Cooperation and Punishment in Public Goods Experiments

By ERNST FEHR AND SIMON GACHTER*

Casual evidence as well as daily experiencesuggest that many people have a strong aversionagainst being the "sucker" in social dilemmasituations. As a consequence, those who coop-erate may be willing to punish free-riding, evenif this is costly for them and even if they cannotexpect future benefits from their punishmentactivities. A main purpose of this paper is toshow experimentally that there is indeed a wide-spread willingness of the cooperators to punishthe free-riders. Our results indicate that thisholds true even if punishment is costly and doesnot provide any material benefits for the pun-isher. In addition, we provide evidence thatfree-riders are punished the more heavily themore they deviate from the cooperation levelsof the cooperators. Potential free-riders, there-fore, can avoid or at least reduce punishment byincreasing their cooperation levels. This, intum, suggests that in the presence of punish-ment opportunities there will be less free riding.Testing this conjecture is the other major aim ofour paper.

* Institute for Empirical Research in Economics, Univer-sity of Zurich, Bliimlisalpstrasse 10, CH-8006 Zurich (e-maii:efehr@iew.unizh.ch; gaechter@iew.unizh.ch: website; http;//www.unizh.ch/iew/grp/fehr/index.html). This paper is part ofthe EU-TMR Research Network ENDEAR (FMRX-CT98-0238). Fehr also acknowledges the hcspilality of the Center forEconomic Studies in Munich and support from the MacArthurEoundation Network on Economic Environments and the Evo-lution of Individual Preferences and Social Noniis. Part of theexperiments are also tinanced by the Swiss National ScienceFoundation under Project No. 1214-051000.97, We gratefullyacknowledge valuable comments by two anonymous referees,seminar participants at the MacArthur Foundation Meeting inStanford, the Workshop in Experimental Economics in Berlin.the ASSA Meeting in New York, the IAREP conference inValencia, the Econometric Society European Meeting in Tou-louse, the ESA meeting in Mannheim, and the European Eco-nomic Association conference in Berlin; and by seminarparticipants at the universities of Basel, Bern, Bonn, Dort-mund. Lausanne, Linz, Munich, Pittsburgh, St. Gallen, andTilburg; and by Richard Beil, Samuel Bowles, Robert Boyd.Martin Brown, Robyn Dawes, Armin Falk, Urs Fischbacher,Herbert Gintis. John Kagel, Georg Kirchsteiger, Serge Kolm,David Laibson. George Lxwwenstein, Tanga McDaniel, JohnMiller, Paul Romer, and Klaus Schmidt. We are particularlygrateful to Urs Fischbacher who did the programming.

For this purpose we conducted a public goodexperiment with and without punishment op-portunities. In the treatment without punishmentopportunities complete free-riding is a dominantstrategy. In the treatment with punishment op-portunities punishing is costly for the punisher.Therefore, purely seltish subjects will neverpunish in a one-shot context. This means that ifthere are only selfish subjects, as is commonlyassumed in economics, the treatment with pun-ishment opportunities should generate the samecontribution behavior as the treatment withoutsuch opportunities. The reason is, of course, thatthe presence of punishment opportunities is ir-relevant for the contribution behavior if there isno punishment. In sharp contrast to this predic-tion we observe vastly different contributions inthe two conditions. In the no-punishment con-dition contributions converge to very low lev-els. In the punishment condition, however,average contribution rates between 50 and 95percent of the endowment can be maintained.

The strong regularities observed in our ex-periments suggest that powerful motives drivethe punishment of free-riders. In our view thismotive is likely to play a role in many socialinteractions, such as industrial disputes, in teamproduction settings, or, quite generally, in themaintenance of social norms. If, for example,striking workers ostracize strike breakers (Hy-wel Francis, 1985) or if. under a piece ratesystem, the violators of production quotas arepunished by those who stick to the norm (e.g.,F. J. Roethlisberger and W. J. Dickson, 1947), itseems likely that similar forces are at work as inour experiments.'

Our work is most akin to the seminal paper

' Francis's (1985 p, 269) description of social ostracism inthe communities of the British miners provides a particularlyvivid example. During die 1984 strike of the miners, whichlasted for several months, he observed the following: 'Toisolate those who supported the 'scab union,' cinemas andshops were boycotted, there were expulsions from footballteams, bands and choirs and 'scabs' were compelled to sing ontheir own in their chapel services. 'Scabs' witnessed their own'death' in communities which no longer accepted them."

950

VOL 90 NO. 4 FEHR AND GACHTER: COOPERATION AND PUNISHMENT

TABLE I—TREATMENT COND[TIONS

981

Without punishment(ten periods)

With punishment(ten periods)

Stranger-treatmentRandom group

composition in each period(Sessions 1-3)

18 groups of size n

18 groups of size n

Partner-freatmentGroup composition

constant across periods(Sessions 4 and 5)

10 groups of size n

10 groups of size n

by Elinor Oslrom et al. (1992). These authorsallowed for costly punishment in a repeatedcommon pool resource game. However, intheir experiments the same group of subjectsinteracted for an ex ante unknown number ofperiods, and subjects could develop an indi-vidual reputation. Hence, there were materialincentives for cooperation and for punish-ment. To rule out such material incentives weeliminated all possibilities for individual rep-utation formation and implemented treatmentconditions with an ex ante known finite hori-zon. In addition, we also had treatments inwhich the group composition changed ran-domly from period to period, and treatmentsin which subjects met only once.

