immediately drops defectors without checking
for additional information.
Recommendations Affect Partner
Selection in Both Conditions
Since comparing requests is difficult because they
have different meanings in the two conditions, the
critical comparison is whether the requested recommendations actually have an impact on the selection
of partners and on the likelihood of cooperation in
either condition. Table 3 reports estimation of the
impact of recommendations on the decision to
propose play with a given partner.7 The dependent
variable Proposal to play equals 1 if the subject
chooses to propose to that partner and 0 otherwise.
Since each of the 14 subjects can propose to 13 other
subjects, there are 143135182 directed dyadic observations in each session, or 728 dyads for the four
sessions for each condition. Because the same dyad is
7
We found no substantively important differences in the main
conclusions reported here for several different specifications of
the regression model. We will discuss the minor differences in the
appropriate sections below.
observed in each of 20 periods, random effects on the
728 dyads are included in the model.
The two recommendation variables count the
number of positive and of negative recommendations
that the proposer actually sees about the potential
partner. We also include variables representing the
history of play within a given dyad, which control for
experiential reputation and provide a comparison of
the impact of recommendations versus personal experience in shaping choices. Separate variables count
occurrences of each of the seven possible outcomes
for the prior four periods as seen by the subjects
during the experiment. The most common outcome
of No proposal from either subject provides the
excluded baseline category, and the remaining six
outcomes listed in Table 3 are labeled from the
perspective of the subject (ego) whose decision is being
analyzed. For example, if the other player (alter)
proposed four periods ago when ego did not, then
both players proposed and mutually cooperated in the
next three periods, Alter proposed would equal 1,
Reward (CC) would equal 3, and the remaining
variables would equal zero for the dyad in the current
period.
To reflect the increasing cost per game as more
games are added, we include a quadratic representation of the number of games played by the subject in
the last period. Finally, we include Period to account
comparing local and central institutions
T ABLE 3
11
The Impact of Recommendations and Past Outcomes on Proposals to Play
Baseline
Variables
INFORMATION
Positive Recommendation
Negative Recommendation
PAST FOUR OUTCOMES
Reward (CC)
Temptation (DC)
Punishment (DD)
Sucker (CD)
Ego proposed
Alter proposed
CONTROLS
Number of Games
Number of Games (sq)
Period
Final Period
Constant
Coefficient
Local
s.e.
Central
Coefficient
s.e.
Coefficient
s.e.
.81
2.63
.073
.117
.72
2.13
.092
.060
1.47
.94
.55
.10
.46
.16
.058
.057
.029
.050
.026
.024
1.31
.53
.49
.06
.50
.06
.041
.057
.035
.054
.029
.027
1.35
.66
.56
.19
.32
.21
.060
.056
.026
.053
.027
.025
2.24
.03
2.05
.56
2.23
.032
.004
.004
.101
.061
2.11
.005
2.04
.54
2.22
.030
.003
.005
.111
.064
2.11
.009
2.05
.43
2.36
.029
.003
.004
.099
.060
Notes: The dependent variable is the dummy variable proposal. Columns report random effects logit estimates for proposals, based on
14560 dyad observations (4 sessions of 20 rounds with 14 players each having 13 potential partners). Random effects are calculated
based on the 728 dyad groupings.
for any trends over time and a Final period dummy to
account for the endgame effect.
Table 3 indicates that recommendations significantly affect the proposals to play for subjects in both
local and central conditions, although the negative
effects are considerably larger in the local condition.
With all variables at their mean, we calculated the
overall average probability of proposing to be 66% in
the local condition and 54% in the central condition.
One additional positive recommendation at this point
increases the likelihood of proposing by 18 percentage
points in both conditions, to 84% and 72%, respectively. A negative recommendation decreases the likelihood by 14% in local but by only 3% in the central
condition. Thus recommendations have a slightly
greater impact on selecting partners in the local
condition, particularly negative recommendations.
Past experiences have very similar effects in all
conditions, suggesting that the basic strategies change
little with respect to personal experience even when
recommendations are available. All past conditions
increase the probability of a proposal over the omitted category that signifies no contacts, indicating
some inertia in all links.8 In all conditions, the reward
8
Remember that even the punishment payoff provides a positive
payoff as long as the subject plays less than five games in a given
period, which accounts for the positive impact compared with no
game at all.
payoff for mutual cooperation provides the greatest
increase in probability, followed by the temptation
and punishment payoff and the statistically insignificant sucker payoff.