Our work is also related to the interestingstudy of David Hirshleifer and Eric Rasmusen(1989) who show that, if there are opportunitiesfor ostracizing noncooperators, rational egoistscan maintain cooperation for 7 - 1 periods ina 7-period prisoner's dilemma. In this modelostracizing noncooperators is part ot asubgame-perfect equilibrium and thus rationalfor selfish group members. This feature distin-guishes the preceding model from our experi-mental setup. In our experiments cooperation orpunishment can never be part of a subgame-perfect equilibrium if rationafity and selfishnessare common knowledge. We deliberately de-signed our experiments in this way to examinewhether people punish free-riders even if it isagainst their material self-interest.

I. The Experimental Design

A. Basic Design

Our overall design consists of a public goodexperiment with four treatment conditions (see

Table 1). There is a "Stranger"-treatment withand without punishment opportunities and a"Partner"-treatment with and without punish-ment opportunities. In the Partner-treatment thesame group of n = 4 subjects plays a finitelyrepeated pubHc good game for ten periods, thatis, the group composition does not changeacross periods. Ten groups of size n ^ 4 par-ticipated in the Partner-treatment. In contrast, inthe Stranger-treatment the total number of par-ticipants in an experimental session, iV = 24, israndomly partitioned into smaller groups of sizen ^ 4 in each of the ten periods. Thus, thegroup composition in the Stranger-treatment israndomly changed from period to period. Thetreatment without punishment opportunitiesserves as a control for the treatment with pun-ishment opportunities. In a given session of theStranger-treatment the same N subjects play tenperiods in the punishment and ten periods in theno-punishment condifion. Similarly, in a ses-sion of the Partner-treatment all groups of size nplay the punishment and the no-punishmentcondition. This has the advantage that, in addi-tion to across-subject comparisons, we can make

' Instructions are included in the long version of thispaper which can be downloaded from our websiie (http;//www.unizh.ch/iew/grp/fehr/index.html). The whole experi-ment was framed in neutral terms.

^ Noie that in the Partner-treatment the probability ofbeing rematched with the same three people in the nextperiod is 100 percent, whereas in the Stranger-treatment it isless than 0.05 percent. We also conducted experiments inwhich the probabilily of meeting the same subjects in futureperiods was exactly zero. Because of space constraints wedo not present the results of these experiments. Contribu-tions as well as punishment behavior in these perfect one-shot experiments are not significantly different fromcontributions and behavior in our Stranger-treatment.Hence, the Stranger-treatment represents a good approxi-mation to perfect one-shot experiments.

982 THE AMERICAN ECONOMIC REVIEW SEPTEMBER 2000

within-subject comparisons of cooperation lev-els, which have much more statistical power. InSessions 1-3 we implemented the Stranger-treatment, whereas in Sessions 4 and 5 we im-plemented the Partner-treatment. In Sessions 1and 2 suhjects first play ten periods in the pun-ishment condition and then ten periods in theno-punishment condition. To test for spillovereffects across conditions the no-punishmentcondition is conducted first in Session 3. InSession 4, which implemented the Partner-treatment, we start with the punishment con-dition, whereas Session 5 begins with the no-punishment condition.

B. Payo_ffs

In the following we first describe the payoffsin the treatments without punishment. In eachperiod each of the n subjects in a group receivesan endowment o( y tokens. A subject can eitherkeep these tokens for him- or herself or invest g,tokens (0 < g. < _y) into a project. The deci-sions about g. are made simultaneously. Themonetary payoff for each subject / in the groupis given by

(1)

0<a<\ <na

in each period, where a is the marginal percapita retum from a contribution to the publicgood. The total payoff from the no-punishmentcondition is the sum of the period-payoffs, asgiven in (1), over all ten periods. Note that (1)implies that full free-riding (g^ = 0) is a dom-inant strategy in the stage game. This followsfrom dtrj/dg^ = -I + a < 0. However, theaggregate payoff S"= , TT,' is maximized if eachgroup member fully cooperates (g, = y) be-cause d S"=, TTJ/dgj = -\ + na > 0.

The major difference between the no-punish-ment and the punishment conditions is the ad-dition of a second decision stage after thesimultaneous contribution decision in each pe-riod. At the second stage, subjects are given theopportunity to simultaneously punish each otherafter they are informed about the individual

contributions of the other group members.Group membery can punish group member ( byassigning so-called punishment points pj to i.For each punishment point assigned to i thefirst-stage payoff of /, -nj, is reduced by 10percent. However, the first-stage payoff of sub-ject / can never be reduced below zero. There-fore, the number of payoff-effective punishmentpoints imposed on subject /, P', is given byP' = min(2j^,pj, 10). The cost of punishmentfor subject / from punishing other subjects isgiven by Ij^- c(p';). where c(p'-) is strictlyincreasing in p'-. The pecuniary payoff of sub-ject /, 77., from both stages of the punishmenttreatment can therefore be written as

(2)

The total payoff from the punishment conditionis the sum of the period-payoffs, as given in (2),over all ten periods.

C. Parameters and Information Conditions

The experiment is conducted in a computer-ized laboratory where subjects anonymously in-teract with each other.' No subject is everinformed about the identity of the other groupmembers. In all treatment conditions the en-dowment is given by y = 20, groups are of sizen = 4, the marginal payoff of the public goodis fixed at <3 = 0.4, and the number of partic-ipants in a session is A' = 24.^ Table 2 showsthe feasible punishment levels and the associ-ated cost for the punisher. In each period subject/ can assign up to ten punishment points ^ toeach group member y, j - 1, ... , 4, j i^ i.

In all treatment conditions subjects arepublicly informed that the condition lastsexactly for ten periods. When subjects playthe first treatment condition in a session theydo not know that a session consists of twoconditions. After period ten of the first treat-ment condition in a session they are informedthat there will be a "new experiment" and

•" For conducting the experiments we used the experi-mental software "z-Tree" developed by Urs Fischbacher(1998).

' An exception is Session 4 where only N = \b subjectsshowed up.

VOL. 90 NO. 4 EEHR AND GACHTER: COOPERATION AND PUNISHMENT

TABLE 2—PUNISHMENT LEVELS AND ASSOCIATED COSTS FOR THE PUNISHING SUBJECT

983

Punishment f)oints p',

Costs of punishment

^•(/^•)

0

0

1 2

I 2

3

4

4

6

5

9

6

12

7

16

8

20

9

25

10

30

that this experiment will again last exactlyfor ten periods. They are also informed thatthe experiment will then be definitelyfinished.

In the no-punishment conditions the payofffunction (1) and the parameter values of y, n.N. and a are common knowledge. At the end ofeach period subjects in each group are informedabout the total contribution S gj to the project intheir group.

In the punishment conditions the payofffunction (2) and Table 2, in addition to >•, n,N, and a. are common knowledge. Further-more, after the contribution stage subjects arealso informed about the whole vector of indi-vidual contributions in their group. To pre-vent the possibility of individual reputationformation across periods in the Partner-treat-ment each subject's own contribution is al-ways listed in the first column of his or hercomputer screen and the remaining three sub-jects' contributions are randomly listed in thesecond, third, or fourth column, respectively.Thus, subject / does not have the informationto construct a link between individual contri-butions of subject; across periods. Therefore,subject j cannot develop a reputation for aparticular individual contribution behavior.This design feature also rules out that / pun-ishes ;• in period t for contribution decisionstaken in period /' < /. Subjects are neitherinformed about the individual punishment ac-tivities of the other group members, nor dothey know the aggregate punishment imposedon other group members. They know onlytheir own punishment activities and the ag-gregate punishments imposed on them by theother group members.

II. Predictions

To have an unambiguous reference predic-tion it is useful to shortly state the implica-tions of the standard approach to the publicgood games of Table 1. If the rationality and

the selfishness of all subjects is commonknowledge, and if subjects apply the back-ward induction logic, the equilibrium predic-tion with regard to g, for each of the four cellsin Table 1 is identical—in all four treatmentconditions all subjects will contribute nothingto the public good in all periods. This is mosttransparent in the Stranger-treatment withoutpunishment. This condition consists of a se-quence of ten (almost pure) one-shot games.In each one-shot game the players' dominantstrategy is to free ride fully. Applying thefamiliar backward induction argument to thePartner-treatment without punishment givesus the same prediction.

In the Stranger-treatment with punishmentthe situation is slightly more complicated be-cause each one-shot game now consists oftwo stages. It is clear that a rational moneymaximizer will never punish at the secondstage because this is costly for the player.Since rational players will recognize thatnobody will punish at the second stage, theexistence of the punishment stage does notchange the behavioral incentives at the firststage relative to the Stranger-treatment with-out punishment. As a consequence, every-body will choose g == 0 at stage one. Forthe same reasons as in the Stranger-treatmentrational subjects in the Partner-treatment withpunishment will choose gj = 0 and p^^ = 0for all j in the final period. By applying thefamiliar backward induction argument we thusarrive at the prediction that g, = 0 and /?^ ^ 0for all; will be chosen by al! subjects in all periodsof the Partner-treatment with punishment.

There is already a lot of evidence for publicgood games like our no-punishment condition.For these games it is well known that coopera-tion strongly deteriorates over time and reachesrather low levels in the final period (John O.Ledyard, 1995). In a recent meta-study Fehr andKlaus M. Schmidt (1999) surveyed 12 differentpublic good experiments without punishmentwhere full free-riding is a dominant strategy in

984 THE AMERICAN ECONOMIC REVIEW SEPTEMBER 2000

the Stage game. During the first periods of theseexperiments average and median contributionlevels varied between 40 and 60 percent of theendowment. However, in the final period 73percent of all individuals (N ^ 1042) chosegj = 0 and many of the remaining playerschose gy close to zero. In view of thesefacts there can be little doubt that in the no-punishment condition subjects are not able toachieve stable cooperation. Therefore, a mainobjective of our experiment is to see whethersubjects are capable of achieving and maintain-ing cooperation in the punishment condition.

In our view, the fact that at the beginning ofthe no-punishment condition one regularly ob-serves relatively high cooperation rates, sug-gests that not all people are driven by pureself-interest. We conjecture that, in addition topurely selfish subjects, there is a nonnegligiblenumber of subjects who are (i) conditionallycooperative and (ii) willing to engage in thecostly punishment of free-riders. This conjec-ture is based on evidence from many otherexperimental games. Trust- or gift-exchangegames (Fehr et al.. 1993; Joyce Berg et al.,1995) indicate that many subjects are condition-ally cooperative, that is, they are willing tocooperate to some extent if others cooperate,too. Bilateral ultimatum and contract enforce-ment games (e.g., Alvin E. Roth. 1995; Fehr etal., 1997) indicate that many subjects are will-ing to punish behavior that is perceived as un-fair. In our public goods context fairness issuesare likely to play a prominent role, too. Webelieve, in particular, that subjects strongly dis-like being the "sucker." that is, being those whocooperate while other group members free ride.This aversion against being the "sucker" mightwell trigger a willingness to punish tree-riders.In fact, recently developed theories of equityand fairness (e.g., Fehr and Schmidt, 1999) pre-dict that free-riders will face credible punish-ment threats, which induces them to cooperate.