With other variables at their mean value, one
additional period of mutual cooperation increases the
likelihood of a renewed proposal by 29% and 33%
respectively in the local and central condition. One
additional positive recommendation, on the other
hand, increases the likelihood by 18%, so two positive
recommendations would slightly outweigh the effect
of mutual cooperation. Positive recommendations
have a greater impact than any past condition other
than mutual cooperation on the probability of proposing. Mutual cooperation and positive recommendations play the expected role in all conditions, while
negative recommendations play an unexpectedly minor role, particularly in the central condition where
vengeful reporting was anticipated to produce primarily negative recommendations.
Positive Recommendations Increase
Cooperation Mainly in Local Condition
The impact of recommendations on cooperation is
potentially more critical than the impact on partner
selection for the success of a reputation mechanism.
Table 4 presents the logistic regression results for the
12
T ABLE 4
t.k. ahn, justin esarey, and john t. scholz
The Impact of Recommendations and Past Outcomes on Cooperation
Baseline
Variables
INFORMATION
Positive Recommendation
Negative Recommendation
PAST FOUR OUTCOMES
Reward (CC)
Sucker (CD)
Temptation (DC)
Punishment (DD)
Ego proposed
Alter proposed
CONTROLS
Period
Final Period
Constant
Censored N
Uncensored N
Coefficient
Local
s.e.
Central
Coefficient
s.e.
Coefficient
s.e.
.39
2.20
.070
.129
.07
2.08
.037
.062
.41
2.06
2.31
2.54
2.09
2.03
.064
.064
.063
.045
.035
.036
.32
2.08
2.35
2.56
2.06
.02
.055
.062
.060
.049
.041
.039
.35
.04
2.16
2.51
2.06
2.08
.078
.066
.065
.044
.038
.040
2.01
21.58
.50
9276
4556
.007
.189
.164
2.02
21.41
.65
8546
5286
.008
.144
.144
2.01
21.02
.48
9438
4394
.008
.218
.198
Notes: The dependent variable is the dummy variable cooperation. Columns report probit estimates with Heckman selection procedure
for cooperation and robust standard errors clustered by dyad. Selection model results are not reported, as explained in text. Wald Chi
squared test of independent equations rejects independence at p , .01 for all equations. Robust standard errors are reported in the ‘‘s.e.’’
columns.
dyadic cooperation variable that equals 1 when the
subject cooperates, 0 when the subject defects and is
undefined when no game is played. Because the data
is censored by the proposal process that determines
when a game is played, we use a Heckman procedure
with a first-stage selection equation analyzing the
joint probability of proposing, based on a similar set
of variables as reported in Table 3. The test of independence of selection and cooperation supports the
need to include the selection equation (p , .0002),
although we do not report the selection equation in
Table 4 since findings for the joint probability just
repeat Table 3 analysis in a less interpretable way.9
Table 4 reveals a difference in the use of recommendations across conditions: positive recommendations have a strong, positive impact on cooperation
in the local but a lesser effect in the central condition.
In the local treatment, positive recommendations
have a strong, statistically significant influence on
cooperation (t55.55); with all other variables at their
means, an increase from 0 positive recommendations
to 1 causes a 14.3 percentage point increase in the
probability of cooperation. In the central treatment,
9
In this selection equation, the dependent variable is 1 if both
players in a dyad propose to each other. The right-hand side of
the selection equation includes variables from both partners in
the dyad for positive recommendations, negative recommendations, number of games, and number of games squared.
the effect of positive recommendations is less statistically significant (t51.85), and the same one-unit
increase causes only a 2.7 percentage point increase
in the probability of cooperation. This effect remains
even after adjustments for the selection effect in
which highly recommended players are more likely
to receive proposals and presumably play more games.
Negative recommendations have no significant effect
in either condition; the only observable punishment
strategy appears to operate by refusing to play, not by
defecting, and even the refusal to play appears to be
marginal in the central condition.
As with proposals, past outcomes have similar impacts on cooperation across all conditions, although
here only reward payoffs significantly increase the likelihood of cooperation. Temptation and punishment
outcomes significantly decrease the likelihood of cooperation, with punishment having the largest negative
impact in each condition. The sucker’s payoff has no
significant effect, perhaps in part because the likelihood
of continuing a relationship is lowest for this outcome.
The significant impacts of both positive recommendations and reward outcomes in the local condition provide additional evidence that information
provision is intertwined with game strategy as a
means of maintaining cooperative relationships. Even
controlling for past mutual cooperation, recommendations from a trusted partner provide additional
comparing local and central institutions
incentives to cooperate. This result is particularly
striking given the more dominant impact of negative
rather than positive recommendations on final bid
price in eBay auctions (Standifird 2001), and the
pervasive finding that ‘‘[b]ad reputations are easy to
acquire but difficult to lose, whereas good reputations
are difficult to acquire but easy to lose’’ (Baumeister
et al. 2001, 344). Avoiding known evils appears to be
the dominant motivation in larger, more anonymous
settings, whereas finding trustworthy partners seems
to be the dominant motivation for providing and
using information when local exchange relations
provide the means to obtain recommendations.10
Summary: Local Mechanisms Can Resolve
the Information Provision Dilemma
Taken as a whole, the results support the argument
that the local condition resolves the information provision dilemma better than does the central condition. Information is apparently used in the central
condition primarily to find cooperative partners,
since most requests are targeted at new potential
partners. But the relatively low provision of information is apparently not sufficient to boost levels of
cooperation, since positive recommendations induce
only a slightly greater propensity to cooperate. Indeed, the sparse information provided in the central
condition may itself trigger a less cooperative approach by subjects who provide recommendations
but are disappointed when others fail to do so. The
failure to even match the performance of the baseline
condition underscores the argument that inadequately designed reputation systems may actually
decrease levels of cooperation (cf. Grinnell 1931).
The local condition, on the other hand, achieves
higher levels of information provision, particularly
for dyads experiencing mutual cooperation. Both
positive and negative recommendations from one’s
partner affect partner selection. Positive recommendations from one partner directly affect cooperation
with other partners as well, further reinforcing cooperation within the overlapping networks of mutual
cooperation. Of course, the local condition performance still falls far short of the social optimum
benchmark. This suboptimal performance, combined
with the very poor performance of the central
condition, raises the question of whether complete
10
Note also that the positive coefficients for positive recommendations as well as for past CC outcomes reflect the dominance of CC
conditions when recommendations are sought; both coefficients are
consistent with the strategy of reinforcing cooperation through
seeking, providing, and reporting on cooperative outcomes.
13
elimination of the information provision dilemma is
necessary to achieve optimal performance.
Can an Omniscient Agency
Do Better?
To test the potential advantage of central reputation
mechanisms that do not depend on voluntary information reporting, we ran two additional sessions
in which full information was provided on request.
Subjects in this ‘‘central truthful condition’’ can
request to see the number of A and B choices of
any other player in the past four periods, so subjects
know exactly what kind of information to expect.
The performance comparison of the two central
conditions indicates that the omniscient agency
dramatically outperforms the original central condition on all of the previously reported measures.
However, the omniscient agency only slightly outperforms the local condition in terms of average
levels of cooperation (61.8% to 58.4%) and mutual
cooperation (2.7 to 2.5 mutual cooperation outcomes
per subject per period). Average earnings are actually
lower (179 to 184 ECUs per period), and none of
these differences are statistically significant.
As expected, information requests in the truthful
condition increase by 75% over the initial central
condition. Somewhat more surprisingly, selection
and cooperation strategies are not much different
than those analyzed for the original central condition;
we do not report the regression results because of
the similarity. The impact per recommendation on
selection is actually lower in the central truthful
condition, but the larger number of available recommendations apparently accounts for the much higher
observed number of links.
In short, even with the complete elimination of
the information provision dilemma, central mechanisms fall far short of social optimality and do not
improve upon the local condition. One clue about
the disappointing performance is suggested by the
variability in results across the two additional sessions
reported in Table 5, which reports aggregate statistics
per subject per period for all sessions. The last central
truthful session dramatically outperforms the best
session in all other conditions, while the first central
truthful session did little better than the modal
baseline session. The same variability is observable
in the initial central condition, with sessions 3 and 4
providing over twice the earnings of sessions 1 and 2.