III. Experimental Results

In total, we have observations from 112 sub-jects. Each subject participated in only one ofthe five experimental sessions. All sessionswere held in January and February 1996 at theUniversity of Zurich (Switzerland). Subjectswere students from many different fields (ex-

cept economics). They were recruited via lettersthat were mailed to their private addresses. Withthis procedure we wanted to maximize thechances that subjects do not know each other.An experimental session lasted about two hoursand subjects earned on average 41 Swiss francs(about US $32 at the time), including a show-upfee of 15 Swiss francs.

A. The Impact of Punishment Opportunitiesin the Stranger-Treatment

If subjects believe that in the presence ofpunishment opportunities free-riding faces nocredible threat we should observe no differencesin contributions across treatments. In sharp con-trast to this prediction we can report the follow-ing result.

RESULT I: The existence of punishment op-portunities causes a large rise in the averagecontribution level in the Stranger-treatment. Onaverage, contribution rates amount to 58 per-cent of the endowment.

Support for Result I is presented in Table3. In columns 2 and 3 of Table 3 we report themean contribution over all ten periods in thethree sessions of the Stranger-treatment. Thetable reveals that in the punishment conditionsubjects contribute between two and four timesmore than in the no-punishment condition. Anonparametric Wilcoxon matched pairs testshows that this difference in contributions issignificant at all conventional significance lev-els (p < 0.0001). This result clearly refutesthe hypothesis of the standard approach thatpunishment opportunities are behaviorally irrel-evant at the contribution stage of the game.

Next we turn to the evolution of contributionsover time. Remember that one of the most robustbehavioral regularities in sequences of one-shotpublic good games, like our Stranger-treatmentwithout punishment, is that contribufions dropover time to very low levels. Our next resultprovides information as to whether punishmentopportunities can prevent such a fall in contribu-tions.

RESULT 2: In the no-punishment condition ofthe Stranger-treatment average contributionsconverge close to full free-riding over time. In

VOL 90 NO. 4 EEHR AND GACHTER: COOPERATION AND PUNISHMENT 985

TABLE 3—MEAN CONTRIBUTIONS IN THE STRANGER-TREATMENT

Sessions

1

2

3

Mean

Mean contribuiion

Withoutpunishmeniopportunity

2.7(5.2)4.0

(5.7)4.5

(6.0)

3.7(5.7)

in all periods

Withpunishmentopportunity

10.9(6.1)12.9(6.4)10.7(4.9)

11.5(5.9)

Mean contributionperiods

Withoutpunishmentopportunity

1.3(4.3)2.3

(4.3)2.0

(3.8)

1.9(4.1)

in the final

Withpunishmentopportunity

9.8(6.8)14.3(5.0)13.1(4.0)

12.3(5.6)

Notes: Numbers in parentheses are standard deviations. Participants of Sessions 1 and 2 firstplayed the treatment with punishmeni opportunities and then the one without such opportu-nities. Participants of Session 3 played in the reverse order.

contrast, in the punishment condition averagecontributions do not decrease or even increaseover time.

Support for Result 2 comes from Table 3 andFigures lA and IB. Colutnns 4 atid 5 of Table3 show that, in each session, in the final periodof the no-punishment condition average contri-butions vary belvt-een 1.3 and 2.3 tokens.^ Incontrast, in the punishment condition averagecontributions vary between 9.8 and 14.3 tokensin period ten. Thus, in the final period of thepunishment condition the average contributionis between 6 and 7.5 times higher than in theno-punishment condition. Moreover, a compar-ison of column 3 with column 5 of Table 3 re-veals that in the punishment condition theaverage contribution in period ten is higher orroughly the same as in all periods.

Figures IA and IB depict the evolution ofaverage contributions over time in both condi-tions. Figure IA shows the results of Sessions 1and 2. in which subjects had to play the pun-ishment condition first. Whereas the averagecontribution is stabilized around 12 tokens inthe punishment condition, there is immediately

^ Note that in the following the term "final period" isalways used to indicate the la.st period in a given treatmentcondition and not only period 20 in a given session. Thus,for example, in Figure IA the tenth period is the final periodof the punishment condition.

a significant drop in contributions in period 11.^This decrease in the no-punishment conditioncontinues until period 18 in which the averagecontribution stabilizes slightly below 2 tokens.Figure IB shows the results of Session 3, inwhich subjects played the no-punishment con-dition first, ln our view Figure IB reveals aneven more remarkable fact. Whereas averagecontributions in the no-punishment conditionconverge again toward 2 tokens they immedi-ately jump upward in period 11 and continue torise until they reach 13 tokens in period 20. Thisindicates that the existence of punishment op-portunities triggers the effectiveness of forcesthat completely remove the drawing power ofthe equilibrium with complete free-riding. Inview of this evidence it is difficult to escape theconclusion that any model which predicts fullfree riding is unambiguously rejected.

Results 1 and 2 deal only with average con-tributions. We are also interested, however, inthe behavioral regularities at the individuallevel and how they are affected by the punish-ment opportunity. Result 3 summarizes the be-havioral regularities in this regard.

RESULT 3: ln the Stranger-treatment withpunishment no stable behavioral regularity

' The null hypothesis that average contributions are thesame in period 10 and 11 can be rejected on the basis of aWilcoxon signed-ranks test {p = 0.0012).

986 THE AMERICAN ECONOMIC REVIEW SEPTEMBER 2000

20

18

16 -

•with punishment

•without punishment

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

Periods

FIGURE 1A. AVERAGE CONTRIBUTIONS OVER TIME IN THE STRANGER-TREATMENT (SESSIONS 1 AND 2)

20

- O - without punishment

-•— with punishment

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

IB. AVERAGE CONTRIBUTIONS OVER TIME IN THE STRANGER-TREATMENT (SESSION 3)

regarding individual contributions emerges,whereas in the no-punishment condition fullfree-riding emerges as the focal individualaction.