In the baseline condition, only one outlier session
exceeded the others by 60%, and in the local
14
T ABLE 5
t.k. ahn, justin esarey, and john t. scholz
Aggregate Statistics per Subject per Period for All Sessions
Treatment
No Information
Average for 4 Sessions
Local Information
Average for 4 Sessions
Central Information
Average for 4 Sessions
Central Truthful
Session
Number of Link
Number of CC Link
Percent Cooperate
Net Earning
1
2
3
4
4.1
4.0
4.5
4.2
4.2
4.2
4.4
5.4
5.4
4.9
3.2
3.8
4.9
4.3
4.0
3.2
5.2
4.2
1.1
1.0
2.6
1.0
1.4
2.2
2.2
3.4
2.3
2.5
0.3
0.1
2.2
1.4
1.0
1.3
4.1
2.7
37.3
36.4
64.8
37.2
43.9
63.1
55.5
64.6
50.6
58.4
26.5
12.3
51.9
43.3
33.7
46.2
76.6
61.8
122.7
122.7
192.9
121.3
139.9
170.5
169.6
219.0
177.2
184.1
77.2
67.8
173.3
139.6
114.5
123.3
235.3
179.3
1
2
3
4
1
2
3
4
1
2
Average for 2 Sessions
condition the outlier was less than 30% higher. The
number of sessions is not sufficient to draw clear
conclusions, but these patterns at least suggest that
the greater availability of information in the central
condition may increase the likelihood that initial
expectations about cooperation will not be met, leading conditional cooperators to give up on others too
soon. Outcomes for the central condition may therefore be much more contingent on initial settings than
for the local condition. Slower information diffusion
through clustered, overlapping linkages and stronger
links between mutual cooperation and information
provision provides a greater chance of reshaping the
initial distribution of behavioral propensities in the
local condition, and hence a more consistent performance across sessions.
Conclusion
Our experiment has demonstrated the robust ability
of local, emergent reputation mechanisms to enhance
cooperation and to match or outperform centralized
reputation mechanisms, at least in small group
settings. The mechanism is very simple, requiring
only that exchange partners in the underlying dilemma can ask their partners about others. Evidence
supports our conjecture that the success of this
mechanism involves the ability to tightly couple
cooperation in the underlying game with the provi-
sion of information, a feature that is not present in
either of the central mechanisms. Although performance falls far short of the socially optimal fullcooperation benchmark, the improved performance
over the baseline condition with no shared information
does illustrate the role of reputation in the development of a cooperative civil society. Direct participation in self-organizing local reputation systems, like
direct participation in other localized institutions,
appears to be a critical component in the development of mutual trust (Putnam 1993) and effective
institutions (Ostrom and Nagendra 2006).
The central reputation mechanism proves to be
considerably more fragile and less effective than
expected, despite the potential advantages of broadly
communicating reputational information to everyone
regardless of their network of contacts and thus
encouraging new relationships to form between distant subjects that do not share a mutual contact.
Although most information in our experiment is
truthful, the central mechanism receives less information than the local condition and comparatively
fewer people retrieve the information that is there;
the mechanism even fails to increase cooperation
over the no-information baseline. Nor is the central
mechanism’s problem limited solely to information
provision, since even the full-information condition
performs no better on average than the local condition
and shows considerable variability across sessions.
The problem of information credibility requires considerable attention (Brown and Morgan 2006), which
comparing local and central institutions
might explain why a successful reputation institution
like eBay employs a large security staff and collaborates closely with governmental law enforcement to
detect and punish cheaters. The observed fragility of
our central mechanism suggests that operationallysuccessful central reputation institutions deserve
considerable credit for resolving credibility and provision dilemmas that are not widely appreciated.
Centralized reputation institutions are likely to
gain in importance in contemporary society to the
extent that interactions increasingly expand to
‘‘strangers’’ outside the scope of existing local networks. Cooperation within the increasingly global
scope of exchange and ever-expanding scale of policy
arenas is likely to depend increasingly on large-scale
reputation mechanisms to signal relevant information about trustworthy partners for economic, political, and other types of exchange. If so, the problems
of credibility and incentive-compatible information
provision will pose the greatest challenge for institutions ranging from local partnerships through United
Nations agencies as they attempt to enhance reputational effects and stimulate cooperative resolution of
global challenges.
By focusing on voluntary exchange, voluntary
information provision, and alternative reputation
mechanisms, our study demonstrates that selection
effects and information provision dilemmas are perhaps as important in the study of real-world reputation institutions as are the cooperative decisions that
receive most formal analysis. Experiments like ours
provide a useful tool for investigating features of the
complex institutions we hope to understand that go
beyond currently available formal analyses, providing
guidance and empirical justification for developing
potentially useful approaches. For example, the coupling of positive recommendations, mutual cooperation, and sustained linkages in the local condition
suggests that an extended cooperation-punishment
strategy deserves further investigation in repeated
games allowing partner selection. By establishing an
empirical foundation of institutional performance
along the local-central dimension, our conjectures and
observed patterns of behavior suggest new directions
to be developed.