A first indication for the absence of a behav-ioral standard in the punishment condition isprovided in Table 3. The table shows that thestandard deviation of individual contributions isquite large in each session. Moreover, the stan-dard deviation in the final period is roughly thesame as in all periods together. This indicates

that the variability of contributions does not de-crease over time. The decisive evidence for Result3, however, comes from Figure 2, which providesinformation about the relative frequency of indi-vidual choices in the final periods of bothStranger-treatments. In the no-punishment con-dition the overwhelming majority (75 percent)of subjects chose g, = 0 in the final period.Thus, full free-riding clearly emerges as thebehavioral regularity in this condition. In con-trast, in the punishment condition individualchoices are scattered over the whole strategy

VOL. 90 NO. 4 FEHR AND GACHTER: COOPERATION AND PUNISHMENT 987

0 1 2

Contributions

with punishmentwithout punishment

18 19 20

FIGURE 2. DiSTRiBUTioN OF CONTRIBUTIONS IN THE FINAL PERIODS OF THE STRANGHt-TREATMENT

WITH AND WITHOUT PUNISHMENT

space in the final period. Although the relativefrequency of 12, 15, and 20 tokens is higherthan that of other contribution levels, even themost frequent choice (g^ = 15) reaches a fre-quency of only 14 percent. Thus, subjects in thepunishment condition were not able to coordi-nate on a specific contribution level differentfrom gi = 0.

B. The Impact of Punishment Opportunities inthe Partner-Treatment

As in the Stranger-treatments our first resultin the Partner-treatments relates to average con-tributions over all periods.

RESULT 4: The existence of punishment op-portunities also causes a large rise in the aver-age contribution level in the Partner-treatment.

Table 4 provides the relevant support forResult 4. A comparison of column 2 and col-umn 3 shows that all ten groups have substan-tially higher average contributions in thepunishment condition. Therefore, the differenceis highly significant (p = 0.0026) according toa nonparametric Wilcoxon matched-pairs testwith group averages as observations.

On average, subjects contribute between 1.5times (group 2) and 4.3 times (group 9) more inthe punishment condition. Thus, punishmentopportunities are again highly effective in rais-

ing average contributions. With regard to theevolution of average contributions over time thedata support the following result.

RESULT 5: In the no-punishment condition ofthe Partner-treatment average contributionsconverge toward full free-riding, whereas in thepunishment condition they increase and con-verge toward full cooperation.

Again Table 4 provides a first indication. Itshows that in the no-punlshment condition theaverage contribution is only slightly above 3tokens in the final period. In sharp contrast, theaverage contribution is above 18 tokens in thepunishment condition. In five of the ten groupsall subjects chose the maximum cooperation of20 in the final period of the punishment condi-tion. Further three groups exhibit average con-tributions of 19.3 or 19.5 tokens, respectively. Aparticularly remarkable fact represents the finalperiod experience of group 9. Whereas alt sub-jects chose full defection (g, = 0) in the no-punishment condition all subjects chose fullcooperation (g, = 20) in the punishmentcondition.

Figures 3A and 3B show the evolution ofaverage contributions over time. Irrespective ofwhether subjects play the punishment conditionat the beginning or after the no-punishmentcondition, their average contributions in the fi-nal period are considerably higher than in the

THE AMERICAN ECONOMIC REVIEW SEPTEMBER 2000

TABLE 4—MEAN CONTRIBUTIONS IN THE PARTNER-TREATMENTS

Groups

1

2

3

4

5

6

7

8

9

10

Mean

Mean contributions iti all

Withoutpunishmentopportutiity

7.0(6.3)10.6(8.5)6.7

(7.8)5.1

(6.3)6.4

(7.2)7.9

(5.7)7.4

(7.1)10.0(6.6)3.9

(5.9)10.0(6.6)

7.5(6.8)

periods

Withpunishmentopportunity

17.5(4.3)16.4(5.2)18.4(3.6)12.1(7.1)14.3(7.0)19.0(2.8)19.0(3.4)17.2(4.3)17.0(5.0)19.0(2.1)

17.0(4.5)

Mean contributionsperiods

Withoutpunishmentopportunity

S.8(5.1)1.0

(1.4)63

(9.5)1.3

(2.5)1.8

(2.9)3.5i^m(2.9)5.0

(6.0)

ao(m5.0(8.0)

3.2(4.4)

in the final

Withpunishmentopportunity

19.5(1.0)19.3(1.5)20.0(0.0)13.5(8.5)10.5

(11.0)20.0(0.0)20.0(0.0)20.0(0.0)20.0(0.0)19.5(1.0)

18.2(2.3)

Notes: Numbers in parentheses are standard deviations. Groups 1-4 (Session 4) first playedthe punishment condition and then the no-punishment condition. Groups 5-10 (Session 5)played in the reverse order.

first period of the punishment condition. Theopposite is true in the no-punishment treatment.Moreover, at the switch points between thetreatments there is a large gap in contributionsin favor of the punishment condition. This in-dicates that the removal or the introduction ofpunishment opportunities immediately affectscontribution behavior.^ Thus, Table 4 and Fig-ures 3A and 3B show that—in the Partner-treatment—punishment opportunities not onlyoverturn the downward trend observed in doz-ens of no-punishment treatments; they also

In Session 4 and in Session 5 average contributions inperiod 11 are significantly different from contributions inperiod 10 [Wilcoxon signed-ranks tests, p = 0.05 (Session4) and /J = 0.027 (Session 5)]. It is particularly remarkableIhat in Session 5 contributions in period 11 are even higherthan in period 1 (Wilcoxon signed-ranks test, p = 0.028).AU six groups of Session 5 contribute more in period 11than in period 1.

show that punishment opportunities rendereight of ten groups capable of achieving almostfull cooperation, although—according to thestandard approach^):/// defection is the uniquesubgame perfect equilibrium.