Acknowledgments
The authors thank Jon Bendor, Elinor Ostrom, Tasos
Kalandrakis, Daniel Diermeier, Werner Güth, Gary
Miller, and other participants to the 2005 annual
meeting of American Political Science Association
15
Meeting and the second Workshop on Social Dilemmas held at Max Plank Institute, Jena, for the useful
comments. The research was partially supported by
NSF grant SES-0519459. T.K. Ahn thanks the support
of the Korean Research Foundation through grant
KRF 2008-321-B00031.
Manuscript submitted 7 January 2008
Manuscript accepted for publication 16 August 2008
References
Andreoni, James, and John H. Miller. 1993. ‘‘Rational Cooperation in the Finitely Repeated Prisoner’s Dilemma: Experimental Evidence.’’ The Economic Journal 103: 570–85.
Axelrod, Robert. 1984. The Evolution of Cooperation. New York:
Basic Books.
Baumeister, Roy F., Ellen Bratslavsky, Catrin Finkenauer, and
Kathleen D. Vohs. 2001. ‘‘Bad Is Stronger Than Good’’ Review
of General Psychology 5 (4): 323–70.
Bendor, Jonathan, and Piotr Swistak. 1997. ‘‘The Evolutionary
Stability of Cooperation.’’ American Political Science Review
91 (2): 290–307.
Bianco, William T, and Robert H. Bates. 1990. ‘‘Cooperation by
Design: Leadership, Structure, and Collective Dilemmas.’’
American Political Science Review 84 (1): 133–47.
Bohnet, Iris, Bruno S. Frey, and Steffen Huck. 2001. ‘‘More Order
with Less Law: On Contract Enforcement, Trust, and Crowding.’’ American Political Science Review 95 (1): 131–44.
Boone, R. Thomas, and Michael Macy. 1999. ‘‘Unlocking the
Doors of the Prisoner’s Dilemma: Dependence, Selectivity,
and Cooperation.’’ Social Psychology Quarterly 62 (1): 32–52.
Boyd, Robert, and Jeffrey Lorberbaum. 1987. ‘‘No Pure Strategy
is Evolutionarily Stable in the Repeated Prisoner’s Dilemma
Game.’’ Nature 327: 58–59.
Boyd, Robert, and Peter Richerson. 1989. ‘‘The Evolution of
Indirect Reciprocity.’’ Social Networks 11: 213–36.
Brown, Jennifer and John Morgan. 2006. ‘‘Reputation in Online
Auctions: The Market for Trust.’’ California Management
Review 49 (1): 61–81.
Coleman, James S. 1987. ‘‘Norms as Social Capital.’’ In Economic
Imperialism: The Economic Approach Applied Outside the Field
of Economics, eds. Gerard Radnitzky and Peter Bernholz. New
York: Paragon House Publishers, 133–55.
Dal Bo, Pedro. 2005. ‘‘Cooperation under the Shadow of the
Future: Experimental Evidence from Infinitely Repeated
Games.’’ American Economic Review 95 (5): 1591–1604.
Fischbacher, Urs. 2007. z-Tree: Zurich Toolbox for Ready-Made
Economic Experiments, Experimental Economics 10 (2), 171–78.
Friedman, James W. 1971. ‘‘A Non-Cooperative Equilibrium for
Supergames.’’ Review of Economic Studies 38: 1–12.
Fudenberg, Drew, and Eric Maskin. 1986. ‘‘The Folk Theorem
in Repeated Games with Discounting or with Incomplete
Information.’’ Econometrica 54 (3): 533–54.
Granovetter, Mark S. 1985. ‘‘Economic Action and Social
Structure: The Problem of Embeddedness.’’ American Journal
of Sociology 91 (3): 481–510.
Greif, Avner. 1989. ‘‘Reputation and Coalitions in Medieval
Trade.’’ Journal of Economic History 49: 857–82.
16
Greif, Avner, Paul Milgrom, and Barry R. Weingast. 1994.
‘‘Coordination, Commitment, and Enforcement: The Case
of the Merchant Guild.’’ Journal of Political Economy 102 (4):
745–76.