A major purpose of the Partner-treatmentwith punishment is to enhance the possibilitiesfor implicit coordination. We conjectured thatthis might enable subjects to converge toward abehavioral standard different from g^ == 0. Re-sult 6 shows that this is indeed the case.

RESULT 6: In the Partner-treatment with pun-ishment, full cooperation emerges as the domi-nant behavioral standard for individualcontributions, whereas in the absence of pun-ishment opportunities full free-riding is the fo-cal action.

Evidence for Result 6 is given by Figure4, which shows the relative frequency of indi-

VOL. 90 NO. 4 FEHR AND GACHTER: COOPERATION AND PUNISHMENT 989

20

4 -.

2 --

0

• with punishment

• without punishment

H h H h

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

Periods

FIGURE 3A. AVERAGE CONTRIBUTIONS OVER TIME IN THE PARTNER-TREATMENT {SESSION 4)

without punishment

with punishment

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

Periods

FIGURE 3B. AVERAGE CONTRIBUTIONS OVER TIME IN THE PARTNER-TREATMENT (SESSION 5)

vidual contributions in the final periods of thePartner-treatments. In the punishment condition82.5 percent of the subjects contribute thewhole endowment, whereas 53 percent of thesame subjects free ride fully in the final periodof the no-punishment condition. Moreover, inthe no-punishment condition the majority ofcontributions is rather close to gj = 0. Themessage of Figure 4 seems so unambiguous thatit requires little further comment.

C. Why Do Punishment Opportunities Rai.seContributions?

If there are indeed subjects who are willing topunish free-riding and if their existence is an-ticipated by at least some potential free-riders,we should observe that punishment opportuni-ties have an immediate impact on contributions.Figures 1 and 3 show that this is indeed thecase. After the introduction of punishment

990 THE AMERICAN ECONOMIC REVIEW SEPTEMBER 2000

with punishmentwithout punishment

Contributions 19 20

FIGURE 4. DISTRIBUTION OF CONTRIBUTIONS IN THE FINAL PERIODS OF THE PARTNER-TREATMENT

wrra AND WITHOUT PUNISHMENT

opportunities in Session 3 (see Figure IB) andSession 5 (see Figure 3B) there is an immediateincrease in contributions. Moreover, after theremoval of punishment opportunities in Ses-sions 1 and 2 (see Figure 1 A) and Session 4 (seeFigure 3A) contributions immediately drop toconsiderably lower levels. This suggests thatpotential free-riders are indeed disciplined inthe punishment condition. A more detailed lookat the regularities of actual punishments pro-vides further support for this view.

RESULT 7: In the Stranger- and the Partner-treatment a subject is more heavily punished themore his or her contribution falls below theaverage contribution of other group members.Contributions above the average are punishedmuch less and do not elicit a systematic punish-ment response.

Figure 5 and Table 5 provide evidence forResult 7. In Figure 5 we have depicted theaverage punishment levels as a function of neg-ative and positive deviations from the others'average contribution in the group. For example,a subject in the Partner-treatment, who contrib-uted between 14 and 20 tokens less than theaverage, received on average 6.8 punishmentpoints from the other group members. The num-bers above the bars indicate the relative fre-quency of observations in the differentdeviation intervals.

Figure 5 shows that in both treatments nega-tive deviations from the average are stronglypunished. Moreover, in the domain of negativedeviations (i.e., in the three intervals below-2) , the relation between punishment and de-viations is clearly negatively sloped. The figurealso indicates that there is a large drop in pun-ishments if an individual's contribution is closeto the average (i.e., in the interval [-2, +2]).^Finally, the figure suggests that positive devia-tions are much less punished and that the size ofthe positive deviation has only a weak impacton the punishment activities by other groupmembers.'"

Figure 5 also provides further support for the emer-gence of a common behavioral standard for individual con-trihutions in the Panner- btil not in the Stranger-treatment.Note that 57 percent of all the individual contributions in thePartner-treatment are in the interval [-2. +2], whereas only26 percent are in this interval in the Stranger-treatment.

' One might ask why individuals with positive devia-tions get punished at all. According to a postexperimentalquestionnaire there are five potential reasons for this, (i)Random error. Since individuals can err on only one side atthe punishment stage (i.e., rewarding others was not possi-ble), each error shows up as a positive punishment, (ii)Subjects with very high individual contributions may viewothers' contributions as too low, even if they are above theaverage, (iii) Subjects may want to eam more than others(i.e., they punish, even if others cooperate, to achieve arelative advantage), (iv) Spiteful revenge. Free-riding sub-jects punish the cooperators because they expect to getpunished by them, (v) Blind revenge. Subjects who get

VOL 90 NO. 4 FEHR AND GACHTER: COOPERATION AND PUNISHMENT 991

[-20,-14) [-14,-8) [-8.-2) [-2.21 (2,8| (8,14] (14,20]

Deviation from the average contribution of the other group members

FIGURE 5. RECEIVED PUNISHMENT POINTS FOR DEVIATIONS FROM OTHERS' AVERAGE CONTRIBUTION

TABLE 5—DETERMINANTS OF GETTING PUNISHED: REGRESSION RESULTS

Independent variables

Constant

Others' average contribution

Absolute negative deviation

Positive deviation

Dependent variable:

Stranger-treatment

2.7363***(0.0485)

-0 .0735***(0.0239)0.2428* ••

(0.0325)-0 .0147(0.0264)

N = 720F[14. 705] = 39.0***Adjusted R^ = 0.43DW = 1.96

received punishment points

Partner-treatment

0.9881(0.6797)

-0.0108(0.0457)0.4168***

(0.0510)-0.0357(0.0355)

N = 400F[21 , 378] = 41 .3***Adjusted R~ - 0.68DW = 1.89

Notes: Standard errors are in parentheses. * denotes signiticance at the 10-percent level, ** atthe 5-percent level, and *** at the 1-percent level. To control for time and matching groups,the regression model also contains period dummies and dummies tor matching groups (i.e.,session dummies in the Stranger-treatment and dummies for each independent group in thePartner-treatment). Results are corrected for heteroskedasticity. Tobit estimation.s yield sim-ilar results.

To provide formal statistical evidence forResult 7 we also conducted a regression anal-ysis of punishment behavior. Table 5 containsthe model and the ordinary least-squares

punished in ; - 1 may assume that punishment was mainlyexerted by the cooperators. By punishing cooperators in /they may take revenge. Note that by doing this they maypunish the wrong target, hecause our design rules out thepossibility of identifying individual contribution histories.

(OLS) regressions separately for the Stranger-treatment and the Partner-treatment. We alsoconducted Tobit regressions with the samevariables. Yet, since they are similar to theOLS estimates we do not report them explic-itly. The dependent variable is "received pun-ishment points" of a subject and theindependent variables comprise "others' av-erage contribution" and the variables "posi-tive deviation" and "absolute negativedeviation," respectively. Figure 5 suggests

992 THE AMERICAN ECONOMIC REVIEW SEPTEMBER 2000

that positive and negative deviations from theothers" average contribution elicit differentpunishment responses. These variables aretherefore included as separate regressors. Thevariable "absolute negative deviation" is theabsolute value of the actual deviation of asubject's contribution from the others' aver-age in case that his or her own contribution isbelow the average. This variable is zero if thesubject's own contribution is equal to orabove the others" average. The variable "pos-itive deviation" is constructed analogously.To model time effects, we included perioddummies in the regression. The model alsoincludes session dummies in the Stranger-treatment and group dummies in the Partner-treatment to control for fixed effects [seeManfred Konigstein (1997)].

The results in Table 5 support the evidencefrom Figure 5. In both treatments the coeffi-cient of the "absolute negative deviation" ispositive and highly significant; thus, the morean individual's contribution falls short of theaverage the more that individual gets pun-ished. In contrast, the size of the positivedeviation has no significant impact on the sizeof the punishment. It is interesting that in thePartner-treatment it is only the negative devi-ation that affects punishment levels systemat-ically, whereas the level of the others'average contribution has no significant im-pact. The low value and the insignificance ofthe coefficient on "others" average contribu-tion" suggests that only deviations from theaverage were punished. This may be taken asevidence that in the Partner-treatment sub-jects quickly established a common groupstandard that did not change over time. If,instead, there would have been subjects whowanted to raise, say, the group standard, oneshould observe that a given negative devia-tion from the average is punished less thehigher that average is. This is exactly what weobserve in the Stranger-treatment in whichthe coefficient on "others' average contribu-tion'" is negative. The fact that there weresubjects in the Stranger-treatment whowanted to raise the group standard is consis-tent with previous evidence which shows thatsubjects in the Stranger-treatment could notestablish a common behavioral standard.

The pattem of punishment indicated by

Figure 5 and Table 5 shows that free-riderscan escape or at least reduce the receivedpunishment substantially by increasing theircontributions relative to the other group mem-bers. The response of subjects who actuallywere punished suggests that they understoodthis. In the Partner-treatment we observed 125sanctions against subjects who contributedless than their endowment. In 89 percent ofthese cases the punished subject increased g^immediately in the next period with an aver-age increase of 4.6 tokens. In the Stranger-treatment we have 368 such cases. In 78percent of these cases g, increased in the nextperiod by an average of 3.8 tokens. Thesenumbers suggest that actual sanctions wererather effective in immediately changing thebehavior of the sanctioned subjects. Subjectsseemed to have had a clear understanding ofwhy they were punished and how they shouldrespond to the punishment.

D. Payoff Consequences of Punishment

A major effect of the punishment opportunity isthat it reduces the payoff of those with a relativelyhigh propensity to free ride. In the following wecall those subjects "free-riders"' who chose g, ~ 0in more than five periods of the no-punishmenttreatment. Twenty percent of subjects in the Part-ner-treatment and 53 percent in the Stranger-treat-ment obey this definition of a free-rider. In theStranger-treatment with punishment opportunitiesthe overall payoff of the free-riders is reduced by24 percent relative to the no-punishment condi-tion; in the Partner-treatment the payoff reductionis 16 percent. This payoff reduction is driven bytwo sources. First, free-riders are punished moreheavily and second, they contribute more to theproject in the punishment condition. On average,free riders raise their contributions between 10 and12 tokens (i.e., by 50 to 60 percent of their en-dowment), relative to the no-puni.shment condi-tion. However, there is also a force that worksagainst the payoff reduction for free riders becausethe other subjects (the "nonfree-riders") also con-tribute more in the punishment condition. Thislimits the payoff reduction for the free-riders.