Grinnell, Flint. 1931. ‘‘The Better Business Bureaus.’’ American
Journal of Police Science 2 (3): 195–201.
Janssen, Marco A. 2008. ‘‘Evolution of Cooperation in a Oneshot Prisoner’s Dilemma Based on Recognition of Trustworthy and Untrustworthy Agents.’’ Journal of Economic
Behavior and Organization 65: 458–71.
Kandori, Michihiro. 1992. ‘‘Social Norms and Community
Enforcement.’’ Review of Economic Studies 59: 63–80.
Kreps, David, Paul Milgrom, John Roberts, and Robert Wilson.
1982. ‘‘Rational Cooperation in the Finitely Repeated Prisoners’
Dilemma.’’ Journal of Economic Theory 27 (2): 245–52.
Lorberbaum, Jeffrey. 1994. ‘‘No Strategy is Evolutionarily Stable
in the Repeated Prisoner’s Dilemma.’’ Journal of Theoretical
Biology 168: 117–30.
Maggi, Giovanni. 1999. ‘‘The Role of Multilateral Institutions in
International Trade Cooperation.’’ American Economic Review
89 (1): 190–214.
Malaga, Ross A. 2001. ‘‘Web-Based Reputation Management
System: Problems and Suggested Solutions.’’ Electronic Commerce Research 1: 403–17.
Milgrom, Paul R., Douglass C. North, and Barry R. Weingast.
1990. ‘‘The Role of Institutions in the Revival of Trade: The
Law Merchant, Private Judges, and the Champagne Fairs.’’
Economics & Politics 2 (1): 1–23.
Nowak, Martin A., and Karl Sigmund. 1998. ‘‘Evolution of
Indirect Reciprocity by Image Scoring.’’ Nature 393: 573–77.
Orbell, John, and Robyn Dawes. 1993. ‘‘Social Welfare, Cooperator’s Advantage, and the Option of Not Playing the Game.’’
American Sociological Review 58 (6): 787–800.
Ostrom, Elinor. 1990. Governing the Commons: The Evolution of
Institutions for Collective Action. Cambridge: Cambridge University Press.
Ostrom, Elinor, and Harini Nagendra. 2006. Insights on Linking
Forests, Trees, and People from the Air, on the Ground, and
t.k. ahn, justin esarey, and john t. scholz
in the Laboratory. Proceedings of the National Academy of
Sciences 103 (51): 19224–231. www.pnas.org. (May 2007).
Ostrom, Elinor, James Walker, and Roy Gardner. 1992. ‘‘Covenants With and Without a Sword: Self-Governance is Possible.’’
American Political Science Review 86 (2): 404–17.
Putnam, Robert D. 1993. Making Democracy Work: Civic Traditions in Modern Italy. Princeton, NJ: Princeton University
Press.
Radaev, Vadim. 2004. ‘‘Coping with Distrust in Emerging
Russian Markets.’’ In Distrust, ed. Russell Hardin. New York:
Russell Sage, 233–48.
Raub, Werner and Jeroen Weesie. 1990. ‘‘Reputation and Efficiency in Social Interactions: An Example of Network Effects.’’
American Journal of Sociology 96 (3): 625–54.
Resnick, Paul, Zeckhauser Richard, Swanson John, and Kate
Lockwood. 2006. ‘‘The Value of Reputation on eBay: A
Controlled Experiment.’’ Experimental Economics 9 (2): 79–101.
Rubinstein, Ariel. 1979. ‘‘Equilibrium in Supergames with the
Overtaking Criterion.’’ Journal of Economic Theory 21: 1–9.
Standifird, Stephen S. 2001. ‘‘Reputation and E-commerce: eBay
Auctions and the Asymmetrical Impact of Positive and
Negative Ratings.’’ Journal of Management 27: 279–95.
Taylor, Michael. 1976. Anarchy and Cooperation. New York: Wiley.
Tullock, Gordon. 1985. ‘‘Adam Smith and the Prisoner’s Dilemma.’’ Quarterly Journal of Economics 100: 1073–81.
Ule, Aljiz. 2006. Exclusion and Cooperation in Networks. Amsterdam, the Netherlands: Tinbergen Institute of Research.
T.K. Ahn is associate professor of public administration, Korea University, Seoul, Korea 136-701.
Justin Esarey is assistant professor of political science,
Emory University, Atlanta, Georgia 30322. John
T. Scholz is eppes professor of political science,
Florida State University, Tallahassee, Fl 32306.