What are the aggregate payoff consequences ofthe punishment condition? To examine thisquestion we compute the difference in the averagegroup payoff between the punishment and the no-

VOL. 90 NO. 4 FEHR AND GACHTER: COOPERATION AND PUNISHMENT 993

punishment condition and divide this difference bythe average group payoff of the ntvpunishment con-dition. This gives us the relative payoff gain of thepunishment condition. Result 8 summarizes the evo-lution of the relative payoff gain for both the Partner-and the Stranger-treatment.

RESULT 8: In both the Stranger- and thePartner-treatment the punishment opportu-nity initially causes a relative payoff loss. Yet,toward the end there is a relative payoff gainin both treatments. In particular, in theStranger-treatment the relative payoff gain ofthe punishment condition is positive in thelast two periods, whereas in the Partner-treatment it is positive from period 4 onward.In the final period the relative payoff gain isroughly 20 percent in the Partner-treatmentand W percent in the Stranger-treatment.

The temporal pattem of relative payoff gainsresults from two sources; (i) In the Partner-treatment, in particular, contributions are lowerin the early periods of the punishment conditionthan during the later periods and this causedmuch more punishment activities in the earlyperiods, (ii) Contributions gradually declineover time in the no-punishment condition.Taken together. Result 8 suggests that the pres-ence of punishment opportunities eventuallyleads to pecuniary efficiency gains. To achievethese gains, however, it is necessary to establishthe full credibility of tbe punishment threat byactual punishments.

IV. Conclusion

This paper provides evidence that spontaneousand uncoordinated punishment activities giverise to heavy punishment of free-riders. In tbeStranger-treatment this punishment occurs, al-though it is costly and provides no future privatebenefits for the punishers. The more an individualnegatively deviates from the contributions of theother group members, the heavier the punishment.Recently developed models of equity and reci-procity predict the widespread punishment offree-riders. Punishment is. however, clearly incon-sistent with mtxlels of pure altruism or wami-glowaltruism (e.g., James AndreonI, 1990) because analtruistic person never uses a costly option toreduce other subjects' payoffs. The apparent will-

ingness to punish constitutes a credible threat forpotential free riders and causes a large increase incooperation levels: very high or even full co-operation can be achieved and maintained inthe punishment condition, whereas the samesubjects converge toward full defection in theno-punishment condition.

In our view punishment of free-riding alsoplays an important role in real life. It seems,for example, rather likely that—under teamproduction—shirking workers elicit strongdisapproval among their peers, and that strike-breaking workers face the spontaneous hostilityof their striking colleagues. The enormous im-pact of the punishment opportunities on contri-butions in our experiment suggests that aneglect of the widespread willingness to punishfree-riders faces the serious risk of makingwrong predictions and, hence, giving wrongnormative advice. Institutional and social struc-tures that, theoretically, trigger the same behav-iors in the absence of the willingness to punishmay cause vastly different behaviors if the will-ingness to punish is taken into account.

REFERENCES

Andreoni, James. "Impure Altruism and Dona-tions to Public Goods: A Theory of WarmGlow Giving?" Economic Journal, June1990. !00i40\), pp. 464-77.

Berg, Joyce; Dickhaut, John and McCabe, Kevin."Trust Reciprocity and Social History,"Games and Economic Behavior, July 1995,I0{\). pp. 122-42.

Fehr, Ernst; Kirchsteiger, Georg and Riedl, Arno."Does Fairness Prevent Market Clearing? AnExperimental Investigation." Quarterly Jour-nal of Economics, May 1993, pp. 437-60.

Fehr, Ernst; Gachter, Simon and Kirchsteiger,Georg. "Reciprocity as a Contract Enforce-ment Device—^Experimental Evidence."Econometrica, July 1997, 65(4), pp. 833-60.

Fehr, Ernst and Schmidt, Klaus M. "A Theory ofFaimess, Competition, and Cooperation."Quarterly Journal of Economics, August1999, !!4(3),pp. 817-68.

Fischbacher, Urs. "z-Tree: Zurich Toolbox forReadymade Economic Experiments. Instruc-tions for Experimenters."' Mimeo, Universityof Zurich, 1998.

994 THE AMERICAN ECONOMIC REVIEW SEPTEMBER 2000

Francis, Hywel. "The Law, Oral Tradition andthe Mining Community."' Journal of Law andSociety. Winter 1985, 12(3), pp. 267-71.

Hirshleifer, David and Rasmusen, Eric. "Coopera-tion in a Repeated Prisoners' Dilemma withOstracism." Journal of Economic Behavior andOrganization, Augast 1989, 72(1), pp. 87-106.

Konigstein, Manfred. "Measuring Treatment Ef-fects in Experimental Cross-Sectional TimeSeries." Mimeo, Humboldt-University, Ber-lin, 1997.

Ledyard, John O. "Public Goods: A Survey ofExperimental Research," in John H. Kageland Aivin E. Roth, eds.. Handbook of exper-imental economics. Princeton: Princeton Uni-versity Press, 1995, pp. 111-94.

Ostrom, Elinor; Walker, James and Gardner,Roy. "Covenants With and Without a Sword:Self-Governance is Possible," American Po-litical Science Review, June 1992, 86(2), pp.404-17.

Roethlisherger, F. J. and Dickson, W. J. Man-agement and the worker: An account of aresearch program conducted by the West-ern Electric Company, Hawthorne Works.Cambridge, MA: Harvard University Press,1947.

Roth, Alvin E. "Bargaining Experiments," inJohn H. Kagel and Alvin E. Roth, eds..Handbook of experimental economics.Princeton: Princeton University Press,1995, pp, 253-348.